Multi-Site Solar Irradiance Forecasting Using Spatiotemporal Deep Learning Networks with Sequential Imagery
Year: 2026
Type: conference-paper
Abstract: Constructing accurate spatiotemporal correlations is a challenging task in the very short-term prediction of solar irradiance across multiple distributed sites. While all-sky imagers have significantly improved forecasting capabilities, most existing image-based deep learning models are strictly limited to single-site predictions. Furthermore, current multi-site methods often rely on computationally heavy graph matrices and require dense, expensive sensor networks. To move beyond these single-point limitations, this paper proposes a streamlined spatiotemporal model based on a Convolutional Long Short-Term Memory (CNN-LSTM) network explicitly designed for joint multi-site forecasting using a single optical source. In this approach, a CNN module processes sequential infrared all-sky images acquired from a single centralized location to capture real-time macroscopic cloud trajectories. These unified regional spatial features are then fused with distinct local weather data sequences (historical irradiance, temperature, and humidity) collected from multiple distributed sites. The combined multimodal data is fed into a 2-layer LSTM to predict nonlinear temporal dynamics and output simultaneous Global Horizontal Irradiance (GHI) forecasts for all target locations. To assess the predictive performance, 2 distributed solar radiation observation stations were selected for evaluation. Traditional single-site persistence models were used as baselines. Compared with the baselines, the proposed centralized-image approach significantly reduces the root mean square error (RMSE) and enhances the Forecast Skill (FS). The results compellingly demonstrate that leveraging a single infrared sky imager combined with distributed local meteorological data can substantially improve multi-point predictive accuracy without the need for complex graph-based architectures.
Institution: Politecnico di Milano
Language: English
Field: Energy
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7202210376