A deep learning based hybrid method for hourly solar radiation forecasting

Solar radiation forecasting is a key technology to improve the control and scheduling performance of photovoltaic power plants. In this paper, a deep learning based hybrid method for 1-hour ahead Global Horizontal Irradiance (GHI) forecasting is proposed. Specifically, a deep learning based clustering method, deep time-series clustering, is adopted to group the GHI time series data into multiple clusters to better identify its irregular patterns and thus providing a better clustering performance. Then, the Feature Attention Deep Forecasting (FADF) deep neural network is built for each cluster to generate the GHI forecasts. The developed FADF dynamically allocates different importance to different features and utilizes the weighted features to forecast the next hour GHI. The solar forecasting performance of the proposed method is evaluated with the National Solar Radiation Database. Simulation results show that the proposed method yields the most accurate solar forecasting among the smart persistence and state-of-the-art models. The proposed method reduces the root mean square error as compared to the smart persistence by 11.88% and 12.65% for the Itupiranga and Ocala dataset, respectively.

Automated summary

Solar radiation forecasting is critical in improving the performance of photovoltaic power plants, and a deep learning based hybrid method for 1-hour ahead Global Horizontal Irradiance (GHI) forecasting is proposed in this study. By utilizing deep time-series clustering and Feature Attention Deep Forecasting (FADF) deep neural network, the developed method achieves more accurate solar forecasting compared to existing models.