Professional Documents
Culture Documents
Manlapaz 2023 IOP Conf. Ser. Earth Environ. Sci. 1199 012006
Manlapaz 2023 IOP Conf. Ser. Earth Environ. Sci. 1199 012006
Manlapaz 2023 IOP Conf. Ser. Earth Environ. Sci. 1199 012006
Environmental Science
Abstract. Distributed generation represents a paradigm shift from the traditional electric grid
to localized generation of electric power along with the capability of incorporating renewable
energy (RE) sources into the energy mix. Responding to the need for sufficient analysis,
simulation, and study of feasibility of distributed generation, this study aims to design a
hybrid energy system for a university building and analyze its economic benefits. The viability
of existing load forecasting methodologies for energy systems is also presented in this paper.
The energy system design was determined through predictive modeling of the load profile of
the building using historical data and optimization using machine learning methods, namely
auto-regressive integrated moving average (ARIMA) and long short-term memory (LSTM).
Simulations were run in HOMER Pro. Results show that a grid-connected solar photovoltaic
(PV) system installed on the roof coupled with an energy storage system (ESS) will provide
the most economic benefits because it yields a reduced cost of energy (COE) per kilowatt-hour
(kWh) for the building. This study shows that in line with efforts to transition towards clean
energy, hybrid energy systems using RE not only have economic benefits, but can also ensure
energy security and environmental sustainability.
1. Introduction
With affordable and clean energy being part of the United Nations’ sustainable development
goals, countries seek to increase renewable energy (RE) generation in their energy mix by 2030
[1]. Distributed energy is a growing segment of the energy industry, representing a paradigm
shift from remote central power plants toward more localized energy generation—especially in
cities, communities, and campuses. Integrating renewable energy sources in the form of solar
photovoltaic (PV) systems is seen as a possible and economic solution to tackle the issues related
to climate change. System sizing and modeling can be done using historical load profiles. In the
event of unavailability of real-time data or the inability of measuring actual energy consumption,
Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
2nd ASEAN International Conference on Energy and Environment IOP Publishing
IOP Conf. Series: Earth and Environmental Science 1199 (2023) 012006 doi:10.1088/1755-1315/1199/1/012006
it is possible to simulate the desired load profile by using historical data through predictive
modeling [2].
This paper aims to simulate a grid-tied PV system in a university building using HOMER
Pro, a software capable of performing techno-economic and sensitivity analyses to determine the
most optimal distributed generation system and component sizing based on the historical power
consumption and the predictive models resulting from the historical data of the building. In this
study, predictive modeling was developed using two load forecasting methods: auto-regressive
integrated moving average (ARIMA) and long short-term memory (LSTM). Historical data was
used to determine the load profile that served as the basis of the HOMER Pro simulation. This
study can provide insights that contribute towards the goal of clean and affordable energy, as
institutions such as universities transition into greener energy systems.
2. Literature review
2.1. Short-term load forecasting methods for microgrid systems
Different forecasting methodologies exist beyond that of surveying to estimate the load profile
used for microgrid applications. Electric load forecasting (ELF) in particular serves as an
important tool for planning, control, and operation of power systems. Most ELF methods
are used for transmission systems, while some are utilized for distribution systems. Marzooghi
et al. used load forecasting for microgrids to compare the performance of three well-known
short-term load forecasting (STLF) methods, namely seasonal auto-regressive integrated moving
average with exogenous variables (seasonal ARIMAX), neural networks (NN), and wavelet neural
networks (WNN) [2]. STLF methods were used because they are designed to predict the load
one week ahead thereby becoming a significant tool for day-to-day operations of power systems.
The study used a four-year historical data set obtained using gross metering, which
independently recorded the load demand and solar generation of select premises in an electricity
network area. Mean absolute error (MAE), maximum error (ME), and regression (R) between
actual load values and predicted outputs were used to gauge which method has the better
predictive ability. Results shown in Table 1 indicate that the WNNs-based model had the
highest accuracy for STLF. Meanwhile, seasonal ARIMAX and NNs have shortcomings in terms
of predicting load trends in microgrid applications since they failed to appraise the behavior of
interlaced high- and low-frequency loads.
2
2nd ASEAN International Conference on Energy and Environment IOP Publishing
IOP Conf. Series: Earth and Environmental Science 1199 (2023) 012006 doi:10.1088/1755-1315/1199/1/012006
combined and subjected to constraints such as minimum renewable fraction, operating reserve,
wind, and solar renewable output. These combinations include grid only, PV-grid combination
with battery, and PV-grid combination without battery. HOMER Pro can determine which
setup produced the least COE through optimization and sensitivity analyses. Results show that
the PV-grid combination without battery setup yielded the least cost of energy, indicating that
the installation of PV panels are economically beneficial.
