412
1, JANUARY 2021
Predictive Hierarchical Control of Power Flow in
Large-Scale PV Microgrids With Energy Storage
Amir Valibeygi , Sai Akhil Reddy Konakalla , and Raymond de Callafon
Abstract—Large-scale PV microgrids constitute a growing por-
tion of distributed renewable generation in many grids worldwide
and are expected to continue their growth in the foreseeable future.
Managing the operation of such microgrids and in particular their
interaction with the main electricity grid is a challenging task
as it involves controlling large intermittent PV resources. The
addition of auxiliary DERs and energy storage systems to such
microgrids is a promising solution that provides some flexibility
to the operation of these fluctuating PV resources. In this work,
a hierarchical predictive controller is designed to perform daily
scheduling and real-time control of a PV microgrid with energy
storage and auxiliary diesel generation. The controller aims to
increase utilization of available PV resources, reduce the daily
variability of power flow to the main grid, while at the same
time controlling power fluctuations at the point of connection. The
controller is designed to utilize updating load and PV predictions
on a receding time horizon and uses the flexibility of energy storage
and diesel resources to compensate for fluctuations and prediction
errors. A study of multiple operational scenarios demonstrates
successful performance of the proposed controller.
Index Terms—PV microgrid, model predictive control,
distributed energy resources, microgrid control and optimization.
I. INTRODUCTION
TILITY scale solar generation has been attracting increas-
U ing interest in light of various factors. Besides its environ-
mental benefits [1], it facilitates high PV penetration, allowing
the grid to deal with a large controllable aggregated entity to
maintain its balance [2]. Emergence of PV-based community
microgrids where a small community can locally and reliably
generate part of its energy needs while also connected to the
main grid is a thriving application of large-scale PV [3]. As PV
penetration throughout the grid rises, the variability and uncer-
tainty of PV and the ensuing challenge of maintaining balance
between generation and demand throughout the day has become
a central problem that needs to be addressed systematically [4],
[5]. Such intermittency is observed both on a daily time scale as
solar ramps down at nights and up again the next morning and
also on shorter scales with PV variations due to temporary cloud
Manuscript received April 3, 2020; revised May 13, 2020 and June 4, 2020;
accepted June 5, 2020. Date of publication June 10, 2020; date of current version
December 16, 2020. Paper no. TSTE-00357-2020. (Corresponding author: Amir
Valibeygi.)
The authors are with the University of California San Diego, La Jolla, CA
92093 USA (e-mail: avalibey@eng.ucsd.edu; skonakal@ucsd.edu; callafon@
ucsd.edu).
Color versions of one or more of the figures in this article are available online
at https://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TSTE.2020.3001260
1949-3029 © 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
IEEE TRANSACTIONS ON SUSTAINABLE ENERGY, VOL. 12, NO.
cover. Short-term PV power fluctuations, if not compensated by
other resources, can lead to voltage fluctuations that can in turn
trigger protective relays and other automated distribution feeder
equipment, reducing resiliency and reliability [6]. Due to such
intermittency and the uncertainty resulting from it, real-time
adjustment of resource dispatch as more accurate forecast infor-
mation becomes available has been garnering more attention.
Moving towards scheduling generation over shorter time steps
constitutes a paradigm shift from more traditional economic
dispatch frameworks where variations were slower and lower
uncertainty did not necessitate fast updating generation sched-
ule [7], [8].
From the perspective of the main electricity grid, the power
provided or consumed by a solar plant should not exceed the
ability of the grid to safely handle such power profile. A study
by the National Renewable Energy Lab has identified several
requirements for meeting changing grid patterns that include fast
power control, strategic renewable curtailment, new ancillary
services for ramping capacity, and energy storage [9]. For this
reason and due to the limited ramp rate of most non-renewable
generation reserves, power agreements between grid operators
and solar microgrids may impose ramp rate limits on power
flow between the microgrid and the main grid. On the other
hand, solar microgrids may be asked to meet ramping capacity
requirements, enabling them to swiftly ramp up their output as
grid conditions demand as well as reduced power fluctuations
requirements. It is expected that with the current trend of in-
creasing solar penetration, these requirements would be more
strictly enforced in the future [9].
With intermittent solar generation, grid operators start to
incorporate other forms of often non-renewable energy genera-
tion or energy storage to maintain the balance of the grid. PV
curtailment is a measure taken to meet rate limits and maintain
balance but is contrary to environmental and economic goals
of renewable penetration. Diesel generation is another reserve
resource to help maintain balance and meet demand under PV
intermittency, yet is environmentally and economically unfavor-
able. The emergence of large-scale Energy Storage (ES) systems
has proved promising in making up for PV intermittency [10].
Such systems can help mitigate voltage fluctuations due to
PV variations [11] and also meet rate limitations with little
PV curtailment or limited emergency diesel dispatch, thereby
bringing both economic and environmental benefits.
Coordinated control of renewable and non-renewable DERs
together with ES systems is a major topic of interest due to its
far-reaching implications for economic and reliable microgrid
VALIBEYGI et al.: PREDICTIVE HIERARCHICAL CONTROL OF POWER FLOW IN LARGE-SCALE PV MICROGRIDS WITH ENERGY STORAGE 413

