The invention provides a water level prediction method based on a hybrid model of a neural network and local Kalman filtering. An original time sequence array of water level data is collected, and a root-mean-square test method is adopted to determine stability of the original time sequence array; if a stability condition is not met, difference processing is carried out on the original data untila stability test thereof is successful; autocorrelation coefficients and partial autocorrelation coefficients are adopted to determine a water level time sequence lag period after a stable data set is acquired; the original time sequence array is split into a set of training samples of which length is the water level time sequence lag period; the artificial neural network is established and trained, and an artificial neural network model is generated to acquire a water level prediction result of a next day; and the water level prediction result of the next day is corrected by Kalman filtering. According to the method, water level data restoration and short-term prediction can be effectively realized, and prediction precision can be improved.