Abstract: [Purposes] To assess the risk of regional land subsidence and clarify the future development trends of regional land subsidence，it is necessary to predict land subsidence and discuss the effectiveness of various prediction methods. [Methods] In the paper，based on long-term level monitoring data，land subsidence characteristics and its relationships with groundwater level in central urban area of Ningbo were analyzed，then some deep learning models including back propagation(BP) neural network，long short-term memory (LSTM)and attention-mechanism-based long short-term memory (AM-LSTM)were used for land subsidence prediction in the study area，the corresponding prediction errors and prediction results were compared and analyzed. [Findings] The results show that： ①The prediction effect of cumulative settlement data is significantly better than that of settlement rate，and monitoring data volume and prediction error show a roughly linear relationship.②After the groundwater was banned from extraction，more than 80% of groundwater monitoring points in the study research area show weak correlation with ground subsidence，and the prediction error of land subsidence increases when groundwater level factors are taken into account.③After removing individual monitoring points，The average relative errors for BP and LSTM predictions are 8.4%～15.8% and 7.1%～7.7%，respectively.The predicted trends of land subsidence development by BP and LSTM are basically consistent，but the LSTM predictions are more stable.④When the amount of monitoring data and the sliding window size is limited，the prediction error of the AM-LSTM model incorporating attention mechanisms has essentially not decreased. [Conclusions] This study demonstrates that using the deep learning methods for subsidence time series prediction is feasible based on long-term level monitoring data，however，abnormal monitoring point quantities and prediction results vary with the different prediction models.When selecting prediction models，sample verification and error analysis should be considered．

Key words: level monitoring data, land subsidence, deep learning model, prediction error, time series prediction