Self-calibrated hybrid weather forecasters for solar thermal and photovoltaic power plants

Abstract

This study presents two novel ensemble forecasters that combine time delay and recurrent neural networks to predict the global horizontal irradiance (GHI), direct normal irradiance (DNI), and ambient temperature (TMP). The weights of these forecasters are optimized to minimize the summation of squared errors between the actual and the forecasted values of sub-models during the previous forecasting step. The developed models have been benchmarked against typical forecasting models for three forecasting horizons. Afterward, the impact of forecasting errors on the power estimation of photovoltaic (PV) and concentrated solar power (CSP) plants was examined. The two proposed forecasters consistently outperformed persistence ones by up to 88.653%. Compared to the aggregated model, the newly proposed forecasters enhanced the predictions of GHI, DNI, and TMP by up to 39.392, 27.081, and 79.598% in terms of mean absolute error, and up to 21.501, 13.645, and 78.961% in terms of root mean square error. The absolute and relative performances of the proposed forecasters also improved as the forecasting horizon decreased. Finally, the performances of the daily and the instantaneously calibrated models were comparable in both plants. Hence, the daily calibrated model is generally recommended as it requires significantly lower computational costs.