Forecasting of Energy Consumption in the

Philippines Using Machine Learning Algorithms

Erru Torculas, Earl James Rentillo, and Ara Abigail Ambita

Division of Physical Sciences and Mathematics

University of the Philippines Visayas
Miagao, Iloilo, Philippines
{egtorculas, eqrentillo, aeambita}@up.edu.ph

Abstract. In the Philippines, usage of energy has been steadily in-

creasing over the years, however, the advent of the COVID-19 pandemic
brought about unforeseen changes to these parameters. By using machine
learning algorithms, energy predictions can be more properly assessed,
and corresponding measures can be put into place. The Monthly and
Quarterly Market Assessment Report of Wholesale Electricity Spot Mar-
ket (WESM), governed by Philippines Electricity Market Corporation
(PEMC)’s data was analyzed using four machine learning algorithms,
namely, Random Forest, XGBoost, Linear Regression, and Support Vec-
tor Regression (SVR) to determine the best algorithm in predicting en-
ergy consumption within the pre-pandemic and pandemic periods. It
was found that the Pre-pandemic (Period 1) data was most accurately
predicted by the XGBoost model, having a Root Mean Square Error
(RMSE) of 366.691 and Mean Percentage Error (MAPE) of 0.044, while
the Pandemic (Period 2) data was most accurately predicted by the Ran-
dom Forest Model from its RMSE of 687.665 and MAPE of 0.061. While
the poorest performing model for both these periods was the SVR, get-
ting an RMSE of 431.366 and 982.202 to the respective periods. The
results show how developing tree-based predictive models, XGBoost and
Random Forest Models, are significant in forecasting energy consump-
tion in the Philippines, and is therefore also beneficial for future studies
that aim to engage in crafting energy conservation and efficiency policies
for economic growth.

Keywords: Energy consumption · Forecast · Machine learning

1 Introduction
Energy has always played a vital part in modern society. In fact, consumption of
such a resource exhibits a continuous rise to developing economies [1]. It performs
as a societal backbone in that its utilization is deemed essential from fueling
activities, from mundane ones up to crucial operations. Exemplified by sectors
such as residential, commercial, and industrial using electricity as a primary
source for operation, energy consumption is thus generally regarded as an index
of standard of living [2].
2 E. Torculas et al.

In the Philippines, the average energy consumption per person reached a peak
of 5,205 kWh in 2019 and has been relatively increasing in each year prior. Some
notable trends which follow this annual increase include country energy con-
sumption (peaked at 563 TWh in 2019), country electricity generation (peaked
at 108.27 TWh in 2021), and average electricity generation per person (peaked
at 961 KWh in 2019) [3, 4].
These statistics, as authored by Ritchie et al., go to show how studying en-
ergy consumption in the Philippines proves to be relevant, especially that the
population’s access to electricity is at an all-time high at 96.84% [3]. Further-
more, there is a lack of datasets in developing economies to be able to predict
future demands in electricity [1]. This is where tools such as regression analysis
become useful to bridge the gap of such data deficiency [5]. It can be used to
generate predicted data and thus arrive at a plausible conclusion if energy de-
mands would still be met in the years to come [4]. More importantly, knowing
the pattern of consumption and predicting the trend of energy consumption can
work to determine priorities in the process of taking decisions on a sustainable
urban environment for the energy sector in the Philippines, and therefore be
a reasonable metric to improve energy efficiency of industrial and commercial
industries for better policy-making and economic growth.
This paper will assess di↵erent machine learning algorithms in forecasting
time series of energy consumption in the Philippines and compare the accu-
racy of each forecasting. Through this, allowing the researchers to investigate on
which approach is likely to be accurate for future implementation of energy con-
sumption prediction. It would be beneficial to concretizing findings established in
relevant studies, as well as exploring how these build on in more specific contexts
such as that of the Philippines.

