Article 2

Computational Methods for Residential Energy Cost Optimization
in Smart Grids: A Survey

2
MUHAMMAD RAISUL ALAM, MARC ST-HILAIRE, and THOMAS KUNZ, Carleton University
A smart power grid transforms the traditional electric grid into a user-centric, intelligent power network.
The cost-saving potential of smart homes is an excellent motivating factor to involve users in smart grid
operations. To that end, this survey explores the contemporary cost-saving strategies for smart grids from the
users’ perspective. The study shows that optimization methods are the most popular cost-saving techniques
reported in the literature. These methods are used to plan scheduling and power utilization schemes of
household appliances, energy storages, renewables, and other energy generation devices. The survey shows
that trading energy among neighborhoods is one of the effective methods for cost optimization. It also
identifies the prediction methods that are used to forecast energy price, generation, and consumption profiles,
which are required to optimize energy cost in advance. The contributions of this article are threefold. First,
it discusses the computational methods reported in the literature with their significance and limitations.
Second, it identifies the components and their characteristics that may reduce energy cost. Finally, it proposes
a unified cost optimization framework and addresses the challenges that may influence the overall residential
energy cost optimization problem in smart grids.
Categories and Subject Descriptors: G.1.6 [Optimization]: Constrained Optimization, Convex Program-
ming, Global Optimization, Integer Programming, Linear Programming, Nonlinear Programming, Quadratic
Programming Methods, Stochastic Programming, Unconstrained Optimization; G.3 [Probability and
Statistics]: Correlation and Regression Analysis, Markov Processes; I.2.8 [Problem Solving, Control
Methods and Search]: Heuristic Methods, Scheduling; K.6.0 [General]: Economics
General Terms: Algorithms, Performance
Additional Key Words and Phrases: Smart grid, smart homes, survey, cost optimization, demand and
response, microgrid, dynamic price, renewables, energy storage
ACM Reference Format:
Muhammad Raisul Alam, Marc St-Hilaire, and Thomas Kunz. 2016. Computational methods for resid-
ential energy cost optimization in smart grids: A survey. ACM Comput. Surv. 49, 1, Article 2 (April 2016), 34
pages.
DOI: http://dx.doi.org/10.1145/2897165
1. INTRODUCTION
A smart home is an application of ubiquitous computing that is able to provide context-
aware automated or assistive services to users in the form of ambient intelligence,
remote home control, or home automation [Alam et al. 2012]. It incorporates smartness
into the dwellings, provisioning a better control of home appliances to ensure comfort,
safety, and security; improve healthcare; and optimize energy cost [Alam et al. 2012].
In a smart grid, the residential user is considered a prosumer, which means that a
user not only consumes energy but also may produce it. There are different types
Authors’ address: M. R. Alam, M. St-Hilaire, and T. Kunz, Department of Systems and Computer Engi-
neering, 1125 Colonel By Drive, Carleton University, Ottawa, Ontario K1S 5B6 Canada; emails: raisul@sce.
carleton.ca, marc_st_hilaire@carleton.ca, tkunz@sce.carleton.ca.
Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted
without fee provided that copies are not made or distributed for profit or commercial advantage and that
copies show this notice on the first page or initial screen of a display along with the full citation. Copyrights for
components of this work owned by others than ACM must be honored. Abstracting with credit is permitted.
To copy otherwise, to republish, to post on servers, to redistribute to lists, or to use any component of this
work in other works requires prior specific permission and/or a fee. Permissions may be requested from
Publications Dept., ACM, Inc., 2 Penn Plaza, Suite 701, New York, NY 10121-0701 USA, fax +1 (212)
869-0481, or permissions@acm.org.
c 2016 ACM 0360-0300/2016/04-ART2 $15.00
DOI: http://dx.doi.org/10.1145/2897165
ACM Computing Surveys, Vol. 49, No. 1, Article 2, Publication date: April 2016.
2:2 M. R. Alam et al.
of users, such as home or apartment users, building management authorities, hotel

management authorities, and office management authorities. The user may also be
a group of cooperative prosumers who collectively agree to exchange information for
cost reduction. In this article, the terms utility and grid are used interchangeably to
represent the electric or power grid.
Recent advancements of information and communication technology changed the tra-
ditional concept of the power grid. The idea of a smart power grid was mainly initiated
by the utilities. The first smart grid implementation is credited to ENEL Spa in Italy,
which initiated the Telegestore project in 2000 [Rogai 2006; ENEL 2014]. The utility
prefers to optimize its own energy production cost, automate the energy transport and
billing systems, and improve the quality of service (QoS). The northeastern blackout
of 2003 affected 10 million people in Ontario and 40 million people in eight U.S. states
[Mansbridge 2003]. For this reason, governments promote smart grid research because
it will have fault detection, tolerance, and self-healing features to prevent power black-
outs. The smart grid encourages the utilization of renewable energy sources, which
reduces carbon emission—the primary factor of climate change. For example, the Fort
Collins district in Colorado is aiming to reduce its carbon emission by 80% by 2030
[Kumagai 2014]. To achieve this target, it will deploy rooftop solar photovoltaics (PVs),
community-supported solar gardens, wind turbines, thermal and electric storage, mi-
crogrids, and energy efficiency schemes. The recent Fukushima Daiichi nuclear disas-
ter accelerated the adoption of renewable energy sources to the smart grid [IEA 2013].
Therefore, the main stakeholders of the smart grid—the governments, the utility, and
the users—inspired by different motivating factors, are gradually adopting smart grid
technologies as a solution to their own interests.
A smart home in the smart grid offers an optimized energy management solution in
collaboration with the utility and the neighbors. The widespread participation of users
plays an important role in achieving the optimal benefit from the smart grid. Users are
mostly motivated by the cost-saving potential of smart homes. This survey addresses
cost-saving strategies for smart homes from the users’ perspective. A smart home has
diverse appliances, devices, and equipment with different energy consumption profiles.
The user has his own preferences for the appliance operation to ensure an optimized
energy cost as well as to maintain a preferred comfort level. Previous research identified
that optimal rescheduling of household loads according to the energy price is one of
the effective methods of cost minimization. One or multiple storage devices accumulate
cheap and excessive energy to support future energy requirements, avoiding the need
to acquire expensive energy from the grid during periods of high prices (e.g., at peak
loads). However, the usage of such storage devices imposes a cost overhead because of
self-discharging and efficiency loss. It was reported that optimized utilization of storage
devices (e.g., thermal, electrical) impacts energy cost. Instead of simply depending on
the utility, a smart home can generate its own energy utilizing solar panels, wind
turbines, geothermal plants, and other renewable energy sources. The renewables show
stochastic energy generation profiles correlated to the weather. It is evident from the
literature that the usage of renewable sources and advance planning of renewable
energy utilization based on the predicted generation quantity reduces energy cost.
Smart homes in a neighborhood may collectively form a microgrid to facilitate energy
trading among themselves. A previous study claimed that energy trading among the
households by forming a microgrid plays an important role for cost minimization in the
smart grid.
Energy cost is primarily optimized by rescheduling the appliances when energy is
cheap (energy sources could be grid, microgrid, or renewables), provisioning appliance
power, interrupting appliance operation, scheduling storage usage, and energy trading.
However, most existing research has failed to address a comprehensive discussion of
Computational Methods for Residential Energy Cost Optimization in Smart Grids: A Survey 2:3
Fig. 1. Overview of cost-saving strategies for smart homes.
smart homes, considering all important possible aspects for cost optimization. To deal
with all of the preceding smart home features, sophisticated algorithms are required.
To that end, this survey explores cost optimization methods for smart homes that in-
volve diverse energy sources, storage and loads constrained by user preferences, and
electrical properties of the participating components. It also addresses cost minimiza-
tion strategies associated with energy trading among participating households in the
microgrid. It identifies potential methods and algorithms, classifies them according to
research objectives and used techniques, and discusses the significance and limitations
of these proposed solutions.
2. COST-SAVING STRATEGIES
A user implements cost-saving strategies considering the available resources and in-
frastructure. In smart homes, visualization and prediction tools are used to provide
estimates of energy supply, usage, and energy prices. Scheduling household loads and
energy sources according to the energy price is the main theme of cost optimization.
The energy price depends on the utility and the locally available microgrid. The utility
tries to indirectly control the loads in the user premises via a demand response (DR)
program. The DR program primarily specifies an energy price to balance the energy
generation and energy demand. In a microgrid, the price signal fluctuates based on
the energy generation from renewables and the energy demand of participating house-
holds. The user may wish to earn a profit by trading the surplus energy in the energy
market. The attainable profit depends on the nature of the market, market conditions,
and opponents’ behavior. Figure 1 presents an overview of methods used to minimize
energy cost or maximize profit in smart homes.
It is reported that the DR program has significant impact on cost reduction [Siano
2014]. Vardakas et al. [2015] classified the DR methods into three categories, as shown
in Figure 2. The first category is based on control mechanism, including centralized
and distributed control. In the second category, DR methods are classified based on the
motivating factors. The price- and incentive-based DR methods fall into this category.
Finally, the third category is arranged based on the decision variables. Energy con-
sumption can be controlled by scheduling a task operation and energy management
(i.e., reducing the power of a load operation). Recently, an increased trend of optimiza-
tion method development for the DR program has been reported [Vardakas et al. 2015;
Shariatzadeh et al. 2015].
Fig. 2. A classification of DR programs. [Vardakas et al. 2015].
2.1. Energy Use Monitoring and Manual Control

Energy consumption monitoring and manually controlling electrical appliances accord-
ing to user preferences and the energy price are basic ways to reduce energy cost. This
section discusses projects that were developed to create awareness of energy consump-
tion of household appliances. The user can visualize the trends of the overall energy
consumption and learn ways to shed loads manually by using this information.
Pal et al. proposed a set top box (STB) as a gateway between the utility grid, smart
energy meter, and home appliances [Pal et al. 2011]. The STB uses a TV or VGA moni-
tor to display the energy pricing and usage information. Appliances communicate with
the STB via a home area network (HAN), and the utility provider uses the Internet
to receive energy usage information. The energy meter sends the energy consumption
information to the STB, which is analyzed and validated by the data management
system of the STB and stored in a data repository. The developed system did not con-
sider intelligent device scheduling. The user monitors energy consumption of different
appliances and manually schedules them according to the electricity price.
Jahn et al. [2010] developed a real-time energy consumption monitoring system for
smart homes based on the Open Service Gateway initiative (OSGi) [OSGi Alliance
2004] framework. It improved the research done by Pal et al. [2011] by integrating
smart phones into the energy monitoring system. The proposed system used energy-
monitoring sensors, named Plogg, to acquire the energy consumption information of the
associated devices using Zigbee or Bluetooth. The Plogg sensors were integrated with
the Hydra framework [Hydra 2015] using the Plogg Software Development Kit (SDK)
wrapped with a Java Native Interface (JNI). The Hydra framework uses the OSGi
environment to implement operational modules. It communicates with remote OSGi
components using a Simple Object Access Protocol (SOAP)-based Web service interface.
The system collects the energy consumption information of devices with the attached
Ploggs and displays the information on a large screen, such as a TV. It allows the user
to recognize a device and view energy consumption information using a smart phone.
The system used the UbiLense architecture for this purpose. The mobile phone sends
the device image to an image recognition server, which provides the Plogg ID of the
object. The phone uses the ID to display the energy consumption. The user can control
(turn on and off) the devices through a mobile phone. The authors argued that repre-
senting the energy consumption in watt-hours is hard to conceive by the general public
because most of them are not familiar with these energy concepts. Hence, the proposed
system presented the energy consumption by converting it to monetary values. It also
displays the predicted annual energy cost of the selected appliance. The system can
accommodate heterogeneous communication protocols, appliances, and energy sensors.

