Abstract
Slope deformation, a key factor affecting slope stability, has complexity and uncertainty. It is crucial for early warning of slope instability disasters to master the future development law of slope deformation. In this paper, a model for point prediction and probability analysis of slope deformation based on DeepAR deep learning algorithm is proposed. In addition, considering the noise problem of slope measurement data, a Gaussian-filter (GF) algorithm is used to reduce the noise of the data, and the final prediction model is the hybrid GF-DeepAR model. Firstly, the noise reduction effect of the GF algorithm is analyzed relying on two actual slope engineering cases, and the DeepAR point prediction based on the original data is also compared with the GF-DeepAR prediction based on the noise reduction data. Secondly, to verify the point prediction performance of the proposed model, it is compared with three typical point prediction models, namely, GF-LSTM, GF-XGBoost, and GF-SVR. Finally, a probability analysis framework for slope deformation is proposed based on the DeepAR algorithm characteristics, and the probability prediction performance of the GF-DeepAR model is compared with that of the GF-GPR and GF-LSTMQR models to further validate the superiority of the GF-DeepAR model. The results of the study show that: 1) The best noise reduction is achieved at the C1 and D2 sites with a standard deviation σ of 0.5. The corresponding SNR and MSE values are 34.91 (0.030) and 35.62 (0.674), respectively. 2) A comparison before and after noise reduction reveals that the R2 values for the C1 and D2 measurement points increased by 0.081 and 0.070, respectively. Additionally, the MAE decreased from 0.079 to 0.639, and the MAPE decreased from 0.737% to 0.912%. 3) The prediction intervals constructed by the GF-DeepAR model can effectively envelop the actual slope deformation curves, and the PICP in both C1 and D1 is 100%. 4) Whether it is point prediction or probability prediction, the GF-DeepAR model excels at extracting feature information from slope deformation sequences characterized by randomness and complexity. It conducts predictions with high accuracy and reliability, indicating superior performance compared to other models. The results of the study can provide a reference for the theory of slope deformation prediction, and can also provide a reference for similar projects.