Predicting the water-yield properties of K₂ limestones based on multivariate LSTM neural network: A case study of the Poli mining area in Yangquan

The Taiyuan Formation K₂ limestones are the main aquifer in the upper part of the No.15 coal seam in the Poli mining area, Yangquan City. Therefore, determining the water yield properties of K₂ limestones is critical to the safe mining of coal seams in the upper and lower formations. To determine the exact distribution of areas with high water-yield properties of the K₂ limestones, this study determined the accurate spatial distribution of K₂ limestones using the conventional wave impedance inversion firstly. Then, nine optimal seismic attributes were selected using the Pearson correlation coefficient method and the cross-validation method in stepwise regression in order to form the training data. By introducing the long short-term memory (LSTM) neural network, which is applicable for processing time-series data and is capable of capturing the correlation with log curves, this study established a multivariate LSTM neural network-based intelligent model for apparent resistivity prediction (also referred to as the multivariate LSTM-based prediction model). The purpose is to accurately predict the apparent resistivity of the study area and further obtain the water yield properties of K₂ limestones. Moreover, this study established the mapping relationship between resistivity log curves of the well locations and the seismic attributes of near-well seismic traces using the conventional multivariate regression algorithm and the multivariate LSTM-based prediction model, respectively. Finally, the multivariate LSTM-based prediction model trained using the data on the well locations were extended to the areas without wells to obtain the apparent resistivity volume of the whole study area. Subsequently, the areas with high water yield properties in the limestones were delineated according to the apparent resistivity values, as well as the development of the geological structures and collapse columns in the mining area. As shown by the test results of actual data, compared to the conventional multivariate regression algorithm, the multivariate LSTM-based prediction model yielded smaller prediction errors and higher correlation coefficients with logs. Therefore, the multivariate LSTM-based prediction model can accurately predict the apparent resistivity of a survey area and is of high application value in predicting the water-yield properties of coal-bearing strata.