Ensemble spatiotemporal forecasting of solar irradiation using variational Bayesian convolutional gate recurrent unit network

Solar irradiation prediction is of vital important to improve solar energy utilization. In recent years, many researches on solar irradiation prediction have been arisen. However, most forecasting model are based only on time series without considering the temporal and spatial variations of the solar energy, which hinders the progress of solar irradiation prediction. In this paper, we embed solar energy and meteorological data from multiple sites into a spatial grid and focus on the spatiotemporal solar irradiation prediction problem. An ensemble spatiotemporal deep learning model is proposed for solving the problem. The proposed model contains a convolutional operator in both the input-to-state and state-to-state transitions of the Gate Recurrent Unit, which makes it particularly suitable for spatiotemporal forecasting problems. Moreover, variational inference is employed in this deep learning model in order to quantify the uncertainty of the prediction. A real-world test case with a spatial region is used to illustrate the full potential of the proposed model. Four state-of-the-art deep learning models are considered for comparison. The experimental results demonstrate that the proposed model significantly outperforms other models in term of three widely used evaluation criteria. Furthermore, the uncertainty estimation is given and it demonstrates that the proposed model is able to provide an effective uncertainty estimation for the prediction.