Abstract: Accurate forecast of water level is an effective measure to prevent flood disasters. Under the background of the continuous development of in-depth learning, a combined hydrological time series forecast model based on convolutional neural network and Markov chain is proposed. The model solves the problems that the existing algorithms do not consider the spatial correlation between stations, multi-dimensional input will increase the complexity of data reconstruction in feature extraction, and the single model only considers the linear part of water level time series without considering the non-linear part, which leads to the low forecast accuracy. Firstly, the combined model uses convolutional neural network to train water level time series and rainfall time series to predict future water level and calculates residual series with original time series. Then, the residual forecast results obtained by Markov chain training residual series and the value of convolution neural network forecast are added together to get the final result. Experiments show that this method can achieve better forecast accuracy than the existing algorithms.

Key words: hydrological time series, spatial correlation, forecasting, Convolutional Neural Network(CNN), Markov Chain(MC)