An efficient stacked recurrent broad learning scheme for PV cluster power forecasting.

Abstract

Precise power prediction for photovoltaic (PV) clusters is more conducive to resource optimization and scheduling compared to single-site prediction. This paper proposes a novel stacked recurrent broad learning system (Stack-RBLS) for PV cluster power prediction, specifically designed for resource-constrained training devices. The proposed scheme simultaneously reduces training time and improves prediction accuracy. Specifically, firstly an improved data cleaning method with linear time complexity using sliding time window (STW)-based isolation forest algorithm and k-nearest neighbors (IFA-KNN) is proposed to ensure that the forecasting relies on correct data. Based on the lightweight broad learning model, this proposed scheme then utilizes RBLS to learn temporal features and BLSs to fit residuals. The residual-stacked design enhances fitting performance by propagating sequences between layers when applied to time series tasks. To simplify the training process, a validation set is built for Sequential Least Squares Programming (SLSQP) optimization, eliminating the need for manual tuning of regularization parameters. Finally, the effectiveness and superiority of the proposed method are validated using real-world PV cluster datasets.