Deep Learning Shield Attitude Prediction Model Based on Grey Correlation Analysis

Objective A stacking ensemble prediction model based on long short-term memory (LSTM) and support vector regression (SVR) is proposed to address the issue of shield attitude deviation from the tunnel design axis.Methods The grey correlation analysis was employed to remove the rolling angle parameters with low grey correlation and then denoise them using discrete wavelet transform (DWT). The DWT–LSTM–SVR stacking ensemble prediction model was obtained by optimally weighting the processed data after prediction by two single models separately. The cutter horizontal displacement, shield tail horizontal displacement, cutter vertical displacement, and shield tail vertical displacement were used as the output variables of the prediction model. In addition, 22 tunneling parameters, such as jack pressure, cutterhead torque, cutterhead speed, tunneling speed, pitch angle, primary jack stroke, and soil chamber pressure, along with five stratigraphic parameters, including natural unit weight, dry density, and specific gravity, were used as input variables. The DWT–LSTM–SVR model was applied to the A3 bidding section of the water resources allocation project in the Pearl River Delta. A total of 410 datasets were collected from the A3 bidding section. These 410 sample datasets were divided into a training set and a test set in a 9:1 ratio. The data from rings 968～1 336 formed the training set for the combined model, while the data from rings 1 337～1 377 constituted the test set.Results and Discussions The results showed that the errors between the predicted and actual values of the cutter horizontal displacement, shield tail horizontal displacement, cutter vertical displacement, and shield tail vertical displacement shield attitude parameters of the DWT–LSTM–SVR model were smaller than those of the DWT–LSTM model, DWT–SVR model, and AWT–LSTM–SVR model. This finding indicated that the model effectively combined the single models and that eliminating parameters with low correlation to the cutter horizontal displacement, shield tail horizontal displacement, cutter vertical displacement, and shield tail vertical displacement shield attitude parameters improved the model’s prediction accuracy. In addition, the evaluation metrics of the DWT–LSTM–SVR model met the construction error requirements, with the best prediction achieved for the shield tail displacement values. The MAPE of the cutter horizontal displacement was 0.061, the MAPE of the cutter vertical displacement was 0.066, the MAPE of the shield tail horizontal displacement was 0.028, and the MAPE of the shield tail vertical displacement was 0.011. The <italic>R</italic>² of the cutter horizontal displacement was 0.996, the <italic>R</italic>² of the cutter vertical displacement was 0.960, the <italic>R</italic>² of the shield tail horizontal displacement was 0.986, and the <italic>R</italic>² of the shield tail vertical displacement was 0.985. These results indicated that the combined DWT–LSTM–SVR prediction model met the model design requirements. Finally, different datasets were established to analyze the sensitivity of data quantity to the prediction accuracy of the stacking ensemble model. It was found that the greater the data quantity, the higher the prediction accuracy of the stacking ensemble model. As the number of datasets decreased, the prediction results of the DWT–LSTM–SVR model began to show large deviations, especially in the last 10 rings.Conclusions The results indicated that the model’s prediction for the first 15 rings is less sensitive to changes in the dataset, whereas the prediction for the last 10 rings is more sensitive to dataset variations. The DWT–LSTM–SVR model can serve as a reference for the advance adjustment of other shield attitudes.