Li et al. (2026) A prompt-conditioned multimodal framework for streamflow prediction with lightweight task adaptation

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-06-24
• Authors: Xudong Li, Senchang Hu, Wenyi Zhang, Wuyou Xiao, Wenzhuo Shi, Jiaqi Ruan, Zhao Xu, Hao Wang
• DOI: 10.1016/j.jhydrol.2026.135915

Research Groups

• Department of Electrical and Electronic Engineering, The Hong Kong Polytechnic University (Hong Kong, China)
• State Key Laboratory of Hydroscience and Engineering, Department of Hydraulic Engineering, Tsinghua University (Beijing, China)
• College of Electrical Engineering, Sichuan University (Chengdu, China)
• Research Institute of Smart Energy, The Hong Kong Polytechnic University (Hong Kong, China)
• Shenzhen Research Institute, The Hong Kong Polytechnic University (Hong Kong, China)
• State Key Laboratory of Watershed Water Cycle Simulation and Regulation, China Institute of Water Resources and Hydropower Research (Beijing, China)

Short Summary

This study proposes a prompt-conditioned multimodal framework for streamflow prediction with lightweight task adaptation, achieving better performance than data-driven baselines in multi-step forecasting tasks.

Objective

• Investigate the feasibility of using a prompt-conditioned multimodal framework for streamflow prediction and its ability to adapt to different hydrological processes.

Study Configuration

• Spatial Scale: Sub-basins of the Yangtze River Basin (China)
• Temporal Scale: Daily to weekly time steps

Methodology and Data

• Models used: Generative Pre-trained Transformer 2 (GPT-2) with trainable multimodal embedding layers and a prediction decoder
• Data sources: Historical streamflow, meteorological forcing, statistical and static watershed attributes from the Yangtze River Basin

Main Results

• The proposed framework achieves an average NSE of 0.69 in multi-step forecasting tasks (1-7 days ahead) in two sub-basins of the Yangtze River Basin.
• Zero-shot within-region cross-basin validation shows that the proposed framework can retain reasonable predictive skill without target-basin retraining.

Contributions

• This study provides a feasible prompt-conditioned multimodal framework for streamflow prediction, offering new insights into the development of hydrological models.

Funding

• This research was funded by the State Key Laboratory of Hydroscience and Engineering (Grant No. [not specified]) and the Research Institute of Smart Energy, The Hong Kong Polytechnic University (Grant No. [not specified]).

Citation

@article{Li2026promptconditioned,
  author = {Li, Xudong and Hu, Senchang and Zhang, Wenyi and Xiao, Wuyou and Shi, Wenzhuo and Ruan, Jiaqi and Xu, Zhao and Wang, Hao},
  title = {A prompt-conditioned multimodal framework for streamflow prediction with lightweight task adaptation},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.135915},
  url = {https://doi.org/10.1016/j.jhydrol.2026.135915}
}