An adaptive multi-modeling approach to solar nowcasting

• We describe a multi-modeling approach to solar nowcasting.
• Best predictions are selected from diverse forecasting models.
• A classification model trained on forecasting results governs model selection.
• An application to ground measurement data from Doha, Qatar, is demonstrated.
• Skill score improvements of 44.92% over the baseline, and 19.06% over the best model.

The ability to forecast solar irradiance in near-real time (nowcasting) is crucial in managing the integration of solar energy in power grids. This paper focuses on minute-by-minute forecasts of the normalized clearness index, a measure of global horizontal irradiation, within a fifteen steps-ahead temporal horizon, using data collected with a radiometric station in Doha, Qatar, for the period January–December 2014. We describe a novel multi-modeling approach to solar forecasting that uses supervised classification of forecasting evaluation results from diverse stochastic models to select the best predictions, according to their expected superiority in terms of lower error rate. The hypothesis that such a multi-modeling approach rivals the performance of any single forecasting model is tested with reference to two autoregressive models, of order 3 and 11 respectively, a support vector regression model, and a persistence model which provide the baseline for solar prediction. The advantages of the proposed approach are demonstrated in an experimental evaluation where its application with these four models shows a relative skill score improvement of 44.92% over the baseline model, and 19.06% over the best performing model (autoregressive of order 11).