Smart Renewable Energy Management System for

Consumer Applications
Justin Schlauwitz∗ , Petr Musilek∗†
∗ Department of Electrical and Computer Engineering, University of Alberta, Edmonton AB, Canada
† Faculty of Electrical Engineering and Computer Science, VSB-Technical University, Ostrava, Czech Republic

Abstract—Due to the benefits of Smart Grid Technology and

renewable energy systems, consumer demand for a controller
design that allows effective integration of these two systems has
been growing over the last few years. This article proposes a new
controller designed with the objective of increasing efficiency and
reliability of a consumer’s renewable energy system. It describes
the system model design and data selection, as well as the con-
troller considerations and design method. The model subsystems
are designed either as raw real world data, formulated models
based on real world data as reasonable approximations, or as
models that allow for proper testing of the system under relevant
abnormal conditions and limitations. The controller is designed
while considering the necessary flexibility to accommodate a Fig. 1. System Model Diagram
wide variety of subsystems and the need for correct behavior
for normal and abnormal states. The applied methods include
frequency pattern analysis, forecasting, probability estimation,
payoff estimation, expected value calculation, and neural-state- simulation environment is then used to design the controller.
space modeling. For the analysis, a comparison to alternative non- The effectiveness of the controller is finally evaluated using
controlled methods has been conducted to determine the relative publicly available household demand and generation data.
performance of the controller in terms of energy efficiency,
reliability, and monetary savings. The results of the analysis show II. S YSTEM M ODEL
that the controller did not reduce energy losses, but is reliable
with respect to performance under blackout and cyber-attack The system considered in this study is composed of a
conditions, and offers at least an 8% savings on energy expenses renewable energy source and a battery bank. It is connected to
and as much as 25% of the expected profit from immediately a power grid operated under assumption of residential owners
selling excess renewable energy.
being the primary target for electricity sales, and driven by
I. I NTRODUCTION typical household energy demands. As household energy data
was readily available1 , the requirements for data set selection
Smart Grid Technology is becoming increasingly popular, was primarily based on completeness of information, data set
as it has made the aging power systems more reliable and size, sampling frequency, and data format.
efficient by incorporating more sophisticated controllers and The simulator was built by connecting subsystem models
integrating renewable energy. Many renewable energy sources for household energy, solar energy, battery energy, and smart
are intermittent, and thus their introduction have prompted grid power and corresponding data paths as shown in Fig. 1.
development of more advanced, smarter controllers. In the Each of the connecting lines applies a conversion efficiency to
area of consumer demand, the introduction of smart meters simulate line loss between the subsystems. The energy flow
allows distribution companies to set prices based on supply through each line is regulated by the controller (regulators
and demand. Corresponding price fluctuations allows load Reg1 to Reg4 and switch LCK respectively). The structure
scheduling to smooth power consumption rates over time. of the network was designed to minimize the number of
With the boom of customer-scale renewable energy and regulators while not being overly restrictive to the dynamics
energy storage devices, an alternative approach gains on and natural flow of the system.
importance: a system that would appropriately manage the
The simulation was mathematically designed in such a
multifaceted energy exchanges at the consumer site. Such a
way that solar energy and household energy values drive the
controller must effectively manage renewable energy sources,
system. The battery bank, which is restricted by its limited
smart grid power supply, and battery banks in response to the
capacity discharge/charge rate, acts as a damper to the total
consumer demand and regardless of the circumstances.
flow of energy into and out of the system. Finally, the power
To design and analyze such an advanced controller, this line acts as a sink that absorbs excess and deficit energy flow
contribution introduces a simulator to test the controller to
ensure its proper functionality. Based on the identification 1 Source: [Link]
of important system characteristics and general approach, the use-data