operation. Model Predictive Control (MPC) is a commonly

used paradigm for optimization of such microgrids often on
a daily optimization horizon. As such, power flow model of
the microgrid together with predictions of future load and PV
are utilized to solve an optimization problem that computes
optimal DER dispatches over the span of one day. The first
instance of the optimal solution is implemented and the solution
is updated at each new time update as new predictions become
available. Some major contributions can be found in [12]–[15].
Stochastic MPC for management of microgrids with load and
renewable uncertainties has been studied in [16]. Other works
have proposed hierarchical MPC structures for power man-
agement of hybrid microgrids where one MPC layer runs at
a slow time scale trying to optimize DER dispatches and the
other at a faster time scale, trying to reduce the deviation of
power from the reference provided by the slow layer between
each two updates of the slower layer [17], [18]. These works
formulate an optimization problem on the basis of predicted
generation and demand and propose a model predictive approach
that recomputes optimal DER setpoints at each update. The work
of [19] implements a single-layer distributed MPC architecture
at two different time-steps of 5 mins and 1 hour and compares its
computational efficiency with centralized MPC approaches for
problems involving a large number of generator resources. The
proposed controller performs real-time adjustment of generation
schedules as more forecast information becomes available. Such
existing MPC approaches often operate with a fast layer update
rate in the order of minutes. Such updates rates are computation-
ally manageable and can provide adjustments to a decent extent,
but are still unable to react to high frequency fluctuations in
seconds time scales. The computational burden of implementing
MPC at seconds time scales is a major hindrance to the adoption
of such approaches.
Fig. 1. One line diagram of the considered PV microgrid.
2) Successful implementation of second and sub-second pre-
dictive control has allowed us to address the problems of
power flow control and energy optimization within the
same framework and systematically embed operational
constraints needed at both levels.
3) The proposed approach utilizes predictions of load and
PV at two different time scales and horizon lengths, each
serving one MPC level.
II. POWER FLOW MODELING AND OPTIMIZATION
We consider a PV-based microgrid as shown in Fig. 1 with
DERs including solar generation (PV), backup diesel genera-
tion (DG), a set of energy storage (ES) systems indexed by
E = {1, . . ., NES } and local loads, connected to the main elec-
tricity grid. The following equations describe energy balance
of the microgrid as well as charge variation of the ES systems
considering their efficiencies. We assume unit power factor for
the PV system with no reactive power injection by the inverters
and hence focus on real power in this model.
  
t
PLoad = PPt CC + PPt V + t
PES i ,D
− PES
t
i ,C
+ PDG
t

Building on the vast literature on optimal and model predictive ESi ∈E

control of microgrids, this work is motivated by the prospect t+1
EES i
= EES
t
i
+ (PES
t
η − PES
i ,C i
t
i ,D
/ηi )ΔT, ∀ESi ∈ E
of two impending paradigms in control of PV microgrids with
(1)
energy storage, i.e. ramp rate control and fast power control.
We thereby propose a systematic method to first compute best where PLoadt
is the microgrid’s power demand, PPt CC indicates
achievable power flow profile with the main grid that maximizes the power flow between the microgrid and the main grid, PPt V
PV utilization while observing the variability and ramp limits is the power from solar resources, and PDG t
is emergency diesel
of power flow throughout the day, and then implements the generator power, all at step t. Additionally, PES t t
, PES
i ,D i ,C
computed power flow profile by real-time dispatch of DERs indicate storage i’s average discharging and charging power
and despite the fluctuations and variations of load and PV from during step t, respectively. ηi is charge/discharge efficiency of
their predicted values. To this purpose, we propose a hierarchical ES i. EESt
is the charge level of ES i at step t and its state of
i
MPC structure at two different update rates where the outer charge is computed as SoCES t
i
= EES
t
i
/CESi where CESi is
loop performs daily energy optimization and computes SoC the energy capacity of ES i. Finally, ΔT is the length of each
and PCC power references for the inner loop on the basis of time step. Limitations on power output of the ES systems are
predicted load and PV, at an update rate of 15 minutes. The described as
inner loop, on the other hand, aims to achieve these references
by coordinated dispatch of the available DERs at seconds time 0 ≤ PES
t
i ,D
≤ δES
t
P max
i ,D ESi

scale and additionally makes decisions for on-demand use of 0 ≤ PES

t
i ,C
≤ δES
t
P max
i ,C ESi
emergency conventional generator. The main contributions of
this work can be summarized as
t
δES i ,D
+ δES
t
i ,C
≤1
1) We have proposed a low-level controller that uses a pre- t
δES i ,D
t
, δES i ,C
∈ {0, 1}, ESi ∈ E (2)
dictive approach to compensate for PV fluctuations and
t t
improve voltage stability. where δES and
i ,D
δES
are binary variables introduced to
i ,C
ensure only either charge or discharge can happen at a time.
414 IEEE TRANSACTIONS ON SUSTAINABLE ENERGY, VOL. 12, NO. 1, JANUARY 2021

