Physics-guided GRU model for stable multi-horizon displacement forecasting

Tools

Parasyris, Apostolos and Stankovic, Lina and Stankovic, Vladimir (2026) Physics-guided GRU model for stable multi-horizon displacement forecasting. In: IEEE 7th International Conference in Electronic Engineering & Information Technology 2026, 2026-06-03 - 2026-06-05. (In Press)

[thumbnail of Parasyris-etal-2026-IEEE-EEITE-Physics-guided-GRU-model-for-stable-multi-horizon-displacement-forecasting] Text. Filename: Parasyris-etal-2026-IEEE-EEITE-Physics-guided-GRU-model-for-stable-multi-horizon-displacement-forecasting.pdf
Accepted Author Manuscript
Restricted to Repository staff only until 1 January 2099.
Download (961kB) | Request a copy

Abstract

Accurate forecasting of displacement time series is critical for the early detection of slope instability or reliable probabilistic slope stability assessment. However, long-horizon prediction of displacement evolution remains challenging due to noise in environmental recordings, and the accumulation of forecasting errors in multi-step predictions. In this study, a physics-guided deep learning framework is proposed for multi-horizon displacement forecasting based on Gated Recurrent Units (GRU). In particular, the model is designed to predict displacement increments over multiple forecasting horizons while reconstructing displacement recursively in predefined windows of accumulation. To improve the stability of predictions, several physics-based constraints are integrated into the training loss function. These include cumulative consistency, temporal shape regularization, and ordered horizon constraints that ensure physically meaningful displacement evolution across prediction horizons. Experimental results indicate that the model achieves significant reductions in prediction error for horizons beyond 10 days while maintaining high predictive performance for short-term increments. Finally, it is shown that the integration of physics-based constraints improves forecast stability and reduces the prediction of inconsistent responses during recursive reconstruction.

ORCID iDs

Parasyris, Apostolos, Stankovic, Lina ORCID logoORCID: https://orcid.org/0000-0002-8112-1976 and Stankovic, Vladimir ORCID logoORCID: https://orcid.org/0000-0002-1075-2420;
CORE (COnnecting REpositories)