978-1-5090-0623-6/16/$31.00 2016
c IEEE 2006
under the condition that the energy needed from the line is battery is possible and could damage the battery bank, a
available. Should the system fail to supply the required energy lockout mechanism is incorporated to allow the system and/or
to the household during a blackout, it would be indicated by controller to protect the battery from damage due to depletion
the system trying to satisfy the equations by drawing energy by preventing more energy from being drawn off the battery
off the line while the blackout is still in effect. by other subsystems.
A. Solar Model C. Smart Grid Model
Solar photo-voltaic (PV) was selected as the form of renew- The smart grid model is designed to accommodate a time-
able energy as it is better suited for residential installations and varying pricing method, blackouts, and price modifying cyber-
more reliable than wind. Energy availability data is generated attacks. For the price model, a 7 day, 5 minute Market Clearing
using a mathematical model because it allowed application Price data set from the Ontario Hourly Energy Price3 was
of solar energy data with only small changes regardless of chosen and scaled to the units used in the simulation. It was
the start time, sample length, and sampling frequency of the then applied in a looped format to account for the difference
household energy data set. The developed simple mathematical in data sampling time span and simulation run time. This
model was compared to a sample data set2 and found to be its data gives a more realistic price but not the overhead charges
reasonable approximation as the cyclical, seasonal, and noise which are an additional expense to the energy price. This
data where graphically similar. The model takes the following was rectified by considering the buying and selling prices
form separately, charging the consumer slightly more for buying
3.124 × (t + θ) energy than they could make by selling an equal amount at
EC = ECmax × sin( ) (1) the same point in time.
12 × r
For the power supply from the grid, a finite supply of up
cos( 3.124×(t+φ+θ)
365×24×r )+1 to 4 MWh is available. This allows emulation of blackouts
ES = −ESmax × (2)
2 that would be triggered when the supply limit is exceeded.
rnd In consideration of the significance of a blackouts effect on
EI = −N × ( ) (3) the potential performance of the controller, they have been
|rnd| + 100
incorporated into the design as the most important adverse
Esolar in = EC + ES + EI (4) condition. Both blackouts and cyber-attacks are implemented
where ECmax is the maximum possible cyclical energy that as rare4 , random occurrences. Blackouts have an unknown,
can be produced, ESmax is the maximum seasonal damping random duration and visible state, while cyber-attacks have a
over one year period, N is the largest noise injection possible, known duration of one period but invisible state.
rnd is a uniform random number within the interval of [0,100],
’φ’ is the offset from the start of the year, ’θ’ is the offset from III. C ONTROLLER D ESIGN
the start of the day needed to match the starting time stamp The controller (shown in Fig.2) was designed based on the
of the household energy data using the units of the iterator t, five control outputs, the expected system model characteristics,
and r is the ratio of one hour divided by the sample rate of and the useful information that could be easily acquired from
the data set. As the solar panel subsystem is a fixed source of the system itself. The controller’s inputs were selected as the
energy, only positive values are transmitted, which could be energy entering and leaving the system (the solar, battery,
absorbed by the household, battery, or power line. household, and grid power), the battery bank’s level of charge,
the price per Wh (the purchase and sale price), and whether
B. Battery Bank Model
or not there is a blackout affecting the grid. For the sake
The battery bank model uses a piece-wise linear sigmoid of simplifying the setup and allowing quick adjustments to
function to account for the limited capacity Ebatt and a fixed different dataset sample rates, the simulated controller also
polarity. The linear relation (y(t) = x(t) + y(t − 1)) between has the iteration and sample-rate-per-hour included as inputs.
battery flow and storage was used as a simple relation to In real world applications the iteration would be internally
allow battery flow losses to be a constant value for all battery generated and the sample-rate-per-hour would be set before
levels. The limit to how quickly a battery can be charged or running the controller.
discharged at a given time are accounted for with δM ax during The objective for the controller design was to effectively
the calculation of the flow into and out of the battery, as is manage the energy in the system for the consumers benefit by
the battery leakage over time (Refer to Table I for the battery minimizing losses. The losses considered are line loss caused
bank’s selected constants). The limitations of the battery bank by exchanges of energy between subsystems, loss of utility
result in a limited operational range of the subsystem, and (customer satisfaction) due to the inconvenience of blackouts,
attempts to exceed these limits result in additional losses. In monetary losses due to inefficient transactions, and loss of
a real world application, such attempts could cause damage
to the system. In consideration that full depletion of the 3 Source: [Link]/Pages/Power-Data/[Link]#
4 The definition of rare is loosely used as it is desirable to have adverse
2 Source: [Link]/[Link] conditions occur at least once on average for each simulation of 500 hours.

