A multiscale long short-term memory model with attention mechanism for improving monthly precipitation prediction

Abstract In this study, a multiscale long short-term memory model with attention mechanism (MLSTM-AM) is proposed to improve the accuracy of monthly precipitation forecasting. In the MLSTM-AM model, a trous wavelet transform (ATWT) is first employed to decompose the standardized monthly precipitation anomaly and climate indices into a certain number of subseries with different time scales. Then at each of time scales, a long short-term memory (LSTM) model is constructed to predict the precipitation anomaly subseries by coupling with partial information (PI) algorithm for selecting the model inputs and attention mechanism (AM) for optimizing the model structure. Finally, all the predicted subseries are summed, and then the summed series is transformed as the monthly precipitation prediction. The proposed MLSTM-AM model is examined with 129 stations over the Yangtze River basin and compared with the MLSTM, LSTM and multiple linear regression (MLR). The MLSTM is formed by coupling the LSTM with the ATWT. For comparison, the inputs of all the models are identified by the PI algorithm. Results show that the prediction of all the models is more skillful in the western region than in the eastern region of the basin, which is strongly correlated with the spatial distribution of the variability of precipitation. The LSTM outperforms the MLR in terms of two performance evaluation indicators at most of stations. The MLSTM and MLSTM-AM improve monthly precipitation forecasts at all of 129 stations over the basin compared to the LSTM, and the MLSTM-AM provides the best prediction performance among all these models.