Charge level of the ES systems should remain within upper second term aims at minimizing the variance of power flow. The
and lower bounds to respect capacity limits and avoid deep ratio of parameters k1H and k2H defines the relative weight of
discharges. the objectives and can be tuned to obtain the desired trade-off.
min max This optimization should minimize the above cost by scheduling
SoCES i
CESi ≤ EES
t
i
≤ SoCES i
CESi , ESi ∈ E (3)
existing DERs subject to their availability and constraints. As
Additionally, limitations on power output of PV and diesel we plan no DG dispatch at this level, the following additional
generator (if the DG is on), can be described as constraint is enforced.
0 ≤ PPt V ≤ PPmax
V
t
PDG =0 (6)
min max
PDG ≤ PDG
t
≤ PDG (4) Future load and maximum available PV are assumed to be
min / max predicted up to the scheduling horizon that is the end of the
where PPmaxV is the maximum available PV and PDG are day. Basic prediction schemes for this purpose will be presented
lower/upper DG power limits at each time step t. These equations in Section III.
constitute the basis for power flow and energy models governing t t|t t+1|t T H −1|t
the microgrid and will be later used at two control levels. P̂Load  (PLoad , PLoad , . . ., PLoad )
The overall objective of the control system is to regulate the max,t max,t|t max,t+1|t max,T H −1|t
microgrid’s power flow with the main grid (PP CC ). In particular, P̂P V  (PP V , PP V , . . ., PP V ) (7)
the control system aims to a) compute daily PCC power profile Additionally, constraints on the allowable rate of change and
characterized by high PV utilization and low variability through- upper and lower levels of PCC power throughout the day are
out the day despite load and solar uncertainties and b) implement imposed as
the computed power flow schedule and reduce disturbances by
CC − PP CC
PPt+1 t
fast control of the available DERs. Considering the different min max
RP CC ≤ ≤ RP
time scales and control requirements concerning each of the ΔT CC

above objectives, they are divided between two separate layers PPmin max
CC ≤ PP CC ≤ PP CC
t
(8)
of control, i.e. daily energy scheduling (slow updates) and fast
power control (fast updates). The scheduling layer will be in Constraints on ES charge level at the beginning and end of the
charge of forward-looking energy optimization and considers a day are defined as
horizon up to the end of the day while the inner layer considers t=0
SoCES = SoCES
t=T
, ESi ∈ E (9)
a short horizon in the future, typically a few seconds. i i

The resulting optimization for daily power scheduling can be

A. Daily Power Scheduling
From an economic standpoint, a renewable-based microgrid
connected to the utility grid aims to increase utilization of its
renewable resources resulting in minimizing its power import (or
maximizing its export for that matter) with the main grid. On the
other hand, it is desirable to reduce the variability of power flow
at PCC over a day, as favored by the grid operators to better align
daily demand and generation in the face of increasing renewable
integration. The scheduling level aims to leverage optimized ES
dispatch and, if needed, PV curtailment to reduce PCC power
flow variability while increasing PV utilization. However, the
scheduling at this layer assumes no diesel generator dispatch
as diesel dispatch is reserved as an emergency measure only
for compensation of large deviations of load and PV from
the predictions, to be decided by the lower level controller in
real-time, as will be discussed in the next section. With limited
storage capacity, the two objectives of increasing PV utilization
and decreasing PCC power variability are conflicting; hence, a
trade-off between them may be favorable. Accordingly, the cost
of the optimization is defined as a weighted sum of total PCC
power flow over 24 hours and its variance during this period.
H H
T −1 T −1
JH = k1H PPk CC + k2H (PPk CC − P̄P CC )2
k=t k=t
The first term in the cost functions serves to minimize power
import from the main grid (maximize PV utilization) while the
formulated as
High-level Optimization:
min JH (PLoad , PESi ,D,C , PP V )
PESi ,D,C ,PP V
subject to (1)–(4), (6)–(9)
B. Fast Power Control
To achieve the desired PCC power flow computed by the
scheduling layer, reduce fluctuations of power flow at PCC
resulting from PV (and load) variations, and enable fast response
to power setpoints imposed externally, another control layer is
devised that operates at a faster update rate; providing power
setpoints to each DER. The update rate is chosen in the range
0.5 s–4 s, comparable with renewable output variation time
scales. The hybrid and constrained nature of the inputs as well as
interoperability with the high-level optimization present motiva-
tions for adopting a finite-horizon predictive control scheme for
this layer. The controller would take its high-level references
(PCC power and SoC) from the power scheduling layer and
solves a new optimization at each step to compute optimal DER
dispatches. To address load and PV deviations from their pre-
(5) dicted values, the optimization at this layer achieves a trade-off
between small deviation from SoC reference computed by the
high-level optimization, deviation from reference PCC profile,
and non-renewable DER (DG) dispatch. Different from the daily
VALIBEYGI et al.: PREDICTIVE HIERARCHICAL CONTROL OF POWER FLOW IN LARGE-SCALE PV MICROGRIDS WITH ENERGY STORAGE 415