2 Literature Review

In a study by Shin and Woo [6], For estimating energy usage in Korea, three ma-
chine learning methods were compared: Random Forest (RF), XGBoost (XGB),
and Long Short-Term Memory (LSTM). The researchers concluded that machine
learning can be used to anticipate energy use, albeit classic econometric methods
beat machine learning in some circumstances. Machine learning demonstrated
advantages in dealing with unexpectedly irregular time series data [6].
Similarly, Rambabu et. al [7] exhaustively delved into the prediction and
analysis of household energy consumption through machine learning algorithms
for energy management. Their paper focused on predicting household energy
consumption where models are trained by using various machine learning al-
gorithms such as Linear Regression, Lasso Regression, Random Forest, Extra
Tree Regressor, XG Boost, etc. It must be noted that patterns of household
energy consumption are observed by the constant changing of di↵erent factors
namely, temperature, humidity, an hour of the day, etc. The researchers’ find-
ings suggested that tree-based models give the best results among the rest of the
machine-learning approaches used [7]. The evaluation metric used was R square
 Forecasting Energy Consumption in the Philippines 3

which can be utilized to gauge how much variance in the dependent variable can
be predicted.
A case study in Malaysia also discussed energy consumption prediction by
using three methodologies of machine learning, specifically, Support Vector Ma-
chine, Artificial Neural Network, and k-Nearest Neighbor [8]. These approaches
were proposed for the algorithm of the predictive model.The paper explored a
great insight into real-life applications where the researchers used two tenants
from a commercial building as proponents of their case study. The metrics of
evaluation used in the paper are compared based on RMSE, NRMSE, and MAPE
metrics [8].
These studies show how machine learning algorithms can be an indispens-
able tool for forecasting power consumption to be able to achieve sustainable
and e↵ective systems in their respective contexts. However, testing the pre-
dictive models they have created into the context of the Philippines may not
show consistent results due to the disparity and nature of the data available for
analysis. For Shin and Woo [6], three machine learning algorithms were more
extensive, being modeled from January 1997 to June 2021. Rambabu et. al [7],
on the other hand, utilized a more continuous dataset on power consumption for
analysis, that being measured at 10-minute intervals over 4.5 months in terms
of house temperatures and humidity. Similarly, the dataset of Salam et al. [9]
for machine learning algorithms for power consumption prediction in Tetouan,
Morocco followed the same interval collection from January 1, 2017 to December
31, 2017, measured in KW.
Thus, it is important to note that data limitations arise as an additional
limitation to model energy demand of developing countries, one of which is
the Philippines. In fact, aside from data limitations, Bhattacharyya and Tim-
ilsina [10] posited that these countries face such a challenge due to institutional
capacity and the specific characteristics of their energy systems as well. In their
study, they have compared di↵erent energy demand forecasting models and their
criteria, including type, purpose, approach, and geographical coverage, among
many others. These models varied in their capabilities and coverage, where some
were more suitable for general analysis while others were more complex and spe-
cific.
These go to show that failure to address these challenges can yield inaccu-
rate results, ultimately misguiding policy recommendations. Therefore, there is
a need to improve energy demand modeling tools and institutional capacities
in developing countries, one that is tailor-made to address the consumption be-
havior of a country given its income group, location, and more importantly, the
data available to analyze and model energy consumption. Hence, closing the
energy consumption gap through e↵ective policies is critical for a sustainable
and successful future. Policymakers may minimize climate change, reduce envi-
ronmental impact, and promote sustainable development by prioritizing energy
efficiency and renewable sources. Diversifying the energy mix and investing in
resilient infrastructure improves energy security by reducing risks and assuring
a consistent supply especially in developing countries like the Philippines.
4 E. Torculas et al.

3 Data and Methodology

3.1 Data

The data used throughout this project is consolidated data from 2014 to the
second quarter of 2022 of the Monthly and Quarterly Market Assessment Report
of the Wholesale Electricity Spot Market (WESM) governed by the Philippine
Electricity Market Corporation (PEMC) [11].
Considering the caveats, the dataset was mostly curated from the summary
reports of WESM. So, the total energy consumption data for 2016 is only avail-
able through a bar graph without any numeric label accompanying the bar in
each month. The dataset also utilized two features which are the Date, specified
by month and year, and Total Energy Consumption (in GWh).

