A multi-step ahead global solar radiation prediction method using an attention-based transformer model with an interpretable mechanism

The conventional multi-step ahead solar radiation prediction method ignores the time-dependence of a future solar radiation time series. Therefore, according to sequence-to-sequence (seq2seq) model theory, this paper proposes the seq2seq long- and short-term memory model (seq2seq-LSTM), the seq2seq-LSTM model with an attention mechanism (seq2seq-at-LSTM), and a transformer model, which consists only of the attention mechanism. The hourly global solar radiation data between 2016 and 2018 from Shaanxi, China, is used to train and validate the models. The results show that the introduction of the attention mechanism can effectively improve the prediction accuracy of the seq2seq-LSTM model. However, the model is still not very good at capturing the long-distance dependence of the solar radiation time series due to the inherent properties of LSTM. In comparison, the transformer model, which is based entirely on the attention mechanism, performs much better at capturing the long-distance dependence of the solar radiation time series. Furthermore, as the number of time-steps increases, the performance of the solar radiation prediction decreases relatively smoothly and slowly. The obtained average coefficient of determination, root mean square error (RMSE), relative RMSE, and mean bias error are 0.9788, 72.91 W/m2, 25.25%, and 38.35 W/m2, respectively. In addition, the average skill score of the transformer model is around 44.9%, which is 20.54% higher than that of the seq2seq-at-LSTM model and about 40.84% higher than that of the seq2seq-LSTM model. Besides, the use of the attention mechanism can explain the improved prediction compared to other models. This model developed in this study could also be used for predictions in other fields, such as wind energy predictions and building energy predictions.