Construction and application of a TCN-LSTM-SVM-based time series prediction model for water inflow in coal seam roofs

Background The accurate prediction of mine water inflow plays a significant role in the prevention and control of water hazards and the safe and efficient production in coal mines.

Methods To construct a prediction model of water inflow in mines under threat of water hazards from extremely thick sandstones aquifer in West China, this study investigated a typical mine (also referred to as the studied mine) severely affected by such water hazards in the Binchang mining area of Shaanxi Province. The correlation between the mining footage and water inflow of the mining face was selected as the characteristic variable for the time series prediction of mine water inflow. Accordingly,this study proposed a prediction model for water inflow along the mining face in the studied mine based on the temporal convolutional network (TCN), long short-term memory (LSTM), and support vector machine (SVM)—the TCN-LSTM-SVM model. First, by raw data processing using the TCN framework, this model extracted the dependency between mining footage and water inflow and its dynamic characteristics. Subsequently, the extracted characteristics were output to the LSTM-SVM model to further capture the time series relationship between mining footage and water inflow and its characteristics.

Results The training and prediction results indicate that the TCN-LSTM-SVM model yielded mean absolute errors ( E_\mathrmMA ) ranging from 56.02 m³/h to 129.89 m³/h, mean absolute percentage errors ( E_\mathrmMAP ) from 3 % to 7 %, root mean square errors ( E_\mathrmRMS ) from 82.60 m³/h to 162.61 m³/h, and coefficients of determination ( R^2 ) from 0.81 to 0.98 based on the training, validation, and test sets. This model exhibited more accurate prediction results compared to the commonly used prediction models like backpropagation neural network (BPNN), random forest (RF), and Transformer while avoiding excessive errors produced by most of these models on the validation and test sets. The results indicate that the TCN-LSTM-SVM model integrated the parallel processing advantages and multi-scale feature extraction capacity of the TCN model while also enjoying the excellent prediction performance and generalization capability of the LSTM-SVM model. Compared to previously developed models, the TCN-LSTM-SVM model demonstrated certain superiority and applicability in the prediction of water inflow along the mining face in the studied mine.

Conclusions The results of this study provide a new approach to water inflow prediction for mines with similar geological conditions to those in the Binchang mining area. Therefore, this study holds practical implications for water inflow prediction and water prevention and control in mining faces with similar geological conditions to those in the studied mine.