Karst springs in north China are important natural resources with multiple attributes of landscape, culture and tourism, which play important roles in the development of local economy and society. Monitoring data of precipitation, groundwater withdrawal and artificial ecological recharge of Baotu Spring from 2016 to 2018 are collected for modelling purposes in order to achieve more accurate prediction results of dynamic changes of the karst spring. Meanwhile, 6 types of BP neural network and genetic algorithm optimized BP neural network are established to predict the water level of the Baotu Spring, and then the prediction results are compared and evaluated. The results show that the BP neural network optimized by genetic algorithm can improve the prediction stability, reduce the maximum iteration of neural network and save a great deal of calculation cost, compared to BP neural network. The GA-BP(LM) neural network which adopts Levenberg-Marquardt training method is more suitable for the prediction of karst spring water level due to its advantages of stable performance, low calculation cost and small prediction error. This research can provide references for the design and implementation of protection measures for karst springs in north China.