Improving the efficiency of forecasting seismic activity during mining of coal reserves in mines using neural network algorithms

A new approach is proposed to the real-time forecasting of seismicity during mining operations in underground mines. It is based on an original approach to forecasting of non-stationary irregular time series using a classifier based on a neural network with the SWIN Transformer architecture. The main task solved is to develop a system to convert standard seismic data on coordinates and magnitudes of events into an image that can be analyzed by a neural network. The forecast is represented as the probability that no event with energy higher than the set limit for the class will occur within the forecast interval. Three classes of events, i.e. with high, medium or low energy, are introduced, the boundaries between which are defined by statistical characteristics of the current data magnitudes. The forecast interval is defined by the number of future events. Using data from one of the mines as a case study, the properties of coordinate and magnitude distributions are assessed, the accuracy of the model depending on the compilation parameters of the dataset is evaluated, and examples of forecasts made using historical data are provided.