Seasonal Dynamics in Energy Forecasting: A Deep Learning Approach for Student Residences

This study presents a comprehensive analysis of various machine-learning models for energy forecasting across different seasons in students' residential settings. Our research investigates the performance of baseline models, including LSTM, GRU, state-of-the-art forecasting models, including Autoregressive feedforward neural network, Transformer, and hybrid approaches, by evaluating their ability to predict energy usage with high accuracy. Seasonal patterns and the occurrence of anomalous events such as vacations, meteorological shifts, unpredictable nature of human activity leading to sudden spikes and drops in energy consumption are particular challenges. The results demonstrate that no single model uniformly outperforms others across all seasons, which shows the necessity for season-specific model selection or design. Our proposed HyperNetLSTM and MiniAutoEncXGBoost models are delivering robustness in learning to seasonal variations especially in capturing sudden changes in consumption during the summer. The research contributes to the field of energy forecasting by the significance of seasonal insight and model behavior.