3.2 Methodology

Time Period Analysis. The analysis is divided into two, namely: Pre-pandemic
and Pandemic periods, additionally labeled as Period 1 and Period 2, respec-
tively. The separation of time periods serves as the basis of the efficiency of the
models when practiced on a sudden shift of regulations brought by the height of
the pandemic [6].
The pre-pandemic period is composed of a rather linear growth in terms of
energy consumption and saw a rise from a starting point of 4422 GWh (Jan
2014) to an endpoint of 6224 GWh (Mar 2020) and a peak point of 7697 GWh
(Jun 2019). Following this, Figure Period 1 training and test data where it made
use of “January 2014 to March 2020” in which the training data started from
“January 2014 to June 2018”, while the rest was the testing data “July 2018 to
March 2020”.
Furthermore, it is evident that the data from April 2020, which was the first
month of lockdown and in turn, led to a staggering decrease in energy consump-
tion (3760 GWh from the previous month’s 6224 GWh) as non-essential build-
ings were cut o↵ to minimize losses, amongst other reasons. Power generation
was drastically reduced due to slowdown in terms of economical manufacturing
production, and the energy consumption both in commercial and infrastructure
sector by 2.0% [6]. Hence, Period 2 or the Pandemic period used data spanning
from “January 2014 to June 2022”, with “January 2014 to March 2020” data
being used as training data, the period wherein it witnessed the drastic change
in energy consumption. Meanwhile, the testing data used “April 2020 to June
2022”.

Machine Learning Models. To evaluate the data concerning the scope of

the project, four machine learning algorithms were run and compared with each
other for accuracy. The models that were implemented in this paper are: Random
Forest (RF) Model, XGBoost, Linear Regression, and lastly, Support Vector
Regression (SVR).
 Forecasting Energy Consumption in the Philippines 5

Random Forest (RF) can be applied to time series forecasting by converting

the time series data into a format suitable for supervised learning and using
a specific method called walk-forward validation to evaluate the model. This
is necessary because using k-fold cross validation on the model would produce
overly optimistic results [6, 7]. In fact, RF has already been utilized by facility
managers for further monitoring and improving of their respectve buildings’
energy efficiency [12].
Meanwhile, XGBoost has been shown to be e↵ective in a variety of tasks,
including time series forecasting. Due to its popularity in regression problems,
it is said to hasten performance in which it utilizes parallel processing. Not only
that, but it also performs well in small datasets and averts overfitting [6, 7].
Linear regression is a simple but powerful technique for forecasting. It is
e↵ective because it makes a strong assumption about the relationship between
the input variables and the output variable, which allows it to make reliable
predictions even when the data is noisy or there are missing values. In fact,
linear regression is fast and easy to implement, making it a popular choice for
many forecasting tasks [7, 14].
Lastly, Support Vector Regression (SVR) is a type of support vector machine
(SVM) that is often used for forecasting because it can handle data with multiple
features and can make predictions for continuous target variables [8, 12, 14]. In
the case of regression, the classes are continuous, and the goal is to find the
hyperplane that best fits the data. Not only that, SVR models can also be
trained relatively quickly, making them efficient for use in forecasting tasks.
These models were selected because they yield a significant analysis and
meaningful information in a predictive approach based from the previous pub-
lished related works [6–8]. Not only that, this paper revolves on finding the accu-
rate model to implement in terms of forecasting energy consumption, therefore
regression is advantageous in capturing associations and relationships between
forecast variable of interest and predictor variables [13, 14].

3.3 Evaluating Forecast Accuracy

Since the result of the focus of the project is regression analysis, in the con-
text of prediction driven models, the most widely adopted reliability analysis
indicators are Root Mean Square Error (RMSE) and Mean Absolute Percentage
Error (MAPE). RMSE measures the di↵erences between the predicted and ac-
tual values and therefore a means to measure the quality of fit between the actual
data and predicted model. It is preferred over the standard Mean Square Error
(MSE) since it is a smaller value and can be compared more straightforwardly.
Furthermore, MAPE is one of the metrics of evaluation used in this paper since
it relatively measures how accurate the forecast system is. RMSE equation is
shown in Equation (1), while the equation of MAPE is shown in Equation (2)
r Pn
i=1 (x1,i x2,i )2
RM SE = (1)
n
6 E. Torculas et al.

