You are on page 1of 5

2020 IEEE the 3rd International Conference on Electronics and Communication Engineering

Regional Electricity Sales Forecasting Research Based on Big Data Application


Service Platform
2020 IEEE 3rd International Conference on Electronics and Communication Engineering (ICECE) | 978-1-7281-9258-1/20/$31.00 ©2020 IEEE | DOI: 10.1109/ICECE51594.2020.9352886

Cui Qi1, Sun Mingyue1, Mi Na1, Wang Honggang1, Jian Yanhong1, Zhu Jing 2
1
Big Data Center of State Grid Corporation of China, Beijing, China
2
Beijing China-Power Information Technology Co., LTD, Beijing, China
e-mail: wangqian_20202020@163.com

Abstract—Regional monthly electricity sales forecast is an supply departments are required to explore new monthly
important basis for regional power grid planning and electricity sales forecasting models to improve monthly
construction, evaluation of regional economic development and electricity sales’ Forecast accuracy [4].
operation, and protection of residents' lives. It is also an Big data application service platform is a unified
important work of regional power regulation and management, platform for company-level power big data resources, data
decision-making of power generation and purchase, product display, sharing, collaboration, communication,
improvement of power supply equipment utilization rate and transaction and capacity sharing services. The standardized
deepening of power system reform. Based on the current business data and algorithm models in the platform can
situation of power supply enterprise information development,
provide support for the electricity forecasting model. The
distribution network business status and characteristics, this
prediction method based on a single model is easily
paper analyzes the factors affecting electricity sales. According
to the characteristics of annual changes in electricity sales and interfered by a variety of influencing factors. Choosing a
data quality factors, the recurrent neural network model is reasonable model for combined forecasting is of great
selected based on the big data application service platform. significance for improving the forecast of electricity sales.
The long short term memory neural network model performs There are some problems in the actual situation of the
multi-step multivariate prediction on time series, and uses the current grid operation, such as: the widespread use of
attention mechanism to combine two independent models for dispatch automation technology, the accumulation of own
prediction. Experiments conducted on the historical electricity data, but the data matching of related departments outside the
sales data set of a power supply company show that compared system is poor; with the rapid economic development and the
with traditional machine learning methods, this method has continuous emergence of new technologies, the scale and
advantages in accuracy and efficiency. complexity of power grids are increasing day by day, and
various uncertain factors are increasing. The unified and
Keywords-Recurrent neural network; Long Short-Term standardized data is required to ensure data quality in
Memory neural network; regional monthly electricity sales; electricity sales forecasting, and a variety of forecasting
electricity sales forecast; big data application service platform methods are used to improve the accuracy of forecasting [5].
The structure of this paper is as follows: First, it
I. INTRODUCTION introduces the mixing forecasting model architecture, then
Monthly electricity sales forecast mainly refers to the gives the combined modeling process, and presents historical
comprehensive forecast and calculation of electricity sales, sample data analysis and neural network model prediction
electricity demand, electricity consumption situation, results, and finally summarizes and prospects.
electricity consumption, etc. in the future period in II. MIXING FORECASTING MODEL ARCHITECTURE
combination with historical data and social factors. The
accurate forecast of monthly electricity sales is a key part of The cyclic neural network uses historical electricity and
distribution network planning, and is a necessary means to its related influencing factors to model, and predicts the
realize economic comprehensive energy services. It is future electricity from the future values of these influencing
helpful for electricity regulation and management, power factors. Its advantage is that it can consider the inherent
generation and purchase decision-making [1], improve the influencing factors related to electricity. The prediction
utilization rate of power supply equipment, and deepen the process is simple, but the disadvantage is that it requires a
power system Reform [2, 3]. The monthly electricity sales large number of samples. Data and the selection of
data is affected by many factors. It not only changes influencing factors [6]; time series is only based on historical
periodically with residents and production activities, but also electricity data for modeling and forecasting. The advantage
changes randomly due to various environmental factors such is that it requires less sample data. The disadvantage is that it
as temperature, precipitation, humidity, holidays, and market does not consider the internal influencing factors related to
transactions. With the advancement of the State Grid electricity, and only has a uniform change of electricity
Corporation’s business, big data, data analysis, data science sequence of higher prediction accuracy [7]. In summary, in
and data-related innovative technologies have begun to be order to achieve accurate forecasting of electricity sales, an
promoted within the power grid. Objectively, various power optimized combination system that considers historical data