It was tested using a PC (as the backend), two appliances (emulated using tablet PCs),
and a smart phone. The system takes a maximum of 2 seconds to read a Plogg sensor,
which will become less interactive if the number of smart devices increases.
Weiss and Guinard [2010] presented an energy monitoring system based on a REST-
ful architecture. The proposed system used Plogg smart electric plugs to collect energy
consumption information of the attached appliances through a Bluetooth network. A
Bluetooth-enabled gateway collects the information from the Plogg energy sensors.
The gateway is also a Web server that provides URLs to enable the RESTful archi-
tecture. The clients use the HTTP protocol to obtain services from the gateway. Two
types of user interfaces had been developed for the clients: a Web interface and a mo-
bile interface. The interfaces show real-time energy consumption and energy usage
trends of the appliances. A user can control the appliances through the user interface.
A prototype was deployed at a private office where real-time energy consumption in-
formation was displayed on a large screen, which enables users to experiment with
the system. The Ploggs were used to monitor a fridge, a kettle, several printers, a file
server, and computers. The gateway communicates with the Ploggs every 30 seconds;
however, for explicit user requests, the communication occurs instantly. The primary
goal of this project was to develop awareness of energy consumption and energy waste.
The authors reported that technical terms should be represented using more abstract
terminologies for better understanding by the general (i.e., nontechnical) public. It
also reported that when the initial curiosity of the users was satisfied, they started
using the system less. The authors suggested to present long-term trends of energy
consumption in the user interface to motivate users to use the system more frequently.
To convince the user to reduce energy consumption, a system can use psychological
concepts such as goal setting, budgeting, and social impact (energy saving among
friends).
If the user decides to use renewable energy sources, energy storage, and microgrid,
the overall system will become more complex. In this case, manually controlling the
loads and sources is a tedious task that may not provide optimal cost reduction. The
household requires an automated intelligent system to gain the maximum benefit of
the underlying technologies. The remaining portion of this article focuses on automated
load and source control.
2.2. Energy Price Negotiation and Prediction

In the smart grid, the grid utilizes dynamic electricity prices to indirectly control
household appliances. It tries to shed some load when the energy demand is high
and increase the energy utilization when demand is low. The user should know the
energy price in advance to efficiently reschedule the loads, renewable sources, and
storages. The projects discussed in this section proposed methods for price prediction
and negotiation in smart grids.
2.2.1. Statistical Correlation. Mohsenian-Rad and Leon-Garcia [2010] proposed an en-
ergy price prediction model for smart grids. The project goal was to develop a method
that has low computational complexity without compromising the prediction accuracy.
The model works for utility grids that use an inclining block rate (IBR) strategy to
determine the electricity price. In IBR pricing, the marginal price increases if the total
consumption increases above a threshold level. The authors presented an exploratory
study on the IBR pricing information of Illinois Power Company from January 2007
to December 2009 [Ameren Corporation 2012]. Based on the statistical analysis, the
authors argued that the current price of electricity depends on the prices of yesterday,
the day before yesterday, and the same day of the past week. The paper developed a
model based on the correlation of these three prices and three different coefficients
to control the weight of these prices. The model was trained with the hourly prices
of the last week of November 2009, and the prediction accuracy was compared to the
real price on December 1, 2009. It was reported that the prediction error is 13% on
average if different coefficients are used for different days of the week. Using the same
coefficients for every day, the error rate was 17% on average.
2.2.2. Convex Optimization. Hovgaard et al. [2013] presented a price prediction model as
a least square optimization problem and predicted the energy price by minimizing the
objective function. The paper used historical hourly data to calculate the average daily
price variation for each month of the year. The price was sampled every hour of a day
(24 data points per day). A smooth baseline had been computed with 1-year data points
by using linear interpolation of two adjacent months. The model predicts the energy
price by using the residual of the training data to predict the future residuals and
adding these to the baseline of the corresponding time window. A positive parameter
was used to minimize the validation error. The same algorithm was also used to predict
the weather temperature. It was trained using electricity prices from January 1, 2007,
to December 31, 2009, and tested with the prices for the entire 2010 year. The price data
range was from 20 EUR/MWh to 100 EUR/MWh, with an average of 46 EUR/MWh.
For the computed baseline, the MAE was 13.2 EUR/MWh. Results showed that the
proposed predictor can minimize the prediction errors from the baseline. It is also
identified that predicting more future hours reduces the accuracy. In other words, the
error is low when the model is predicting the price of the next hour, and it increases
when it tries to predict for additional hours.
Samadi et al. [2010] proposed a real-time pricing (RTP) algorithm that maximizes
aggregated utility and welfare (difference between utility and cost) of the participating
users. The paper assumed that each user is connected to other users and the utility
through bi-directional communication links. The users and the utility interact with
each other by exchanging control messages related to energy consumption and elec-
tricity price. The utility is expressed as a quadratic equation that is a function of energy
consumption and a time-varying user-dependent parameter. The cost function is also
a quadratic function that depends on the power generation capacity of the utility. The
proposed convex optimization model tries to maximize the overall difference between
the utility of all users and their costs. The model is expressed using a Lagrangian dual
function that is further decomposed into two separate objective functions. An objective
function is used by the individual user to maximize welfare. The other objective func-
tion is utilized by the utility to maximize the profit. Each individual user receives a
proposed energy price from the utility, and the user sends back the consumption value
to the utility. The utility receives the consumption value from all users and resends
an updated price based on the total energy consumption and total energy generation
capacity. The process iteratively continues for each time slot until it maximizes the
total welfare of the participating parties. The authors compared the method to a fixed
price scenario that did not consider the welfare of the users. Results showed that it
improved the overall welfare of all users. The proposed model ensures that the user can
negotiate with the utility to pay a competitive price for the consumed energy. Instead
of acting autonomously, the utility cooperates with the users to ensure the maximum
benefit for the whole community.
2.2.3. Service-Oriented Architecture. Pagani and Aiello [2015] presented a service-
oriented architecture (SOA) to simulate dynamic price generation in the smart grid
[Pagani and Aiello 2015]. The authors considered the day-ahead prices advertised by
the wholesale market as the retail energy prices for the end user. The research as-
sumed that there is a possibility to have multiple energy providers and considered
the renewable energy sources (PV and wind) as additional energy providers. The en-
ergy prices for the renewables are formulated as a function of investment, operation,
and maintenance cost. The authors proposed a simulation model using the energy
price (wholesale price as retail price), forecasted renewable energy generation capacity
based on weather information, and renewable energy prices to evaluate the impact
of dynamic prices on cost savings. The authors pointed out the fact that in a current
situation, when the dynamic price for the end user is rarely available, considering the
wholesale energy price as the retail price is closer to the actual scenario.
2.2.4. Summary. The energy price is highly related to the hours of the day [Mohsenian-
Rad and Leon-Garcia 2010]. It has a correlation with the previous day and the same day
the previous week. It also depends on the month of the year that actually represents
the seasons. In winter or summer, the water and space heating and cooling system
consume a major portion of energy. The operating hour of these loads has a significant
influence on the energy price. It is expected that electric vehicles (EVs) will appear as
a major load in a smart grid. From the current research, it is not obvious how EVs will
influence the future energy price.
The energy price has been estimated from the utility point of view. Usually, the
energy generation capacity of the utility is stable. Therefore, the supply of energy in
the market is fixed. The price signal is utilized to change the energy demand. In a smart
grid, the users can collectively form a microgrid to sell energy between them. In this
situation, the supply of energy will change according to weather and user preferences.
Moreover, the energy supply will also be influenced by market price and demand. In
this situation, the price will depend on more correlated parameters. The proposed price
prediction algorithms did not consider this situation.
Pagani and Aiello [2015] proposed a dynamic energy price prediction method based
on a wholesale energy price. Most of the utilities are still using time of use (TOU)
rates. In TOU, the utility defines the on-peak, shoulder, and off-peak energy prices in
advance for specific months or seasons. Sometimes the energy price depends on the
energy consumption of the user (e.g., IBR). If the total energy consumption exceeds a
threshold value, the user has to pay an additional price. Therefore, the energy price is
typically known in advance in these situations. In the current grids, for most utilities,
the user does not require prediction algorithms to estimate the energy price. However,
it is recommended that the utility should widely use a dynamic energy price to engage
residential users in DR program [Shen et al. 2014].
2.3. Renewable Energy Generation Prediction
The smart grid promotes the integration of renewable energy sources with smart
homes. The renewable sources harvest energy from natural resources, such as biomass,
biogas, solar power, wind power, and geothermal power. The presence of environmental
uncertainty on renewable energy sources has a significant impact on the energy supply
in smart homes. This section presents an overview of the methods used to forecast
renewable energy generation in smart homes.
2.3.1. Artificial Neural Network. Mellit et al. [2010] used a multilayer perception (MLP)
model to forecast solar irradiance. An MLP model is a feed-forward ANN that utilizes
the back-propagation technique to train the network. The proposed MLP is trained us-
ing the values of hourly solar irradiance and air temperature to forecast 24-hour solar
irradiance. The project used practical data collected in Trieste, Italy. A reference PV
cell was used to measure solar irradiance, and a temperature sensor (PT 100) was in-
stalled near the PV array plane to measure air temperature. A 10-fold cross-validation
was used to train and test the proposed approach. The efficiency of the proposed
model was evaluated by measuring different statistical parameters (e.g., mean absolute
error (MAE), mean bias error (MBE), root mean square error (RMSE), and correlation
coefficient) between the measured and forecasted power of the installed PV array. It is
reported that correlation coefficients were more than 98% for sunny days and around
95% for cloudy days. It was also found that the computing time of the proposed ap-
proach is high because of the required number for iterations. The authors suggested
that if a genetic algorithm (GA) is used to train the ANN, it will reduce the number
of iterations (because of the simpler MLP architecture) and the solution time. They
also emphasized on using large datasets (of more than 1 year) collected in a specific
geographic location, which will result in more accurate prediction.
Chitsaz et al. [2015] proposed a wind power forecast model using a Wavelet neural
network (WNN) with Morlet wavelets function. The model predicts wind power for the
next 6 hours based on the past power generation profile. It did not consider meteoro-
logical parameters like wind speed, power, direction, and other climate variables. An
improved clonal selection algorithm (ICSA) is used as the training and testing strat-
egy. The authors argued that the probability distribution of the wind power prediction
error is not a Gaussian distribution, so the maximum correntropy criterion (MCC)
method is a more appropriate procedure than the mean square error (MSE) to esti-
mate the forecast error. Therefore, they used MCC to calculate the accuracy of the
proposed model. The authors also presented the error rate in terms of MSE because
it is widely used in forecasting models. It used practical wind power generation data
collected in Alberta, Canada. The proposed method is compared to radial basis func-
tion (RBF), MLP, WNN with MSE, and persistence methods. Results show that the
proposed method outperforms the others. The article reported that for the proposed
WNN, the ICSA training strategy performs better than the clonal selection algorithm
(CSA), differential evolution (DE), particle swarm optimization (PSO), and simulated
annealing (SA).
Ciabattoni et al. [2012] proposed an online solar irradiance prediction method using
an ANN The proposed ANN uses an RBF as the activation function. The learning
method dynamically adds and prunes neurons in the model based on the minimal
resource allocating network (MRAN) strategy. The network parameters are updated
by an extended Kalman filter. The proposed online model is compared to the same
model trained with offline data, a classical RBF ANN, and actual solar irradiance. The
prediction accuracies were evaluated by comparing RMSEs and standard deviations,
which shows that the proposed model outperforms the others.
2.3.2. Support Vector Machines. Shi et al. [2011] argued that the PV power forecasting
models should be classified based on weather conditions and that a model specific to
a certain weather condition produces better results. Therefore, the authors used four
support vector machine (SVM) PV power forecasting models for four types of weather
conditions: cloudy, foggy, sunny, and rainy. They collected power outputs of a PV station
(20kW) in southern China for a 10-month period in 15-minute intervals and used this
information to train and test the proposed model. The performance of the SVM models
were evaluated using statistical parameters such as RMSE andmean relative error
(MRE). Their results showed that the average forecasting errors of the proposed SVM
models were 8.64% in MRE and 2.10MW in RMSE. It was found that the foggy and
the sunny SVM models outperform the others because of insufficient data to train the
SVM models for cloudy and rainy weather conditions.
2.3.3. Hybrid Models. Dong et al. [2015] proposed a hybrid solar irradiance prediction
model based on self-organizing maps (SOMs), support vector regression (SVR), and
PSO. SOM is applied to separate the entire input data into different regions based on
similar data characteristics. A tree-structured architecture is utilized to identify the re-
gions because the exact number of regions are unknown before the training procedure.
Each node of the tree represents a distinct region with specific characteristics. Each
end-node dataset was used to train an SVR with the parameters obtained from the PSO.
The performance of the SVR model strongly depends on three user-defined parameters,
namely the width parameter of the Gaussian kernel function, the deviation, and the
trade-off coefficient. The PSO is used to determine these three parameters of the SVR
model. Two datasets of solar irradiance time series are obtained from different geo-
graphic locations. The first dataset (17 years of data) was collected in South Park, Col-
orado, and the second dataset (3 years of data) was obtained from the Solar Energy Re-
search Institute of Singapore (SERIS). The performance of the proposed hybrid model
was evaluated by two different types of statistical measurement techniques: the nor-
malized mean bias error (nMBE) and the normalized root mean square error (nRMSE).
The average error was measured every hour from the average 1-hour actual and pre-
dicted average solar irradiance. The model was compared to four prediction methods:
random walk (RW), simple exponential smoothing (SES), linear exponential smooth-
ing (LES), and the autoregressive integrated moving average (ARIMA) model. Results
show that the proposed hybrid model has better prediction accuracy than others.
2.3.4. Summary. It is obvious from the literature that the performance of the predic-
tion algorithms depends on two main criteria. First, the accuracy of the forecasting
algorithm depends on the method or technique used to classify the input data into
similar regions based on the characteristics. The classification criterion can be based
on weather condition [Shi et al. 2011], or it can be an unsupervised learning method
that can identify the unique clusters with unknown characteristics [Dong et al. 2015].
Second, the accuracy of the forecasting method depends on the performance of the
prediction method applied to each classified region. The accuracy of the forecasting
algorithm is measured using statistical error estimation methods such as MSE, MRE,
MRSE, nRMSE, and nMBE. It is reported that SVM shows better performance than
ANN considering the solution time and result, although its performance strongly de-
pends on the amount of training data [Shi et al. 2011].
2.4. Energy Use Prediction