3. Methodology
3.1. Predictive modeling
The predictive models for both ARIMA and LSTM were developed using a raw data set
containing monthly energy consumption of the university building in kWh from the years 2016
to 2019. Fig. 1 shows the flowchart of the development ARIMA model. In this model, the data
set was preprocessed such that the first 36 months were allotted for the training data set while
the remaining 12 months were assigned to the testing data set. To make sure that the data
set acted independently from the typical seasonal behavior observed in annual energy use, the
stationarity of the data set was checked. If the data set is not stationary, that means that it
is still under the influence of seasonal behavior and thus it was transformed into a stationary
data set using differencing. Once the data set is stationary, the best ARIMA parameters were
identified using the grid searching method which determines the best values that the ARIMA
model can output based on the input data set. The resulting ARIMA model obtained through
the grid searching method were evaluated using performance metrics such as root mean square
method (RMSE) to determine the best ARIMA model to predict the year-long load profile of
the university building.
Utilizing the Keras library [5], the LSTM model on the other hand was developed by first
3
2nd ASEAN International Conference on Energy and Environment IOP Publishing
IOP Conf. Series: Earth and Environmental Science 1199 (2023) 012006 doi:10.1088/1755-1315/1199/1/012006
Table 2. Performance indicators used in comparing the accuracy of LSTM and ARIMA.
Performance
Meaning Equation Where
indicator
n = number of samples
Root mean
q P
1 n 2
RMSE n (f
i=0 i − o i ) fi = forecast values
squared error
oi = actual values
n = number of times the
Mean absolute 1 P n At −Ft summation iteration happens
MAPE n t=1 At
percentage error At = actual values
Ft = forecast values
RSS = Sum of squares
Coefficient of
R2 1 − RSS
T SS of residuals
determination
T SS = total sum of squares
normalizing the same raw data set used in the development of the ARIMA model, resulting in
a data set with values ranging from zero to one. Similar to the ARIMA model, the data set was
again preprocessed such that the first 36 months were allotted for the training data set while the
12 remaining months were assigned to the testing data set. It is important to normalize the data
set since LSTMs are sensitive to the scale of the input data such as in cases where the LSTM
uses either the sigmoid or tanh activation function [6]. Since LSTM is a variant of recurrent
neural networks, the LSTM weights were initialized and subsequently used in the training of the
neural network. A notable hyperparameter in developing the LSTM model is the the number
of epochs. This refers to the number of iterations that the LSTM model carries out and the
entire training data set is being processed per iteration [7]. The number of epochs was set to 50,
100, 300, 500, and 1,000, respectively, to determine which LSTM model has the best fit to the
testing data set. Once training the data was completed, results were denormalized, producing
an LSTM model that predicted a year-long load profile of the university building. Fig. 2 shows
the flowchart of the development of the LSTM model.
After the models were obtained, the root mean squared error (RMSE), mean absolute
percentage error (MAPE), and the coefficient of determination (R2 ) were used to identify the
most appropriate model among the LSTM and ARIMA models to be used for the microgrid
simulation. The RMSE defines how far the predicted values are from the actual values on
average. The MAPE represents the prediction accuracy as a ratio of the forecasting method
[6]. Lastly, R2 describes how well the actual data is replicated by the model according to the
proportion of total variation of results from the model [8]. The equations for all three metrics
are shown in Table 2.
4
2nd ASEAN International Conference on Energy and Environment IOP Publishing
IOP Conf. Series: Earth and Environmental Science 1199 (2023) 012006 doi:10.1088/1755-1315/1199/1/012006
the monthly average solar global horizontal irradiance (GHI) data of the university building. The
data was imported from an NREL database found in HOMER Pro.
Figure 4. Solar GHI data at the location of the university building (source: NREL, HOMER
Pro).
