Integrating Exogenous Variables Into Cross-Attention Transformers for Robust 15-Minute-Ahead Solar Power Prediction

This research introduces a Transformer-based forecasting framework based on TimeXer, which has been developed to forecast solar energy production over brief intervals of 15 minutes. The primary objective of this study is to enhance both the precision and efficacy of solar power prediction through the adept utilization of meteorological data. The proposed model is systematically evaluated against three widely adopted deep learning baselines, including Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and the Vanilla Transformer, under a unified experimental protocol. Experiments are conducted on a multi-plant dataset comprising eight utility-scale photovoltaic (PV) power stations with exogenous meteorological variables.Results demonstrate that TimeXer achieves superior and more generalizable forecasting accuracy in terms of Mean Absolute Error (MAE) and Normalized Mean Absolute Percentage Error (MAPE), while maintaining competitive Root Mean Squared Error (RMSE) performance. It exhibited stability and reliability even in scenarios where various types of input data were either introduced or removed, thereby illustrating its capacity to discern significant patterns within weather data. The incorporation of extended historical data significantly enhanced prediction accuracy, particularly for larger solar power installations. TimeXer operates by independently processing solar power generation data alongside external meteorological variables while permitting their interaction, thus enabling the model to comprehend intricate relationships and produce more accurate forecasts. Additional analyses, including feature-adjustment experiments, data-size sensitivity evaluation, computational efficiency assessment, and residual distribution analysis, further confirm the robustness and practical applicability of the proposed approach. This study provides a comprehensive and reproducible benchmark for 15-minute-ahead solar power forecasting and demonstrates the effectiveness of exogenous-aware cross-attention mechanisms in improving short-term PV prediction accuracy.