optimization, the cost of this problem is defined as a weighted models used for forecasting non-stationary time-series data [20].
sum of the above three costs. A regular ARIMA model is represented as ARIMA(p, d, q)
L L where p is order of the auto-regressive part, d is the number of
−1
t+H −1
t+H
JL = k1L (PDG
k
)2 + k2L (PPk CC − PPk CC,ref )2 differences needed for stationarity, and q is order of the moving
k=t k=t average part of the model.
L
1) Short-Term Prediction: Short term prediction is intended
 −1
t+H
to forecast load and PV over the horizon of the low-level
+ k3L (SoCES
k
− SoCES
k
i ,ref
)2 (10)
i controller; hence in time scales of seconds. Past load and PV
ESi ∈E k=t
data with 1-second sampling rate is used to train two different
Given the relatively fast update rate of this loop, ramp rate limits ARIMA models for each of load and PV. Model orders are de-
of the dispatched DERs should be considered as they could cided by examining auto-correlation function (ACF) and partial
become limiting factors in computing optimal DER setpoints. auto-correlation function (PACF) plots of load and PV data and
Hence, in addition to the energy balance (1) and power con- by cross validation. ACF plot describes the correlation of an
straints (2, 3, 4) mentioned before, the following constraints observation with its values at prior time steps while PACF plot in-
applies to enforce the rate limitation of each DER’s power dicates such correlation with the effect of intervening time steps
output. removed or equivalently the correlation of the residual at a time
   
 P t+1 −P t   P t+1 −P t  with its lags. Using such plots as also explained in [20], the model
 ESi ,D ESi ,D   ESi ,C ESi ,C 
orders are chosen to be (pshort = 8, dshort = 1, qPshort = 0) for
  ≤ RESi ,D ,   ≤ RESi ,C PV PV V
 ΔT   ΔT  PV predictor and (pLoad = 15, dLoad = 1, qLoad
short short short
= 0) for load
 t+1   t+1  predictor. The models are utilized at each new time step to predict
 PP V − PPt V   t 
  ≤ RP V ,  PDG − PDG  ≤ RDG (11) over a horizon of length H L using last pshort measurements
 ΔT   ΔT 
of actual PV/load values. The models are updated by retraining
where R(.) indicate ramp limits of different DERs. Another with new data every 10 minutes. The mean squared model errors
operational constraint is the minimum on-time and off-time of averaged over a rolling window of H L time steps, and computed
the diesel generator. Once the diesel turns on, it should stay on over 24 hours with the load and PV profiles of the next section
for at least a minimum duration of time to avoid frequent startup are found to be 4.3 kW and 1.6 kW with normalized values of
and shutdown behavior. The same is also true when the DG turns 0.04% and 0.07% respectively.
off, it should remain in that state for a minimum period. These 2) Daily Prediction: Daily load prediction is performed by
t
conditions can be formulated by defining a binary variable δDG ARIMA models using historical load data up to the current time
indicating on-off state of the diesel generator for training. For this purpose, load data with 15-min sampling
min rate is used to train two different models, one for normal week-
t+1
δDG − δDG
t
≤ δDG
τ
∀τ ∈ {t + 1, . . ., t + TDG,on }
days and another for weekends or days with lower demand
min
t−1
δDG − δDG
t
≤ 1 − δDG
τ
∀τ ∈ {t + 1, . . ., t + TDG,of f}
pattern. This is needed due to the difference observed in load
(12) patterns for these two classes of days. The trained models are
used to make predictions up to the end of the day. Finally,
Similar to the high-level optimization, the optimal solution to the predictions are updated every 15 minutes using the latest
this problem requires knowledge of load and solar generation for available load data. The mean squared model error averaged
the short-term optimization horizon (few seconds). The resulting over a window up to the end of the day and computed over 30
short-term, fast updating optimization can be formulated as days is found to be 89.4 kW (normalized 2.98%).
Low-level Optimization: As recent sky condition is responsible for PV generation over
a day, daily solar prediction can be performed using dedicated
min JL (PLoad , PESi ,D,C , PP V , PDG ) solar forecasting methods or web-based forecasting services for
PESi ,D,C ,PP V ,PDG
the considered region. A detailed review of applicable solar
subject to (1)–(4), (11), (12) forecasting techniques can be found in [21]. In this work, data
from a solar forecasting technique is used together with the PV
It should be noted that although the two optimizations are
system characterization based on historical data. In this way,
concerned with the same overall energy flow model, they seek
at the beginning of each day and at recurring steps along the
different objectives, each pertaining to its respective time scale
day, solar energy generation is predicted to obtain P̂Pmax
V (t|k) for
and respective constraints. Altogether, the overall operation of
k = 0 : N − 1 and t = k : N − 1 with N = 96 (prediction up
these two loops leads to realization of overall control objectives
to the end of the day). The mean squared model error averaged
of the system.
over a window up to the end of the day and computed over 30
days is found to be 52.0 kW with a normalized value of 2.80%.
III. LOAD AND PV PREDICTION
In order to find solutions to the above optimization problems,
IV. MPC FORMULATION AND ONLINE SOLUTION
load and PV prediction must be provided to both control levels.
Auto-Regressive Integrated Moving Average (ARIMA) models The MPC controller is implemented by posing the two op-
are used for this purpose. ARIMA refers to a general class of timizations introduced in Section II as finite horizon optimal
416 IEEE TRANSACTIONS ON SUSTAINABLE ENERGY, VOL. 12, NO. 1, JANUARY 2021