n
1 X x1,i x2,i
M AP E = ⇥ 100 (2)
n i=1 x1,i

4 Results and Discussion

The group utilized the Sci-kit learn package to build Random Forest, XGBoost,
Linear Regression, and Support Vector Regression models. With that, the final
model was chosen based on the lowest root mean squared error (RMSE) value.
Table 1 summarizes the metric of evaluation for the forecasting. It compares
the test data RMSE and MAPE values of the machine learning models for two
di↵erent periods, Period 1 and 2.
Accordingly, the machine learning model that yielded the lowest RMSE for
Period 1 is the XGBoost model with a value of 366.691. Whereas, for Period 2,
the Random Forest model was regarded as the final model since it has a RMSE
of 687.665 and it has the least RMSE among the models implemented. Based
on the empirical data that was presented, the Support Vector Regression (SVR)
model accumulated the highest RMSE for both periods with values of 431.366
(Period 1) and 982.202 (Period 2) which in this case will not be regarded as the
final model because of it overfitting of data.

Table 1: Performance of the models by period

ML Models
Metric RF XGBoost Linear Reg SVR
Period 1 RMSE 422.737 366.691 411.578 431.366
MAPE 0.050 0.044 0.047 0.050

Period 2 RMSE 687.665 692.077 935.880 982.202

MAPE 0.061 0.061 0.123 0.131

Lewis [15] categorized the accuracy of the forecast based on the predictive
models’ MAPE values. Through this, the paper utilized Table 2 interpretation of
forecasting accuracy. In Period 1, all the models showed a high accurate forecast
with their MAPE values of 0.05 (5%), 0.044 (4.4%), 0.047 (4.7%), and 0.05 (5%),
for RF, XGBoost, Linear Regression, and SVR, respectively. Comparatively, the
Period 2 models forecast returned only good accuracy for both Linear Regression
and SVR interpreting that both models were not as high as RF and XGBoost’s
with 0.061 (6.1%). This goes to show that the latter models can be identified
as more accurate models for Period 2 than the former with 0.123 (12.3%) and
0.131 (13.1%) Linear Regression and SVR, respectively.
Overall, the metrics of evaluation showed the machine learning models that
yielded the least RMSE value are the XGBoost and Random Forest model,
Period 1 and 2, respectively. Both of these models were deemed Highly Accurate
 Forecasting Energy Consumption in the Philippines 7

Table 2: Interpretation of MAPE Results for Forecasting Accuracy.

MAPE-value Accuracy of Forecast
Less than 10% Highly Accurate Forecast
11% to 20% Good Forecast
21% to 50% Reasonable Forecast
More than 51% Inaccurate Forecast

in terms of their forecasting ability. On the other hand, Linear Regression and
Support Vector Regression (SVR) gained the highest RMSE amongst the models
in Period 1 and 2. Not only that, but both of these models forecasting accuracy
are simply Good Forecast as opposed to the aforementioned models with high
accuracy.

4.1 Philippine Energy Consumption Model Forecasts

(a) (b)

(c) (d)

Fig. 1: Random forest model comparison by period with prediction error.

As shown in Figure 1, both periods were compared in the Random Forest

model in which they are accompanied by their respective prediction errors. Fig-
ures 1a and 1b show actual values forecasted against the predicted values. Period
8 E. Torculas et al.

1 depicts that the predicted values were much lower than the actual values from
April 2019 to March 2020.
Whereas, the prediction in Period 2 slightly forecasted the actual values
between January 2021 until March 2022. April 2021 to September 2021 were
accurately predicted. The data of the Random Forest model can further be
observed through the prediction errors wherein the residual were quite loose
for Period 1 as compared to Period 2 which are compact given that there is
one outlier due to sudden decrease of energy consumption during the start of
lockdowns.
Figure 2 illustrates the XGBoost model comparison of Period 1 and 2 with
their respective prediction errors. Figures 2a and 2b show the actual values fore-
casted against the predicted values in which Period 1 depicts that the predicted
values were much lower than the actual values from April 2019 to October 2020.

(a) (b)

(c) (d)

Fig. 2: XGBoost model comparison by period with prediction error.

