Spatiotemporal Forecasting of Solar and Wind Energy Production: A Robust Deep Learning Model with Attention Framework

Abstract

The variability in the spatiotemporal distribution of power generation is a significant challenge for accurately predicting renewable energy production patterns. Furthermore, numerous forms of unforeseen data contamination degrade the precision of forecasts since superfluous data points adversely affect the regression model. In this context, a novel robust deep learning model, termed the Convolutional Neural Network-Bidirectional Long Short-Term Memory model with spatiotemporal attention mechanism (CNN-BiLSTM-STA), is developed in this study. The suggested model integrates the feature extraction expertise of CNNs with the sequence modeling proficiency of BiLSTM networks to capture spatial linkages and temporal interdependence adeptly. Moreover, the integrated spatiotemporal attention mechanism selectively focuses on significant spatial regions and time steps to enhance the prediction of spatiotemporal sequences of time-resolved grid data. The proposed architecture allows plant proprietors and system operators to obtain accurate predictions across extensive spatiotemporal patterns by eliminating the necessity for individual model fitting for each site/horizon or an additional data preprocessing phase before training. In addition, the Correntropy-based training criterion is employed to ensure the robustness of the recommended method against various types of data contamination, including data incompletion, Gaussian noises, outliers, and a mixed combination of disturbances. Furthermore, the Partial Reinforcement Optimization technique is applied to optimize the hyperparameters of the proposed model. The suggested framework incorporates numerous photovoltaic installations in Arizona and wind power installations in Texas to provide concurrent forecasts for multiple periods. The efficacy of the suggested forecasting model is evaluated by comparing it with three state-of-the-art methods. Numerical findings demonstrate that the proposed model surpasses other methods by successfully integrating spatial and temporal characteristics.