Displacement prediction at the Ridracoli dam foundation: comparison between statistical and machine learning approaches

Among civil engineering works of geotechnical interest, dams represent a particularly significant case, as even during normal operation they are subjected to a wide range of loading conditions. These arise not only from variations in the reservoir level but also from thermal fluctuations and changes in pore pressure regimes. Continuous monitoring data from existing dams constitute a fundamental resource, as they allow assessment of the material characterisation performed during the design stage and promote a deeper understanding of the overall deformation behaviour of the structure at the scale of the entire system. This paper compares predictions obtained using advanced statistical tools and machine learning techniques applied to displacement data recorded by selected instruments installed in the rock mass underlying the Ridracoli arch-gravity dam, adopted as a case study. The models were calibrated over a multi-year observation period and subsequently applied, in a predictive setting, to more recent years. A multistep statistical approach, derived from the classical hydrostatic-seasonaltime regression model and sequentially incorporating further predictors for time-dependent effects, is compared with two machine-learning techniques, the Boosted Regression Trees and the Random Forests, both based on decision trees. The results are evaluated in terms of predictive accuracy, robustness and interpretability. The comparison highlights the strengths and limitations of each approach when applied to displacement monitoring data of a concrete dam.