⎡ ⎤
t
PES 1 ,D
⎢ t ⎥
⎢PES1 ,C ⎥
⎢ ⎥
⎢P t ⎥
⎢ ES2 ,D ⎥
⎢ t ⎥
  ⎢PES2 ,C ⎥  
⎢ ⎥
t
EES ⎢ δES
t ⎥ P t
xt = 1
, ut = ⎢
⎢ δt
1 ,D ⎥ , wt =
⎥
Load
t
EES ⎢ ES1 ,C ⎥ PPmax,t
V
2
⎢ t ⎥
⎢ δES2 ,D ⎥
⎢ ⎥
⎢ δt ⎥
⎢ ES2 ,C ⎥
Fig. 2. Schematic of the two-layer MPC control structure and the predictor. ⎢ t ⎥
⎣ PDG ⎦
PPt V
 
1 0
control problems that are solved at two separate times scales and A= , E = 02×2
are constructed in a cascaded fashion as shown in Fig. 2. The 0 1
outer layer operates at a slower rate and is henceforth referred  
−ΔT H /η1 η1 ΔT H 0 0 01×6
to as High-Level Predictive Controller (HLPC) while the inner B=
fast layer is referred as Low-Level Predictive Controller (LLPC). 0 0 −ΔT H /η2 η2 ΔT H 01×6
The outputs of the HLPC are provided as references to the LLPC.
and the constraints are
These references include PCC reference power profile and SoC
max t max t
of the ES systems. The optimal solution to each problem is 0 ≤ ut1 ≤ PES 1
u5 , 0 ≤ ut2 ≤ PES 1
u6
updated at its respective sampling instant using the current state max t max t
of the system as the initial state and only the first instance of the 0 ≤ ut3 ≤ PES 2
u7 , 0 ≤ ut4 ≤ PES 2
u8
computed control is implemented. ut5 + ut6 ≤ 1, ut7 + ut8 ≤ 1, ut5 , ut6 , ut7 , ut8 ∈ {0, 1}
ut9 = 0, 0 ≤ ut10 ≤ w1t
A. High Level Predictive Controller (HLPC) - Slow Layer    
min max
SoCES C1 SoCES C1
The slow layer comprises daily power scheduling and con- 1
≤x ≤
t 1
min max
cerns more forward-looking power optimization up to the end SoCES 2
C2 SoCES 2
C2
of the day. The objective of this layer is computing a PCC power
The output of this layer y = [EES t t
EES PPt CC ]T is pro-
setpoint that meets certain requirements as outlined in Sec- 1 2
vided as reference to the lower level and is computed as yk =
tion II-A and its corresponding SoC profile for the ES systems.
Cxk + Duk + F wk where
These power and SoC setpoints are then passed as references
 T  T
to the inner loop. The problem is posed as a model predictive 1 0 0 0 0 1
control problem with a finite horizon starting at the current C= ,F =
time step and lasting up to the end of day, with step size ΔT H 0 1 0 0 0 0
(= 15 mins). Hence the length of the horizon decreases by one ⎡ ⎤
01×10
with each time update. New predictions are obtained at each new ⎢ ⎥
update based on which a new optimal solution is computed for D=⎣ 01×10 ⎦
the optimization horizon. Only the first instance of the resulting −1 1 −1 1 01×4 −1 −1
input sequence is implemented and the solution is recomputed ŷ3 (t + 1|k) − ŷ3 (t|k)
min max
at the new time update. The HLPC problem is formulated in RP CC ≤ ≤ RP
ΔT CC
state-space as
PPmin max
CC ≤ ŷ3 (t|k) ≤ PP CC , t ∈ {k, k + 1, . . ., N − 1}
H
T −1
min J(xk , uk , wk ) B. Low Level Predictive Controller (LLPC) - Fast Layer
ut ,ut+1 ,...,uT H −1
k=t In contrast to the upper layer, the length of the moving horizon
subject to xk+1 = Axk + Buk + Ewk (H L ) is constant in this layer and the horizon only shifts one
step at each new sampling time. This layer aims to achieve the
(xk , uk ) ∈ X × U dual objective of following the SoC and PCC power references
computed by the scheduling layer, and is granted the additional
where control freedom of DG dispatch, however the amount of this
dispatch is penalized in the cost function. Additionally, the
J = k1H PPt CC + k2H (PPt CC − P̄P CC )2 constraints (4) and (12) should be enforced which together with
the cost function (10) make the optimization a mixed integer
T H = 96, ΔT H = 15 min quadratic program. Solving such MPC problem at each time
VALIBEYGI et al.: PREDICTIVE HIERARCHICAL CONTROL OF POWER FLOW IN LARGE-SCALE PV MICROGRIDS WITH ENERGY STORAGE 417

