Short-Term Spatio-Temporal Solar Irradiance Forecasting using Multi-Resolution Deep Learning Models

Abstract

Accurate solar generation forecasting is critical for ensuring power system reliability, economics, and effectiveness and controlling the supply-demand balance. This research offers novel multi-branch spatio-temporal forecasting models to improve forecasting accuracy and minimize forecasting errors. The first step is to build temporal models employing advanced deep learning architectures, such as Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and GRU with Attention (AttGRU). Next, spatio-temporal solar forecasting models are constructed. A novel multi-branch Attentive Gated Recurrent Residual network (ResAttGRU) consisting of multiple branches of residual networks (ResNet), GRU, and the attention mechanism is introduced. The proposed multi-branch ResAttGRU is capable of modeling data at various resolutions, extracting hierarchical features, and capturing short- and long-term dependencies. Moreover, this network also presents a strong multi-time-scale representative, while GRUs can exploit temporal information at less computational cost than the popular LSTM. The novelty of the developed architecture is in the utilization of multiple convolutional-based branches to learn multi-time-scale features jointly, accelerate the learning process, and reduce overfitting. This dissertation also compares the multi-branch ResAttGRU networks with state-of-the-art deep learning methods using 18 years of NSRDB data at 12 solar sites. The proposed multi-branch ResAttGRU requires 7.1% fewer parameters than multi-branch residual LSTM (ResLSTM) while achieving similar average RMSE, MAE, and R-squared values. Finally, to effectively model spatial correlation among neighboring solar sites as well as to alleviate performance degradation due to overfitting of conventional neural networks, a spatio-temporal framework comprised of concatenated multi-branch Residual network and Transformer (ResTrans) is developed. Numerical results indicate that the multi-branch ResTrans structure achieves the highest forecasting accuracy, with an average RMSE of 0.049 ( W/m^2 ), an average MAE of 0.031 (W/m^2 ), and a R^2 coefficient of 97%.