Water Inflow Forecasting Based on Visual MODFLOW and GS-SARIMA-LSTM Methods

Abstract

Mine water inflow is a significant safety concern in coal mine operations. Accurately predicting the volume of mine water inflow is vital for ensuring mine safety and environmental protection. This study focused on the Laohutai mining area in Liaoning, China, to reduce the reliance on hydrogeological parameters in the mine water inflow prediction process. An integrated approach combining grid search (GS) with the Seasonal Autoregressive Integrated Moving Average (SARIMA) and Long Short-Term Memory (LSTM) model was proposed, and its results were compared with Visual MODFLOW. The grid search was used to optimize the SARIMA model, modeling the linear component of nine years of water inflow data, with the remaining six months of data used for model validation. Subsequently, the prediction residuals from the SARIMA model were input into the LSTM model to capture the nonlinear features in the data and enhance the generalization capability and stability of the LSTM model by introducing Dropout, EarlyStopping, and the Adam optimizer. This model effectively handles long-term trends and seasonal fluctuations in the data while overcoming limitations in capturing periodicity and trends in complex time series data. The results indicated that the GC-SARIMA-LSTM model performs better than the Visual MODFLOW numerical simulation software in predicting mine water inflow. Therefore, without hydrogeological parameters, the GC-SARIMA-LSTM model can serve as an effective tool for short-term prediction, advancing the application of deep learning in coal mine water inflow forecasting and providing reliable technical support for mine water hazard prevention.