However, Period 2’s prediction forecasted the actual values between April
2021 to September 2021. The data were closely tight from June 2019 until June
2022. Observing the prediction errors, in Figures 2c and 2d, the residual were
quite loose for Period 1 as compared to Period 2 which are more compact given
that there is one outlier due to sudden decrease of energy consumption during
the start of lockdowns.
Figure 3 illustrates the Linear Regression model comparison of Period 1 and
2 with their respective prediction errors. Based on the graph, Figures 2a and 2b,
 Forecasting Energy Consumption in the Philippines 9

it showed a great disparity between the actual values and the predicted values
in Period 1. The start of pre-pandemic, August 2018 to November 2018, it was
fairly forecasted. However, it took a great turn on the following months wherein
the prediction were incongruent to the actual. The analysis can also extend to
Period 2’s forecast since the graph showed no similarities to the actual values.
Importantly, Figures 3c and 3d showed the prediction of error relatively a↵ected
the forecast since the predicted values were not consistent with the actual values.

(a) (b)

(c) (d)

Fig. 3: Linear Regression model comparison by period with prediction error.

Figure 4 illustrates the Support Vector Regression model comparison of Pe-

riod 1 and 2 with their respective prediction errors. Figures 4a and 4b, same
with the previous model, Linear Regression, it also showed a great disparity be-
tween the actual values and the predicted values in Period 1. The forecasting was
incongruent with the actual wherein there is somewhat a disconnect using the
model in terms of actualizing the predicted values with the actual. The analysis
can also extend to Period 2’s forecast since the graph showed no similarities to
the actual values. With that, comparing the SVR and Linear Regression models,
both were quite similar in their forecast.
Figures 1 - 4 compare the predicted values from machine learning models
with the actual values and the predicted values from the optimal model for each
time period.
Figures 5a and 5b summarize the featured machine learning models forecast
using the predicted values against the actual values for Period 1 and Period 2.
It can be seen that the forecasting ability of the models varied and there was
a noticeable di↵erence in the predicted values when tracking the post-rebound
10 E. Torculas et al.

rise. In the first period, the XGBoost model performed the best by following
similar trend intervals. In the second period, the optimal model, which was the
Random Forest (RF) model, produced predictions that were almost the same
as the actual values by only a margin with XGBoost. In their respective cases,
both the XGBoost and Random Forest (RF) models show high adaptability
to forecasting values given the nonlinearity of the dataset, as modeled by the
January 2019 to April 2019 data in Period 1 and the sudden decline of energy
consumption in April 2020 as seen in Period 2.

(a) (b)

(c) (d)

Fig. 4: SVR model comparison by period with prediction error.

The two weakest predictive models, the Linear Regression Model and Support
Vector Regression, demonstrate comparable patterns in Periods 1 and 2. Despite
the fact that their projected values are closely clustered, they do not perform as
well as the XGBoost and RF models.
In time-series problem-solving, tree-based predictive models such as RF and
XGBoost outperform other models. Tree-based models, as opposed to linear
regression and SVR (linear), may successfully capture non-linear correlations
between variables. While SVR can include nonlinearity through the use of kernel
functions, it frequently involves a tradeo↵ between efficiency and complexity,
leading in longer training times and possibly overfitting. XGBoost and RF, on
the other hand, provide quick training, precise results, and scalability, making
them ideal for the task at hand. Furthermore, tree-based approaches have the
advantage of being able to handle missing data, with XGBoost automatically
inputting missing values using defaults or column means/medians.
 Forecasting Energy Consumption in the Philippines 11

(a)

(b)

