Actual rainfall forecast is critical to the management and allocation of
water resources. In recent years, deep learning has been proved to be
superior to traditional forecasting methods when predicting rainfall
time series with high temporal and spatial variability. In this study,
the discrete wavelet transform (DWT) and two typical deep learning
approaches, namely long-short term memory (LSTM) and dilated causal
convolutional neural network (DCCNN), are integrated innovatively and
the hybrid model (DWT-CLSTM-DCCNN) is used for monthly rainfall
forecasting for the first time. Monthly rainfall time series of four
major cities in China (Beijing, Tianjin, Chongqing and Guangzhou) are
used as the dataset of DWT-CLSTM-DCCNN. Firstly, two methods of sample
construction are used to train DWT-CLSTM-DCCNN and their effects on the
model performance are analyzed. Then, LSTM and the dilated causal
convolutional network (DCCNN) are established as the benchmark models,
and their forecast accuracy is compared with that of DWT-CLSTM-DCCNN.
From the results of the evaluation criteria such as mean absolute error
(MAE), root mean squared error (RMSE) and Nash-Sutcliffe model
efficiency coefficient (NSE) as well as the fitting curve for forecasted
rainfall, it can be concluded that the DWT-CLSTM-DCCNN developed in this
study outperforms the benchmark models in model accuracy, peak and
mutational rainfall capturing ability. Compared with the previous
studies, DWT-CLSTM-DCCNN is proven to be better peak capture and more
suitable for long-term rainfall forecasting.