Radiation multi-feature integration unit with dynamic spatial dependency: A model for coastal solar irradiance interval forecasting

With the rapid expansion of photovoltaic installations in coastal areas, accurate solar irradiance interval forecasting has become essential for ensuring stable power system operation. However, land-sea breeze systems in coastal regions cause pronounced hourly-scale variations in temperature, relative humidity, and cloud cover, creating complex spatiotemporal radiation distributions that pose significant challenges for existing forecasting methods. To address these challenges, this study proposes the Radiation Multi-feature Integration Unit with Dynamic Spatial Dependencies (RADSD) model, which incorporates two core components: (1) a dynamic spatial dependency aggregation method that adaptively adjusts inter-station spatial dependency weights based on feature similarity analysis, overcoming the limitations of fixed topological structures in capturing dynamically varying coastal meteorological conditions; and (2) a Radiation Multi-feature Integration Unit (RAMI) that integrates temporal encoding, heuristic solar position encoding, multi-scale feature extraction, and adaptive gating mechanisms to capture complex radiation temporal patterns specific to coastal environments. The model employs quantile regression to construct prediction intervals at specified confidence levels. Experiments conducted using data from Ningbo demonstrate that RADSD achieves narrower prediction intervals than baseline models while maintaining coverage rates at nominal confidence levels. Ablation studies confirm the necessity of both the Dynamic Graph Convolutional Network (DGCN) and RAMI components. Furthermore, generalization experiments in Qingdao and Wenchang verify the model's strong adaptability across diverse coastal climates. The proposed model provides reliable uncertainty quantification for solar irradiance in coastal areas, offering practical data support for photovoltaic power system scheduling and solar energy resource assessment. Unlike existing approaches that rely on static spatial structures, RADSD dynamically captures evolving inter-station spatial dependencies driven by coastal meteorological variations, while RAMI is specifically designed to extract radiation-relevant temporal features unique to coastal environments, representing a novel integration of adaptive spatial modeling and domain-specific temporal feature extraction for coastal solar irradiance interval forecasting.