Improving Solar Radiation Forecasting in Cloudy Conditions by Integrating Satellite Observations

Solar radiation forecasting is the basis of building a robust solar power system. Most ground-based forecasting methods are unable to consider the impact of cloud changes on future solar radiation. To alleviate this limitation, this study develops a hybrid network which relies on a convolutional neural network to extract cloud motion patterns from time series of satellite observations and a long short-term memory neural network to establish the relationship between future solar radiation and cloud information, as well as antecedent measurements. We carefully select the optimal scales to consider the spatial and temporal correlations of solar radiation and design test experiments at ten stations to check the model performance in various climate zones. The results demonstrate that the solar radiation forecasting accuracy is considerably improved, particularly in cloudy conditions, compared with purely ground-based models. The maximum magnitude of improvements reaches up to 50 W/m 2 (15%) in terms of the (relative) root mean squared error (RMSE) for 1 h ahead forecasts. The network achieves superior forecasts with correlation coefficients varying from 0.96 at 1 h ahead to 0.85 at 6 h ahead. Forecast errors are related to cloud regimes, of which the cloud amount leads to a maximum relative RMSE difference of about 50% with an additional 5% from cloud variability. This study ascertains that multi-source data fusion contributes to a better simulation of cloud impacts and a combination of different deep learning techniques enables more reliable forecasts of solar radiation. In addition, multi-step forecasts with a low latency make the advance planning and management of solar energy possible in practical applications.