978-1-7281-9258-1/20/$31.00 ©2020 IEEE 229

Authorized licensed use limited to: Pontificia Universidad Catolica Del Peru. Downloaded on June 11,2022 at 19:27:04 UTC from IEEE Xplore. Restrictions apply.
and influencing factors should be established on the premise stationary model input data are usually called time series.
of ensuring data quality. Fig.1 shows the overall architecture Examples of time series include time-varying temperature
of the combined forecasting model. values, time-varying stock prices, or forecasting electricity
sales in a certain area. The input to this type of model is a
time series signal, defined by observations made in
chronological order. Some existing studies have pointed out
that for a certain regional power grid, its population and
economic scale are determined. Due to the inertial thrust of
power use, without the influence of special factors, its
electricity sales in a particular month will increase year by
year and present a similar S-shaped curve. And this trend is
more obvious under the condition of larger data samples,
which are shown in Fig.2 [8, 9].
The general form of the recurrent neural network RNN
shown in Fig.3 is:
𝑂𝑡 = 𝑓(ℎ𝑡 ; 𝜃) (1)
Figure 1. Overall framework of Mixing forecasting model
where 𝑂𝑡 is the system output, 𝜃 is the given weight and
deviation, and the output of the RNN at time t is a function
III. MIXING MODEL PROCESS of the state of the hidden layer h. And a
1) Recurrent neural network model ℎ𝑡 = 𝑔(ℎ𝑡−1 , 𝑥 𝑡 ; 𝜃) (2)
In the past ten years, the research of machine learning
algorithms and their applications has received a lot of where ℎ𝑡 at time t is equal to the function at the previous
attention. In addition to audiovisual applications, such as time h, which is the input of a given θ at time t(x).
image recognition or speech recognition, another mainstream 2) Long Short Term Memory neural network model
field is sequence-to-sequence applications, such as machine With the development of economy, the trend component
translation or stock price prediction. The former mostly uses sequence of monthly electricity sales tends to keep
convolutional neural networks, and the latter mostly uses increasing gradually over time [10]. Based on its obvious
recurrent neural networks. trend characteristics, the long short term memory neural
network model is selected to predict it.
Due to the differences in climate every year, monthly
electricity sales tend to show different seasonal cycle
components. The multiplication model can decompose the
monthly electricity sales sequence into the product of the
monthly sales trend component sequence, seasonal cycle
component sequence and random component sequence [11].
Figure 2. Growth trend of regional monthly electricity sales The specific expression is as follows:
𝐷 = 𝐷− 𝑄 × 𝐷− 𝑋 × 𝐷− 𝑆 (3)
where D is the monthly electricity sales sequence; D-Q, D-X,
D-S represent the monthly electricity sales trend component
sequence, seasonal cycle component sequence and random
component sequence, respectively.
One certain research object will often go through a
process of generation, development and the end. The
different quantitative features of the research object over
time are arranged in chronological order, and the sequence
formed is called Time series of research objects. Long short
Figure 3. Recurrent neural network structure
term memory LSTM refers to the long- and short-term
Common machine learning models use some memory storage network used in the field of deep learning.
observations (samples/examples) as input features, but there The difference from the recurrent neural network is that
is no time dimension in the data. Time series forecasting LSTM can learn long-term dependence, especially in
models are models that can predict future values based on sequence prediction problems. The long short term memory
previously observed values. Time series forecasting has been network has a feedback connection, that is, in addition to a
widely used for non-stationary data. Non-stationary data is single data point, it can also process the entire data sequence.
called its statistical properties (such as mean and standard The long short term memory model shown in Fig.4 has a
deviation). Data that changes over time is not constant, but storage unit called cell state at its core, which keeps the state
data that measures these changes over time. These non- over time. Information can be added to or removed from the

