Short Time Power Load Probabilistic Forecasting Based on Constrained Parallel-LSTM Neural Network Quantile Regression Mode

Abstract

The probabilistic forecasting can accurately quantify the uncertainty of the future power load and provide comprehensive prediction information for the operation decision of power system. Addressing the temporal characteristics of power load and the overlap between the adjacent quantile forecasts of the existing quantile regression methods, this paper proposes a short time power load probabilistic forecasting method based on the constrained parallel long short-term memory (LSTM) neural network quantile regression model. This method integrates the LSTM and the quantile regression to generate multiple load quantile forecasts in parallel. Furthermore, it ensures the rationality of the quantile forecasts through adding a combination layer considering the constraints between the adjacent quantile forecasts. The results of an actual example show that the proposed method not only has higher prediction efficiency but also can obtain more reasonable quantile forecasts compared with the existing load probabilistic forecasting methods