HA-MQFNet: A Horizon-Adaptive Multi-Quantile Forecasting Network for Day-Ahead Probabilistic Load Forecasting

Day-ahead probabilistic load forecasting provides both load estimates and uncertainty information for power system operation. However, existing methods often use one shared fusion strategy for all future time steps, which limits their ability to model horizon-dependent information. In addition, discrete quantile outputs may cause quantile crossing, and it is still difficult to balance quantile accuracy, interval reliability, and interval sharpness. To address these issues, this paper proposes HA-MQFNet, a Horizon-Adaptive Multi-Quantile Forecasting Network. The proposed method combines branch-wise feature encoding, horizon-adaptive fusion, monotone quantile output, and joint optimization in one end-to-end framework. Experimental results show that HA-MQFNet achieves the best overall performance among the compared methods, with a PICP close to the nominal coverage level, the lowest Pinball loss and Winkler score, and zero quantile crossing rate.