Multi-Branch ResNet-Transformer Based Deep Hybrid Approach for Short-term Spatio-Temporal Solar Irradiance Forecasting

Motivated by existing challenges on how to effectively model the spatial correlation and similar time-varying patterns among neighboring solar sites as well as alleviate performance degradation due to overfitting of conventional neural networks such as long short-term memory (LSTM) caused by long temporal inputs, we propose a spatio-temporal framework comprised of concatenated multi-branch residual networks and transformers (ResTrans). Using 17 years of data from 12 different solar sites in Philadelphia, USA, we demonstrate that the proposed approach exhibits enhanced performance compared to hybrid benchmark architectures such as single-branch ResTrans and multi-branch ResNet-LSTM (ResLSTM), single-branch ResLSTM, and CNN-LSTM.