Electricity demand depends on the energy usage profile of the users. Energy usage
profiles provide an estimation of the household loads to the optimization and trading
algorithms so that they can yield informed solutions. This section presents an overview
of the projects that proposed energy usage prediction models.
2.4.1. Machine Learning. Bapat et al. [2011] used a smart plug, named jPlug, which
is able to measure the power usage of an appliance and is also able to control the
device. To analyze the device usage pattern, the power consumptions were monitored
according to predetermined time slots. Researchers used clustering and the frequent
item-set mining subroutine of WEKA [WEKA 2014] to derive the device usage profile.
2.4.2. Semantic Web. Cheong et al. [2011] proposed an ontology to predict the energy
saving states of smart appliances The proposed prototype utilized mobile phones and a
home server (or a cloud server) to acquire and process information. To predict the next
state of an appliance, the server tracks the power states (e.g., unplugged, power-off,
power-save, power-on) of the appliances and the residents’ location. A smart controller
is presumed to be attached to the appliance, which transmits the power states and
execute commands. The research team identified 174 keywords and 345 classes related
to smart homes. They employed Web Ontology Language–Description Logic (OWL-
DL) to model different classes and their relationships. The server executes concept
rules that are based on Semantic Web Rule Language (SWRL) to predict a high-level
situation when a collection of primitive contexts (e.g., user location, power level) is
provided as input. Then, the appliances’ energy-saving states are determined using
the application rules. The proposed research discussed an ontology-based classification
and prediction method that helps to predict energy usage in a smart home.
2.4.3. Evolutionary Algorithm. Liao [2012] utilized the WNN algorithm to forecast air-
conditioning load in smart grids. The improvement differential evolution algorithm
(IDEA) was used to improve the performance of WNN by optimizing the input pa-
rameters. The proposed IDEA-WNN method utilizes indoor and outdoor temperature,
indoor and outdoor relative humidity, solar radiation strength, air-conditioning load,
and precipitation information as input parameters. The parameter values of WNN are
represented with the chromosome coding of IDEA. IDEA applies mutation, crossover,
and selection processes to achieve the optimal parameters for the WNN algorithm.
The optimized parameters were used as input parameters of the WNN algorithm to
predict the air-conditioning load. The model was simulated with data obtained from
a 20-story commercial building. The system was trained with 7 months of data and
tested with 1 month of data. Finally, it was used to forecast the energy consumed by
the air conditioner for 2 months. The author represented the prediction accuracy using
the average mean absolute percentage error (MAPE) value. The MAPE is a measure
of accuracy computed from the actual and forecasted values. The proposed methods
achieved minimum MAPE for load forecasting of the air-conditioning system compared
to the ANN, particle swarm optimization artificial neural network (PSO-ANN), ant
colony optimization artificial neural network (ACO-ANN), genetic algorithm artificial
neural network (GA-ANN), and evolutionary programming artificial neural network
(EP-ANN).
2.4.4. Markov Process. Ardakanian et al. [2011] derived reference load models for
households using a k-state Markov process. For this purpose, the authors analyzed
the power consumption of 20 homes for 4 months. The homes were categorized into
four classes according to a classification used by a local utility that is based on the
home heating system. Using visual observation, the researchers classified a day into
three periods based on the electricity demand: on-peak, off-peak and mid-peak. There-
fore, they constructed 3 reference models (one for each period) for each class, for a total
of 12 reference models. The k x k transition rate matrix captures the state transition
probabilities. Another matrix is defined using the k-means clustering algorithm, which
represents the centroids of k clusters. The proposed load model is a continuous-time
Markov model that is defined by a tuple of these two matrices. The derived models
were validated by comparing the traces that were generated by the proposed models
with real load distributions.
2.4.5. Summary. Energy usage prediction is an important prerequisite to achieve an
optimal cost-effective scheduling of appliances. The proposed methods quantify the en-
ergy consumption per appliance and the total load of the scheduling time horizon. In
fact, the proposed algorithms are trying to indirectly predict the human activities in
smart homes using the pattern of electrical appliance usage. However, user behavior
depends on other contexts as well (e.g., location, day of the week, weather, and corre-
lation between the tasks). Hence, an effective algorithm should consider the impact of
these parameters on energy consumption.
2.5. Load Scheduling and Power Shifting

Manual appliance control requires human intervention, which is not efficient for com-
plex systems like smart homes. This section explores automated residential energy opti-
mization methods proposed for the smart grid. The methods are based on two main con-
cepts. The first technique is avoiding an appliance operation when energy is expensive.
It could be implemented by delaying an appliance operation. The second technique

may be applicable if the appliance operation cannot be delayed. In this case, the algo-
rithms try to operate an appliance at a lower energy consumption level. The following
sections review the proposed ideas on cost minimization by rescheduling the appliance
operation or operating appliances at lower power levels. The ideas are categorized by
the optimization technique used.
2.5.1. Linear Programming. Mohsenian-Rad and Leon-Garcia [2010] proposed a mixed

integer linear programming (MILP) model to develop an optimal scheduling method
for home appliances. The proposed optimization model is formulated to reduce energy
cost as well as waiting cost, which has been defined as disutility. The disutility is
modeled as a function of delay and power consumption. The delay cost depends on
the user preference, which is adjustable by a control parameter. The model considers
the energy price as a function of hourly load. This price function calculates the price
from the total power consumed by all appliances at a certain time slot. The model
implemented IBR as the energy price. In IBR, if the load exceeds a certain threshold
level, then the price is increased by a fixed value and vice versa. The model also
implemented interruptible and uninterruptible appliances. Simulation results showed
that the proposed scheduling model and price prediction algorithm significantly reduce
the energy cost and peak-to-average energy demand ratios.
Similarly, Bapat et al. [2011] proposed Yupik, a system that uses sensing, analytics,
and optimization to schedule home appliances for cost minimization. An integer linear
programing (ILP) model is used to formulate the scheduling problem. The objective
function of the proposed ILP model is a function of the energy and the inconvenience
costs. If an appliance is deferrable, its operation can be delayed, and if reducible, it
can be operated in reduced power states. The inconvenience cost is calculated from
the distance between the current time and the preferable earlier time. The model can
control uninterruptible and interruptible devices. The authors used CPLEX to solve
the ILP. They collected 4 weeks of energy usage information from a home. The data
consists of the usage pattern of a power strip, a musical system, and a TV. The power
strip was considered a virtual device because the connected device was unknown. The
authors showed that different values of arbitrary control variables provide different
savings and inconvenience costs that directly reflect scheduling of the devices.
De Angelis et al. [2013] proposed an MILP algorithm for optimal task scheduling
under dynamic electrical and thermal constraints. The model divided the task into
subtasks (e.g., washing machine cycles) and can control the order of consecutive tasks.
It can also interrupt the operations of the noisy appliances while the user is sleeping.
The authors addressed both schedulable and nonschedulable appliances. The heating
or cooling load is expressed by the law of thermodynamics, which utilizes the thermal
resistance of the household to calculate temperature variation. The objective function
optimizes the cost using both renewable sources and a storage device. The model con-
sidered the electric grid primarily as an energy consumer, not as an energy source.
If the model does not find any feasible solution, it recalculates the cost by discarding
the lowest priority load until the best solution is obtained. The simulated results were
evaluated using static and dynamic optimization approaches. The newly added load is
considered as an interrupt, and the optimizer reschedules the scheduling based on the
new scenario. The authors claimed that the overall cost was reduced to almost 50% by
using the proposed MILP optimization. However, additional cost optimization is possi-
ble by using or storing the grid electricity when the price is lower than the renewable
energy production cost. The proposed model did not consider this.
Zhu et al. [2012] proposed a demand-side management method using an MILP
technique. The appliances are arranged into nonshiftable, time-shiftable, and
power-shiftable categories according to their power requirement and operational peri-

ods. The objective function is defined to minimize the hourly load. The time-shiftable
appliances, which may have different power consumption patterns, are controlled by bi-
nary vectors. The power-shiftable appliances are limited to operate between maximum
and minimum power. Nonshiftable appliances have only one power level, which must
be maintained during the execution time slots. The proposed model had been simulated
using seven appliances from three categories in an individual household and a small
neighborhood area with multiple households to measure the optimal scheduling power
requirement. Results showed that the model is able to schedule the optimal operation
time for the time-shiftable appliances and optimal power for the power-shiftable ap-
pliances while maintaining the power consumption patterns of individual appliance
and the user preference. The authors did not consider the disutility caused by delaying
the appliance operation by the scheduler. They also did not discuss interruptible and
uninterruptible appliances. Although the model focused on hourly load optimization, it
can be used for cost optimization by slightly modifying the constraints and the objective
function.
Conejo et al. [2010] presented an optimization model to minimize energy cost using
linear programming (LP). The model proposed a linear consumer utility function that is
expressed using a monetary value. Initially, a fixed monetary value (41.5 €/MWh) was
used as the consumer utility. The objective function tries to schedule more loads in the
time slots that have lower prices than the consumer utility. A significant contribution
of this model is that it included the forecasted energy price boundaries (predicted
maximum price and other price parameters bounded by minimum and maximum price
difference) as the parameters of the objective function. The price bounds (maximum and
minimum price) are obtained using the ARIMA model proposed in Conejo et al. [2005]
with a 95% confidence level. Results showed that the inclusion of price parameters in
the objective function provides 16.22% higher utility (cost savings) than simply using a
forecasted price in the model. Unlike other models, it does not optimize the energy cost
and load for the next 24 hours in advance. It assumed that the energy supplier provides
the energy price before the beginning of every hour (10 minutes in advance), and it
computes the load for each hour based on the current price, future load, and forecasted
price parameters. From the results, it is obvious that the model is scheduling load
into lower energy price time slots. However, it is not clear how the load is delayed for
different appliance types, (e.g., interruptible vs. uninterruptible devices). The model
does not consider the user discomfort that is created by delaying the tasks.
Hubert and Grijalva [2011] modeled the heating, ventilating, and air-conditioning
(HVAC) system by the law of thermodynamics, which formulated the change in room
temperature as a function of HVAC heating rate and rate of losing the heat due to the
external weather. The mass of the air in the room and the specific heat of the air are used
to measure the heat exchange. The rate of energy change of the HVAC is expressed
as a function of temperature difference between the HVAC and the room, specific
heat, and airflow rate (assumed to be constant). The rate of energy loss due to the
external temperature was modeled by temperature difference and thermal resistance.
The thermal resistance can be calculated at first based on geometrical properties and
then be improved over time based on both room and outside temperature histories.
Constraints were applied to limit the range and the maximum variation rate of the
HVAC internal temperature. The authors proposed an optimization system that was
formulated as an MILP model. The model did not consider inconvenience cost for
delaying an appliance operation. The room temperature is maintained between the
user-defined comfort levels, which does not create inconvenience.
Marwan et al. [2012] developed a DR model to control an air-conditioning system
for cost minimization. The authors proposed an LP model to calculate the energy cost
from the air-conditioner power rating and energy prices. A constraint is enforced to
maintain the room temperature at a comfort level that is modeled using the laws of
thermodynamics. The authors investigated the effect of price spikes on the objective
function. The energy price is considered a constant all the time except during spike
periods. The price increases to a certain fixed value when it spikes. Three scenarios
had been considered to evaluate the proposed model. A system with no price spikes
did not show the precooling characteristics of the model. The system with price spikes
precooled the room when the energy was cheap. When the model was considered with
a certain probability of price spikes, it showed the same precooling properties as the
deterministic price spikes scenario. However, the higher the risk of price spikes, the
more precooling occurs. Precooling minimizes the total energy cost by reducing energy
consumption when the energy is expensive.
2.5.2. Convex Programming. Hovgaard et al. [2013] proposed a convex programming

(CP) method to minimize the total energy cost of a commercial multizone refrigerator.
The authors used a modified version of model predictive control (MPC) to formulate the
cost function. MPC is an advanced process control method that is normally used in pro-
cess industries such as chemical plants and oil refineries. The objective function tries
to minimize the energy cost and the temperature constraint violation cost. The con-
straints are convex inequalities that are related to energy usage by the compressor and
outdoor and indoor temperatures. The objective function is nonconvex and was solved
using the sequential convex programming (SCP) method. The SCP method divides
the nonconvex objective function into a sequence of approximated convex functions.
The CVXGEN [Mattingley et al. 2011] tool was used to solve these convex quadratic
functions. The proposed model showed a clear tendency to apply more cooling when
the energy price is cheap and vice versa. The authors compared the proposed model
with the thermostat control system. For this purpose, the same system was simulated
using conventional thermostat control policies. The conventional thermostat is com-
pletely unaware of the energy price, so it cannot be operated according to a DR system.
Both models were simulated using 1 year of historical data. Results showed that the
proposed model can reduce cost by 30% compared to a conventional thermostat-based
system.
Tsui and Chan [2012] studied a CP DR optimization model to optimize energy cost in
smart homes. The authors categorized household loads into four groups: schedule-based
appliances with uninterruptible load (SA-ULs), schedule-based appliances with inter-
ruptible load (SA-ILs), battery-assisted appliances (BAs), and model-based appliances
(MAs). The schedule-based appliances can be delayed. The task may be interruptible or
uninterruptible. The BAs have an internal battery that can be used as energy sources.
The MAs (e.g., heating system) can be represented by physical models. The schedulable
appliances were modeled using binary variables. The authors argued that the usage
of binary variables makes the model a convex mixed integer nonlinear programming
(MINLP) problem, which is an issue when the solution time is of great concern. To avoid
the binary variables, the authors used L1 regularization techniques to transform the
model from a convex MINLP to a standard CP problem. The binary variables related to
the constraints of SA-IL and SA-UL appliances were relaxed using real numbers and L1
regularization. The objective function uses a quadratic price model and dissatisfaction
costs to optimize the total cost of the model. The dissatisfaction function is different for
each type of load. For example, the dissatisfaction cost related to the air conditioner is a
function of temperature, whereas for a BA, it is related to the operational performance
of that appliance. The objective function is convex, and the inequality constraints are
also convex. Hence, the model was solved by a MATLAB-based convex optimization
package named CVX [CVX 2014]. The authors used an MILP model (which considers
only schedulable appliances) as a benchmark to compare the performance of the pro-

