Enhancing streamflow forecasting accuracy through reprogrammed large language models
Abstract
Accurate streamflow forecasting is essential for hydrological planning, disaster preparedness, sustainable agriculture, and flood management, but remains challenging due to uncertainties in climate conditions and complex data handling. Traditional models and recent machine learning approaches, like LSTMs [3] and transformers [1], have improved prediction accuracy, but often require location-specific designs. This study utilizes the Time-LLM framework [2], which repro- grams raw streamflow data into text-based prototypes and employs prompt-as-prefix to guide data transformation for LLMs shows in Figure 1. The model is fine-tuned to generate 24-hour forecasts. Time-LLM’s performance is compared with models like FEDformer, LSTM, Persistence, Lasso, SVR, and PGR on datasets including the Waterman dataset, M ̄anoa Stream, and 125 USGS sta- tions. Using metrics like RMSE, NSE, and Correlation, the study shows Time-LLM excels in stable environments, while models like FEDformer are effective in complex scenarios. The results highlight the potential of LLMs for accurate forecasting, with room for model refinement.
Primary Faculty Mentor Name
Samuel Chevalier
Status
Graduate
Student College
College of Engineering and Mathematical Sciences
Program/Major
Computer Science
Primary Research Category
Engineering and Math Science
Enhancing streamflow forecasting accuracy through reprogrammed large language models
Accurate streamflow forecasting is essential for hydrological planning, disaster preparedness, sustainable agriculture, and flood management, but remains challenging due to uncertainties in climate conditions and complex data handling. Traditional models and recent machine learning approaches, like LSTMs [3] and transformers [1], have improved prediction accuracy, but often require location-specific designs. This study utilizes the Time-LLM framework [2], which repro- grams raw streamflow data into text-based prototypes and employs prompt-as-prefix to guide data transformation for LLMs shows in Figure 1. The model is fine-tuned to generate 24-hour forecasts. Time-LLM’s performance is compared with models like FEDformer, LSTM, Persistence, Lasso, SVR, and PGR on datasets including the Waterman dataset, M ̄anoa Stream, and 125 USGS sta- tions. Using metrics like RMSE, NSE, and Correlation, the study shows Time-LLM excels in stable environments, while models like FEDformer are effective in complex scenarios. The results highlight the potential of LLMs for accurate forecasting, with room for model refinement.