Optimal Adaptive Prediction Intervals for Electricity Load Forecasting in Distribution Systems via Reinforcement Learning

Prediction intervals (PIs) offer an effective tool for quantifying uncertainty of loads in distribution systems. The traditional central PIs cannot adapt well to skewed distributions, and their offline training fashion is vulnerable the unforeseen change future load patterns. Therefore, we propose optimal PI estimation approach, which online adaptive different data distributions by adaptively determining symmetric or asymmetric probability proportion pairs quantiles PIs’ bounds. It relies on learning ability reinforcement (RL) integrate two tasks, i.e., selection quantile predictions, both are modeled neural networks. As such, quality quantiles-formed can guide process pairs, forms a closed loop improve quality. Furthermore, efficiency forecasts, prioritized experience replay (PER) strategy proposed regression processes. Case studies net demonstrate that method better compared with method. Compared offline-trained methods, it obtains more robust against concept drift.