Synchrosqueezed Transform Based Multicondense Residual Network for Ultra-Short-Term Solar Power Forecasting

Abstract

This article introduces an efficient multistep nonparametric residual network for improved solar power forecasting. The proposed architecture first incorporates a synchrosqueezing transform to extract high-resolution time-frequency coefficients of solar power inputs in their respective time-frequency scales. An improved residual network, a Multicondense Residual Network (M-cDRN), integrating multiresidual network and condense network techniques to predict solar power coefficients, is proposed. M-cDRN addresses challenges of overfitting and vanishing gradients in residual networks. A quantile regression network is employed to generate quantiles with different proportions. The model's efficacy is validated using real-time datasets from two geographical locations, showing significant improvements in mean square error: 75.36% for sunny, 21.74% for partially sunny, and 35.42% for cloudy days.