2016 IEEE Congress on Evolutionary Computation (CEC) 2007

 Fig. 2. Controller Block Diagram

stability due to cyber-attacks. Due to the variety of available information on when the largest amount of solar energy was
subsystems, the controller needs to account for the facts that usually produced and when the largest amount of household
• not all renewable energy sources have the same produc- energy was usually consumed in the past can be extracted from
tion rates; the remaining cyclical components. This historical information
• battery banks will likely be have different capacities; is used to predict future behavior.
• consumers will have different rates and patterns of con- 1) Forecasting: Prediction of future states is an important
sumption; and task as it is the deciding factor of whether the controller
• none of subsystem conditions can be considered fixed for should act immediately or wait for a time when an action
a given consumer over time. will likely be more beneficial [2]. The controller multiplies the
Possible solutions have been identified using applications of cleaned solar energy frequency pattern by the estimated sale
signal processing, probability, game theory, human psychol- price frequency pattern, and multiplies the estimated buy price
ogy, economics, and control systems. These approaches are frequency pattern by the cleaned household energy frequency
briefly outlined in the following subsections. pattern. Solar energy can be viewed as surplus that could be
stored or sold. However, if sold immediately when the price
A. Frequency Domain Analysis is not at or near its projected maximum value within the day
The frequency domain analysis [1] is often used in signal (based on the understanding that the energy price cycles are
processing as it can provide a pattern that is, for the most part, daily), the maximum profit could not be gained. Similarly, the
time invariant outside non-periodic changes. In addition, its household energy consumption can considered a deficit that
operational equivalent is greatly simplified in comparison with could be temporarily accommodated from the battery bank or
the time domain convolution. This operation has been imple- immediately offset by buying energy. Then, if the price is not
mented through discrete cosine transform (DCT). As opposed at or near its projected minimum value, the maximum savings
to the fast Fourier transform (FFT), DCT concentrates more could not be achieved. As buying and selling are opposing
signal energy into the low frequency values. The resulting fre- actions, they are combined as a difference into one profit
quency pattern of the price data can be easily compared to the pattern. This way, the most opportune times to sell are when
average pattern. The error value resulting from this comparison the time domain value of the profit pattern is most positive
can be used as a simple measure of probability whether or not while the best times to buy are when the time domain values
a cyber-attack has occurred. With the resulting information, of the profit pattern are most negative. However, due to the
the controller can quickly reconstruct an approximation of the fact that charge and discharge rates of the battery bank are
price per Wh supplied by the grid regardless of whether or limited, the exchange of energy needs to be performed over
not the price data was modified. This may compromise the an interval of time centered around the best time to allow for
accuracy of the data, but in turn, increases the reliability of the an optimal total flow.
system by acting on known patterns instead of possible false
data injection. When applying this procedure to the renewable B. Expected Values
energy and consumer demand, the controller can easily remove Expected value calculations are indispensable as they are
trends, low frequency bias and high frequency noise. The necessary for the controller to select one of three actions

2008 2016 IEEE Congress on Evolutionary Computation (CEC)

