INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Special Issue | Volume XIV, Issue XIII, October 2025
www.ijltemas.in Page 134
Comparative Analysis of Machine Learning Algorithms for Energy
Consumption Forecasting
Sonali Nemade*, Ashwini Patil, Deepashree Mehendale, Reshma Masurekar
Department of Computer Science, Dr. D. Y. Patil Arts, Commerce and Science College, Pimpri, Pune 18, Maharashtra,
India
DOI: https://doi.org/10.51583/IJLTEMAS.2025.1413SP029
Received: 26 June 2025; Accepted: 30 June 2025; Published: 24 October 2025
Abstract: Forecasting energy use has become a crucial component of contemporary smart grid systems, allowing stakeholders to
guarantee system dependability, cost effectiveness, and energy efficiency. For the integration of intermittent renewable energy
sources, load balancing, and real-time energy management, the capacity to predict power demand is essential. The use and
relative effectiveness of five supervised machine learning algorithms Linear Regression, Decision Tree, Random Forest,
XGBoost, and Gradient Boosting for predicting short-term building-level energy consumption are examined in this work. In order
to train and evaluate models, we carried out a thorough preprocessing and feature engineering procedure using a large dataset that
included operational, meteorological, and temporal variables.
Each model was assessed using three key performance metrics: mean absolute error (MAE), root mean square error (RMSE), and
coefficient of determination (R²). Among the models tested, Gradient Boosting achieved the highest accuracy, with an MAE of
575 kWh, RMSE of 851 kWh, and an R² of 0.949, outperforming both traditional and advanced ensemble models.
Our results highlight how well boosting strategies work for energy forecasting jobs and how crucial it is to choose models
according to deployment restrictions and data properties. The knowledge gained from this research can help designers create
responsive, scalable, and intelligent energy forecasting systems that are appropriate for smart infrastructure.
Keywords: Energy forecasting, smart grid, Machine learning, Gradient Boosting, Random Forest, Time series
I. Introduction
With increasing global energy demands and the transition toward decentralized power systems, accurate energy consumption
forecasting has become essential. Effective forecasting ensures reliability in supply, facilitates energy conservation and supports
demand-side management. In the era of smart grids, energy providers and consumers rely heavily on real-time analytics and
predictive models to anticipate fluctuations in energy demand.
Energy forecasting plays a crucial role not only in daily operations of utilities but also in long-term infrastructure planning, cost
optimization, and environmental protection. The increasing penetration of renewable energy sources such as solar and wind
which is inherently variable adds further complexity to energy load management. Consequently, forecasting systems must adapt
to incorporate weather conditions, consumer behavior, and temporal trends.
Traditional statistical approaches often assume linearity and stationarity, making them insufficient for modeling the dynamic and
nonlinear nature of real-world energy consumption. They typically fail to accommodate irregular consumption patterns caused by
holidays, changes in occupancy, or abrupt weather events. In contrast, machine learning (ML) offers scalable and flexible
frameworks capable of capturing these intricate relationships without strict assumptions about data distributions.
This study aims to explore the application of several well-established ML models for short-term load forecasting in a commercial
building setting, focusing specifically on summer peak periods when energy consumption is typically at its highest. By
benchmarking multiple algorithms, we intend to provide insights into their strengths and trade-offs and to identify the most
promising model for practical deployment in energy management systems. With increasing global energy demands and the
transition toward decentralized power systems, accurate energy consumption forecasting has become essential. Effective
forecasting ensures reliability in supply, facilitates energy conservation and supports demand-side management. Traditional
statistical approaches often assume linearity and stationarity, making them insufficient for modeling the dynamic and nonlinear
nature of real-world energy consumption. Machine learning (ML), in contrast, provides flexible, data-driven techniques capable
of uncovering hidden patterns within complex, high-dimensional datasets. This paper explores the use of several ML models for
predicting energy consumption in commercial buildings during summer peak periods.
II. Literature Review:
Traditional models like ARIMA (Autoregressive Integrated Moving Average) and exponential smoothing have historically been
used for energy forecasting due to their simplicity and interpretability. These methods rely on past trends and seasonality to make
future predictions, assuming that patterns repeat over time. While suitable for stationary time series, these models struggle to
account for dynamic, real-time changes in energy consumption due to varying user behavior, irregular external influences such as
weather, and policy interventions.
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Special Issue | Volume XIV, Issue XIII, October 2025
www.ijltemas.in Page 135
More recent research has turned toward machine learning (ML) techniques for their ability to capture nonlinear patterns and
complex feature interactions. Ensemble methods like Random Forest (RF) and Gradient Boosting Machines (GBM) have
demonstrated robustness and accuracy, especially when trained on high-dimensional datasets. Breiman (2001) introduced
Random Forests as a way to improve prediction accuracy by averaging multiple decision trees trained on bootstrapped data
subsets. Similarly, Friedman (2001) developed the Gradient Boosting framework, which sequentially builds models to minimize
residual errors.
