Multi-step photovoltaic power forecasting using transformer and recurrent neural networks

Abstract

Affordable and clean energy is an important UN sustainable development goal. Solar energy is more difficult to control than fossil fuels, highlighting the need for accurate solar power forecasts. This study develops three variants of the transformer networks, called PVTransNet, for a multi-step day-ahead photovoltaic power forecasting. The transformer networks use historical solar power generation, weather observation, weather forecast and solar geometry data as input to effectively predict next-day hourly power generation. In particular, the third variant model combines long short-term memory (LSTM) to transformer networks to supplement weather forecasts from the weather station. The combined model, PVTransNet-EDR, outperformed individual LSTM and other transformer models in the experiments conducted on data from two photovoltaic power plants. The model performed 48.3 % better, in mean absolute error, than simple LSTM in the power forecasting. Accurate solar power forecasting model is expected to be utilized for efficient energy storage and microgrid management, effective energy supply policy, and optimal plant location selection.