Abstract: The effect of the traditional groundwater level prediction model is not ideal due to the lack of considering the temporal and spatial variation rules and the relevant influencing factors of groundwater level.The authors used the long and short term memory neural network (LSTM), adopted the method of multivariable input,and constructed the groundwater level prediction model. Taking the monitoring Well J1 in Daiyue district of Tai'an City as an example, the groundwater level dynamic monitoring data from 2001 to 2014 and the relevant influencing factor data were used to predict the groundwater level from 2015 to 2016 by using the multivariable LSTM neural network, and further compared with the single variable LSTM neural network model and BP neural network model. The research results show that the BP neural network prediction model can only make the corresponding prediction according to the change of the influencing variables without considering the change rule of timing sequence, so the prediction error is large (2.399 3). The single-variable LSTM model only takes the groundwater level as the variable input with no considering the influence of relevant factors, and the prediction error is 2.102 2. The prediction accuracy of the LSTM neural network prediction model based on multivariable input is significantly higher than that of the univariate LSTM neural network and BP neural network models, and the prediction root mean square error is only 1.919 1, which successfully verifies the accuracy of the multivariable LSTM neural network groundwater level prediction model.

Key words: groundwater level prediction, long short-term memory, multivariate