Prediction of Underground Mine’s Surface Subsidence using a Recursive Multi-Step Forecasting Model with an Artificial Neural Network

Abstract

The subsidence of the surface due to mining activities is a significant issue in mining areas. Therefore, predicting surface subsidence is a necessary task to ensure safety and production efficiency. This article applied an Artificial Neural Network (ANN) model to predict surface subsidence resulting from underground mining operations in the Mong Dương mine. The ANN model proposed in this research uses a recursive multi-step forecasting model, where the predicted value at the previous step is added to the time series to forecast the next value. The experimental dataset consists of 12 monitoring cycles over 24 months, with a 2-month interval, divided into a training set containing the first 9 measurement cycles and a test set containing the last 3 cycles. First, the K-fold cross-validation method is applied to the training set to determine the best parameters for the model. Then, these parameters are used to predict surface subsidence for the values in the test set. The prediction error depends on the time gap between the last measurement cycle and the forecasting cycle. The relative errors in the tenth cycle for the four points are 0.9%, -1.7%, -1.7%, and 1.4%. These error values increase to 1.4%, -1.8%, -1.8%, and -1.7% in the eleventh cycle and further to 2.0%, -2.2%, -2.2%, and 2.5% in the twelfth cycle. The absolute errors are determined to be small, within the range of 20 mm. These results demonstrate that the proposed method and ANN model are suitable for the time-series monitoring data in mining areas.