Microclimate prediction for sandy photovoltaic power plants using a spatio-temporal graph convolutional network with environmental covariates

The deployment of photovoltaic (PV) power plants in sandy regions alters the local microclimate. Accurate microclimate prediction is essential for optimizing short-term operations and maintenance plans. Compared with traditional statistical and numerical models, neural networks are suited to capture the complex nonlinear dependencies in sandy-region PV environments. Graph neural networks (GNNs) extend neural networks to process data defined on graph domains, enabling them to confront high-dimensional feature redundancy, attenuated temporal dependencies, and pronounced spatial heterogeneity. Accordingly, we propose a spatio-temporal graph convolutional network with environmental covariates (EC-STGCN) for station-level microclimate prediction at meteorological stations. Experimental results demonstrate that EC-STGCN achieves accurate high-resolution spatio-temporal predictions of temperature, relative humidity (RH), and particulate matter (PM 10 ) and shows improved predictive performance relative to conventional GNNs, while covariate-level analysis shows that individual environmental covariates contribute heterogeneously to these predictions.