Answers are grounded in this paper and cite their sources.
High-frequency earth-rockfill dam displacement forecasting faces incomplete monitoring and delayed displacement responses. We develop an observation-preserving workflow that separates displacement responses from exogenous hydro-thermal and seepage variables. In the data pipeline, incomplete monitoring histories are completed to form model inputs and a filtered structural-response target, while the original observation mask restricts supervision and evaluation to originally observed future target positions. We propose DamXer, a lag-aware multi-point forecasting model that jointly predicts 89 horizontal-displacement channels over a 96-step (8-day) horizon. A shared temporal encoder learns channel-specific response states from recent displacement histories, and Response-State-Guided Lag Attention uses their pooled state to query an environmental lag-feature bank constructed from longer hydraulic, seepage, and thermal histories. In experiments, DamXer is evaluated against six baselines over a forecasting interval containing 8,400 timestamps sampled every 2 h. Across five random seeds, DamXer achieves a standardized-space test MSE of 0.179 ± 0.006, 5.04% lower than that of PatchTST, and reduces pooled physical-unit RMSE by 3.86% in the matched comparison. Controlled analyses show that mixed-range lag coverage improves accuracy and that environmental retrieval contributes complementary scale-normalized information. A zero-lookahead forecast-origin evaluation further confirms that the model can generate forecasts using pre-origin information alone. The anonymized monitoring dataset, source code, and configurations are publicly available at https://github.com/murasame612/damXer .
No takes yet. Share an insight, caveat, or question.
Zhu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: