Multi-Branch ResNet-Transformer for Short-Term Spatio-Temporal Solar Irradiance Forecasting
Saeedeh Ziyabari, Zhenyu Zhao, Liang Du, Saroj Biswas
Abstract
The increasing penetration of solar generation into power grids has promoted the need for accurate and reliable short-term solar irradiance forecasting. Existing methods utilizing advanced deep learning architectures have shown advanced performance compared to conventional time-series analytical techniques but in general encountered shortcomings in modeling spatial correlations among neighboring solar generation sites, exploring the similarity of long-term, time-varying patterns, and alleviating overfitting issues in convolutional and recurrent neural networks, such as the popular Long Short-term Memory (LSTM). To effectively but yet reliably tackle these challenges in the existing literature, this article proposes a spatio-temporal framework consisting of a multi-branch hybrid Residual network and the <monospace xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Transformer</monospace> architecture (ResTrans). The proposed framework has been tested on two groups' real-world data containing 17 years-long data from different solar sites in Philadelphia, USA, including 12 and 18 locations, respectively. Compared to other hybrid benchmark architectures, including single-branch ResTrans and multi-branch ResNet-LSTM (ResLSTM), single-branch ResLSTM, and CNN-LSTM, the proposed multi-branch ResTrans achieves the highest forecasting accuracy with an average RMSE of 0.049 (W/m <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> ), an average MAE of 0.031 (W/m <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> ), and an <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$R^{2}$</tex-math></inline-formula> coefficient of 97%.