 • buy, where pin is the probability given by the simple calculation
• sell, or of the likelihood that a given event has or will occur, and
• hold, individual terms a1 , a2 , b1 and b2 are defined as follows
given that the environment could be in or could enter one of r1
a1 = 1 −
four states r2
• normal, r4
a2 =
• blackout, r3
• cyber-attack, or
2 × r3
• blackout and cyber-attack. b1 =
r2 × (1 + r1 × r4 )
Since the controller could easily determine the probabilities
and payoffs for each of the twelve possible outcomes using 2 × r2
b2 =
existing values, simple penalties, and simplified probability r3 × (1 + r1 × r4 )
models, it is reasonable to apply principles derived from Terms a1 and a2 describe the relative consistency of the infor-
the Game Theory [3] by breaking the game down into sub- mation from the probabilistic model, and b1 and b2 describe
games for each possible action. Each sub-game is then played the distance between the average and maximum recorded
using the probability of the environment entering each state value conditioned by the threshold and standard deviation. The
to determine the expected value for each action the controller additional terms for the derived formula are, standard deviation
could make. By selecting the action with the largest expected (r1 or r4 ), average (r2 or r3 ), maximum value (r3 or r2 ), and
value, the controller can then achieve the maximum gain given threshold (r4 or r1 ). The specific order depends on what the
the circumstances. However, this requires that the controller probability and additional terms describe (i.e. the occurrence
has some derived understanding of the probability and payoffs or absence of the given event).
of certain events within the environment. When the output exceeds the range of 0% to 100% proba-
1) Probability Estimation: The estimation of naturally oc- bility, it is sufficient to say that anything more than 100% is
curring probability of rare and random events cannot be equivalent to 100% and anything less than 0% is equivalent to
done without large amounts of data and processing [4]. 0%. To match the valid range of probabilities between 0 and 1,
However, using some statistical information and/or applying it is necessary to represent values outside the range with the
some round-a-bout psychology methods, a vague measure of appropriate limits. In addition to calculating the probability
probability can be acquired. Measuring the probability of a concerning blackouts, it is possible to reverse the simple
cyber-attack is trivial as it could be considered as the absolute probabilistic model to calculate how much more (or less)
percent error between the average pattern and the current energy is needed to be comfortable regarding the survivability
pattern in the price data since a cyber-attack can only be of a blackout.
measured as a fluctuation deviating from the normal pattern. 3) Pay-offs: Based on the principles of Game Theory [3],
For measuring the probability of a blackout, Murphy’s Law [5] the profitability of each choice needs to be determined for
is applied based on the understanding that the lower the battery each possible state that could occur. For the controller, it was
level, the more likely a blackout will affect the consumer. determined that there are four possible states the environment
Although the derived probability values could exceed the valid could select and three possible actions the controller could
probability range, when attitude-towards-risk [3] is applied, the take, resulting in twelve payoffs that need to be calculated.
value could still be considered valid as a measure exceeding Under normal conditions, the payoffs are the expected profits
the certainty of the outcome. This does not mean that the for each action with some penalty. The added penalty is for
controller is certain of the outcome when told that the outcome having the battery bank energy deviate from the value deemed
is guaranteed; it simply means that certainty values outside of to be minimally sufficient for surviving a blackout. Under
the 0% to 100% range could still produce computationally abnormal conditions, the payoffs of the normal condition is re-
valid probabilities after considering the controller’s attitude- used while adding additional penalties for sub-optimal actions
towards-risk (i.e. if the input claims that the probability that based on insight of the potential consequences.
something will occur is 120% it may be interpreted as being
40% when considering attitude-towards-risk). C. Neural-State-Space Hybrid Model
2) Attitude Towards Risk: From understanding how human The final decision from the game is then fed into a hybrid-
rationality can affect the choices an individual makes, the state-space model to generate the best corresponding controller
application of attitude-towards-risk was selected as a valid output state. This model was adopted to gain the desired
method of augmenting the basic understanding of probability characteristics present in both state-space models [7], [8] and
into a more reliable form. Additionally, attitude-towards-risk neural networks. The state-space model allows for a simple
provides an insight into the gambling tendencies [6] of the structure with easily defined weights to determine the output
controller and the reliability of the probabilistic model used values, but lacks a restricted range which could reduce the
pin − a1 pin − a2 performance needed without rigorous testing. The (unipolar)
patr = + + 0.5 (5) neural network provides the desired fixed output range, but it
|pin − a1 | + b1 |pin − a2 | + b2