XGBoost (Extreme Gradient Boosting), proposed by Chen and Guestrin (2016), extended GBM with advanced regularization,
efficient parallel computation, and handling of missing data, making it a top-performing method in many data science
competitions. In the context of energy forecasting, these tree-based models provide not only high accuracy but also interpretable
outputs through feature importance metrics.
Although deep learning methods (e.g., LSTM, Transformer architectures) show promising results, they require larger datasets,
extensive tuning, and longer training times. Therefore, in this work, we focus on ML models that strike a balance between
performance, interpretability, and computational efficiency.
Problem Statement:
Accurately predicting daily energy consumption has become essential for energy providers, policymakers, and infrastructure
planners due to the increase in demand for electricity throughout the summer. Better demand-side management and resource
allocation are made possible by accurate energy forecasting, which also helps prevent grid overloads and unforeseen peak
occurrences that could result in blackouts or expensive operating expenses.
In order to predict daily energy consumption and identify possible peak power demand days throughout the summer, this project
intends to create a machine learning-based forecasting model using historical temperature data, activity levels, and previous
energy usage trends. The goal is to develop a predictive model that can effectively generalize and assist in data-driven energy
management decisions by examining a dataset that contains comprehensive temperature readings (T1–T10), energy measures, and
user activity.
Objective:
To understand and preprocess the provided dataset (summer_peaks.csv), which includes temperature variables (T1–
T10), daily energy consumption, peak power indicators, and activity levels.
To explore the influence of temperature, activity, and temporal features (like day of the week) on daily energy usage.
To design and train multiple machine learning models (as demonstrated in the provided Jupyter notebook) to predict
energy consumption, including algorithms like Linear Regression, Random Forest, and XGBoost.
To compare model performance using evaluation metrics such as Mean Absolute Error (MAE), Root Mean Squared
Error (RMSE), and R² score to identify the most effective forecasting approach.
III. Methodology:
This study uses machine learning techniques to forecast daily energy use over the summer using an organized, data-driven
methodology. Ten temperature sensor readings (T1–T10), total energy consumption, peak power characteristics, a binary peak
day indication, activity levels, and date-related data are all included in the summer_peaks.csv dataset, which serves as the basis
for the analysis.
The dataset is first preprocessed using appropriate imputation techniques to address missing values, especially in the later
temperature variables (T6–T10). Time-based patterns are captured by extracting temporal information, including the day of the
week, and converting the date field to the correct date time format. In order to simplify the input space while maintaining
significant variance, the 10 temperature values are also combined to calculate an average temperature feature.
Statistical plots and summary statistics are used in exploratory data analysis (EDA) to find relationships between temperature,
activity, and energy consumption. By adding additional variables like average temperature and weekday/weekend classification
and eliminating superfluous or strongly correlated features, feature engineering is used to improve model input and lessen over
fitting.
After that, a number of machine learning models such as XGBoost Regressor, Random Forest Regressor, and Linear Regression
are created and trained. Eighty percent of the dataset is used to train these Python-implemented models, with the remaining
twenty percent set aside for testing. Standard regression indicators, such as Mean Absolute Error (MAE), Root Mean Squared
Error (RMSE), and the Coefficient of Determination (R2 Score), are used to assess the efficacy of the model. The model chosen
for predicting is the one that performs the best across these metrics. Lastly, future energy usage is predicted using the chosen
model. To determine which input factors, have the greatest impact on energy consumption, feature importance analysis is
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Special Issue | Volume XIV, Issue XIII, October 2025
www.ijltemas.in Page 136
performed. Particular focus is placed on finding and forecasting days with peak demand. This makes it possible for the suggested
system to be a practical tool for proactive load management and energy demand predictions during summertime peak demand.
Figure1: Energy Forecasting Methodology
Dataset:
589 records with 18 columns make up the summer_peaks.csv dataset, which collects daily information on temperature, energy
use and peak electricity demand. With the date in the day column and the day of the week (where 1 denotes Monday and 7
Sunday) in the wd column, each row corresponds to a single day. Ten variables, designated T1 through T10, collect temperature
data; these values most likely come from various sensors or at various times of the day. These temperature data, which are given
in degrees Celsius, shed light on the daily fluctuations in temperature.
The dataset contains energy-related parameters in addition to temperature. Peak power captures the maximum power usage seen
during the day, whereas the energy column documents the entire energy consumption for the day. The system's peak power usage
duration is indicated by the peak duration field, while the peak intensity field indicates how strong or severe the peak event was.
A Boolean flag in this is peak column indicates if a peak event happened that day (True or False).
Lastly, the dataset has an activity column, a numerical number that can represent the degree of human activity or the intensity of
appliance use on a particular day. This dataset is generally well-suited for examining trends in energy use, comprehending how
temperature and activity levels affect power demand, and creating models to anticipate or control summertime energy peaks.
