AECF-UC: An Adaptive Electricity Consumption Forecast Approach for Universal Environments

The development of an accurate electricity consumption forecast model is crucial for stable operation and intelligent management of power systems. Traditional methods often overlook user heterogeneity and lack measures to address concept drift caused by distribution changes in electricity data over time. We propose an adaptive electricity consumption probability forecasting method tailored to universal environments. The method includes a nonmonotonic correlation elimination-based recursive feature selection that adaptively determines the optimal feature combination. Our model incorporates a joint loss function combining point and probability forecasting evaluations to accurately quantify online batch errors. It also features a buffer to store batch data showing pattern changes and dynamically adjusts weights to counteract concept drift. We validated our method, adaptive electricity consumption forecast for universal environments (AECF-UC), against 7 mainstream methods using a multi-environment dataset. Comparative and ablation experiments show that AECF-UC outperforms others, achieving average RMSE and pinball scores of 0.3041 and 0.0567 respectively, with the joint loss method improving prediction accuracy by about 6% over the single-loss method. These results indicate that the proposed method exhibits certain advantages in universality and adaptability.