230

Authorized licensed use limited to: Pontificia Universidad Catolica Del Peru. Downloaded on June 11,2022 at 19:27:04 UTC from IEEE Xplore. Restrictions apply.
unit state in the LSTM, and is controlled by gates, which sum them. The calculated value is used as the seasonal
optionally allow information to flow in and out of the unit, component of current monthly electricity sales. The specific
including a point-wise multiplication and an S-shaped nerve expression is as follows:
to assist the mechanism Network layer. 𝐷− 𝑋𝑖,𝑗 = 𝛼𝐷− 𝑋𝑖−1,𝑗 + 𝛼(𝛼 − 1)𝐷− 𝑋𝑖−2,𝑗 + 𝛼(𝛼 −
1)𝑛−1 𝐷− 𝑋𝑖−𝑛,𝑗 (13)
where D- Xi,j is the seasonal component of monthly electricity
sales in the jth month of the i-th year; is the weighting
coefficient. It can be known from the literature that if the
seasonal component of monthly electricity sales fluctuates
little, is selected between 0.1 and 0.5, and if there is a rapid
and obvious trend of change, it is selected between 0.6 and
0.8. The average value of the historical random component
Figure 4. LSTM algorithm flow chart
over the same period is adopted, and its value is used as the
According to the basic principles of the LSTM model, random component of the current monthly electricity sales.
the LSTM(p,q,d) model of the monthly electricity sales trend The specific expression is as follows:
component sequence D_Q can be obtained as follows: 1
𝐷− 𝑆𝑖,𝑗 = �𝐷− 𝑆𝑖−1,𝑗 + 𝐷− 𝑆𝑖−2,𝑗 + ⋯ + 𝐷− 𝑆𝑖−𝑛,𝑗 � (14)
𝑛
𝑔𝑡 = tanh�𝑊𝑥𝑔 𝐷− 𝑄𝑡 + 𝑊𝑟𝑔 𝑟𝑡−1 + 𝑏𝑔 � (4)
where𝐷− 𝑆𝑖,𝑗 is the random component of monthly electricity
𝑖𝑡 = 𝜎(𝑊𝑥𝑖 𝐷− 𝑄𝑡 + 𝑊𝑟𝑖 𝑟𝑡−1 + 𝑏𝑖 ) (5) sales in the jth month of the ith year [12, 13].
𝑓𝑡 = 𝜎�𝑊𝑥𝑓 𝐷− 𝑄𝑡 + 𝑊𝑟𝑓 𝑟𝑡−1 + 𝑏𝑓 � (6)
𝑐𝑡 = 𝑖𝑡 ⊙ 𝑔𝑡 + 𝑓𝑡 ⊙ 𝑐𝑡−1 (7)
𝑜𝑡 = 𝜎(𝑊𝑥𝑜 𝐷− 𝑄𝑡 + 𝑊𝑟𝑜 𝑟𝑡−1 + 𝑏0 ) (8)
ℎ𝑡 = (tanh(𝑐𝑡 ), 𝑥𝑡 ) (9)
𝑚𝑡 = 𝑜𝑡 ⊙ (𝑊Res1 ℎ𝑡 ) (10)
𝑦𝑡 = 𝑊𝑟𝑝 𝑚𝑡 (11)
𝑟𝑡 = 𝑦𝑡 (1: 𝑛𝑟 ) (12)
where D- Q t is the monthly electricity sales trend component
sequence, 𝑥𝑡 and 𝑦𝑡 are the current input and output; 𝑔𝑡 is the Figure 5. Flow chart of LSTM time series algorithm
processed input, which combines the current input and the
information output at the previous time; 𝑐𝑡 is the state value Combined with the big data application service platform,
of the cell, used to save historical information ; 𝑖𝑡 , 𝑓𝑡 and 𝑜𝑡 the LSTM time series algorithm forecasting steps shown in
are input gates, forget gates and output gates respectively, Fig.5 are: collect regional electricity sales related data
which control how many or which processed inputs are through the platform, collect the regional total electricity
selected, how many or which state values are forgotten or sales, industry classification electricity consumption, change
which state values at the previous moment are selected, and trends, etc. Use the big data application service platform to
how many or which state activation values are selected as decompose the trend of time series data and judge whether
outputs; 𝜎 is the sigmoid function; 𝑊represents the monthly the time series data is stable. If the data is stable, use the
electricity sales trend component sequence, 𝜎 is the bias LSTM time series model to calculate the degree of influence
vector; ⊙ represents the corresponding multiplication of the of different factors on the change in electricity sales; if the
elements of the two vectors. In the formula. 𝑖𝑡 , 𝑓𝑡 , 𝑜𝑡 , 𝑔𝑡 , 𝑐𝑡 data is not stable, you need to decompose the multiplication
and 𝑦𝑡 have the same dimensions; 𝑊𝑟𝑝 is the projection model, use the multiplication model to decompose the time
matrix, 𝑟𝑡 = 𝑦𝑡 (1: 𝑛𝑟 ) indicates that 𝑟𝑡 is the first 𝑛𝑟 series data, and decompose it into different trends Factors,
seasonal factors and random factors. Bring the final data into
elements of 𝑦𝑡 , and 𝑊𝑝 is the projection matrix used to adjust
the regression equation of the exponential smoothing method,
the input dimensions, If the input and output dimensions are calculate the predicted value of electricity sales, and compare
equal, 𝑊𝑝 is the unit matrix. ℎ𝑡 is the original vector and the it with the actual observation value to calculate the error rate.
vector after the input is spliced, 𝑊Res1 is the dimension
adjusted projection matrix, 𝑖𝑡 , 𝑓𝑡 , 𝑔𝑡 and 𝑚𝑡 have the same IV. EXPERIMENTAL RESULTS
dimensions. This paper first uses the recurrent neural network model
The periodic component sequence of monthly electricity in the big data application service platform to verify the
sales is based on the principle of near large, far small, to give monthly electricity sales sample data from 2012 to 2018, and
certain weights to the historical seasonal components and