IV. Result Analysis and Performance:
The performance of several machine learning models used for energy consumption predictions is extensively investigated in this
section. Historical data on energy usage, together with related characteristics like temperature, humidity, time of day, and
calendar factors, was used to train and evaluate the models. Evaluating each model's predictive power, accuracy, and robustness
over various time periods was the main objective.
The forecasting ability of the machine learning model utilized in this study is shown graphically in the "Actual vs. Predicted
Energy Consumption" scatter plot. A single model forecast compared to the actual recorded energy consumption for the same
instance is represented by each point on the graph.
Figure 2: Actual Vs Predicted Energy Consumption
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Special Issue | Volume XIV, Issue XIII, October 2025
www.ijltemas.in Page 137
Three common regression metrics were used to assess the machine learning models' forecasting accuracy:
Mean Absolute Error, or MAE, calculates the average size of prediction mistakes without taking direction into account. Root
Mean Squared Error, or RMSE, is more sensitive to outliers and penalizes greater errors more severely than MAE.
The percentage of the variance in the dependent variable that can be predicted from the independent variables is represented by
the R2 Score (Coefficient of Determination). A better fit is indicated by a higher number that is nearer 1.
Table1: Accuracy of Models
MAE RMSE R2
Linear Regression 714.359172 925.431599 0.939948
Decision Tree 769.793220 1501.457350 0.841925
Random Forest 615.304034 941.653462 0.937824
XGBoost 591.655873 881.032733 0.945572
Gradient Boosting 575.412316 850.587164 0.949269
Gradient Boosting performed best on all three metrics, with the highest R2 score (0.9493) and the lowest MAE (575.41) and
RMSE (850.59). This suggests it is the most accurate and trustworthy model among those tested for projecting energy use.
Furthermore, Random Forest performed well, striking a compromise between model interpretability and error minimization. Its
lower MAE suggests superior overall consistency even if its RMSE is slightly larger than Linear Regression. Linear regression
showed greater MAE and RMSE but did very well with an R2 of 0.9399. Non-linear patterns in the data probably caused it
problems. With the lowest R2 score and the highest RMSE (1501.46), the Decision Tree performed the worst, indicating that it
might be over fitting or not generalizing effectively on test data, while being simple to understand.
Additionally, XGBoost outperformed other models, demonstrating its resilience in managing intricate patterns and feature
interactions, albeit with somewhat larger errors than Gradient Boosting.
V. Conclusion:
A comparative examination of many machine learning techniques for energy consumption forecasting was reported in this study.
Planning for a sustainable power system, cost optimization, and effective energy management all depend on accurate energy
forecasts. Using historical energy usage data, the study deployed several models Linear Regression, Decision Tree, Random
Forest, XGBoost, and Gradient Boosting and assessed each model's performance using MAE, RMSE, and R2 metrics.
According to the experimental findings, ensemble learning models in particular, Gradient Boosting and XGBoost perform
noticeably better than conventional models in terms of accuracy and robustness. With the lowest prediction error and the best R2
score (0.949), gradient boosting proved to be highly effective at identifying complex trends and nonlinear correlations in the
behavior of energy use.
This study concludes that machine learning, particularly sophisticated ensemble approaches, provides a dependable and
expandable solution for forecasting energy use. Utility companies, building managers, and smart grid systems can use these
insights to help them make well-informed decisions about energy conservation, load balancing, and infrastructure design.
References
1. Hyndman, R. J. & Athanasopoulos, G. Forecasting: Principles and Practice, 3rd ed., OTexts, 2021.
2. Breiman, L. “Random Forests,” Mach. Learn. 45, 5-32 (2001).
3. Friedman, J. H. “Greedy Function Approximation: A Gradient Boosting Machine,” Ann. Stat. 29, 1189-1232 (2001).
4. Chen, T. & Guestrin, C. “XGBoost: A Scalable Tree Boosting System,” Proc. KDD 2016, 785-794.
5. Li, K. et al. “Short-Term Load Forecasting in Smart Grids: A Combined Deep Learning Approach,” Energy 188, 116–119
(2019).
6. Lu, M. H. L., Ser, Y. C., Selvachandran, G., Thong, P. H., Cuong, L., Son, L. H., et al. (2022). A comparative study of
forecasting electricity consumption using machine learning models. Mathematics, 10(8), 1329.
7. Bilal, M., Kim, H., Fayaz, M., & Pawar, P. (2022). Comparative analysis of time series forecasting approaches for
household electricity consumption prediction. arXiv.
8. Alawadi, S., Mera, D., Fernández-Delgado, M., et al. (2022). A comparison of machine learning algorithms for forecasting
indoor temperature in smart buildings. Energy Systems, 13, 689–705.