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

Area of Science:

• Building energy systems
• Computational intelligence
• Sustainable energy

Background:

• Effective building energy prediction is crucial for energy efficiency and grid reliability.
• Machine learning (ML), especially deep learning (DL), is widely used for power demand forecasting.
• Current evaluations often overlook model interpretability and computational costs, hindering real-world application.

Purpose of the Study:

• To conduct a multi-perspective evaluation of ML models for building energy prediction.
• To analyze prediction accuracy, interpretability (global/local), and computational efficiency.
• To provide guidance for selecting appropriate ML algorithms for load forecasting.

Main Methods:

• Comparative analysis of three popular DL models (RNN, GRU, LSTM) and three tree-based models (Random Forest, XGBoost, LightGBM).
• Evaluation metrics included prediction accuracy (CV-RMSE), interpretability (feature importance, model structure visualization), and computational efficiency.
• Models were assessed across different power demand levels.

Main Results:

• Model performance varies with power demand levels; tree-based models are competitive with DL models at lower power usage.
• Past power usage and time-related features are key predictors; tree-based models offer clearer feature significance insights.
• DL models provide interpretability through hidden state visualization, while tree-based models offer intuitive decision rules.

Conclusions:

• A multi-perspective evaluation is essential for selecting ML models in load forecasting.
• Trade-offs between accuracy, interpretability, and computational efficiency should guide model selection.
• This study offers practical insights for applying ML to building energy prediction.