2016 IEEE Congress on Evolutionary Computation (CEC) 2009

 that, with reasonable frequency, allows assessment of the con-
troller’s behavior under adverse conditions. The battery bank
parameters were selected as an initial guess to the minimum
capacity that could be used without being too small to supply
the energy needed over a reasonably long blackout period5 ,
and the initial conditions were selected to give the system
Fig. 3. Neural-State-Space Hybrid Model
a reasonable starting allowance without being too favorable.
The conversion efficiencies were selected to approximately
would have also been necessary to select a suitable learning match the actual losses expected when converting between
method for determining all the weights and select a topology two different types of energy67 . The controller parameters
for the best performance [9]. In the first stage of the model, were set to allow a suitable memory size to include the useful
the weights/terms were initialized in a way that prevents frequencies, battery lockout threshold, cyber attack probability
conflicting or irrational output states based on the previous threshold, and high frequency cutoff level.
internal values and state-space inputs. Given that the scale of
the system could not be allowed to affect the performance of TABLE I
I.C. S AND CONSTANTS
the controller, it was determined that normalizing the energy
flow values and separating positive and negative flow would Solar Value Units
be a more reliable [10]. This allows control outputs to be Area 8 m2
Production 305 W h/m2
determined by relative values of flow into and out of the Grid – –
system. A second stage was then applied as a function strictly limit 4 MWh
based on the inputs as a way of suppressing the stimulus of pblackout 0.2 %
pattack 0.1 %
the first stage, allowing for more refined actions and transitions Controller – –
and preventing instability due to overreacting to the inputs. The Memory length 48 hrs
controller output is generated as a unipolar outcome between Battery Lockout 10 %
Energy noise filter 10 1/hrs
0 and 1 from the results of the hybrid-state-space model. Fig. ATK Threshold 50 %
3 shows the hybrid model topological equivalent to the state- Battery Bank – –
space equation where the input array u affects the internal state Ebatt 32545 Wh
ΔM ax 3254.5 W h/Δt
x through the matrix B and the output state through matrix Leakage 0.32545 W h/Δt
−D. Internally, the state x affects the next state through matrix EI.C. 16272.5 Wh
A and determines the output through an identity matrix C. Due ΔEI.C. 0 W h/Δt
Conversion efficiency – –
to the unipolar nature of the neurons the outputs of x and y are DC ⇒ DC 93 %
held between 0 and 1, preventing undefined outputs, drifting, DC ⇒ AC 95 %
and damped/delayed response. AC ⇒ DC 60 %
AC ⇒ AC 99 %
IV. S IMULATION
During simulations, all data thought to be useful for ei-
ther aggregated simulations, re-running of simulations with B. Results and Analysis
identical conditions, or for later analysis, was logged for The results have been analyzed using a simple comparison
convenience. As there was a need to store data as a part of the proposed controlled system with three other system
of the controller’s memory, the controller logged its data setups: (1) Traditional power system without renewable or
separately and updated its memory with each iteration while stored energy, (2) Basic system with renewable energy without
being restricted to a maximum size. For ease of access, all the storage, and (3) Isolated system (with renewable and stored
data was stored in csv file format. The simulation completed energy, but not connected to a power grid). The comparisons
1000 iterations of data, was run on the Octave 4.0.0 GUI on a have been performed in terms of system energy losses (mostly
Windows 10 OS (Intel core i7-4710HQ @ 2.50Ghz and 16Gb internal line losses), energy deficiency (insufficient power
RAM), and was repeated 100 times to get the final analytical for household demand), and cash flow (monetary gains and
results, requiring approximately one minute per simulation losses). For better visual comparison, these values have been
run. normalized based on the largest magnitude for each method to
A. Selection of Initial Conditions and Constants give a percentage of the largest value for given comparison.
This analysis has been repeated while changing the solar panel
The initial conditions and constants were selected as a
reasonable approximation for the initial state and capacity of 5 Derived from Canada’s household electricity consumption at:
the system and its corresponding subsystems. The solar panel [Link]
6 AC to DC source: [Link]
was chosen to be of a reasonable physical size and quality.
inverters/
The grid parameters were selected so as to not cause the 7 DC to DC source: [Link]/id/93-efficient-DC-to-DC-
simulation to be too pessimistic but still provide a simulation Converter/

2010 2016 IEEE Congress on Evolutionary Computation (CEC)

