Dynamic monitoring data of groundwater level is an important basis for understanding the current situation of groundwater development and for the utilization and planning of sustainable exploitation. Dynamic monitoring data of groundwater level are typical spatio-temporal sequence data, which have the characteristics of non-linearity and strong spatio-temporal correlation. The trend of dynamic change of groundwater level is the key factor for the optimal allocation of groundwater resources. However, most of the existing groundwater level prediction models are insufficient in considering temporal and spatial factors and their spatio-temporal correlation. Therefore, construction of a space–time prediction model of groundwater level considering space–time factors and improving the prediction accuracy of groundwater level dynamic changes is of considerable theoretical and practical importance for the sustainable development of groundwater resources utilization. Based on the analysis of spatial–temporal characteristics of groundwater level of the pore confined aquifer II of Changwu area in the Yangtze River Delta region of China, the wavelet transform method was used to remove the noise in the original data, and the K-nearest neighbor (KNN) method was used to calculate the water level. The spatial–temporal dataset and the long short-term memory (LSTM) were reconstructed by screening the spatial correlation of the monitoring wells in the study area. A spatio-temporal KNN-LSTM prediction model for groundwater level considering spatio-temporal factors was also constructed. The reliability and accuracy of KNN-LSTM, LSTM, support vector regression (SVR), and autoregressive integrated moving average (ARIMA) model were evaluated by a cross-validation algorithm. Results showed that the prediction accuracy of KNN-LSTM is 20.68%, 46.54%, and 55.34% higher than that of the other single prediction models (LSTM, SVR, and ARIMA, respectively).

The traditional advancing front technique algorithm encounters many problems due to the complex geometric characteristics of the front edge shape. These problems include poor quality, a slow algorithm, low robustness, and the inability of the mesh unit to converge. To address these problems, an optimized adaptive triangular mesh dynamic generation algorithm called R-TIN is proposed and applied to 3D engineering geological modeling in this study. Firstly, all the shapes involved in advancing the front edge inward were classified into four types, and then the optimal triangular unit was constructed by using the candidate mesh point heuristic algorithm. Then, the robustness of this algorithm could be maintained by the graded concession of the included angle threshold in the adjacent front-line segments. Finally, based on 160 engineering geological boreholes in the study area, the 3D engineering geological model was constructed and the accuracy and visualization effect of the overall geological model have been greatly improved, which can better present the spatial distribution of strata and lithological characteristics. At the same time, this algorithm can be used in geoscience information services to support the regional or national exploration of resources and energy, sustainable development and utilization, environmental protection and the prevention of geological disasters.

The over-exploitation of groundwater has led to a significant drop in groundwater levels, which may lead to a series of geological disasters and ecological environmental problems such as ground subsidence and ground cracks. Therefore, through studying the dynamic change characteristics of groundwater, we can grasp the dynamic changes in groundwater level over time and invert the hydrogeological parameters, which provides an important basis for the management of groundwater resources. In this study, the confined aquifer III groundwater between 2005 and 2014 in Yancheng City was selected as the research object, and the Back Propagation (BP) neural network, Spatial-temporal Auto Regressive and Moving Average (STARMA) model, and BP-STARMA model were used to predict the spatial and temporal evolution trends of groundwater. In order to compare the prediction effectiveness of the BP-STARMA model, the fitting and prediction accuracies of the three models were measured from the perspectives of time and space. The results of the Relative Squared Error (RSE), Normal Mean Squared Error (NMSE), Root-Mean-Squared Error (RMSE), and Mean Absolute Error (MAE) were used to assess the robustness of the BP-STARMA model. The results showed that the fitting of the RMSE of BP-STARMA model was reduced by 39.92%, 38.35%, 30.25%, 31.55%, and 13.57% compared with the STARMA model, and by 22.2%, 8.7%, 15.9%, 28.5%, and 4.42% compared with the BP neural network model, respectively. Collectively, this shows that the BP-STARMA model has a better spatiotemporal prediction of groundwater level than the STARMA and BP neural network models, is more applicable to spatially continuous time-discrete spatiotemporal sequences, and is more applicable to spatiotemporal sequences that respond to natural geographic phenomena.