A Short-Term Telephone Traffic Forecasting Method for Power Grid Customer Service via Ensemble Learning Using GRU Model with Correntropy Loss

To address the challenges of nonlinearity, strong temporal dependence, and accuracy degradation caused by sudden disturbances in power grid customer service telephone traffic forecasting, this paper proposes a novel forecasting method based on an ensemble model pairing Gated Recurrent Unit (GRU) with Correntropy loss (CL) (called EnsCL-GRU). First, to overcome the sensitivity of the traditional Mean Squared Error (MSE) loss to abnormal spikes and its difficulty in capturing the overall trend consistency of the sequence, a CL is introduced as the loss function for the GRU model. This loss function calculates the normalized Correntropy coefficient between the predicted sequence and the true sequence in the time-delay domain, guiding the model to focus on the overall shape matching of the time series data rather than point-wise error fitting. Furthermore, the gated memory mechanism of the GRU can capture long-term dependencies in the time series, while the CL constrains the consistency of the predicted dynamic trends from the sequence level. This preserves the GRU’s temporal modeling capability while enhancing the model’s response accuracy to sudden disturbances and trend changes. Second, to improve the generalization ability of a single GRU model, an ensemble strategy is employed to train multiple CL-enhanced GRU base models serially. By adaptively adjusting sample weights, the fitting capability for difficult samples (such as telephone traffic spikes) is improved, further improving the model’s robustness. Finally, Bayesian optimization is introduced to automatically search for the optimal hyperparameters of the ensemble model, efficiently approximating the global optimal configuration within a limited number of evaluations. Experimental results demonstrate that the proposed method outperforms traditional approaches. Specifically, compared with the standard GRU model, the proposed method reduces MAPE from 29.15% to 22.61%. It also consistently outperforms the ensemble baseline EnsGRU, achieving a MAPE reduction of 4.73 percentage points. The results indicate that the proposed model significantly improves forecasting accuracy and robustness, particularly under scenarios with nonlinear fluctuations and sudden disturbances, providing reliable support for optimal resource allocation in power grid customer service systems.