Enhanced Machine Learning Models with GEMED Environmental Data Integration for Short-Term Energy Forecasting in Smart Campus Environments

Abstract

This research takes a deeper look at the accuracy of forecasting energy consumption using both traditional and new machine learning techniques and compares these techniques with newer machine learning methods for use within higher education. The study specifically provides new insights into energy usage forecasting on campus using the Green Environmental Monitoring and Energy Data (GEMED) dataset, where synthetic data for energy usage forecasting is employed in conjunction with real GEMED data, together with several well-known energy usage and environmental sustainability algorithms. Seven energy usage forecasting models were applied, with a focus on a multi-dimensional evaluation framework for the selection of these models (standard evaluation metrics, directional accuracy, significance tests of differences, and uncertainty quantification), with the addition of unusual measures with respect to energy usage, the environmental impact of indoor air quality (IAQ), and thermal comfort constraints. The XGBoost model had the highest performance on the GEMED data (R² = 0.895, RMSE = 28.31 kW). The attention-based hybrid model exhibited enhanced feature interpretability, indicating that indoor temperature, occupancy patterns, and temporal features were the key determinants of energy consumption. Linear regression showed surprisingly robust results on the synthetic data, highlighting the importance of feature engineering. Rather than simplifying environmental inputs or treating constraints after the fact, the proposed Enhanced Data-Based Smart Campus Energy Forecasting (DB-SCEF) framework includes environmental constraints as part of the forecast pipeline and allows for reproducible, practitioner-focused approaches to uncertainty-aware, interpretable, and sustainable forecasts that support data-driven energy management systems.