Short-term load probabilistic forecasting based on quantile regression convolutional neural network and Epanechnikov kernel density estimation
Hui He, Junting Pan, Nanyan Lu, Bo Chen, Runhai Jiao
Abstract
Electricity load forecasting plays an indispensable role in the electric power systems. However, its characteristics of uncertainty and complexity are hard to handle. This paper proposes a probabilistic load forecasting approach named QRCNN-E. Specifically, the deep convolutional neural network is applied to model the non-linear relationship with the electricity load and its influencing factors. By replacing the traditional loss function with pinball loss, the network can eventually forecast loads in quantiles. Then, kernel density estimation takes quantile forecasts as inputs and produces deterministic and probabilistic results. Case studies on GEFCom2014 show that the proposed method presents better performance than other cutting-edge models.