Manifold Learning Based Quantile Regression Method for Short-Term Power Load Probability Density Forecasting

Abstract

Short term load forecasting is important for power system planning, operation, and power trading. In this paper, a quantile regression method based on manifold learning is proposed for the short-term power load probability density forecasting. Considering the good performance, the local linear embedding (LLE) method (i.e., one excellent manifold learning method) is used to reduce the dimension of the data, which can extract the underlying factors of power load to improve the forecasting accuracy and significantly reduce the complexity. Then we use a BI-LSTM network to better analyze the time correlation of the low-dimensional data, so as to better forecast the power load under different quantiles. Considering the good performance of the Gaussian kernel function, we use the Gaussian kernel smoothing method to derive the power load probability density. Simulation results on actual power load data validate that the proposed model can provide a better performance than other quantile regression methods.