Fig. 4. Power Line Flow (Blue), Solar Energy (Violet), Household Energy Fig. 6. Grid data: Line Energy (Red), Blackout (Blue) and Attack (Red), and
(Green) Price (Green)

Fig. 5. Battery Level (red) and System Losses (Blue)

Fig. 7. Regulators: Reg1 (Blue), Reg2 (Red), Reg3 (Violet), Reg4 (Green)

area to observe how the controller’s performance is affected

by the renewable energy generation scale.
The simulation show that the controller properly reacts to
the conditions while ignoring the cyber-attacks (Fig. 4 shows
the Energy exchange for each source/drain). The system also
retains enough energy to survive all blackouts that occurred
during the simulation period. Based on Fig. 5, which shows
results specifically for the household consumption data used
in the simulation, a battery bank capacity of approximately
20,000 Wh would be sufficient to survive a blackout and
accommodate the additional energy stored for the purpose of
selling at a later time. From Fig. 7, it can be seen that the
controller is properly switching between states in accordance Fig. 8. Attitude Towards Risk of Cyber-attack
to the best actions that should be taken for the corresponding
environmental conditions of the time shown in Fig. 6. Fig.
8 shows that the error-based probabilistic model chosen to that the constant exchanges by the controller caused much
identify the probability of a cyber-attack is reliable due to the larger amounts of energy loss than any other system. However,
near piece-wise linear input-output relation. It is also shown it never allows a blackout to affect the consumer and is still
by the large regions of certainty, that the error-based model is able to reduce the monetary losses even when there is no
more sensitive than needed. Fig. 9 indicates that the controller renewable energy available. From Table III, it can be seen
is extremely confident that the system can survive a blackout that just using the controller with a battery bank would give
based on the large region of certainty in avoiding a blackout 8% savings for energy costs despite the increased energy
(indicated as an output value of zero). consumption due to line losses. From Table IV, it is understood
The performance of the controlled system relative to the that the line losses increase because there is more energy
alternative uncontrolled systems used in the analysis shows moving within the system. Although the monetary savings

2016 IEEE Congress on Evolutionary Computation (CEC) 2011

 TABLE IV
R ENEWABLE E NERGY M AXIMUM IS SET TO 2440 W H FOR 500 HOURS

Set up ELoss [kWh] EDef icit [MWh] Cost [$]

(1) 14.721 0.010 4.6077
(2) 4.781 0.014 23.580
(3) 36.123 140.10 0
(4) 300.250 0 15.035
[%] [%] [%]
(1) 4.9 0.0072 20
(2) 1.6 0.0097 100
(3) 12 100 0
(4) 100 0 64

TABLE V
Fig. 9. Attitude Towards Risk of a Blackout R ENEWABLE E NERGY M AXIMUM IS SET TO 4880 W H FOR 500 HOURS

Set up ELoss [kWh] EDef icit [MWh] Cost [$]

(1) 26.689 0.012 -14.199
are not as good as for the system without battery, there is an (2) – – –
obvious improvement in the cost reduction as compared to the (3) 58.665 44.955 0
traditional system. Comparing Tables V and VII, it is apparent (4) 534.6 0 7.2961
[%] [%] [%]
that there is a cross-over point where the controller could
(1) 5.0 0.027 -60
achieve net energy cost of approximately zero. Re-running the (2) – – 100
analysis with an estimate that a 6100 Wh renewable energy (3) 11 100 0
source would be close to the cross-over point, Table VI was (4) 100 0 31
acquired. From this information, it can be estimated that a
renewable energy source with a maximum output of 6200 Wh
TABLE VI
would result in a net cost of approximately 0$ per Wh for R ENEWABLE E NERGY M AXIMUM IS SET TO 6100 W H FOR 500 HOURS
this household. With the data from Table VI and Table VII
as well as the known non-linear trend of the cost for the Set up ELoss [kWh] EDef icit [MWh] Cost [$]
(1) 32.757 0.0116 -23.487
proposed controller, it can be estimated that the monetary gain (2) 4.7814 0.0169 23.458
of the controlled system approaches 25% of the gain from the (3) 70.367 1.7426 0
renewable system without a battery. (4) 586.35 0 -0.56654
[%] [%] [%]
(1) 5.6 0.67 -100
TABLE II (2) 0.82 0.97 100
S ETUP L EGEND (3) 12 100 0
(4) 100 0 -2.4
(1) No Battery (capacity = 0)
(2) Traditional setup (ESol = 0 & capacity is 0)
(3) No Power Line
(4) Proposed Controller TABLE VII
R ENEWABLE E NERGY M AXIMUM IS SET TO 7320 W H FOR 500 HOURS

