In order to further leverage the advantages of real-time monitoring with global navigation satellite system (GNSS) and explore the latent features and hidden information in time-series data,enhancing the accuracy of ground subsidence prediction,the paper proposed a hybrid prediction method of surface subsidence deformation named complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN)-convolutional neural networks (CNN)-bi-directional long short-term memory (BiLSTM):focusing on a large coal mine working face and an industrial square area in northern Anhui as the validation sites,a comparative analysis of stable and critical monitoring area data morphologies was carried out;then CEEMDAN was utilized to reconstruct elevation data components at monitoring stations,and the CNN model was applied to extract implicit information from these components;finally,a BiLSTM model was constructed to achieve the short-term predictions for subsidence monitoring points.Experimental results showed that,compared to traditional CNN and long short-term memory models,the proposed hybrid model could efficiently reduce the prediction errors;specifically,the reduction of mean absolute percentage error (MAPE) would range from 40％ to 90％,and root mean square (RMS) error from 52％ to 87％;in general,the performance of the proposed model could exhibit superior capabilities in capturing spatiotemporal features and generalization,providing a more accurate and reliable solution for short-term prediction of GNSS time-series data.
维普
万方数据