Improved bidirectional long short-term memory network-based short-term forecasting of photovoltaic power for different seasonal types and weather factors

Abstract

Current photovoltaic (PV) power forecasts have not rigorously investigated the intrinsic characteristics of PV data clustering associated with various seasonal weather types to explore the potential for enhanced predictive accuracy. To address this issue, a short-term prediction method that correlates seasonal weather patterns with improved bi-directional long and short-term memory network (BiLSTM) modelling is proposed. Firstly, an improved k-means clustering algorithm is employed to categorize PV data according to each season, thereby enabling an in-depth analysis of PV characteristics under distinct seasonal weather conditions. Using a variational modal decomposition (VMD) algorithm for data decomposition, the dimensionality is then reduced using a kernel principal component analysis (KPCA) and this minimizes data redundancy. An improved bidirectional long and short-term memory network (BiLSTM) model is also deployed, and this aims to comprehensively incorporate the temporal characteristics of the data. Finally, the simulation results demonstrate that the forecast accuracy of the proposed model produces improvements of up to 58.2 %, 41.3 %, and 35.4 % over the CNN, BiLSTM, and VMD-KPCA-BiLSTM models, respectively.