Solar irradiance prediction using reinforcement learning pre-trained with limited historical data

Abstract

Accurate day-ahead forecasting of solar irradiance is crucial for maintaining a steady power supply and minimizing energy losses. To date, various solar irradiance prediction models have been established, but these typically require extensive weather data collected over long periods within the area of prediction or consistent updates using field measurements. This research introduces a reinforcement learning-based model capable of long-term solar irradiance prediction, even in areas with limited accumulated data. Our proposed model can forecast solar radiation for more than a year using just two weeks of solar radiation learning and readily available weather forecasts. It demonstrated a promising performance, with an annual average CVRMSE error of 7.0%, which is a more optimized predictive performance than the 12.8% CVRMSE yielded by the existing LSTM-based Reference model constructed by adding out-of-atmosphere solar radiation input values.