Set up ELoss [kWh] EDef icit [MWh] Cost [$]

(1) 38.85 0.0102 -33.326
TABLE III (2) – 0.0150 –
R ENEWABLE E NERGY M AXIMUM IS SET TO 0 W H FOR 500 HOURS (3) 82.179 0.1998 0
(4) 630.9 0 -8.0606
Set up ELoss [kWh] EDef icit [MWh] Cost [$] [%] [%] [%]
(2) – – – (1) 6.2 5.1 -100
(3) 23.668 235.42 0 (2) 0.76 7.5 71
(4) 48.091 0 21.476 (3) 13 100 0
[%] [%] [%] (4) 100 0 -24
(2) – – 100
(3) 49 100 0
(4) 100 0 92

rely on renewable energy and that the large number of repeated

In the first setup in Table III is omitted because there simulation results produced similar values.
is no renewable energy generation and thus the results are
identical to second setup. The values for the second setup were V. D ISCUSSION
partially omitted in Tables III, V, and VII because the results
were nearly identical to the first simulation values shown in Some comparisons have been made to other proposed
Table IV (values of Table VI are shown as an example). This controllers for relative performance. Using system model (1)
similarity is due to the fact that the traditional setup did not in Table II as a reference point, the proposed controller has

2012 2016 IEEE Congress on Evolutionary Computation (CEC)