posed model. Results showed that the proposed model, most of the time, provides a
solution that usually falls within 1% deviation of the optimal solution. The authors
argued that even without rounding the relaxed real variables to the nearest binary
number, the model finds the same schedule as the binary MILP solution.
2.5.3. Dynamic Programming. Kowahl and Kuh [2010] extended the work of Livengood
and Larson [2009], which proposed a dynamic programming algorithm to optimize cost
and maximize user comfort. Dynamic programming algorithm is based on mainly six
components: stages, states, decisions, decision rules, state transition rules, and cost
function. At each stage, the algorithm executes the decisions based on the states. The
states are changed based on state transition rules. These rules consider the current
states and the decision at the previous stages. The optimal decision at a stage is
determined by the potential decisions in that stage, as well as the current and future
stage costs. The authors presented mathematical formulations to model the grid price,
indoor temperature, outdoor temperature, load, battery state, and wind speed. The total
cost is a function of energy cost and comfort costs. The comfort cost is applicable to the
cooling (or heating) device, which depends on the indoor temperature. The optimal
decision is made for every state using dynamic programming assuming a 24-hour
scheduling horizon. The model was then improved by using softmax (a reinforcement
learning algorithm) with a neighborhood update model. Instead of discrete states with
known transition probabilities, softmax provides solutions to an unknown transition
model and infinite number of possible states. Results showed that the softmax approach
performs similar and sometimes better than dynamic programming depending on the
training time.
2.5.4. Game Theory. Atzeni et al. [2013] formulated the load optimization problem us-
ing a game-theoretical approach. The authors proposed a noncooperative game where
each active player competes against the others given the energy loads, energy produc-
tion, and storage strategies to minimize energy cost. These individual game strategies
impact all users and lead to a Nash equilibrium where all users are unilaterally sat-
isfied. A distributed algorithm based on the proximal decomposition was developed
to be executed in the smart meters, which allows computing of optimal strategies.
The authors presented a simulation study utilizing practical cost functions. Results
showed that resulting demand curve from the optimization had flattened the energy
consumption, which eventually reduced the energy cost.
Mohsenian-Rad et al. [2010] presented a game-theoretic model for load scheduling in
smart homes. The proposed architecture requires connectivity among the smart meters
of the households to share hourly load information. Each smart meter is equipped
with an automatic energy consumption scheduler (ECS) that interacts automatically
with other smart meters by running a distributed algorithm. The proposed distributed
algorithm is formulated as an energy consumption game among users, in which each
user tries to maximize his payoff by proposing a load schedule. The payoffs function
ensures Nash equilibrium (i.e., all users pay their own minimum cost to the utility).
The cost function used in this model is convex, and therefore the proposed model is
convex. It was solved by using the IPM algorithm [Boyd and Vandenberghe 2004].
Results show that the proposed model can minimize the energy cost and reduce the
peak to average ratio (PAR). However, it did not consider user discomfort created by
delaying an appliance. Hence, it is not able to impose user delay preferences specific to
an appliance.
2.5.5. Reinforcement Learning. O’Neill et al. [2010] proposed a novel algorithm—the
consumer automated energy management system (CAES) to optimize electricity usage
for a DR system. The model was developed for a single residence. The system state at a
certain time is represented using pending energy backlog, average pending workload,
user reservation vector, and pricing sequence. For a given energy control policy and
system state, CAES calculates the cost from the energy cost and the disutility cost.
A control variable is used to maintain the ratio between the monetary cost and the
disutility cost. The performance of the system is calculated from the total discounted
period cost over an infinite horizon. The model was solved by a Q learning algorithm
to minimize cost. The numerical simulation showed that CAES reduces energy cost by
16% to 40% with respect to a system that does not have any energy price information.
2.5.6. Breadth First Search. Georgievski et al. [2012] proposed a system to monitor and
control electrical appliances in a building to save energy costs. The model defined the
appliance energy requirement constraints by grouping those under five policies: repeat,
total, multiple, strict, and pattern policy. The repeat policy defines the constraints for
an appliance that runs periodically. It depends on the specific time. The total policy
limits the total energy consumption of an appliance regardless of the specific time.
If a device is required multiple times but does not follow any periodic cycles, then it
is defined under the multiple policy. Nonschedulable urgent tasks are considered in
the strict policy. The pattern policy defines the power consumption levels of a device
according to different internal states. The sleep policy can be enforced on any devices
when its operation is not required. The system collects the device’s usage pattern,
renewable energy generation quantity, and grid energy prices from different utilities
using a RESTful architecture. The gathered information is processed by a breadth
first search optimization algorithm to achieve the optimal cost while maintaining the
enforced policies. The system performance was evaluated for 4 weeks in an office at
the University of Groningen. Results showed that the proposed model can save up
to 50% of cost and 15% of energy. The authors evaluated the inconvenience created
by the system by performing a survey on user experiences. The majority of the users
were unaware of the experiment. The survey report showed that it created negligible
dissatisfaction to the occupants. The proposed model did not consider the preference of
the user. It considered only common office appliances (e.g., laptops, computers, printers)
for optimization. It did not include the lighting, cooling, or heating systems, which are
significant components for cost savings and might affect visual and thermal comfort.
2.5.7. Summary. Load scheduling and power shifting are the most efficient ways to
reduce energy costs in smart homes. The introduction of dynamic energy prices requires
frequent and effective human involvement to control the household load, which is not
convenient and unlikely to be widely adopted. Failure to ensure the participation of
users in load control will ultimately yield an inefficient system. The methods presented
in this section provide automated control of the appliances based on energy price and
user comfort.
Based on user preference, the appliances are primarily divided into two types:
shiftable and nonshiftable. The nonshiftable appliances cannot be delayed but may
be operated at reduced power. Rescheduling is applicable to the shiftable devices that
can be delayed to relatively lower energy cost time periods. The delay preference for this
type of appliances depends on user comfort. The shiftable appliances can also be oper-
ated at reduced power. User preference related to specific appliances is generally main-
tained by the constraints. It is also possible to maintain a preference matrix containing
a rating for each appliance. A user can also prioritize the tasks to ensure early execu-
tion of important jobs. Graditi et al. [2015] proposed six control logics to attain user-
defined effects. The logics are comfort (maximize comfort), economy (minimize cost),
emergency (in case of grid failure), energy (control energy consumption), power (en-
force a threshold power level), and thermal storage (store energy, e.g., air-conditioning
system). Siano et. al. [2013] presented a comparison between the comfort and economy
mode of operation.
Based on the power consumption profile, an appliance can be interruptible or unin-
terruptible. An interruptible device can be turned off in the middle of its operation and
resumed later. An uninterruptible appliance must be allowed to continue its operation
until the task is finished. Both types of appliances may require different levels of power
during the task execution based on internal operational states. These appliance-specific
properties are generally enforced by constraints.
Load scheduling and power shifting create discomfort to the user by delaying the task
or reducing the task quality (by reducing power). A few methods considered this disu-
tility [Mohsenian-Rad and Leon-Garcia 2010; Bapat et al. 2011]. Some research consid-
ered disutility created by temperature variation [Kowahl and Kuh 2010]. Performing a
survey on participating users is a method to measure this disutility [Georgievski et al.
2012]. The survey is more appropriate than other hypothetical assumptions because
it is based on human evaluation. These survey results are available at the end of the
experiments. The proposed research did not consider how these survey results can be
included in the system so that the model will reflect the disutility of the user. Another
method to measure disutility is to apply a penalty whenever an appliance operation is
delayed or operated at reduced power [Mohsenian-Rad and Leon-Garcia 2010; Bapat
et al. 2011]. The models introduce a disutility cost in addition to the energy cost. The
efficiency of this type of model lies on the effective calculation of the disutility cost. The
energy cost is a monetary value. Therefore, the disutility cost should be an equivalent
monetary value. In economic theory, disutility is expressed using an approximation
function. In the proposed models, the disutility is expressed as a function of power or
delay or both. However, disutility varies according to users, appliances, and situations.
To adjust this variation, different user-defined scaling parameters are used for different
users and appliances [Mohsenian-Rad and Leon-Garcia 2010]. The previous research
rarely considered situations when no feasible solution of the problem is achievable. De
Angelis et al. [2013] proposed a solution of the problem by discarding lower-priority
loads.
2.6. Energy Storage Scheduling and Power Optimization