step with sampling time of 1 sec is computationally expensive

and might even become impossible if smaller update times
is required. Therefore in this section we devise a technique
of turning this problem into a regular quadratic program by
exploiting the minimum on-off property of the diesel generator.
Assuming the LLPC layer is being solved at step t with a
horizon of the next H L steps and with the DG being off at the
beginning of the horizon, condition (12) on minimum on-time
implies that switching from off to on can happen at most at
only one of the next H L steps assuming the minimum off-time Fig. 3. Lowest possible PCC power variability for different PV utilizations
shown for five different initial SoCs. Microgrid specifications are shown in
condition is already satisfied by time t.1 As such switching Table I.
may happens at each of the next H L steps, we can decompose
this MIQP into H L different quadratic programs each of which
account for switching from off to on at one of the next H L steps.  t+1   t+1 
 u9 − ut9   t 
These QPs are all solved at step t and the one with the lowest   ≤ RDG ,  u10 − u10  ≤ RP V
 ΔT L   ΔT L 
cost turns out to be the optimal solution to the original MIQP.
   
This technique allows us to solve a limited number of regular min
SoCES C SoC max
C
1 1
quadratic programs instead of a MIQP, making it computation- 1
min
≤x ≤
t ES
max
1

SoCES C2 SoCES C2
ally more tractable for real-time implementation provided the 2 2

size of the horizon H L is not too large. Indexing these QPs by

h where 1 ≤ h ≤ H L , the LLPC loop can be formulated as V. RESULTS AND DISCUSSION
L To quantify PV power utilization and PCC power variability,
−1
t+H
min Jh (xk , uk , wk ) we define
u0 ,u1 ,...,un  24hr t
t=0 PP V
k=t
PV Utilization =  24hr max,t
× 100
subject to xk+1 = Axk + Buk + Ewk t=0 PP V
(xk , uk ) ∈ X × U max PPt CC − min PPt CC
PCC Variability = × 100 (13)
with ut , xt , wt , A, B, and E the same as above and max PPt CC
where the former shows the percentage of total available PV
Jh = k1L PDG
t
+ k2L (PPt CC − PPt CC,ref )2
energy that is utilized computed over one day and the latter is

+ k3L (SoCES t
i
− SoCES
t
i ,ref
)2 used as a measure of power flow variability over one day. Fig. 3
ESi ∈E indicates possible values of PV utilization versus PCC variability
achievable by varying the ratio k1H /k2H and for different initial
H L = 10, ΔT L = 1 sec SoCs. The results are obtained with the assumption of perfect
and load and PV prediction and with microgrid specifications of
Table I. It is seen that the ratio of these two parameters is a
P̂P CC (t|k) = ŵ1 (t|k) − ut10 − ut1 + ut2 − ut3 + ut4 − ut9 determinant of the weight given to PV utilization versus PCC
variability. Interestingly, the fact that the curves have a crossing
and the constraints
point indicates that the effect of initial SoC depends on the level
max t max t
0 ≤ ut1 ≤ PES 1
u5 , 0 ≤ ut2 ≤ PES 1
u6 of PV utilization that we are interested in within our design.
max t max t Next,the hierarchical MPC controller and predictor are imple-
0 ≤ ut3 ≤ PES 2
u7 , 0 ≤ ut4 ≤ PES 2
u8 mented in a real-time simulation on an Intel Core i7-7500 U
ut5 , ut6 , ut7 , ut8 ∈ {0, 1} computer in MATLAB. The fast control loop is executed at
1 Hz . Additionally, the controller is tested on a remote HIL
ut9 = u09 t ≤ h, PDG
min max
≤ ut9 ≤ PDG t>h testbench and its execution is proved successful in the networked
communication settings, the results of which will be reported in a
0 ≤ ut10 ≤ P̂Pmax
V (t|k) separate manuscript. Load and PV data are used from PMU data
 t+1   t+1 
 u1 − ut1   t
 ≤ RES ,D ,  u2 − u2  ≤ RES ,C
collected in an existing PV microgrid, downsampled to 1 Hz.

