Missing-Data Tolerant Hybrid Learning Method for Solar Power Forecasting

Abstract

Solar power forecasting is a key task in modern power grid operation, which can be achieved by machine learning-based methods. Due to multiple practical issues, the data may be incomplete, making the existing machine learning models inaccurate or even ineffective. To counteract the missing data problem, this paper proposes a hybrid learning method. Firstly, a super-resolution perception convolutional neural network (SRPCNN) is designed to reconstruct the flawed data with missing data in both random missing and block missing patterns. With the recovered data, an incremental broad learning system (IBLS) is developed as the prediction model. Due to its strong approximation ability, low computational burden, and flexible structure, the IBLS model can be easily and rapidly updated by broadening the network structure to maintain the forecasting accuracy without the need for an entire retraining progress. Therefore, the proposed method is not only missing-data tolerant but also online updatable for accuracy maintenance/enhancement. On an open dataset <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">in Australia,</b> the proposed method is tested and compared with existing methods. The simulation results <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">verify the effectiveness of the proposed method.</b>

Topics & Concepts