Post-Processing of NWP Forecasts Using Kalman Filtering With Operational Constraints for Day-Ahead Solar Power Forecasting in Thailand

Solar power forecasting with a day-ahead horizon has played an important role in the operational planning of generating units in power system operations. We aim to develop a solar power forecasting model suitable for a tropical climate, using Thailand as a model, and hence present a linear recursive regression model as a post-processing step for reducing the errors obtained from the Weather Research and Forecasting (WRF) model. This model consists of submodels, each of which predicts the solar irradiance of a particular time of the day. By using a stepwise regression method, we found that WRF forecasts of irradiance, temperature, relative humidity, and the solar zenith angle were selected as highly relevant inputs of the model. The regression model coefficients are updated according to a Kalman filtering (KF) scheme so that the model can flexibly adapt to fluctuations in the solar irradiance. We then modify the KF update formula to accommodate the limitation in measurement availability at the time of executing the forecasts. The proposed KF formula can be generalized to find the optimal prediction given that the available measurements are mapped by an affine transformation. The obtained results using actual data from a solar rooftop system located in the central region of Thailand showed that the normalized root-mean-square error (NRMSE) of solar power prediction was about 12-13%, which was decreased from the NRMSE of the WRF model by 7-12% on average. This improvement was the best out of similar post-processing methods based on the model output statistics framework.

Automated summary

The study aimed to develop a solar power forecasting model suitable for a tropical climate, using Thailand as a model, and proposed a linear recursive regression model to reduce errors obtained from the WRF model. By utilizing stepwise regression and a Kalman filtering scheme, the forecasting errors of the WRF model were decreased by 7-12% on average, leading to improved prediction accuracy.