Solar radiation prediction with weather scenarios for flexible decision-making in building energy systems

Accurate solar radiation forecasting is essential for buildings, where weather-driven solar variability directly impacts energy management and operational decisions. However, the complex relationships among meteorological parameters make it difficult to generate well-defined and practically usable weather scenarios, thereby limiting the model’s applicability and clarity—particularly in building and distributed energy systems where timely and informed decision-making is essential. This study presents a weather scenario-based solar radiation prediction model that integrates transformation matrices for dimensionless processing, Convolutional Neural Networks for local feature classification, multi-scenario probability generation, decision modules, and Long Short-Term Memory Networks subtasks to capture long-term uncertainties. The model proposes a well-defined set of weather solar radiation scenarios (sixteen patterns) encompassing various weather conditions and develops a unified mathematical description to achieve dimensionless processing of weather scenarios, enabling their applicability to evaluations and calculations in whole year. The results demonstrate that the proposed probabilistic method effectively differentiates between credible and highly uncertain weather solar radiation scenarios, accurately quantifying the size and duration of prediction intervals under uncertain conditions. The model is trained and validated on a weather dataset from Tokyo spanning from 2000 to 2023, maintaining high accuracy with an R² of 0.97. Furthermore, it also provides clear guidance to energy decision-makers, identifying specific scenarios and time periods that require additional measures to address uncertainties.