Time-LLM-based multisource and multihorizon deformation forecasting for dam safety monitoring

Dams are high-consequence infrastructures whose deformation behavior is governed by complex interactions among water level, temperature, rainfall, and structural conditions. Traditional statistical and deep-learning approaches often struggle with nonlinearity, nonstationarity, sparse extreme events, and cross-point heterogeneity. To address these challenges, this study proposes a Time-LLM-based forecasting framework that integrates multisource monitoring data, including displacement from six monitoring lines (L3-L8), reservoir water level, air temperature, and rainfall, into a unified modeling paradigm. By employing prompt engineering and parameter-efficient fine-tuning (LoRA/Adapter), the model incorporates physical priors such as short-term monotonic water-level response and temperature-driven seasonal cycles, enabling physically consistent and interpretable predictions. Experimental results show that Time-LLM significantly outperforms HTT, LSTM, and Informer across overall accuracy, long-horizon extrapolation, extreme-condition robustness, cross-line transferability, and uncertainty quantification. On the test set, MAE is reduced by up to 42% compared with traditional baselines, while 90% prediction intervals achieve both high coverage (PICP = 0.92) and narrow width (MPIW = 0.49). These capabilities support a practical “threshold-lead-time-confidence” early-warning strategy. The study demonstrates that Time-LLM provides a powerful and interpretable pathway for proactive dam-safety risk management and offers a promising direction for integrating engineering knowledge with large-model intelligence.