Applicability of Displacement Prediction Models for Creep‐Type Highway Slopes Based on Residual Information Structure

To improve the selection of displacement prediction models for creep‐type highway slopes with low signal‐to‐noise ratios and limited monitoring samples, this study proposes a model suitability analysis framework based on residual information structure. A GNSS monitoring point on a cutting slope along the Hengyang–Yongzhou expressway was selected as the case study. Based on 123 consecutive days of three‐dimensional daily displacement data in the X, Y, and H directions, gross errors were removed using the 3σ criterion, and high‐frequency noise was reduced using adaptive Kalman filtering. Six prediction models with different structural complexities, namely rolling GM(1,1), ARIMA, BP neural network, SVR, GM(1,1)‐BP, and ARIMA‐SVR, were established and compared. The results show that Kalman filtering reduced the noise standard deviation in the X, Y, and H directions by 66.7%, 54.8%, and 77.0%, respectively. The optimal model varied significantly among directions. ARIMA performed best in the X direction, where the displacement series showed weak trends and near‐stationary fluctuations. Rolling GM(1,1) achieved the highest accuracy in the Y direction, which was dominated by monotonic slow creep. ARIMA‐SVR obtained the best comprehensive performance in the H direction, where elevation solution uncertainty and environmental disturbances caused stronger residual correlation and nonlinearity. Further analysis indicates that first‐order residual autocorrelation can be used as an auxiliary indicator for deciding whether model complexity should be increased. However, final model selection should also consider trend dominance, volatility, and nonlinear characteristics. The proposed framework provides a lightweight and data‐adaptive reference for automated highway slope monitoring and early warning systems.