A multi-level prediction model of concrete dam displacement considering time hysteresis and residual correction

Displacement directly reflects the operational status of dams and exhibits time hysteresis. Moreover, data-driven predictive models fail to account for all physical mechanisms, leaving valuable information uncaptured in residuals. Hence, this study establishes a multi-level displacement prediction model for concrete dams considering the time hysteresis of monitoring data and residual correction. Initially, employing Bayesian optimization gated recurrent units (BO-GRU) and considering multivariate time series (MTS) prediction, a single-level displacement prediction model, MTS-BO-GRU, is developed to capture displacement time hysteresis. Subsequently, acknowledging the chaotic characteristics in residual sequences, a random forest (RF) model is utilized in conjunction with univariate time series (UTS) prediction to correct residual sequences, resulting in the UTS-RF model. The corrected values are then combined with the predicted values of the MTS-BO-GRU model to establish the MTS-BO-GRU+ multi-level displacement prediction model. Finally, employing an in-service concrete dam as a case study, the performance of the proposed multi-level model is compared to validate and evaluate its superiority. Results demonstrate that the MTS-BO-GRU+ model, considering displacement time hysteresis and residual correction, exhibits the best predictive performance. Additionally, MTS prediction effectively captures displacement time hysteresis, while the UTS-RF model efficiently identifies valuable information in residual sequences. This research provides scientific basis and technical support for dam safety monitoring, health service diagnosis, and operational management, offering new insights for structural health monitoring.