Exploiting Inter-Day Weather Dynamics for Improved Day-Ahead Solar Irradiance Forecasting

Accurate day-ahead solar forecasting is essential for grid stability and energy planning. This study introduces a specialized forecasting framework that enhances accuracy by training models on specific day-to-day sky condition transitions. The framework employs a dual-attention transformer model, which captures complex temporal and feature-wise relationships, using a dataset of approximately 5000 daily sequences from three sites in Mongolia (2018–2024). Our core contribution is a specialized training strategy where the dataset is first classified into nine distinct classes based on the sky condition transition from the previous day to the forecast day, such as ‘Clear’ to ‘Partly cloudy’. A dedicated transformer model is then trained for each transitional state, enabling it to become an expert on that specific weather dynamic. This specialized framework is benchmarked against a naive persistence model, a standard, generalized transformer trained on all data and a ‘cluster-then-forecast’ approach. Results show the proposed approach achieves superior performance improvement compared to baseline models (p < 0.001) across all error metrics, demonstrating the value of modeling inter-day weather dynamics. Furthermore, the framework is extended to probabilistic forecasting using quantile regression to generate 80% prediction intervals, providing crucial uncertainty information for operational decision-making in power grids.