5
2nd ASEAN International Conference on Energy and Environment IOP Publishing
IOP Conf. Series: Earth and Environmental Science 1199 (2023) 012006 doi:10.1088/1755-1315/1199/1/012006
6
2nd ASEAN International Conference on Energy and Environment IOP Publishing
IOP Conf. Series: Earth and Environmental Science 1199 (2023) 012006 doi:10.1088/1755-1315/1199/1/012006
energy while simultaneously taking economics into account. An ”ideal” grid-tied PV system
should be able to satisfy the following benchmarks:
• The levelized cost of energy (LCOE) is less than the commercial grid rates provided by
MERALCO, the local electric power distributor. This benchmark was utilized in the study
of Sharma and Kaur [2]. LCOE shows the monthly cost of energy if PV and other renewable
energy sources are installed. The latter will provide useful electrical energy on top of what
the university building would receive from the main utility grid.
• The system yielded the lowest total net present cost (NPC). The NPC of a system is the
present value of all the costs the system incurs over its lifetime, minus the present value
of all the revenue it earns over its lifetime. Costs include capital costs, replacement costs,
operations and maintenance (O&M) costs, fuel costs, emissions penalties, and the costs of
buying power from the grid [11]. Revenues include salvage value and grid sales revenue.
HOMER Pro calculates the total NPC by summing the total discounted cash flows in each
year of the project lifetime. The NPC can be used to compute the annual cost of the entire
grid-tied PV system. This is done by dividing the NPC by the number of years of the
project, which is twenty-five years.
7
2nd ASEAN International Conference on Energy and Environment IOP Publishing
IOP Conf. Series: Earth and Environmental Science 1199 (2023) 012006 doi:10.1088/1755-1315/1199/1/012006
Figure 7. The actual load versus the predicted load using the ARIMA and LSTM models.
Figure 8. Cost of energy from the HOMER Pro simulation of the solar PV and battery system.
8
2nd ASEAN International Conference on Energy and Environment IOP Publishing
IOP Conf. Series: Earth and Environmental Science 1199 (2023) 012006 doi:10.1088/1755-1315/1199/1/012006
5. Conclusion
Machine learning can be used to model the energy consumption of a building using historical
monthly consumption data. Two predictive models, namely ARIMA and LSTM, were used
in this study. The LSTM model achieved higher accuracy than the ARIMA model. The
combination of solar PV and energy storage system resulting from the HOMER Pro simulations
is the optimum energy system for the university building because the predicted cost of energy
is reduced by 10.96% based on the average monthly electricity rate. The proposed distributed
generation system utilizes renewable energy and storage. It is therefore environmentally friendly
and it can provide a certain level of autonomy from the utility grid.
6. Acknowledgments
The authors would like to thank the School of Science and Engineering (SOSE) Sanggunian
Academic Subsidy (SAS) and the University Research Council (URC) of the Ateneo de Manila
University for their support of this study.
References
[1] 2022 Goal 7 — Department of Economic and Social Affairs (United Nations)
[2] Marzooghi H, Emami K, Wolfs P and Holcombe B 2018 Short-term Electric Load Forecasting in Microgrids:
Issues and Challenges 2018 Australasian Universities Power Engineering Conference (AUPEC)
[3] Sharma H and Kaur G 2016 Optimization and simulation of smart grid distributed generation: A case study
of university campus 2016 IEEE Smart Energy Grid Engineering (SEGE)
[4] Hamatwi E, Nyirenda C and Davidson IE 2018 Cost Optimization and Design of a Hybrid Distributed
Generation System for a DC Microgrid 2018 IEEE PES/IAS PowerAfrica
[5] Mpawenimana I, Pegatoquet A, Roy V, Rodriguez L and Belleudy C 2020 A comparative study of LSTM
and ARIMA for energy load prediction with enhanced data preprocessing 2020 IEEE Sensors Applications
Symposium (SAS)
[6] Cui C, He M, Di F, Lu Y, Dai Y and Lv F 2020 Research on Power Load Forecasting Method based on LSTM
model 2020 IEEE 5th Information Technology and Mechatronics Engineering Conference (ITOEC)
[7] Parizad A and Hatziadoniu C 2021 Deep Learning Algorithms and Parallel Distributed Computing Techniques
for High-Resolution Load Forecasting Applying Hyperparameter Optimization IEEE Systems Journal. 16
3758–3769.
[8] 2022 2.5 - The Coefficient of Determination, r-squared — STAT 462.
[9] Elliott D 2001 Philippines Wind Energy Resource Atlas Development Business and Investment Forum for
Renewable Energy and Efficiency in Asia and the Pacific Region
[10] 2022 Google Maps location of PLDT-Convergent Technologies Center, Ateneo de Manila University (Google
Maps)
[11] 2022 Total Net Present Cost (Homer Pro 3.14)
[12] 2022 Rates Archives (Meralco)