231

Authorized licensed use limited to: Pontificia Universidad Catolica Del Peru. Downloaded on June 11,2022 at 19:27:04 UTC from IEEE Xplore. Restrictions apply.
obtain the error situation. Comparing the error, it can be seen seen that both models have advantages in forecasting
that the single cyclic neural network model follows the electricity sales. Therefore, relying on the attention
principle of inertial change of electricity sales, and overall mechanism to generate weight coefficients of the model in
this forecast does not deviate significantly. However, it is the combined model, and the weighted values of the typical
limited to the lack of analysis of influencing factors in the attention mechanism are shown in Tab.1 (model 1 and 2
cyclic neural network model itself, and does not take into refer to the cyclic neural network model and the LSTM time
account factors such as the economy, population, and special series model). After the weight parameters of the two models
events year by year. Therefore, there must be large forecast are weighted by the attention mechanism, the combined
errors. This indicates that the single cyclic neural network model is used to predict the monthly electricity sales of the
model has problems and needs other models are assisted. region in 2019, and the error analysis between the predicted
Taking into account that monthly electricity sales are value and the true value is performed. The results are shown
affected by economic development, industrial structure in Tab.2. It can be seen that the combination model of the
changes, seasonal changes and other factors, it shows attention mechanism is weighted to make predictions, which
obvious trends, seasonal periodicity and randomness. improves the accuracy of model prediction.
Obviously, the seasonal and random components in the
monthly electricity sales will directly affect the prediction V. CONCLUSION
accuracy of the time series model. Based on the long short Based on the data system and algorithm model of the big
term memory neural network time series model built by the data application service platform, this paper attempts to
big data application service platform, the time series is predict the regional monthly electricity sales using the cyclic
decomposed into a trend component sequence, a seasonal neural network-long short-term memory time series
cycle component sequence and a random component combined model fitted by the attention mechanism. The
sequence to predict electricity sales. model fitting results show that the cyclic neural network-
long short-term memory time series combined model
TABLE I. WEIGHT OF EACH MODEL IN MIXING MODEL
weighted by the attention mechanism can better coordinate
the historical data inertia, seasonality and randomness of
Weight Weight monthly electricity sales, and improve the prediction
mouth mouth Accuracy. The constructed attention mechanism combination
Model1 Model2 Model1 Model2 model prediction process can be used for predictive analysis
1 0.498 0.502 7 0.337 0.663
in the big data application service platform, and feedback
based on the results. Relying on the standardized sample data
2 0.566 0.434 8 0.396 0.604 and algorithm model of the big data application service
3 0.489 0.511 9 0.471 0.529 platform, detailed adjustments can be made to the
constructed combination model, which is worthy of
4 0.321 0.679 10 0.392 0.608
promotion at the power marketing level.
5 0.312 0.688 11 0.295 0.705
REFERENCES
6 0.388 0.612 12 0.416 0.584
[1] Huamao, Huang, et al. Method and System for Detecting Utilization
Rate of Power Distribution Network Equipment with Distributed
TABLE II. FORECAST ANALYSIS BASED ON MIXING MODEL Power Supply. 2016.
[2] Zhuangwen, Wu. Portable Fuel Cell Vehicle Power Supply
Mixing forecast Equipment. 2018.
Mixing forecast analysis
analysis [3] Valipour, Mohammad, et al. “Comparison of the ARMA, ARIMA,
mouth mouth and the Autoregressive Artificial Neural Network Models in
Forecast Forecast Forecasting the Monthly Inflow of Dez Dam Reservoir.” Journal of
Error Error Hydrology, vol. 476, 2013, pp. 433–441.
value/100M value/100M
/% /%
kWh kWh [4] Feng, Qi, et al. “Wavelet Analysis-Support Vector Machine Coupled
Models for Monthly Rainfall Forecasting in Arid Regions.” Water
1 10.041 0.18 7 9.288 -0.23 Resources Management, vol. 29, no. 4, 2015, pp. 1049–1065.
2 8.984 0.22 8 11.812 0.34 [5] Wenzhe, Zhang, and Li Yangyang. “Application of the Improved
Grey Model in the Monthly Electricity Consumption Forecasting.”
3 8.385 0.21 9 11.784 -0.59 Science Discovery, vol. 4, no. 1, 2016, p. 1.
4 9.174 0.43 10 12.074 0.38 [6] Wu, C. L., and K. W. Chau. “Data-Driven Models for Monthly
Streamflow Time Series Prediction.” Engineering Applications of
5 8.277 0.54 11 12.642 -0.35 Artificial Intelligence, vol. 23, no. 8, 2010, pp. 1350–1367.
[7] Bo, Zeng, et al. “Research on the Prediction Model of Oscillatory
6 8.544 0.32 12 12.716 0.54 Sequence Based on GM (1, 1) and Its Application in Electricity
Demand Prediction.” Proceedings of 2013 IEEE International
Conference on Grey Systems and Intelligent Services (GSIS), 2013,
This paper collects data through the big data application pp. 88–92.
service platform, and uses two single models in the platform [8] Beutner, R., and Joseph Lozner. “The Relation of Life to Electricity:
to achieve year-to-month verification forecasts. It can be Part IV the Electromotive Action of Homologous Fatty Acids:

