Research on eXplainable artificial intelligence in the CNN-LSTM hybrid model for energy forecasting

Abstract

At present, data-driven approaches have achieved satisfactory results in energy forecasting for building cooling and air conditioning systems, particularly using hybrid deep learning models based on Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM). However, these complex hybrid deep learning models often struggle to offer sufficient explainability, making it difficult for building professionals to understand the models and thereby reducing their trust in them. This study developed a CNN-LSTM hybrid model for predicting energy consumption in building refrigeration and air conditioning systems. To preserve the positional information of input features, the pooling layers in the CNN module were removed. The modified model still achieved high prediction accuracy, with R² = 0.9729. Subsequently, a comparative study of four eXplainable artificial intelligence (XAI) methods was conducted for the hybrid model. Among them, Gradient-weighted Class Activation Mapping (Grad-CAM) and Gradient-weighted Absolute Class Activation Mapping (Grad-Absolute-CAM) provided more reasonable global explanations. Additionally, Grad-Absolute-CAM exhibited superior consistency in local explanations. An ablation study was conducted based on the feature importance rankings provided by the four XAI methods. Through comprehensive comparison, the 11 most influential features for energy consumption prediction were identified. This study partially fills the research gap in explainability of hybrid models and highlights the potential of XAI in feature selection and improving energy forecasting. This has significant implications for developing more explainable, accurate, and efficient prediction models.