a savings of at least 25%8 in uncertain conditions while As with many new, research-based technologies, the pro-
the average savings from the Near-Optimal Model-Based posed controller has a large number of parameters, functions
controller[11] was estimated to be 27%9 in ideal conditions for and initial conditions. These can be optimized to improve the
the model parameters they used. The nature of the proposed performance of the controller and make the simulations more
controller design is similar to scheduling based models[12], realistic. Optimizing internal components of the controller
[13], [14] with the exception that it independently schedules has yet to be explored and could potentially offer major im-
energy exchanges with the grid based on current and past data provements of system performance. One major consideration
instead of shifting the load schedule and/or planning from an is to make better use of the full range of the controllers
upstream load forecast. output values and reduce the repetitive on-off actions of each
regulator. Finally, thorough comparison to other controllers
VI. C ONCLUSION
should be performed to gauge its relative performance.
The proposed controller meets can effectively manage flow
of energy in a customer-installed, grid-connected PV system VII. ACKNOWLEDGMENTS
with energy storage. Additionally, it allows identification of Support for this project was generously provided by the
the minimum system requirements that provide the desired Natural Sciences and Engineering Research Council (NSERC)
results. The system model develops a sufficiently realistic of Canada.
and flexible environment to reveal what is expected and
needed of the controller. The controller design draws from R EFERENCES
several constituent fields of knowledge, effectively integrates [1] Z. Li, M. S. Drew, and J. Liu, Fundamentals of Multimedia, 2nd ed.
Switzerland: Springer, 2014.
corresponding concepts, and successfully applies the resulting [2] B. Bass, Advantages and Disadvantages of Forecasting Methods
system for energy management in the smart grid. The analysis in Production and Operations Management. Houston Chronicle,
results allow for comparison of the controller’s performance Jan 2016. [Online]. Available: [Link]
advantages- disadvantages- forecasting- methods- production- operations-
to several other common system designs and for identification [Link]
of a relation between the controller’s performance and the [3] M. Peterson, An Introduction to Decision Theory. New York: Cam-
scale of renewable energy used. Being aware of the controller’s bridge, 2009.
[4] N. Turner and J. Powel, “Probability, random events, and the
potential benefits and problems, the results of this study will mathematics of gambling,” Tech. Rep., 1998. [Online]. Available:
aid in determining what future course of action will be needed [Link]
to improve the controller’s performance. [Link]
[5] M. laws site, “All the laws of murphy in one place,” Jan 2016. [Online].
Based on the supply and demand principles in economics Available: [Link]
for an Oligopoly, the controller performs extremely well when [6] J. I. Kassinove, “Development of the gambling attitude scales:
there are few people using this type of control system. As Preliminary findings,” Journal of Clinical Psychology, vol. 54, no. 6,
pp. 763–771, 1998. [Online]. Available: [Link]
more people start it, the price would become more heav- 1097-4679(199810)54:6763::AID-JCLP2[Link];2-I
ily influenced by the near synchronous supply and demand [7] R. C. Dorf and R. H. Bishop, Modern Control Systems, 12th ed. Upper
fluctuation generated by the consumers buying and selling at Saddle River, New Jersey 07458: Pearson, 2011.
[8] J. J. Craig, Introduction to Robotics Mechanics and Control, 3rd ed.
approximately the same time. As the supply is generated to Upper Saddle River, New Jersey 07458: Pearson, 2005.
meet or exceed demand, prices during that time would drop; [9] M. T. Hagan, H. B. Demuth, and O. D. Jesús, “An introduction to the
conversely, as the demand is generated to meet or exceed the use of neural networks in control systems,” Tech. Rep., 2016. [Online].
Available: [Link]
excess supply, prices during that interval would increase. In [10] N. N. Lab, “How to standardize data for neural networks,”
the long run, this action can be seen to oppose price peaks Visual Studio Magazine, Jan 2014. [Online]. Available: https://
and troughs and press for a more consistent energy pricing. [Link]/articles/2014/01/01/how-to-standardize-data-
[Link]
A potentially critical deficiency of this controller, which [11] Y. Wang, X. Lin, and M. Pedram, “A near-optimal model-based control
relies on internal memory, is that it cannot correct long algorithm for households equipped with residential photovoltaic power
term cyber-attacks that change the pattern over time. In the generation and energy storage systems,” IEEE Transactions on Sustain-
able Energy, vol. 7, no. 1, pp. 77–86, Jan 2016.
event that the cyber-attack price pattern drastically differs [12] A. Molderink, V. Bakker, M. G. C. Bosman, J. L. Hurink, and G. J. M.
from the actual price pattern and is abruptly stopped, it can Smit, “Management and control of domestic smart grid technology,”
cause sudden and persisting instability because the controller IEEE Transactions on Smart Grid, vol. 1, no. 2, pp. 109–119, Sept
2010.
remembers the wrong pattern. Additionally, if the price data [13] M. B. T. Molina, T. Gafurov, and M. Prodanovi, “Proactive control for
takes on an erratic nature (like a stock market), this controller energy systems in smart buildings,” in Innovative Smart Grid Technolo-
can still be expected to perform reasonably well. However, gies (ISGT Europe), 2011 2nd IEEE PES International Conference and
Exhibition on, Dec 2011, pp. 1–8.
it will not be able to exploit the high frequency values, nor [14] E. Matallanas, M. Castillo-Cagigal, A. Gutirrez, F. Monasterio-Huelin,
would it be able to properly detect cyber-attacks (although in E. Caamao-Martn, D. Masa, and J. Jimnez-Leube, “Neural network
such cases it is expected that a cyber-attack would have little controller for active demand-side management with {PV} energy in
the residential sector,” Applied Energy, vol. 91, no. 1, pp. 90 – 97,
to no effect). 2012. [Online]. Available: [Link]
8 When
pii/S0306261911005630
the solar panel is very small to very large.
9 Average is estimated from Fig. 5 over all points using ’No storage’ and
’Proposed’ data. The specifics of the panel size was not specified.

2016 IEEE Congress on Evolutionary Computation (CEC) 2013