Regional groundwater level prediction is crucial for water resource management, especially in seasonally frozen areas. Accurate predicting groundwater levels during freeze–thaw periods is essential for optimizing water resource allocation and preventing soil salinization. Although deep learning models have been widely employed in groundwater level prediction, they remain black boxes, making it difficult to simultaneously predict groundwater levels and understand the dynamic causes. This study simulated the groundwater level dynamics of 138 monitoring wells in the Songnen Plain, China, using a long short-term memory (LSTM) neural network. The expected gradient (EG) method was applied to interpret LSTM decision principles during different periods, revealing groundwater dynamics mechanisms in seasonally frozen soil areas. The results showed that the LSTM model could accurately simulate daily groundwater level trends, with 81.88 % of monitoring sites achieving NSE above 0.7 on the test set. The EG method revealed that atmospheric precipitation was the primary source of groundwater recharge, while discharge occurred through evaporation, runoff, and artificial extraction, forming three groundwater dynamics types: precipitation infiltration–evaporation, precipitation infiltration–runoff, and extraction. During the freeze–thaw period, groundwater levels in the precipitation infiltration–evaporation type decreased during the freezing period and increased during the thawing period due to water potential gradient changes driving soil–groundwater exchange. In contrast, the precipitation infiltration–runoff and extraction types exhibited continuously increasing and decreasing trends, driven by recovery after extraction and precipitation recharge. Our findings provide essential support for groundwater resource assessment and ecological environmental protection in seasonally frozen soil areas.