Data driven modelling of a complex mining subsidence process

The prediction of subsidence rates and magnitudes is a challenging problem due to the range of complex variables that combine to determine the displacement of the surface. Many subsidence prediction models utilise an approach that involves detailed modelling of mechanical behaviour of strata transferring strain from the underground void to the surface. Such approaches are typically calibrated using subsidence records. Even after this calibration they generally struggle to predict accurately and reliably actual subsidence in virgin terrain. In this paper a model is presented based on an alternative data-driven approach using statistical techniques. This approach utilises past patterns of monitored subsidence to predict future movements at any point in space and time as a consequence of mining activities. Testing of the model proved that 89% of the estimations are between -1.65 mm/year and +1.40 mm/year of the actual subsidence value and 51% of the estimations are between -0.6 mm/year and 0.4 mm/year of the actual subsidence value.