Based on the meteorological data, social data and groundwater depth data of Beijing-Tianjin-Hebei, support vector machine (SVM), recurrent neural network (RNN) and long and short-term memory neural network (LSTM) were constructed to simulate the groundwater depth of 13 cities. The adaptability of three models were evaluated in indicators, such as the determination coefficient, root mean square error, mean absolute percentage error, and Nash coefficient. The results showed that LSTM model performed the best, followed by RNN and SVM. Meanwhile, simulation results of different cities demonstrated the least parameter adjustments and the best adaptability of LSTM model for the groundwater depth simulation, and SVM model parameters had the most parameters adjustments. Three models were applied to 6 randomly selected stations for verification and it was shown that in the shallow groundwater depth simulation in North China, LSTM model had good simulation accuracy and reliability, with strong adaptability. Therefore, it is the first choice in the groundwater depth simulation of the studying area.