Machine Learning-Based Short-Term Load Forecasting for Campus Energy Monitoring: A Case Study of UMPSA Region 3

Abstract

Energy monitoring on campus improves efficiency, cost awareness, and sustainability. The present paper is an exploratory data analysis and machine learning short-term load forecasting study with real monitoring data from Region 3 at Universiti Malaysia Pahang Al-Sultan Abdullah (UMPSA) consisting of FKOM, FTKPM and FTKMA2 and the University Mosque. The dataset includes time-stamped electrical and energy-related parameters such as active power, reactive power, apparent power, voltage, phase currents, power factors, peak energy consumption, off-peak energy consumption, carbon emissions, and energy cost. Exploratory Data Analysis was carried out to investigate the variation of loads, daily energy use, weekday and weekend usage, usage by tariff periods, and electrical operating characteristics. In view of these results, time-based, lag-based, and rolling statistical features were engineered and integrated with lightweight machine learning models (Linear Regression, Regression Trees, Random Forest, and Gradient Boosting) for 15-minute-ahead active power forecasting. Standard regression metrics were calculated and compared to a persistence baseline. The findings show the feasibility of using such models to convert campus energy monitoring data into predictive data for operational planning, tariff-aware energy management, and sustainability monitoring, with the best model achieving RMSE of 0.012, MAE of 0.09, MAPE below 5%, and $\mathrm{R}^{2}$ exceeding 0.93. Compared to the persistence baseline, this represents a reduction in RMSE by $\mathbf{1 8. 6 \%}$, MAE by $\mathbf{1 5. 2 \%}$, and MAPE by $12.4 \%$, while improving $\mathbf{R}^{\mathbf{2}}$ from 0.82 to 0.93, underscoring the statistical significance of the proposed method.