Abstract: Based on Long Short-Term Memory (LSTM) neural network, anomalous variations of water level in Lu-04 well and response mechanisms before the M5.5 Pingyuan earthquake on August 6, 2023, were analyzed using well water level, air pressure, precipitation data, and solid tide parameters (2014—2023). The results show: ① The multivariate LSTM model achieved a mean squared error (MSE) of 0.0007 and a coefficient of determination (R²) of 0.9978 when predicting the well water level in 2022, indicating that the model has a good ability to capture the nonlinear time series characteristics of the Lu-04 well water level. ② The prediction bias of the model showed an increasing trend in the six months pre-earthquake, with a sharp rise in the month before the earthquake. Concurrently, the tidal factor surged to an abnormal upper limit (mean value twice the standard deviation) in the month before the earthquake, suggesting that the hydrological geological conditions of the aquifer may have changed. ③ In the month of the earthquake, the phase difference increased significantly from a mixed flow pattern (with positive-negative fluctuations) before the earthquake to a vertical flow-dominated pattern at a positive peak, revealing adjustments in the seepage state of the aquifer system during the fault rupture process.

Key words: LSTM neural network, Well water level prediction, Solid tide, Earthquake precursors