A Causally Explainable Deep Learning Model with Modular Bayesian Network for Predicting Electric Energy Demand

Abstract

Efficient management of residential power consumption, particularly during peak demand, poses significant challenges. Deep learning models excel in predicting electricity demand but lack of interpretability due to the interdependent nature of electricity data. To overcome this limitation, we propose a novel explanatory model that incorporates modular Bayesian network with deep learning parameters. The proposed method leverages associations among deep learning parameters and provides probabilistic explanation for demand patterns in the four types: global active power increase, decrease, peak, and others. The key idea is to accommodate modular Bayesian networks with association rules that are mined with the Apriori algorithm. This enables probabilistic explanation that can account for the complex relationships of variables in predicting energy demand. We evaluate the effectiveness of the proposed method with the UCI household electric power consumption dataset, comprising 2,075,259 time-series measures over a 4-year period. The method is also compared to the SHAP algorithm, confirming that it outperforms the SHAP algorithm with a cosine similarity of 0.8472 in identifying causal variables with 0.9391. © 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.