Energy storage acts like a load while charging and like a source while discharging.
Storage scheduling has a dynamic impact on cost optimization. This section discusses
proposed optimization approaches that explicitly include energy storage devices in
their formulation.
2.6.1. Linear Programming. De Angelis et al. [2013] modeled energy storage using linear
constraints and proposed an MILP model to optimize energy cost. The current storage
energy is expressed as a function of the previous stored energy level, storage efficiency,
and charging and discharging rates. Several constraints are used to limit the storage
capacity, charging rate, and discharging rate. The model considered EVs as energy
storage. The authors presented solutions using both static and dynamic optimization
approaches. Dynamic optimization re-estimates the optimized cost based on newly
added or removed storage (e.g., EV battery). Simulated results show that this can
dynamically update the cost parameter according to the new storage capacity.
Zhang et al. [2014] utilized a central electrical storage and a thermal storage for a
smart building. The electrical storage is the sum of the storage capacities from each
home. Each home can charge or discharge the storage. However, to use the stored
energy, the respective home should charge it first. Therefore, each home is assumed to
have its own flexible substorage, and the total storage capacity of the whole building is
a predefined constant. Similarly, the central thermal storage had been considered as
accumulated subthermal storage from each household. The authors proposed an MILP
model for the smart building to optimize energy cost. The electrical storage cannot be
charged and discharged simultaneously for the same home. This condition is also true
for the thermal storage. The proposed model considered the storage maintenance cost
as a cost overhead that is shared by all homes. The project used a 4kWh electrical
storage with 95% efficiency, and the maintenance cost was 0.5p/kWh (pence per kWh).
The thermal storage capacity was 6kWh with 98% efficiency, and the maintenance cost
was 0.1p/kWh (pence per kWh). The model was implemented using CPLEX 12.4.0.1 in
GAMS 23.9. Results showed that the proposed cost optimization model can save 24%
to 30% energy cost with cost fairness among the homes.
Hopkins et al. [2012] used energy storage to optimize energy cost using an LP model.
The proposed cost function considers energy selling, buying, and storage loss. A full
storage charge and a full storage discharge have been defined as the storage cycle, and
the life span of the storage is represented by the number of cycles. The cycle cost is a
function of the total cost of the storage and its life span in cycles. The inclusion of stor-
age cycles in the cost function not only optimizes the energy cost but also maximizes the
battery life span. The proposed model had been implemented using the linprog func-
tion of MATLAB. The simulation considered eight T105 deep-cycle batteries. Results
showed that the surplus energy generated at midday is stored in the battery to be used
later when it is more valuable instead of selling it back to the grid. The average daily
savings with renewables and the storage was $1.81; with only the batteries (without
renewables), the average daily cost savings was $0.26.
2.6.2. Genetic Algorithms. Arabali et al. [2013] proposed a method to select the opti-
mal storage capacity for the smart grid. The model also optimizes the usage of energy
storage according to load and energy sources. The cost function uses the storage ca-
pacity with other energy generators to optimize the installation cost. The total capital
cost of the storage system is calculated from four components: energy-related stor-
age cost, power-related storage cost, energy-related balance-of-system (BOS) cost, and
power-related BOS cost [Mosher 2010]. Storage charging is formulated as a function
of hourly self-discharge rate and round-trip efficiency (the ratio between energy recov-
ered and energy input) of the storage. Storage discharging is a function of the hourly
self-discharge rate of the storage. The model used a constraint to maintain the energy
level of the storage between a given maximum and minimum capacity. A GA-based
optimization method is used to estimate the storage capacity considering the load and
the energy generated by the renewable sources. Results showed that if wind energy is
available, the system requires less storage capacity compared to using only PVs. This is
because the wind energy is generated almost every hour of a day. It is also shown that
increasing the load shifting rate (i.e., the percentage of appliances that are shiftable)
decreases the required storage capacity of the system. In this case, energy losses due
to storage efficiency and self-discharging is minimized.
2.6.3. Game Theory. Vytelingum et al. [2010b] proposed a game-theoretic framework
to analyze the effect of household storage on the energy price. The authors considered
all users in the existing market as agents who try to maximize their own profit. The
proposed method tries to determine the Nash equilibrium, which leads to optimum
costs for each agent considering that the agents behave rationally and have complete
information of the market. Initially, each agent computes his own storage strategy that
minimizes its cost. Then, it gradually adapts its storage strategy following the trends of
the market price. The authors presented an empirical study on the UK energy market.
Results showed that the user can save 13% more compared to the current system with
no storage. The savings settled at an equilibrium point when 38% of users adopted
storage. At the equilibrium, the average saving of all users was 8.5%. Interestingly, the
average saving reduces when more than 38% of users adopted storage devices in their
homes. This is because beyond this equilibrium point, additional storage adds more
volatility to already flattened market prices. In this case, a user can save more by not
having any storage.
2.6.4. Summary. There are two types of energy storages: electrical storage and thermal
storage. The space heating and hot water storage of the household can be considered
thermal storages. The EV battery is considered electrical storage when it is plugged
in. The system loses energy during the energy-storing process (e.g., charging, heating).
Each type of storage shows a gradual spontaneous energy loss (e.g., self-discharging,
temperature loss due to external environment).
Energy storage has great potential to flatten the energy demand curve. In practical
scenarios, the energy consumption rate is not the same at different hours of a day. In
off-peak hours, energy demand is low, and in on-peak hours, energy demand is high.
The utility has to ensure the production capability to meet the maximum demand. In
the smart grid, the utility tries to motivate the user to store energy when the demand
is low so that during on-peak hours the user can utilize the storage energy instead
of drawing energy from the grid. As a result, the utility can reduce the initial capital
cost. However, the impact of an increased utilization of energy storage may create a
new trend in energy demand. In this case, an off-peak hour may become an on-peak
hour. In addition, energy price may then be estimated based on the wrong demand
information and may not work as expected. More research is required to analyze the
impact of storage on energy demand and on efficient ways to utilize the storage for cost
optimization.
Energy storage may have an unintended impact on the smart grid. It is evident
that an unplanned installation of storage at the user side may create a new critical
peak energy demand. Sometimes, instead of saving energy cost, storage systems may
increase energy cost [Vytelingum et al. 2010b]. In contrast, some research has reported
that distributed storage may be more beneficial than distributed generators [Poonpun
and Jewell 2008]. Concerned authorities should carefully define the policies related to
storage installation in the smart grid.
2.7. Energy Optimization Using Renewables and Power Generation Devices
This section discusses the impact of renewable energy sources and power-generating
devices or technologies (combined heat and power (CHP) generators, boilers, etc.) on
cost optimization.
2.7.1. Genetic Algorithms. Arabali et al. [2013] proposed a probabilistic GA-based op-
timization using the two-point estimation (2PE) method to minimize PV and wind
generation installation cost and increase energy usage efficiency. The authors used
an HVAC system as the reference load. The 2PE-based analytical method is used to
model the stochastic behavior of wind power output, solar power output, and load
power consumption. Before applying 2PE, 10 years of historical hourly load data, solar
irradiance, and wind speed were clustered into 10 groups using fuzzy C-means (FCM)
clustering. Each cluster is composed of days with similar 24-hour load data, solar ir-
radiance, and wind speed. The model used the probability density functions (PDFs)
of wind and solar power generation to normalize the parameters of 2PE. A weighted
sum of the installation cost of all clusters is used as the fitness function of the GA. The
evolutionary algorithm was repeated until the best chromosome had been identified
as the optimum solution. The proposed GA algorithm started with a population of 200
and continued for 150 generations. The method used 78% crossover and 20% mutation
rate. The scenarios were presented with 10%, 20% 30%, 40%, and 50% shifting of the
deferrable load. Results showed that increased load shifting provides more flexibility
and leads to less excess energy generation. The user can use the proposed model to
estimate the optimized installation cost of PV, wind turbine, and energy storage. Load
shifting may introduce inconvenience to the user by setting the room temperature to
an uncomfortable level. The model did not consider this inconvenience to the user.
Bilil et al. [2014] proposed a multiobjective optimization model to optimize annu-
alized cost when diverse distributed generations (DGs) are used to generate energy.
The model considered the annualized cost of the DGs as a function of the capital cost,
maintenance cost, and replacement cost. The objective function also minimizes the
renewable energy–load disparity (RELD) to balance the generation and loads. The pro-
posed model was solved using a GA. It had been observed that using both solar panels
and wind turbines is obviously better than using the wind turbines alone. It was also
reported that wind generators are more cost effective than Vs.
2.7.2. Linear Programming. De Angelis et al. [2013] addressed the significance of renew-
able energy sources for cost optimization in smart grids. The proposed model limits grid
energy usage to a certain maximum level. The energy demand was primarily serviced
from renewable energy sources. The authors used an MILP model to calculate the op-
timum energy generation capacity of the renewables considering the given load—that
is, it provides an estimate of the capacity of the renewable energy generation devices
that is required to be installed at the user’s premises for optimal cost savings. It did
not consider the dynamic nature of renewable energy sources. The storage capacity in
a household has a potential impact on the cost. In this model, if more storage capacity
is added, the user can minimize cost by selling surplus energy to the grid. However, it
requires an additional capital cost. A model can be developed to address the trade-off
between capital cost and daily or monthly profit.
Zhang et al. [2014] considered multiple homes in a smart building that share common
distributed energy resources (DERs) such as CHP generators and boilers. The building
has a CHP generator with a capacity of 4kWe and 35% electrical efficiency. It is operated
by natural gas at a cost of 2.7p/kWh. The building also has a 24kWth boiler operated
by natural gas. The proposed system is an MILP model that was implemented using
CPLEX 12.4.0.1 in GAMS 23.9. Their results showed that the CHP was providing
constant maximum output of 4kW most of the time, except at night when heat demand
is low. The peak demand had been shifted to the nighttime from evening, and the
peak demand had been reduced by 7kW to 32kW. The results also showed that the
proposed cost optimization model saves 24% to 30% of the energy cost by rescheduling
the sources and the loads with fairly distributed cost among the homes.
Hopkins et al. [2012] proposed an LP model to analyze the impact of DGs on cost
optimization. The cost function considers the energy buying cost from the utility, energy
selling cost to the utility, and storage loss that is calculated from the storage life
span. The renewable sources have been modeled using an energy balance constraint
ensuring that the total generated energy is the same as the total consumed energy. The
simulation considered a 2.24kW solar system, eight T105 deep-cycle batteries, and an
Outbreak GTFX2524 inverter. Results showed that with solar panels (without using
the storage), the proposed solution can save an average of $1.84 per day.
2.7.3. Summary. There are mainly three research questions related to renewables and
energy generation devices. First, a user needs to determine the capacity of the energy
sources in advance to be installed in smart homes. The energy price of a renewable
source is calculated from the capital cost and the life span of the device. Other energy
sources require fuel costs that contribute to the energy prices. The user has to strike a
balance where further capital investment does not pay off. If the user decides to become
an energy seller, the energy demand and distribution capacity limit the production
rate of an individual seller. The researcher should be more concerned about the capital
cost optimization in the smart grid. Second, given that a household is equipped with
the energy sources, the user needs an estimate of the probable energy generation in
advance from the renewables. The prediction algorithms are utilized to forecast energy
generation of the renewables. If the user optimizes the forecasted energy generation
from the renewables with the estimated load profile, it will provide an estimate of how
much energy from other sources (e.g., grid, energy generation devices (if any)) will
be required. Finally, a combination of different energy sources may result in a more
predictable energy generation [Bilil et al. 2014]. Current research trends have not yet
identified the most effective combination of different energy sources.
2.8. Energy Trading

Trading energy in the energy market is an effective way to minimize energy cost. It
is assumed that users know their own energy demand and expected generation in
advance. The energy price and usage preference motivate the user to earn a profit by
selling energy in the open energy market. This section discusses projects on effective
energy trading methods in smart grids.
2.8.1. Particle Swarm Optimization. Ramachandran et al. [2011] proposed a risk-based
(RB) auction strategy for profit maximization while selling energy to the energy market.
The auctioneer agent interacts with a load agent and a DER agent to obtain the bid
and ask prices, respectively. The risk attitude of the trader (buyer and seller) has
been classified into three categories: risk neutral, risk averse, and risk seeking. A
risk-seeking policy means that the trader is looking for high profit, which has a lower
probability of transaction. A risk-averse attitude means that the trader will accept a low
profit margin but the probability of transaction is high. A risk-neutral strategy seeks
a price that maximizes its expected profit. This attitude is represented by a variable
called the risk factor. The RB auction strategy starts trading by using risk-neutral
policy and then gradually updates the risk factor based on the current competitive
equilibrium (CE) price until the bid/ask improvement over the outstanding bid/ask
is less than a threshold level. In the experiment, it was found that an agent trades
very carefully during the first round because its knowledge was limited. After the first
round, the risk factor converges rapidly. The authors used grid energy prices from
the Amsterdam Power Exchange (ApX) [ApX 2014] and assumed energy prices of the
renewable energy sources. Results showed that with the aid of a PSO-based algorithm,
the energy cost of the utility is reduced by 37%. The users share the benefit of this
reduced cost.
Wang and Wang [2012] developed a PSO-based negotiating agent to facilitate en-
ergy trading in the smart grid. The utility grid and the smart building are considered
traders, which can be buyers and sellers. The PSO optimizer indirectly predicts the
opponents’ preference based on its previous prices. The proposed optimizer uses a time
pressure function, which is a function of eagerness to determine the attitude of the
agent. Higher eagerness means that the agent is willing to complete the transaction
by the deadline at any price, whereas lower eagerness means that the agent is moti-
vated to risk losing the deal for the chance of a better price. Eagerness is a combined
effect of long-term eagerness and short-term eagerness. The short-term attitude of the
eagerness function for both the smart building and the utility depends on the previous
offers of the opponents. The long-term eagerness function of the smart building as a
seller depends on the number of recent transactions and the state of charge (SOC)
of the battery. When the smart building is a buyer, the long-term eagerness function
depends on interruptible loads and comfort levels. Long-term eagerness of the utility
is a function of the number of successful recent transactions. Smulated results showed
that the PSO-based agent in the smart building (buyer) can learn the pattern of the
utility (seller) and can mimic the opponent’s behavior. The agent saved 17% cost and
reduced negotiation time by 9% compared to an agent without learning capability.
2.8.2. Auction. Vytelingum et al. [2010a] developed a trading agent based on the
continuous double auction (CDA) strategy. The CDA allows the buyer and seller to con-
tinuously change the energy price during the trading time in a continuously clearing
market. The buyer submits two types of orders. One is the inelastic limit order, which
represents the minimum required energy in the household. The other is the elastic
market order, which represents optional household demand depending on the market
price. The sellers’ orders are elastic price-sensitive market orders. The limit order bids
offered by the buyers cleared immediately (if that amount of energy is available in the
market) by the market because it is an inelastic demand and does not depend on the en-
ergy price. When the orderbook (where all bids and asks are recorded) is changed due to
a new or improved bid or ask, a market clearing algorithm searches for a possible match
where a buying price is more than or equal to a selling price. If a best match is found, a
transaction occurs between the buyer and the seller and the market clears the matched
ask and bid. The auction occurs a day ahead of the actual consumption. The authors
proposed an online balancing policy when the real-time demand exceeds beyond the
traded market amount or supply drops below it. The proposed mechanism uses the
unmatched orders of the orderbook records to rematch the new demand and reduced
supply in real time. The authors also considered the cost of using transmission lines
(resembling the transportation cost of traditional goods), transmission line constraints,
and secure transaction policies. The lower-bound efficiency of the market was evaluated
by a zero intelligence (ZI) [Gode and Sunder 1993] strategy that uses random bid and
ask prices. For this baseline strategy, the market efficiency was 88% to 96%. When the
same scenario was simulated with an improved version of an adaptive aggressive (AA)
[Vytelingum et al. 2008] strategy, it showed 92% to 99% market efficiency. The authors
assumed that the untraded energy would remain on the market to be used at the actual
consumption time, which may not be true all the time. The seller may decide to increase
energy consumption or reduce energy generation because of lower demand. In this case,
balancing the real-time demand with the day-ahead outdated information may not be
feasible.
Ilic el al. [2012] described an energy market, named NOBEL [NOBEL 2013], to
evaluate market-driven DR of electricity trading. The market facilitates electricity
trading in a local network to avoid (or reduce) transportation cost and energy loss.
The model utilized a stock exchange model where energy trade time is defined as
discrete time slots. The orders are maintained in an orderbook, and a matching al-
gorithm searches for the best match to clear the transaction. Old orders get higher
priority over new orders. The orders may have different configurations depending
on the user preference. An order may have to be matched partially or fully. A fully
matched order should be matched exactly; otherwise, it can wait in the orderbook to
be matched later if more orders arrive. The matching algorithm may not let a wait-
ing full match order to block trading. An order may have to be matched immediately,
which will be canceled automatically from the market if the matching is unsuccessful.
The market was implemented in a Java-based local simulator as well as in an online
application server. The trading agent applied the ZI [Gode and Sunder 1993] strat-
egy to bid to the market, which means that it has no capability to learn the market
trends from the previous history. It randomly bids following a maximum sleep time
to facilitate a high-level order matching. The simulator used household load profiles
and a renewable energy source with random weather effects as input parameters. The
results showed that market efficiency drops when the supply meets the demand and
becomes stable if supply exceeds demand, but it never drops below 70%. The minimum
matching rate was 75%, which means that 75% of the total orders were traded in the
market.
Recently, a transactive energy framework was proposed to standardize the archi-
tecture of transactive control [GridWise 2015]. The term transactive energy refers to
techniques for managing the generation, consumption, or flow of electric power system
through the use of economic- or market-based constructs while considering grid relia-
bility constraints [GridWise 2015]. The projects related to transactive energy mainly
considered the underlying architecture to control the energy flow in the smart grid
[Hammerstrom et al. 2007]. The framework use auction mechanisms for energy trad-
ing that are similar to the ones discussed in this section [Huberman and Clearwater
1995].
2.8.3. Game Theory. Capodieci et al. [2011] developed a multiagent system to eval-
uate the energy trading strategy of households. The authors categorized agents into
two types: main agents and auxiliary agents. Main agents are energy producers and
consumers (e.g., prosumers (buy and sell energy), consumers (buy energy), and Gencos
(traditional energy-generating companies)). Auxiliary agents provide information and
mediation to support the behavior of the main agents. The proposed energy trading
strategy uses an auction system. All main agents participate at every round in the
auction. The authors proposed a learning algorithm based on game theory that pre-
dicts the energy price using previous experiences. The authors reported that a learning
algorithm improves the performance of the system.
2.8.4. Summary. From the preceding methodologies, it is obvious that different types
of auctions are the most popular strategies for energy trading in the energy market.
Auction strategies could be improved by using algorithms that predicts the opponents’
behavior. These algorithms use functions that adaptively determine the risk attitude
of users from the previous trading history. The risk attitude of users could be classified
into risk seeking (high profit but higher risk of untraded energy), risk averse (low profit
and lower risk of untraded energy), and risk neutral. The performance of the intelligent
algorithms is compared to the ZI auction strategy, which is based on random bid and
ask prices.
Market orders are classified into two categories based on elasticity: inelastic and
elastic. Inelastic energy demand must be cleared immediately and does not depend on
energy prices. Elastic demands are optional energy requirements of users and could be
delayed or discarded based on market prices. A order could be a fully matched order or
a partially matched order depending on users’ preferences. A full match order must be
matched exactly. A partial match order may be matched with any amount of ask or bid
quantity. A matching algorithm clears transactions based on the matching criteria.
Energy trading takes place in advance before the actual generation. Energy genera-
tion on the user side mostly depends on renewable sources that fluctuate with weather
conditions. Sometimes it may not be possible for the seller to supply the traded amount
of electricity to the buyer. There may also be some issues related to the quality of the
supplied energy in the market. Energy quality refers to an uninterrupted electricity
supply with regulated voltage. There should not be any service disruption, and the
voltage level should maintain the standard level throughout the distribution time. It
is not clear how the microgrid will ensure the QoS between buyers and sellers and how
the current market will deal with any disputes occurring between two trading parties.
3. A UNIFIED OPTIMIZATION FRAMEWORK
This section proposes a unified energy cost optimization framework for smart homes
in a smart grid. Although the cost minimization is not explicitly mentioned in some
works (e.g., the papers on prediction), the outcomes of these research works are integral
Fig. 3. A unified optimization framework including the grid, loads, renewables, storage devices, and
microgrid.
parts of the cost optimization methods. The optimization models rely on the prediction
methods for planning an optimal load and source scheduling strategy that minimizes
the energy cost and maximizes user satisfaction. Hence, the unified framework is a cost
optimization model that strongly depends on the accuracy of the prediction methods
for better results.
To date, most previous research works only considered a partial aspect of the cost
optimization problem. As a result, they failed to analyze the scenarios when inter-
connected components and their properties have to be considered simultaneously. The
proposed framework integrates these partial models into a unified cost optimization
model. It combines load, source, and storage scheduling and energy trading capabilities
of smart homes into a single cost optimization model where each household collabora-
tively determines the energy price for microgrid trading.
Figure 3 describes the input and output parameters of the proposed unified cost
optimization framework. The grid is considered to have an infinite energy supply. The
grid energy prices are either predetermined or predicted. Load profiles include the en-
ergy consumption information of the loads (e.g., predicted power, duration, and time of
operation). It also specifies the characteristics (e.g., interruptibility) of the household
loads. In addition to that, the energy consumption profile depends on user preferences.
The framework uses a forecasted renewable energy generation profile to optimize en-
ergy cost. The storage devices are modeled using the constraints related to energy
conversion efficiency and energy loss. The final outcome of this unified framework is
the minimum energy cost given the predefined or predicted input parameters as well
as the optimal load and source schedule to achieve this minimum cost. Furthermore,
Table I. Nomenclature
Notation Representation Notation Representation
t ∈T Set of time slots xt Observed data
s∈S Set of energy sources zm Correlated m-th parameter
l ∈L Set of loads Dl,k Penalty function related to a load
k ∈K Set of households Ck Total cost of a household
Ps,t,k Source energy price αl,k Earliest start time of a load operation
Es,t,k Energy drawn from a source βl,k End time of a load operation
El,t,k Energy consumption of a load min
Es,t,k Minimum energy limit of a source
total
El,k Total energy consumption of a load max
Es,t,k Maximum energy limit of a source
the microgrid energy price can be another outcome of the framework, which has a
significant impact on cost reduction and fair cost distribution among the households.
Beaudin and Zareipour [2015] presented a comparison of the home energy manage-
ment system based on cost reduction. The study reported that the average cost savings
of the households is 23.1%. We do not agree with the authors’ approach of comparing
different optimization models. There are two reasons for this. First, the optimization
methods can be composed of different constraints. Sometimes a model is simply a sub-
set of another model. A method may show more cost reduction because it is simpler
but less realistic than the others. Therefore, the presented quantitative comparison is
not an appropriate parameter for ranking the methods. Second, the compared meth-
ods used a different set of data, making the results incomparable. We believe that
quantitative comparison among the proposed methods should be performed against
a unified model using the same data. Hence, we proposed a unified framework as a
benchmarking tool. Ranking the cost optimization models based on cost reduction is
an open research problem and is rarely identified in the literature.
3.1. Prediction Methods

