A Multi-step Water Level Prediction Model Based On CNN-LSTM-ATTENTION Combined With Wavelet Transform

In order to improve the ability to predict river water levels, a multivariate multi-step water level prediction model using wavelet transform(WT) and attention mechanism combined with convolutional neural network(CNN) and long short-term memory network(LSTM) was proposed. In this model, the water level data is decomposed and reconstructed through WT by combining the water level characteristics of multiple water level stations and rainfall information. CNN and LSTM are used to obtain local feature information and time series features of the reconstructed hydrological data sequence. Finally, attention mechanism is introduced to assign adaptive weights to the output sequence of LSTM, and LSTM is used to output multi-step predicted water level data. The experimental results show that the water level prediction model incorporating rainfall information has achieved multi-step water level prediction for 96 consecutive hours, and has achieved good results in various evaluation indicators.