Abstract: In view of the nonstationarity and nonlinearity of geomagnetic variation field and the difficulty of physical model prediction, an improved LSTM neural network prediction method is proposed and verified by experiments. Firstly, the VMD method is used to process and analyze the geomagnetic station data, and then the chaos theory is introduced to optimize the sample set according to the chaotic characteristics of geomagnetic variation. Finally, the LSTM network is used as the predictor. The results show that the prediction of the optimized method has good robustness, the average MAE is less than 2nT, R² is more than 0.8, the predicted value is in good agreement with the measured value, and it is widely used in Chinese Mainland.