Weather Embedding Cluster Transformer for Solar Irradiance Prediction

Solar irradiance prediction is a critical component of photovoltaic (PV) power generation systems, as energy planners must adapt to the uncertainty and variability of solar irradiance to maximize solar energy utilization. Recently, Convolutional Neural Networks (CNNs) have shown promising results in solar irradiance forecasting. However, existing methods are limited in their ability to capture multi-scale data features and often fail to account for the impact of weather conditions on solar irradiance. This paper proposes a Weather Embedding Cluster Transformer (WECT) model that integrates weather clustering and similar-day clustering techniques. First, the K-means clustering algorithm is applied to observational data—ground-level relative humidity, ambient temperature, and solar irradiance—to classify weather into four categories. Historical solar irradiance data is then preprocessed into 12-hour slices, embedding both weather and location information. Subsequently, two Transformer modules are constructed: a Temporal Transformer module and a Similar-Day Semantic Transformer module, which respectively capture short- and long-term temporal features and solar irradiance patterns. The self-attention mechanism within these modules identifies frames crucial for prediction, assigning them greater weight within the neural network. Extensive results on three real-world public datasets demonstrate that WECT can reliably predict solar irradiance.