The uncertainty of future energy prices, energy generation, and consumption quantities
limits the efficiency of the optimization framework. Although most utilities currently
use TOU prices for residential users, there is a possibility that dynamic energy prices
(e.g., RTP) will be widely available in the future to take the benefit of the DR program.
The probable quantities of renewable energy (e.g., PV, wind) and household energy
consumption must be available in advance to be used for resource planning.
Prediction algorithms generally depend on three types of input parameters. The
first type of algorithm uses past observed data to forecast the future. For example,
a method can predict the future solar irradiance based on past observed irradiance
information. The second type of forecasting algorithm solely depends on current input
parameters that are correlated to the output. For instance, wind power generation
capacity is strongly correlated with wind speed. Therefore, this approach uses wind
speed to forecast the output wind power. Finally, the third type combines these two ap-
proaches (i.e., forecasted values are determined based on correlated input parameters
and past observed data). Therefore, the prediction methods can be modeled using the
following general formulation. Table I provides a list of notations used for prediction
and optimization methods.
If t is the current time, xt−1 , xt−2 , . . . . . . , xt−n represent the past observed data from
t − n to t − 1 time and z1 , z2 , . . . . , zm represent the m correlated parameters, then the
general formulation for the predicted values for t to t + u future time can be expressed
as
xt+u = f˜(xt−1 , xt−2 , . . . , xt−n, z1 , z2 , . . . , zm). (1)
Table II. Prediction Methods Applied for Energy Price, Generation, and Consumption Forecast
Methods Input Parameters Purpose
MLP Solar irradiance and air temperature To forecast solar irradiance for
[Mellit and Pavan 2010] 24 hours [Mellit and Pavan 2010]
WNN Past observed wind power [Chitsaz To predict wind power for 6 hours
et al. 2015] [Chitsaz et al. 2015]
Indoor and outdoor temperature, To forecast energy consumed by the
indoor and outdoor relative humidity, air conditioner for 2 months [Liao
solar radiation strength, 2012]
air-conditioning load, and
precipitation information [Liao 2012]
SVM Past observed PV power [Shi et al. To forecast PV power based on cloudy,
2011] foggy, sunny, and rainy weather [Shi
et al. 2011]
SOM, SVR, and Past observed solar irradiance [Dong To predict solar irradiance [Dong et al.
PSO et al. 2015] 2015]
Semantic Web Power states (e.g.,unplugged, To predict energy usage [Cheong et al.
power-off, power-save, power-on) of 2011]
appliances and residents’ location
[Cheong et al. 2011]
Markov process Power consumption of households To forecast total load of a household
[Ardakanian et al. 2011] [Ardakanian et al. 2011]
Statistical Energy prices of yesterday, the day To predict 24-hour energy price
correlation before yesterday, and the same day of [Mohsenian-Rad and Leon-Garcia
the past week [Mohsenian-Rad and 2010]
Leon-Garcia 2010]
Convex Historical hourly prices [Hovgaard To predict energy prices [Hovgaard
optimization et al. 2013] et al. 2013]
Proposed energy price by the utility To forecast energy price that
and the user [Samadi et al. 2010] maximizes aggregated utility and
welfare [Samadi et al. 2010]
SOA Day-ahead prices advertised by the Retail RTP prices [Pagani and Aiello
wholesale market [Pagani and Aiello 2015]
2015]
WEKA Device usage pattern To predict the device usage profile
[Bapat et al. 2011]
Here, f̃ is an approximate forecast function to predict the information for the next
t + u times. Table II provides a list of the prediction methods, input parameters, and
purpose of the proposed methods.
3.2. Optimization Methods
Optimization methods use the forecasted or preadvertised energy price, predicted con-
sumption, and generation information to plan the schedule of energy sources and loads
for a certain time frame. The source of energy can be the grid, renewables, storage
devices, and the microgrid. The energy is consumed by the loads, storage devices, and
the microgrid.
Unlike other components, the storage and the microgrid can act as loads and energy
sources as shown in Figures 4 and 5. The energy storage has a close relationship with
all energy sources and loads. The storage can store the surplus energy generated by
renewables to use it when there is a shortage. Furthermore, it can buy and store energy
when the energy price is inexpensive and sell or use it when energy is expensive. The
microgrid shows characteristics similar to the storage. The optimizer needs to decide
the time when it will buy or sell energy. If it decides to sell energy, it has to determine the
energy quantity and price that will maximize profit. If the user is buying energy from
the microgrid at a certain time, the optimizer should reschedule the loads according
Fig. 4. Dual characteristics of energy storage.
Fig. 5. Dual characteristics of a microgrid.
to it. These reciprocal characteristics of the storage devices and microgrid impact the
computational complexity of the optimization framework.
The space heating and cooling system, water heating and cooling system, and electri-
cal storage devices (e.g., EV battery) can be considered energy storage. These storage
devices can preserve energy for future use. A thermal storage can be preheated to the
maximum level (or precooled to a minimum level) to avoid electrical energy consump-
tion when the energy price is high. Similarly, an electrical storage can store energy to
utilize it when energy is costly.
Suppose that the total cost of the first household is denoted by C1 , the total cost of
the second household is denoted by C2 , and so on. In general, the total cost of the k-th
household is

Ck = Ps,t,k Es,t,k + Dl,k. (2)
s∈S,t∈T l∈L
Here, Ps,t,k and Es,t,k are the energy price and energy quantity drawn from source s at
time slot t in k households, respectively. Dl,k is a penalty function related to load l in k
households, which imposes a cost due to delay in operation and/or reduced performance.
The objective is to minimize all cost functions as follows:
min(C1 , C2 , . . . , Ck). (3)
Table III. Constraints for Optimization

Category Constraints
User preferences To maintain the start time, end time, and duration of a load operation.
To specify whether a load is shiftable or nonshiftable, because for some loads,
users do not prefer delaying the operation.
Load To specify whether the appliance operation can be interrupted.
characteristics To maintain the duration of operation.
To maintain energy consumption of the load. A load may have idle or sleep
mode. In sleep mode it consumes energy.
Storage property For the space cooling and heating system, to express the relationship
between the external weather temperature, household window status, and
availability of sunlight in the household.
For the water heating system, to express the relationship with the external
temperature.
For EV, to express the relationship with driving time.
To formulate energy conversion loss during the energy storing process (e.g.,
charging, heating).
To formulate gradual spontaneous energy loss (e.g., self-discharging,
temperature loss due to the external environment).
Sources To ensure that the total energy drawn from the energy sources do not exceed
the available energy.
Objective (3) is a multiobjective optimization problem. Table III presents a list of

applicable constraints for cost optimization. These constraints can be formulated using
the following general constraints related to energy sources and loads.
Energy balance constraints. For a specific household, the total energy consumed by
all loads in a specific time slot should be the same as the total energy supplied by all
sources in that time slot. This relation is expressed in (4):

El,t,k = Es,t,k, (t ∈ T , k ∈ K). (4)
l∈L s∈S
Load energy consumption constraints. A specific load needs to be operated for a

certain duration and consume a certain amount of energy:
βl,k

El,t,k = El,k
total
, (l ∈ L, k ∈ K). (5)
t=αl,k
Source energy limit constraints. An energy source may have a specific amount of
energy in a certain time slot, and the energy drawn from the source should be equal or
less than it:
min
Es,t,k <= Es,t,k <= Es,t,k
max
, (s ∈ S, t ∈ T , k ∈ K). (6)
A storage must have sufficient energy to act as a source. It should not exceed the
maximum storage capacity and should not go below the minimum energy level. A
renewable energy source and microgrid have a limited amount of energy in a certain
max
time slot. The grid is typically considered to have infinite energy (Es,t,k = ∞). If energy
max
is not available from a source at a specific time, it can be considered 0 (Es,t,k = 0 ). In
general, the minimum limit of energy is 0 (Es,k = 0 ) for most energy sources.
min
3.3. Energy Trading