Fig. 5: Machine learning models forecast in Pre-pandemic and Pandemic Period

5 Conclusion

The paper introduced di↵erent machine learning algorithms in forecasting time

series of energy consumption and compared the viability of these models on which
approach has high accuracy to predict these patterns or trends surrounding
energy consumption. Two periods were used in the analysis, pre-pandemic and
pandemic, Period 1 and 2, respectively.
To summarize, this study explored the use of machine learning to forecast
energy consumption and determined the viable model that can be implemented
in analyzing energy forecast in the Philippines. In particular, the XGBoost model
was the most accurate in predicting the first period, while the RF model had
the lowest RMSE in the second period. Thus, these convincingly demonstrate
that tree-based predictive models have a superior advantage over other models in
solving time-series problems. Knowing this, the study also verified the predictive
power of machine learning in the energy market using real data and suggested
that this forecasting model could be used by energy companies and governments
to respond better to changes in energy consumption and improve the reliability
of energy supply and demand data.
There are a few limitations to this research and potential areas for future
study. Firstly, the accuracy of the model’s prediction and analysis can vary de-
pending on the data and variable settings, making it difficult to determine which
approach is superior in all cases. Secondly, the caveats mentioned earlier in the
data summarize how the data regarding energy consumption is lacking here in
the Philippines. In order to build an e↵ective way for forecasting the consump-
12 E. Torculas et al.

tion of energy, we must start in consolidating the energy data in the country.
Thirdly, to further invoke a much higher accuracy, the samples of the data must
be increased or at least the data points should be widened for distinguishing
more models with better predictive approach. Therefore, the future work should
further investigate on time series forecasting algorithms such as Autoregressive
Integrated Moving Average (ARIMA) and Seasonal Trend Decomposition to
enhance the analysis by integrating energy consumption by day or seasons.

References
1. S. Barak and S. Sadegh. “Forecasting energy consumption using ensemble
ARIMA–ANFIS hybrid algorithm”. In: International Journal of Electrical Power
Energy Systems 82 (2016).
2. The National Institute of Open Schooling (NIOS). Importance of Energy in
the Society. Accessed: 2022-11-27. (2012). URL: https://nios.ac.in/media/
documents/333courseE/27B.pdf.
3. H. Ritchie, Roser M., and P. Rosado. Philippines: Energy Country Profile. Ac-
cessed: 2022-11-27. Our World in Data. (2022). URL: https://ourworldindata.
org/energy/country/philippines.
4. M. D. Sahakian. “Understanding household energy consumption patterns: When
“West Is Best” in Metro Manila”. In: Energy Policy 39.2 (2011), pp. 596–602.
5. E. Garcia-Martin et al. How to Measure Energy Consumption in Machine Learning
Algorithms. Lecture Notes in Computer Science. Cham, Switzer- land: Springer,
(2019), pp. 243–255.
6. S. Shin and H. Woo. “Energy Consumption Forecasting in Korea Using Machine
Learning Algorithms”. In: Energies 15.13 (2022).
7. M. Rambabu, N. Ramakrishna, and P.K. Polamarasetty. “Prediction and Analysis
of Household Energy Consumption by Machine Learning Algorithms in Energy
Management”. In: E3S Web of Conferences (2022).
8. M. Siti Birkha Mohd Ali et al. “Analysis of energy consumption and potential
energy savings of an institutional building in Malaysia. In: Alexandria Engineering
Journal 60-1 (2021), pp. 805–820.
9. Salam, A., and Hibaoui, A. E. (2018). Comparison of Machine Learning Al-
gorithms for the Power Consumption Prediction: Case Study of Tetouan city.
2018 6th International Renewable and Sustainable Energy Conference (IRSEC).
doi:10.1109/irsec.2018.8703007
10. Bhattacharyya, S. C., and Timilsina, G. R. Modelling energy demand of developing
countries: Are the specific features adequately captured? Energy Policy, 38(4),
1979–1990. doi:10.1016/j.enpol.2009.11.079. (2010)
11. Philippine Electricity Market Corporation. Monthly Market Assessment Report.
Accessed: 2022-11-27. (2019). URL: https://www.wesm.ph/market-outcomes/
market-assessment-reports/monthly-market-assessment-report.
12. C. Li et al. “Building Energy Consumption Prediction: An Extreme Deep Learning
Approach”. In: Energies 10.10 (2017).
13. A. Mosavi A. Bahmani. Energy Consumption Prediction Using Machine Learn-
ing; A Review. Accessed: 2022-11-27. (2019). URL: https://www.preprints.org/
manuscript/201903.0131/v1.
14. E. Garcia-Martin et al. “Estimation of energy consumption in machine learning”.
In: Journal of Parallel and Distributed Computing 134 (2019), pp. 75–88.