Enhancing building energy performance prediction: A fusion of deep learning and first-principles simulation methods

Abstract

Accurate energy demand prediction is essential for effective control of energy consumption and generation, enabling optimal energy management and reducing emissions. This study introduces a hybrid model integrating a physics-based EnergyPlus simulation with a long short-term memory (LSTM) neural network to enhance energy demand forecasting for buildings. While physics-based models offer computational efficiency, they often result in significant discrepancies between predicted and actual energy demand. Conversely, LSTM models improve prediction accuracy but struggle with peak forecasting. The proposed hybrid approach leverages the strengths of both methods, combining simulation outputs with LSTM predictions to reduce the normalized mean absolute error (NMAE) and improve peak predictions. A feature importance algorithm is incorporated to identify key variables, such as water heater cycle on count, that influence peak values during the LSTM training process. The hybrid model demonstrates a 5% and 20% reduction in normalized root mean square error (NRMSE) compared to the standalone LSTM and simulation models, respectively, during the warmest week of summer. Slight improvements are also observed during winter, with NRMSE reductions of 3% and 11% for the coldest week. Additionally, the hybrid model outperforms the single LSTM in predicting peak energy demand across various thresholds, as evidenced by superior F1-scores, precision and recall. While the model is less effective during certain winter periods, its overall robustness in predicting peaks and valleys across seasonal variations underscores its utility for real-world applications.