The unified framework determines the optimal quantity of energy shortage and surplus
in a specific time slot for each household. If the energy price in the microgrid is less
than the grid energy price, the optimization methods initially try to fulfill the energy
demand by trading energy in the microgrid. Households buy energy from the grid if
the microgrid does not have sufficient energy to satisfy its energy demand.
The presence of energy storage makes the energy demand and supply dynamic. The
energy shortage and demand are not constant for a specific time; rather, they can be
varied between a minimum and maximum value because the storage can preserve the
energy to use it later. Therefore, the storage devices create a complex demand and
supply scenario that has direct impact on cost optimization.
In general, different types of auctions are the most common strategies for energy
trading in a microgrid. The efficiency of the auction algorithm can be improved by
using a prediction algorithm. The prediction algorithm provides a possible estimate of
energy demand and price in the microgrid. The household can maximize its profit by
adjusting its energy supply in the microgrid based on this information.
4. DISCUSSION
The efficiency of the smart grid depends on the effective interaction between the smart
grid components. This section discusses the challenges to optimize cost for smart homes.
It explores the relationship between the cost-optimizing components and strategies. It
also discusses the possible impact of optimization models on the overall system.
4.1. Capital Cost
The storage and renewables capacity in a household have a potential impact on energy
cost. If more storage and renewables are installed, the user can minimize his cost by
selling surplus energy into the grid. However, this requires additional investments. The
initial investments on these systems are hard to evaluate from an economic standpoint.
In fact, in many cases, the initial cost is never recouped [Hopkins and Pahwa 2009].
Ciabattoni et al. [2015] reported that the economical benefit of load shifting is directly
related to the energy generation capacity of PV plants. The authors considered 20 years
to evaluate the cost benefit analysis. This research also reported that load shifting
improves economical benefits from €286 to €757 depending on the number of shiftable
appliances. Their results show that the internal rate of return is 12% for an optimally
sized PV plant.
4.2. Incentives and Loans
From the users’ perspective, financing the initial capital cost is one of the problems
for the implementation of the smart grid. If the market situation is not attractive
enough, the utility or the government can provide incentives or other types of financial
support to accelerate the rapid adoption of the smart grid. The government and existing
financial corporations may introduce attractive financing schemes (e.g., bank loans) for
this purpose. Sauter and Volkery studied the impact of energy policy and reported that
investment in energy efficiency decreases consumer energy demand and reduces energy
expenditure [Sauter and Volkery 2013]. Denmark uses 40% of its tax revenues to fund
energy investment costs, which contribute to 10% of gross added value generated by the
total industrial sector [PricewaterhouseCoopers 2013]. Designing optimal incentives
for such capital investments to achieve the desired generation of distributed energy
remains an open research problem.
4.3. Correlation Between Storage, Renewables, Energy Price, and Trading
If energy is stored, the amount of stored energy reduces because of the energy loss due
to heat and internal resistance. If it is transported through the distribution line, energy
is reduced due to the line resistance. Stored energy reduces automatically because of
self-discharge. Similar to the traditional market, the user would try to sell the energy
as early as possible if the immediate trade seems more profitable. If the user decided
to store energy, this would increase the overall risk because the amount of energy
reduces due to self-discharging according to the trading delay. Previous research did
not analyze the risk and profit margin of storage systems in smart grids.
The energy price also has a correlation with storage capacity. When a seller has
surplus energy, he may wish to store the energy to sell it later at a higher price. But
this situation may create a shortage in the energy supply and may eventually increase
the price. Therefore, the user should sell the surplus energy instead of storing it. This
reciprocal condition could define an equilibrium state of demand and supply in the en-
ergy market that will define the final price of energy. This energy price will depend on
the uncertain state of the energy market. Hence, the user should carefully determine
the energy storage capacity. This aspect of energy trading has not been considered in
previous research. A storage system stabilizes the uncertainty of energy generation
by the renewable sources. Energy generation profiles of the renewable sources deter-
mine the required quantity of renewable energy source installations. More research is
required to analyze scheduling between the load, storage, and energy source.
4.4. New Critical Peak
In general, any devices that can be shifted may shape the daily energy consumption
pattern in new and potentially unpredictable ways. If the load shifting among users
has not been coordinated, it may create a new critical peak in energy demand. This
situation may reduce the benefits from the DR program if the energy price does not
reflect the instantaneous energy demand.
4.5. Computational Complexity
Tsui and Chan [2012] reported that their proposed convex MINLP problem requires
a significant amount of time to solve, which is an important issue when the compu-
tational time is of great concern. To reduce the time complexity, the authors used L1
regularization techniques to transform the model from a convex MINLP to a standard
CP problem. Alam et al. [2014] argued that the cost optimization model that unifies the
components of the smart grids can be computationally very complex. Even for small
problems (few households, few appliances, and a small number of time slots), it can
take several hours to find solutions. This makes the solution approaches unattrac-
tive, as they do not scale to real-world problem sizes. Nor do they allow running of
the optimizer in real time, rescheduling appliances, for example, when new/updated
information about the availability of renewable energy resources is provided.
It is obvious that the total number of appliances, number of time slots, optimization
time horizon, number of participating households in a microgrid, and frequency of
price updates will have a significant impact on the required time to derive a feasible
solution. Most surveyed works use small scenarios with several appliances and 24
time slots, representing the 24 hours in a day, without showing the impact on the
computation time if the problem size increases. Therefore, the provided results do not
reflect the time requirement in real situations. Moreover, some methods are proposed
to be implemented in smart meters, which may have limited computational power and
data storage capability.
4.6. Impact of Time Horizon on the Forecast Methods
Wang et al. [2011] classified wind forecasting methods based on time horizon as shown
in Table IV. This classification is also applicable to solar irradiance forecasting. The
table shows the application of the short- and long-term forecast. In general, a shorter
forecast horizon reduces forecast errors [Wang et al. 2011]. However, when a prediction
method with a shorter forecast horizon is used as an input to an optimization model,
Table IV. Classification of the Forecasting Methods Based on Time Horizon [Wang et al. 2011]
Time Scale Range Application
Immediate short term 8 hours ahead Real-time grid operations
Regulation actions
Short term Day ahead Economic load dispatch planning
Load reasonable decisions
Operational security in the electricity market
Long term Multiple days ahead Maintenance planning
Operation management
Optimal operating cost
the optimization model needs to be re-executed frequently with the updated values to
regenerate the optimal solutions.
4.7. Impact of Uncertainties on Cost Optimization

