Based on the soil temperature and meteorological observation data from the Wudaogou Experimental Station in the Huaibei Plain from 2000 to 2024, the correlation function feature (CFF) method was adopted to systematically analyze the lag effects of meteorological factors, such as air temperature and air pressure, at different soil depths (10, 20, 40, 80, 160, and 320 cm). Using the lag-processed meteorological factors at each depth as inputs, the convolutional neural network (CNN), long short-term memory (LSTM), and CNN-LSTM models were driven to simulate the evolution processes of soil temperature at the six depths, and the applicability and accuracy of the three machine learning models at different depths were evaluated. The results indicate that the optimal lag time of meteorological factors increases with the increase of soil depth. The lag time of shallow soil (10-20 cm) is the shortest (0-3 d); the lag time of middle soil (40-80 cm) is 8-29 d; the lag time of deep soil (160-320 cm) extends to 44-89 d. The simulation performance of the three machine learning models exhibits clear stratified differences at different depths, with the shallow layer generally outperforming the deep layer. The CNN-LSTM model demonstrates higher accuracy and stability at all depths, and its coefficient of determination ( R 2) is generally maintained above 0.94, outperforming the single models. In the deep soil of 320 cm, the R 2 of the CNN-LSTM model is 0.941, which improves by 17% compared with the LSTM model and 45% compared with the CNN model. This reflects the advantages of coupling convolutional feature extraction with temporal memory capability.