Application of a novel deep learning method for electricity theft detection based on explainable artificial intelligence

To address the challenges of weak feature representation, difficult extraction, and insufficient classification accuracy in electricity consumption time-series data for smart grid security monitoring, this paper proposes a temporal convolutional network (TCN) model based on Bayesian optimization. This study innovatively combines deep learning’s feature extraction capability with Bayesian optimization’s parameter tuning strategy, effectively improving traditional TCN models. The proposed Bayesian optimization-enhanced TCN (BO-TCN) integrates the TCN’s powerful temporal feature extraction with BO’s efficient hyperparameter search to achieve optimal model configuration for electricity theft detection. To address the critical issue that deep learning performance is sensitive to hyperparameter settings, Bayesian optimization constructs a Gaussian process model to approximate the objective function and uses acquisition functions for efficient hyperparameter selection, enabling automatic optimization of TCN hyperparameters. This method significantly improves parameter optimization efficiency while maintaining prediction performance. Experimental results on the SGCC dataset demonstrate that the optimized TCN model exhibits excellent generalization ability in electricity theft detection, achieving 98.04% classification accuracy—significantly outperforming LSTM, CNN, and baseline models. The effectiveness of the BO-TCN is verified through comprehensive ablation experiments and post-hoc explainable artificial intelligence analysis using SHAP and LIME techniques. Compared with the NTCN (eliminating the 1 × 1 convolutional residual block), the BO-TCN improves accuracy by ∼1.23%; compared with the ANTCN (eliminating all residual structures), BO-TCN’s accuracy improves by ∼2.27%.