A Multilevel Temporal Convolutional Network Model with Wavelet Decomposition and Boruta Selection for Forecasting Monthly Precipitation

Abstract In this study, a multilevel temporal convolutional network (MTCN) model is proposed for 1-month-ahead forecasting of precipitation. In the MTCN model, à trous wavelet transform (ATWT) is first utilized to decompose the standardized monthly precipitation anomaly and its candidate predictors into their components with the different time scales. Then, at each of the time levels, a temporal convolutional network (TCN) model is built to forecast the precipitation anomaly component by combining with the Boruta selection algorithm (TCN-B) for identifying important model inputs from corresponding predictor components. Finally, the precipitation forecast is achieved by summing all the forecasted anomaly components and applying the inverse transform of the standardized monthly precipitation. The proposed MTCN is tested and compared to the TCN-B and TCN using monthly precipitation at 189 stations in the Yangtze River basin. The TCN-B is formed by coupling the TCN with the Boruta algorithm. The comparison results show that the TCN-B outperforms the TCN, and the MTCN has the best performance among the three models. Compared to the TCN, the MTCN provides a significant improvement for all stations, especially for the eastern stations of the basin. It is also shown that all three models perform better in spring and summer and have the weakest abilities in winter. The MTCN has a great improvement in predicting precipitation of all four seasons compared with the other two models. Additionally, all three models exhibit better prediction performance in the western region than in the eastern region of the basin, which is strongly related to the spatial distribution of precipitation variability.