Enhancing One-Day-Ahead Probabilistic Solar Power Forecast With a Hybrid Transformer-LUBE Model and Missing Data Imputation

Abstract

Forecasting solar photovoltaic (PV) power generation has become important due to solar power's chaotic and intermittent nature. With the development of artificial intelligence (AI) techniques, particularly deep learning algorithms, has proven to be very powerful for renewable power forecasts. This study aims to develop a probabilistic forecasting approach, leveraging the Transformer-Lower Upper Bound Estimation (Transformer-LUBE) hybrid model, enhanced by data processing techniques, for one-day-ahead PV power generation forecasts. This involves integrating an effective imputation technique, modifying the Transformer architecture, and employing a combined post-processing approach. In the pre-processing stage, missing data are imputed using XGBoost in combination with predictive mean matching (PMM) and bootstrapping. The Pearson Correlation Coefficient is then utilized to compute correlations between input features. Unlike the original Transformer architecture, the proposed model incorporates additional normalization and dropout layers between blocks, as well as multi-head attention mechanisms. Finally, a post-processing strategy is implemented based on daily numerical fitting curves along with Gated Recurrent Unit (GRU) to correct residual error biases. The framework is trained using historical measurement data from ten solar farms in Taiwan, as well as Numerical Weather Predictions (NWPs) data from the Taiwan Central Weather Bureau (CWB). The accuracy of the proposed model is evaluated in comparison to other AI models, including Artificial Neural Network (ANN), Long Short-Term Memory (LSTM), GRU, and XGBoost. Through comprehensive experimentation, the proposed framework demonstrates superior accuracy and reliability.