Review of deep learning techniques for power generation prediction of industrial solar photovoltaic plants

Abstract

Varying power generation by industrial solar photovoltaic plants impacts the steadiness of the electric grid which necessitates the prediction of solar power generation accurately. In this study, a comprehensive updated review of standalone and hybrid machine learning techniques for PV power forecasting is presented. Forecasting solar generation is of importance for the sustainability of grid power and also to achieve the UN sustainable development targets by 2030. The comparison of techniques shows that grouping datasets based on input feature similarity, results in higher accuracy. Long-Short Term Memory (LSTM) is found to perform better than other deep learning networks for all time horizons. The Gate Recurrent Unit (GRU), with few trainings, is found to be better for small datasets than LSTM. Based on the more complicated data patterns, a novel architecture of the Deep Learning Network model, with the capability to analyze and forecast is presented considering factors influencing industrial solar power generation. The study is of importance to researchers, solar industry, and electricity distribution companies for sustainable development worldwide.