232

Authorized licensed use limited to: Pontificia Universidad Catolica Del Peru. Downloaded on June 11,2022 at 19:27:04 UTC from IEEE Xplore. Restrictions apply.
Exhaustion as an Electrochemical Sequence of a Chemical Splitting [11] Sak, Hasim, et al. “Long Short-Term Memory Recurrent Neural
of Higher Molecular Compounds.” Protoplasma, vol. 12, no. 1, 1931, Network Architectures for Large Scale Acoustic Modeling.”
pp. 380–393. INTERSPEECH, 2014, pp. 338–342.
[9] Li, Der Chiang, et al. “Detecting Representative Data and Generating [12] Chiu, Jason P. C., and Eric Nichols. “Named Entity Recognition with
Synthetic Samples to Improve Learning Accuracy with Imbalanced Bidirectional LSTM-CNNs.” Transactions of the Association for
Data Sets.” PLOS ONE, vol. 12, no. 8, 2017. Computational Linguistics, vol. 4, no. 1, 2016, pp. 357–370.
[10] Tai, Kai Sheng, et al. “Improved Semantic Representations From [13] Chiu, Jason P. C., and Eric Nichols. “Named Entity Recognition with
Tree-Structured Long Short-Term Memory Networks.” Proceedings Bidirectional LSTM-CNNs.” Transactions of the Association for
of the 53rd Annual Meeting of the Association for Computational Computational Linguistics, vol. 4, no. 1, 2016, pp. 357–370.
Linguistics and the 7th International Joint Conference on Natural [14] Srivastava, Nitish, et al. “Unsupervised Learning of Video
Language Processing (Volume 1: Long Papers), vol. 1, 2015, pp. Representations Using LSTMs.” ArXiv Preprint ArXiv:1502.04681,
1556–1566. 2015.

233

Authorized licensed use limited to: Pontificia Universidad Catolica Del Peru. Downloaded on June 11,2022 at 19:27:04 UTC from IEEE Xplore. Restrictions apply.

You might also like