Machine learning forecasts of short wave radiation from geostationary satellite measurements to optimize solar photovoltaic and concentrated solar power systems

Abstract

In the context of global energy transition and sustainable development, accurate short wave radiation (SWR) forecasting is increasingly vital for enhancing the efficiency and economic viability of solar photovoltaic (PV) and concentrated solar power (CSP) systems. This study presents an innovative machine-learning forecasting model of SWR within the next hour, using multi-band shortwave solar radiation measurements from the geostationary satellite. The model is based on a cloud cover-weighted hybrid model combining the convolutional long short-term memory (ConvLSTM) and Fourier neural operator (FNO) models. During testing, the optimized hybrid model performed better than ERA5 data, reducing the prediction error by 24.14%, the average absolute error by 38.62%, and improving the R² value by 6.4%. In the prediction area (longitude 104–110°, latitude 35–40°), the prediction accuracy in barren and sparsely vegetated areas was 8.13% higher compared to grasslands, indicating future potential for further enhancement through optimized solar power plant site selection. The improved model can reduce power generation losses by 0.067 USD/m² in PV systems through real-time grid regulation and other strategies, and can also prevent daily energy losses of 13.97 kWh/m² in concentrated solar power systems by timely adjusting the heliostats and receiver measures. Under the Shared Socioeconomic Pathways sustainable development scenario (SSP1-2.6), by 2100, the adoption of the hybrid SWR forecasting model is expected to increase power generation by 895.47 TWh compared to the original plan. The proposed hybrid forecasting model significantly improves solar radiation forecast accuracy, enhancing the future development of solar power generation.