Development of a Solar diffuse Irradiance Prediction Model using Existing Solar Decomposition Models and Reinforcement Learning

Abstract

An accurate irradiance prediction by orientation is important for solar systems installed in buildings. Thus, solar decomposition methods are essential. Existing solar decomposition models are site dependent. Thus, elaboration is required for local applications. In this work, we proposed reinforcement learning to improve Watanabe and Reindl models for a target location. For learning, measured solar diffuse irradiance data only for two weeks were required when existing models were employed. Results showed that the proposed model reduced the prediction error compared to the prediction case with existing models. Such an increase in the accuracy was apparent when the proposed method was applied to a relatively inaccurate decomposition model such as the Watanabe model, showing an error reduction by 40%. This was obtained by simple two weeks reinforcement learning for predicting solar diffuse irradiance over one year.