 ΔT L  1  ΔT L  1 Next, five different test scenarios are simulated for a microgrid
 t+1   t+1  with the specifications of Table I to indicate the performance
 u3 − ut3   t
  ≤ RES ,D ,  u4 − u4  ≤ RES ,C of the hierarchical MPC controller throughout the day. These
 ΔT L  2  ΔT L  2
scenarios aim to capture prediction errors, PV fluctuations, and
different weightings for PCC power variability according to
1 This statement is valid under the assumption that H L t
< TDG,of which Table II.
f /on
is highly realistic since HL is in order of seconds while minimum on- and Scenarios 1 and 2 enforce high PV utilization at the expense of
off-time are in order of minutes. high PCC power variation by design of their high level controller
418 IEEE TRANSACTIONS ON SUSTAINABLE ENERGY, VOL. 12, NO. 1, JANUARY 2021

TABLE I
MICROGRID SPECIFICATIONS

Fig. 4. Scenario 1.

TABLE II
SIMULATION SCENARIOS

Fig. 5. Scenario 2.

weightings. Scenarios 3 and 4, on the other hand, enforce low

PCC power variability over one day at the expense of lower
PV utilization. Scenarios 1 and 3 employ conservative load
predictions, as opposed to scenarios 2 and 4. Load overprediction
and underprediction scenarios are realized by using predictions
of the load at the beginning of the day for two different days (in
the 5th and 95th percentile prediction errors) and not update the
predictions as we proceed along the day. In these four scenarios,
solar prediction is assumed to be ideal. Scenario 5, on the other
hand, assumes close to perfect load prediction and allows high
PCC variability but faces a sudden unanticipated drop in the
available solar generation in the middle of the day. In all sce-
narios, predictions are refined as time proceeds and as forecasts Fig. 6. Scenario 3.
are updated. One full day of controller operation is shown in
Figs. 4–8, pertaining to scenarios 1 to 5. The top plot in each
figure shows load and PV predictions updated every 15 minutes.
The HLPC problem is solved and optimal PCC power and ES
SoC references (output vector y) are computed and passed to
the lower level controller (LLPC). The low level controller uses
these references and in turn, solves the LLPC problem every
second to compute the optimal dispatch commands to the DERs.
Under ideal forecast assumption and without disturbances, the
LLPC can ideally implement the references given by the HLPC
and the DER dispatches will be precisely the same as those
computed by the HLPC. However, due to prediction errors in
load and PV fluctuations, the LLPC will attempt to achieve PCC
power setpoint by deviating from ES SoC references as well as Fig. 7. Scenario 4.
VALIBEYGI et al.: PREDICTIVE HIERARCHICAL CONTROL OF POWER FLOW IN LARGE-SCALE PV MICROGRIDS WITH ENERGY STORAGE
Fig. 8. Scenario 5.
dispatching DG. The second plot in each scenario compares net
load of the microgrid if all PV was directly used at the time of
generation versus the PCC power profile realized under the pro-
posed MPC control. The reduce variability and intermittency of
the realized power profile is achieved by coordinated utilization
of PV, energy storage, and diesel generation if necessary. The
third plot in each scenario indicates power dispatch of each DER
at the end of the day resulted from every 15 min computation of
the HLPC and every second computation of the LLPC. As seen
in these plots, scenarios 1 and 2 demonstrate high PV utilization
with relatively high PCC power variation throughout the day.
However, the controller can follow the setpoint prescribed by the
HLPC, smoothing power flow fluctuations due to load and PV
variations. Scenarios 3 and 4, on the other hand, aim to achieve
low variability of PCC power flow throughout the day and
consequently achieve low PV utilization due to limited storage
capacity. In scenarios 1 and 3 (load overprediction), real-time
operation of LLPC does not mandate any diesel dispatch as the
HLPC has accounted for more extreme load and PV conditions
without diesel. Opposite circumstances in scenarios 2 and 4
demand diesel dispatch at parts of the day to maintain balance
and meet PCC and SoC tracking. Scenario 5 aims to simulate
an unanticipated solar generation drop in the middle of the day.
With close to perfect load prediction, this scenario is able to
compensate for the unrealized solar generation and meet the
reference PCC power by slight deviation from SoC reference.
The bottom plots in each scenario indicates corresponding SoC
of ES systems under the computed dispatch. Actual SoC pro-
files follow the reference SoC computed by the HLPC closely
and deviate only to account for load and PV fluctuations. The
ES system with higher efficiency is preferred for short term
charge/discharges. In all scenarios, the resulting PCC power
profile can closely track the profiles generated by the HLPC
