Comparative analysis of deep learning and tree-based models in power demand prediction: Accuracy, interpretability,...

Research and development have demonstrated that effective building energy prediction is significant for enhancing energy efficiency and ensuring grid reliability. Many machine learning (ML) models, particularly deep learning (DL) approaches, are widely used for power or peak demand forecasting. However, evaluating prediction models solely based on accuracy is insufficient, as complex models often suffer from low interpretability and high computational costs, making them difficult to implement in real-world applications. This study proposes a multi-perspective evaluation analysis that includes prediction accuracy (both overall and at different power levels), interpretability (global/local perspectives and model structure), and computational efficiency. Three popular DL models-recurrent neural network, gated recurrent unit, long short-term memory, and three tree-based models-random forecast, extreme gradient boosting, and light gradient boosting machine-are analyzed due to their popularity and high prediction accuracy in the field of power demand prediction. The comparison reveals the following: (1) The best-performing prediction model changes under different power demand levels. In scenarios with lower power usage patterns, tree-based models achieve an average CV-RMSE of 13.62%, which is comparable to the 12.17% average CV-RMSE of DL models. (2) Global and local interpretations indicate that past power use and time-related features are the most important. Tree-based models excel at identifying which specific lagged features are more significant. (3) The DL model behavior can be interpreted by visualizing the hidden state at each layer to reveal how the model captures temporal dynamics across different time steps. However, tree-based models are more intuitive to interpret using straightforward decision rules and structures. This study provides guidance for applying ML algorithms to load forecasting, offering multiple perspectives on model selection trade-offs.