Prediction algorithms are used to estimate the probable amount of generation, con-
sumption, and energy price, which may not always represent the real situation. These
forecasted quantities are used as inputs to the optimization and trading methods.
Whenever these predicted quantities are re-estimated based on more recent (and pre-
sumably more accurate) information, we need to rerun the optimization and trading
methods, which is computationally expensive. The time complexity of this procedure
can be improved if a method incrementally refines the scheduling plan as more data is
available. Uncertainties on forecast of production, energy price, and consumption may
increase energy cost, as it may result in an inefficient scheduling of energy sources
and loads. It may also result in the failure to fulfill the energy supply commitment to a
neighbor. Any drastic change in the environment or inaccurately predicted generation
and consumption quantities may result in an unstable control system that not only
will increase the energy cost but also will create inconvenience at the user level.
5. CONCLUSION
This article explores methods and components of the smart grid that have significant
impact on residential energy cost reduction. It investigates the potentials of the pro-
posed methods and component interactions. A study on the interdependencies between
the components and the algorithms has been presented according to current research
trends. It also discusses limitations of the methods and future research challenges.
It is obvious that the proposed intelligent methods depend on the advancement of
information and communication technologies for implementation. For example, ac-
cessing and operating the household appliances automatically requires stable and
well-accepted Internet of Things (IoT) communication protocols and data formats. A
bi-directional distribution infrastructure is essential for energy trading in the micro-
grid. The true benefit of the smart grid will not be realized without improvement in
these underlying technologies.
The behavior of the components is not clear from the literature if all of the identified
components and features are implemented in an unified single optimization model. The
impact of the components’ properties on cost optimization is an important research
challenge. For example, if the optimization algorithms transform an off-peak hour
into a peak hour, how should the utility define the energy price? If the energy price
is based on the instantaneous demand, could it be reflected back to the residential
energy consumption pattern effectively? Therefore, the time complexity of the unified
optimization model impacts the effective and timely response to any change in the
system.
Smart homes are the building blocks of smart cities and communities. The behavior
of the population and characteristics of the components for specific scenarios have not
yet been identified. Therefore, the component-specific cost optimization profiles and
the impact of the smart grid technologies on user energy consumption and generation
pattern will influence the future direction of smart grid research.
REFERENCES
Muhammad Raisul Alam, Mamun Bin Ibne Reaz, and Mohd Alauddin Mohd Ali. 2012. A review of smart
homes—past, present, and future. IEEE Transactions on Systems, Man, and Cybernetics, Part C: Appli-
cations and Reviews 42, 6, 1190–1203.
Muhammad Raisul Alam, Marc St-Hilaire, and Thomas Kunz. 2014. Optimizing residential energy con-
sumption: The need for Pareto optimality. In Proceedings of the IEEE/CIC International Conference on
Communications in China (ICCC’14). 1–7.
Ameren Corporation. 2012. Day Ahead Pricing Used for Billing RTP and HSS Services. Retrieved March 19,
2012, from https://www2.ameren.com/RetailEnergy/realtimeprices.aspx.
ApX. 2014. ApX Home Page. Retrieved July 4, 2014, from https://www.apxgroup.com.
A. Arabali, M. Ghofrani, M. Etezadi-Amoli, M. S. Fadali, and Y. Baghzouz. 2013. Genetic-algorithm-based
optimization approach for energy management. IEEE Transactions on Power Delivery 22, 1, 162–170.
Omid Ardakanian, Srinivasan Keshav, and Catherine Rosenberg. 2011. Markovian models for home elec-
tricity consumption. In Proceedings of the 2nd ACM SIGCOMM Workshop on Green Networking. 31–36.
Italo Atzeni, Luis G. Ordonez, Gesualdo Scutari, Daniel P. Palomar, and Javier R. Fonollosa. 2013. Demand-
side management via distributed energy generation and storage optimization. IEEE Transactions on
Smart Grid 4, 2, 866–876.
Tanuja Bapat, Neha Sengupta, Sunil Kumar Ghai, Vijay Arya, Yedendra Babu Shrinivasan, and Deva
Seetharam. 2011. User-sensitive scheduling of home appliances. In Proceedings of the 2nd ACM SIG-
COMM Workshop on Green Networking (GreenNets’11). 43–48.
Marc Beaudin and Hamidreza Zareipour. 2015. Home energy management systems: A review of modelling
and complexity. Renewable and Sustainable Energy Reviews 45, 318–335.
Hasnae Bilil, Ghassane Aniba, and Mohamed Maaroufi. 2014. Multiobjective optimization of renewable
energy penetration rate in power systems. Energy Procedia 50, 368–375.
Stephen Boyd and Lieven Vandenberghe. 2004. Convex Optimization. Cambridge University Press.
Nicola Capodieci, Giuliano Andrea Pagani, Giacomo Cabri, and Marco Aiello. 2011. Smart meter aware
domestic energy trading agents. In Proceedings of the 2011 Workshop on E-energy Market Challenge.
ACM, New York, NY, 1–10.
Yun-Gyung Cheong, Yeo-Jin Kim, Seung Yeol Yoo, Hosub Lee, Sunjae Lee, Seung Chul Chae, and Hyun-Jin
Choi. 2011. An ontology-based reasoning approach towards energy-aware smart homes. In Proceedings
of the IEEE Consumer Communications and Networking Conference (CCNC’11). 850–854.
Hamed Chitsaz, Nima Amjady, and Hamidreza Zareipour. 2015. Wind power forecast using wavelet neural
network trained by improved clonal selection algorithm. Energy Conversion and Management 89, 588–
598.
Lucio Ciabattoni, Francesco Ferracuti, Massimo Grisostomi, Gianluca Ippoliti, and Sauro Longhi. 2015.
Fuzzy logic based economical analysis of photovoltaic energy management. Neurocomputing 170, 296–
305.
Lucio Ciabattoni, Massimo Grisostomi, Gianluca Ippoliti, Sauro Longhi, and Emanuele Mainardi. 2012. On
line solar irradiation forecasting by minimal resource allocating networks. In Proceedings of the 20th
Mediterranean Conference on Control and Automation (MED’12). 1506–1511.
Antonio J. Conejo, Javier Contreras, Rosa Espinola, and Miguel A. Plazas. 2005. Forecasting electricity prices
for a day-ahead pool-based electric energy market. International Journal of Forecasting 21, 3, 435–462.
Antonio J. Conejo, Juan M. Morales, and Luis Baringo. 2010. Real-time demand response model. IEEE
Transactions on Smart Grid 1, 3, 236–242.
CVX. 2014. CVX: Matlab Software for Disciplined Convex Programming, Version 2.0 Beta. Retrieved Febru-
ary 7, 2013, from http://cvxr.com/cvx.
Francesco De Angelis, Matteo Boaro, Danilo Fuselli, Stefano Squartini, Francesco Piazza, and Qingla Wei.
2013. Optimal home energy management under dynamic electrical and thermal constraints. IEEE
Transactions on Industrial Informatics 9, 3, 1518–1527.
Zibo Dong, Dazhi Yang, Thomas Reindl, and Wilfred M. Walsh. 2015. A novel hybrid approach based on self-
organizing maps, support vector regression and particle swarm optimization to forecast solar irradiance.
Energy 82, 570–577.
ENEL. 2014. Enel Smart Grid. Retrieved May 3, 2014, from http://www.enel.com/en-GB/innovation/smart_
grids/.
I. Georgievski, V. Degeler, G. A. Pagani, T. A. Nguyen, A. Lazovik, and M. Aiello. 2012. Optimizing energy
costs for offices connected to the smart grid. IEEE Transactions on Smart Grid 3, 4, 2273–2285.
Dhananjay K. Gode and Shyam Sunder. 1993. Allocative efficiency of markets with zero-intelligence traders:
Market as a partial substitute for individual rationality. Journal of Political Economy 101, 1, 119–137.
G. Graditi, M. G. Ippolito, R. Lamedica, A. Piccolo, A. Ruvio, E. Santini, P. Siano, and G. Zizzo. 2015.
Innovative control logics for a rational utilization of electric loads and air-conditioning systems in a
residential building. Energy and Buildings 102, 1–17.
GridWise. 2015. GridWise Transactive Energy Framework Version 1.0. Retrieved October 12, 2015, from
http://www.gridwiseac.org/pdfs/te_framework_report_pnnl-22946.pdf.
D. J. Hammerstrom, R. Ambrosio, J. Brous, T. A. Carlon, D. P. Chassin, J. G. DeSteese, R. T. Guttromson,
G. R. Horst, O. M. Jarvegren, R. Kajfasz, S. Katipamula, L. Kiesling, N. T. Le, P. Michie, T. V. Oliver, R.
G. Pratt, S. Thompson, and M. Yao. 2007. Pacific Northwest GridWise Testbed Demonstration Projects:
Part I. Olympic Peninsula Project. Retrieved October 12, 2015, from https://www.smartgrid.gov/files/
PNNL_GridWise_Testbed_Demo_Projects_Part_I_Olympic_Peninsula_200711.pdf.
Mark Hopkins and Anil Pahwa. 2009. Economic feasibility of owning of a small wind generator. In Proceedings
of the 2009 Sustainability Conference.
Mark D. Hopkins, Anil Pahwa, and Todd Easton. 2012. Intelligent dispatch for distributed renewable re-
sources. IEEE Transactions on Smart Grid 3, 2, 1047–1054.
Tobias Gybel Hovgaard, Stephen Boyd, Lars F. S. Larsen, and John Bagterp Jorgensen. 2013. Nonconvex
model predictive control for commercial refrigeration. International Journal of Control 86, 8, 1349–1366.
Bernardo A. Huberman and Scott H. Clearwater. 1995. A multi-agent system for controlling building envi-
ronments. In Proceedings of the 1st International Conference on Multiagent Systems (ICMA’95). 171–176.
Tanguy Hubert and Santiago Grijalva. 2011. Realizing smart grid benefits requires energy optimization
algorithms at residential level. In Proceedings of the Conference on Innovative Smart Grid Technologies
(ISGT’11). 1–8.
Hydra. 2015. Hydra Home Page. Retrieved September 25, 2015, from http://www.hydramiddleware.eu.
IEA. 2013. Energy Policies of IEA Countries—Germany: 2013 review. Retrieved September 17, 2014, from
http://www.iea.org/countries/membercountries/germany/.
Dejan Ilic, P. Goncalves Da Silva, Stamatis Karnouskos, and Martin Griesemer. 2012. An energy market for
trading electricity in smart grid neighbourhoods. In Proceedings of the 6th IEEE International Conference
on Digital Ecosystems Technologies (DEST’12). 1–6.
Marco Jahn, Marc Jentsch, Christian R. Prause, Ferry Pramudianto, Amro Al-Akkad, and Rene Reiners.
2010. The energy aware smart home. In Proceedings of the 5th International Conference on Future
Information Technology (FutureTech’10). 1–8.
Nathan Kowahl and Anthony Kuh. 2010. Micro-scale smart grid optimization. In Proceedings of the Inter-
national Joint Conference on Neural Networks (IJCNN’10). 1–8.
Jean Kumagai. 2014. The rise of the personal power plant. IEEE Spectrum 51, 6, 54–59.
Gwo-Ching Liao. 2012. Application a novel evolutionary computation algorithm for load forecasting of air
conditioning. In Proceedings of the Asia-Pacific Power and Energy Engineering Conference (APPEEC’12).
1–4.
Daniel Livengood and Richard Larson. 2009. The energy box: Locally automated optimal control of residential
electricity usage. Service Science 1, 1, 1–16.
Peter Mansbridge. 2003. 2003: The Great North America Blackout. Retrieved September 15, 2014, from
http://www.cbc.ca/archives/categories/economy-business/energy/energy-general/the-great-2003-north-
america-blackout.html.
Marwan Marwan, Gerard Ledwich, and Arindam Ghosh. 2012. Smart grid-demand side response model for
optimization air conditioning. In Proceedings of the 22nd Australasian Universities Power Engineering
Conference (AUPEC’12). 1–6.
Jacob Mattingley, Yang Wang, and Stephen Boyd. 2011. Receding horizon control. IEEE Control Systems 31,
3, 52–65.
Adel Mellit and Alessandro Massi Pavan. 2010. A 24-h forecast of solar irradiance using artificial neural
network: Application for performance prediction of a grid-connected PV plant at Trieste, Italy. Solar
Energy 84, 5, 807–821.
Amir-Hamed Mohsenian-Rad and Alberto Leon-Garcia. 2010. Optimal residential load control with price
prediction in real-time electricity pricing environments. IEEE Transactions on Smart Grid 1, 2, 120–
133.
Amir-Hamed Mohsenian-Rad, Vincent W. S. Wong, Juri Jatskevich, Robert Schober, and Alberto Leon-Garcia.
2010. Autonomous demand-side management based on game-theoretic energy consumption scheduling
for the future smart grid. IEEE Transactions on Smart Grid 1, 3, 320–331.
Trannon Mosher. 2010. Economic Valuation of Energy Storage Coupled with Photovoltaics: Current Tech-
nologies and Future Projections. Master’s Thesis. Massachusetts Institute of Technology, Cambridge,
MA.
NOBEL. 2013. Nobel Home Page. Retrieved July 4, 2013, from http://www.ict-nobel.eu/.
Daniel O’Neill, Marco Levorato, Andrea Goldsmith, and Urbashi Mitra. 2010. Residential demand response
using reinforcement learning. In Proceedings of the 1st IEEE International Conference on Smart Grid
Communications (SmartGridComm’10). 409–414.
OSGi Alliance. 2004. OSGi Home Page. Retrieved October 7, 2014, from www.osgi.org.
Giuliano Andrea Pagani and Marco Aiello. 2015. Generating realistic dynamic prices and services for the
smart grid. IEEE Systems Journal 9, 1, 191–198.
Arpan Pal, Chirabrata Bhaumik, Jasma Shukla, and Subrata Kolay. 2011. Energy information gateway
for home. In Proceedings of the 2nd International Conference on Intelligent Systems, Modelling, and
Simulation (ISMS’11). 235–240.
Piyasak Poonpun and Ward T. Jewell. 2008. Analysis of the cost per kilowatt hour to store electricity. IEEE
Transactions on Energy Conversion 23, 2, 529–534.
PricewaterhouseCoopers. 2013. Decarbonisation and the Economy. An Empirical Analysis of the Economic
Impact of Energy and Climate Change Policies in Denmark, Sweden, Germany, UK, and the Nether-
lands. Retrieved March 4, 2016, from http://www.pwc.nl/assets/documents/pwc-decarbonisation-and-
the-economy.pdf.
Bhuvaneswari Ramachandran, Sanjeev K. Srivastava, Chris S. Edrington, and David A. Cartes. 2011. An
intelligent auction scheme for smart grid market using a hybrid immune algorithm. IEEE Transactions
on Industrial Electronics 58, 10, 4603–4612.
Sergio Rogai. 2006. ENEL’s metering system and telegestore project. In Proceedings of the National As-
sociation of Regulatory Utility Commissioners Conference (NARUC’06). http://www.narucmeetings.org/
Presentations/ENEL.pdf.
Pedram Samadi, Amir-Hamed Mohsenian-Rad, Robert Schober, Vincent W. S. Wong, and Juri Jatskevich.
2010. Optimal real-time pricing algorithm based on utility maximization for smart grid. In Proceedings of
the 1st IEEE International Conference on Smart Grid Communications (SmartGridComm’10). 415–420.
Raphael Sauter and Axel Volkery. 2013. Costs and Benefits of Energy Savings. Institute for European Envi-
ronmental Policy (IEEP) for the Coalition of Energy Savings.
Farshid Shariatzadeh, Paras Mandal, and Anurag K. Srivastava. 2015. Demand response for sustainable
energy systems: A review, application and implementation strategy. Renewable and Sustainable Energy
Reviews 45, 343–350.
Bo Shen, Girish Ghatikar, Zeng Lei, Jinkai Li, Greg Wikler, and Phil Martin. 2014. The role of regulatory
reforms, market changes, and technology development to make demand response a viable resource in
meeting energy challenges. Applied Energy 130, 814–823.
Jie Shi, Wei-Jen Lee, Yongqian Liu, Yongping Yang, and Peng Wang. 2011. Forecasting power output of
photovoltaic system based on weather classification and support vector machine. In Proceedings of the
2011 IEEE Industry Applications Society Annual Meeting (IAS’11). 1–6.
Pierluigi Siano. 2014. Demand response and smart grids: A survey. Renewable and Sustainable Energy
Reviews 30, 461–478.
Pierluigi Siano, Giorgio Graditi, Mauro Atrigna, and Antonio Piccolo. 2013. Designing and testing deci-
sion support and energy management systems for smart homes. Journal of Ambient Intelligence and
Humanized Computing 4, 6, 651–661.
K. M. Tsui and S. C. Chan. 2012. Demand response optimization for smart home scheduling under real-time
pricing. IEEE Transactions on Smart Grid 3, 4, 1812–1821.
John S. Vardakas, Nizar Zorba, and Christos V. Verikoukis. 2015. A survey on demand response programs in
smart grids: Pricing methods and optimization algorithms. IEEE Communications Surveys and Tutorials
17, 1, 152–178.
Perukrishnen Vytelingum, Dave Cliff, and Nicholas R. Jennings. 2008. Strategic bidding in continuous
double auctions. Artificial Intelligence 172, 14, 1700–1729.
Perukrishnen Vytelingum, Sarvapali D. Ramchurn, Thomas D. Voice, Alex Rogers, and Nicholas R. Jennings.
2010a. Trading agents for the smart electricity grid. In Proceedings of the 9th International Conference
on Autonomous Agents and Multiagent Systems. 897–904.
Perukrishnen Vytelingum, Thomas D. Voice, Sarvapali D. Ramchurn, Alex Rogers, and Nicholas R.
Jennings. 2010b. Agent-based micro-storage management for the smart grid. In Proceedings of the
9th International Conference on Autonomous Agents and Multiagent Systems. 39–46.
Xiaochen Wang, Peng Guo, and Xiaobin Huang. 2011. A review of wind power forecasting models. Energy
Procedia 12, 770–778.
Zhu Wang and Lingfeng Wang. 2012. Intelligent negotiation agent with learning capability for energy trading
between building and utility grid. In Proceedings of the IEEE Innovative Smart Grid Technologies-Asia
(ISGT Asia). 1–6.
Markus Weiss and Dominique Guinard. 2010. Increasing energy awareness through Web-enabled power
outlets. In Proceedings of the 9th International Conference on Mobile and Ubiquitous Multimedia. 20:1–
20:10.
WEKA. 2014. Data Mining Software in Java. Retrieved May 10, 2014, from http://www.cs.waikato.ac.nz/
ml/weka/.
Di Zhang, Songsong Liu, and Lazaros G. Papageorgiou. 2014. Fair cost distribution among smart homes with
microgrid. Energy Conversion and Management 80, 498–508.
Ziming Zhu, Jie Tang, Sangarapillai Lambotharan, Woon Hau Chin, and Zhong Fan. 2012. An integer linear
programming based optimization for home demand-side management in smart grid. In Proceedings of
the IEEE PES Conference on Innovative Smart Grid Technologies (ISGT’12). 1–5.
Received March 2015; revised January 2016; accepted January 2016
Copyright of ACM Computing Surveys is the property of Association for Computing
Machinery and its content may not be copied or emailed to multiple sites or posted to a
listserv without the copyright holder's express written permission. However, users may print,
download, or email articles for individual use.
Copyright of ACM Computing Surveys is the property of Association for Computing
Machinery and its content may not be copied or emailed to multiple sites or posted to a
listserv without the copyright holder's express written permission. However, users may print,
download, or email articles for individual use.

Article 2

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Article 2

Uploaded by

Copyright:

Available Formats

Computational Methods for Residential Energy Cost Optimization

in Smart Grids: A Survey

of users, such as home or apartment users, building management authorities, hotel

Fig. 1. Overview of cost-saving strategies for smart homes.

Fig. 2. A classification of DR programs. [Vardakas et al. 2015].

2.1. Energy Use Monitoring and Manual Control

accommodate heterogeneous communication protocols, appliances, and energy sensors.

2.2. Energy Price Negotiation and Prediction

2.4. Energy Use Prediction

2.5. Load Scheduling and Power Shifting

It could be implemented by delaying an appliance operation. The second technique

2.5.1. Linear Programming. Mohsenian-Rad and Leon-Garcia [2010] proposed a mixed

power-shiftable categories according to their power requirement and operational peri-

2.5.2. Convex Programming. Hovgaard et al. [2013] proposed a convex programming

only schedulable appliances) as a benchmark to compare the performance of the pro-

2.6. Energy Storage Scheduling and Power Optimization

2.8. Energy Trading

3.1. Prediction Methods

Fig. 4. Dual characteristics of energy storage.

Fig. 5. Dual characteristics of a microgrid.

Table III. Constraints for Optimization

Objective (3) is a multiobjective optimization problem. Table III presents a list of

Load energy consumption constraints. A specific load needs to be operated for a

3.3. Energy Trading

4.7. Impact of Uncertainties on Cost Optimization

Received March 2015; revised January 2016; accepted January 2016

You might also like