Optimal selection of time functions for describing coal mining-induced dynamic subsidence at single surface point using AIC criterion

Abstract

The time function method is one of the most commonly used methods for predicting surface dynamic displacements in coal mine areas. In which, the accuracy and reliability of the predicted displacements, to a large extent, depends on the selected mathematical functions for describing the “S”-typed dynamic subsidence at a single surface point (referred to as time functions). Nearly all of the existing studies primarily improve or introduce “S”-shaped growth functions with a single object to minimizing the fitting residuals between the in-situ monitored and the model-fitted subsidence. Such a strategy, however, would result in “overfitting” (or over-parameterization), thereby increasing the complexity of the constructed time function model and the difficulty of model parameter inversion. To this end, the optimal selection of time functions was analyzed in this paper using two indicators of fitting residual and model complexity, rather than the former one in existing studies. More specifically, time-series subsidence observations at 103 field points in seven coal mining areas with different geological mining conditions were selected to be observation samples for ensuring the applicability of the optimal time function. Then, 12 common “S-shaped” growth models were chosen to candidates, and the theoretical analysis and Akaike information criterion (AIC) were further used to analyze the optimal selection of time function from the chosen 12 “S”-shaped models. The results show that: ① Among the 12 selected models, the mean mis-fitting error of the five four-parameter models is about 3.51 cm, which is obviously smaller than that of the two-parameter Knothe model (14.10 cm), but just slightly smaller than the six three-parameter models (4.78 cm); ② In the view of making a trade-off between fitting residuals and model complexity (assessing by the AIC), the AICs of the six three-parameter models are smaller than those of the four-parameter and two-parameter models.This indicates that the three-parameters models are preferrable to describe the temporal evolution of subsidence at a single point, and the four-parameter and two-parameters models may be over-fitted and under-fitted, respectively; ③ Among the six selected models, the optimal selection of time function is related to the lithology of the overburden rock strata; that is, Hossfeld model, which has not been introduced into the time function method, is preferrable under soft and medium-hard overburden strata, whereas Weibull model is preferrable under hard overburden strata.

Keywords

• surface subsidence
• dynamic subsidence prediction
• time function
• akaike information criterion