despite load and PV uncertainties and fluctuations.
VI. CONCLUSION
A hierarchical predictive controller is proposed for a large-
scale PV microgrid to increase utilization of available PV
resources, reduce the variability of power flow between the
microgrid and the main grid, and control power flow fluctuations
at the point of connection. The controller is designed to utilize
419
load and PV predictions on a receding time horizon and uses the
flexibility of energy storage and diesel resources to compensate
for prediction errors. A study of five different operational scenar-
ios demonstrates successful operation of the proposed controller
under different PV availability and load demand scenarios.
REFERENCES
[1] R. R. Hernandez et al., “Environmental impacts of utility-scale solar
energy,” Renew. Sustain. Energy Rev., vol. 29, pp. 766–779, 2014.
[2] H. Shaker, H. Zareipour, and D. Wood, “Impacts of large-scale wind and
solar power integration on California’s net electrical load,” Renew. Sustain.
Energy Rev., vol. 58, pp. 761–774, 2016.
[3] J. Arkhangelski, P. Siano, A.-T. Mahamadou, and G. Lefebvre, “Evaluating
the economic benefits of a smart-community microgrid with centralized
electrical storage and photovoltaic systems,” Energies, vol. 13, no. 7, 2020,
Art. no. 1764.
[4] L. Bird, M. Milligan, and D. Lew, “Integrating variable renewable energy:
Challenges and solutions,” Nat. Renewable Energy Lab., Golden, CO,
USA, Tech. Rep. NREL/TP-6A20-60451, 2013.
[5] P. Denholm, M. O’Connell, G. Brinkman, and J. Jorgenson, “Overgener-
ation from solar energy in california. a field guide to the duck chart,” Nat.
Renewable Energy Lab., Golden, CO, USA, Tech. Rep. NREL/TP-6A20-
65023, 2015.
[6] J. Kleissl, M. Lave, M. Jamaly, and J.-L. Bosch, “Aggregate solar variabil-
ity,” in Proc. IEEE Power Energy Soc. General Meeting, 2012, pp. 1–3.
[7] R. Khatami, M. Parvania, Y. C. Chen, S. Guggilam, and S. Dhople,
“Dynamics-aware continuous-time economic dispatch: A solution for
optimal frequency regulation,” Proc. 53rd Hawaii Int. Conf. Syst. Sci.,
2020, pp. 3186–3195.
[8] S. A. R. Konakalla, A. Valibeygi, and R. A. de Callafon, “Microgrid
dynamic modeling and islanding control with synchrophasor data,” IEEE
Trans. Smart Grid, vol. 11, no. 1, pp. 905–915, Jan. 2020.
[9] M. Milligan, B. Frew, E. Zhou, and D. J. Arent, “Advancing system flex-
ibility for high penetration renewable integration (Chinese translation),”
Nat. Renewable Energy Lab., Golden, CO, USA, Tech. Rep. NREL/TP-
6A20-64864, 2015.
[10] P. Denholm and R. Margolis, “Energy storage requirements for achieving
50% solar photovoltaic energy penetration in California,” Nat. Renewable
Energy Lab., Golden, CO, USA, Tech. Rep. NREL/TP-6A20-66595, 2016.
[11] A. Woyte, V. Van Thong, R. Belmans, and J. Nijs, “Voltage fluctuations
on distribution level introduced by photovoltaic systems,” IEEE Trans.
Energy Convers., vol. 21, no. 1, pp. 202–209, Mar. 2006.
[12] A. Parisio, E. Rikos, and L. Glielmo, “A model predictive control approach
to microgrid operation optimization,” IEEE Trans. Control Syst. Technol.,
vol. 22, no. 5, pp. 1813–1827, Sep. 2014.
[13] R. Palma-Behnke et al., “A microgrid energy management system based
on the rolling horizon strategy,” IEEE Trans. Smart Grid, vol. 4, no. 2,
pp. 996–1006, Jun. 2013.
[14] S. M. Dawoud, X. Lin, and M. I. Okba, “Hybrid renewable microgrid
optimization techniques: A review,” Renew. Sustain. Energy Rev., vol. 82,
pp. 2039–2052, 2018.
[15] R. A. De Callafon, S. A. R. Konakalla, C. H. Wells, and P. T. Lee,
“Optimizing power contribution of distributed energy resources for real
time power demand scheduling,” US Patent 10,452,032, Oct. 22 2019.
[16] A. Parisio and L. Glielmo, “Stochastic model predictive control for eco-
nomic/environmental operation management of microgrids,” in Proc. Eur.
Control Conf., 2013, pp. 2014–2019.
[17] J. Sachs and O. Sawodny, “A two-stage model predictive control strategy
for economic diesel-PV-battery island microgrid operation in rural areas,”
IEEE Trans. Sustain. Energy, vol. 7, no. 3, pp. 903–913, Jul. 2016.
[18] S. R. Cominesi, M. Farina, L. Giulioni, B. Picasso, and R. Scattolini,
“A two-layer stochastic model predictive control scheme for microgrids,”
IEEE Trans. Control Syst. Technol., vol. 26, no. 1, pp. 1–13, Jan. 2018.
[19] M. A. Velasquez, J. Barreiro-Gomez, N. Quijano, A. I. Cadena, and M.
Shahidehpour, “Intra-hour microgrid economic dispatch based on model
predictive control,” IEEE Trans. Smart Grid, vol. 11, no. 3, pp. 1968–1979,
May 2020.
[20] J. Contreras, R. Espinola, F. J. Nogales, and A. J. Conejo, “ARIMA models
to predict next-day electricity prices,” IEEE Trans. Power Syst., vol. 18,
no. 3, pp. 1014–1020, Aug. 2003.
[21] J. Kleissl, Solar Energy Forecasting and Resource Assessment. New York,
NY, USA: Academic, 2013.