Time series foundation model chronos enhances nitrogen forecasting under data scarcity

• Fine-tuning using observed data improves the foundation model’s water-quality prediction. • The foundation model, Chronos outperforms benchmark models using <60 % of available training data. • Chronos demonstrates superior performance at high-complexity and peak event water-quality stations. • Cross-station transferability is feasible for Chronos but dictated by data complexity. • Foundation models offer a data-efficient paradigm for environmental monitoring. Accurate water quality prediction is critical for water resource management, yet task-specific models are often limited by their reliance on extensive training data. The emergence of large time series foundation models presents a promising yet unexplored solution for water quality forecasting under data scarcity. This study evaluates Chronos, an advanced time series foundation model using the T5 large language model (LLM), against a suite of task-specific benchmark models (i.e., RF, MLP, LSTM, Transformer, XGBoost, LightGBM, KAN, PatchTST, and Crossformer). We evaluated their performances in predicting total nitrogen (TN) at 12 stations in the Huang-Huai-Hai River basin of China under varying data availability ratios (20%-100%) and different forecast ahead times (4-48 hours). Adapting the Chronos model with station-specific data (i.e., fine-tuning) yields superior performance compared to using the model’s general, pre-trained capabilities directly (i.e., zero-shot). Fine-tuned Chronos consistently outperformed benchmark models when trained on less than 60% of available data across all ahead times, though LSTM and PatchTST achieved comparable or superior performance with sufficient training data (≥80%), particularly for medium-to-low complexity sequences. The model showed its greatest performance advantage at high-complexity stations and during peak pollution events, maintaining superiority regardless of data availability. Through transfer learning, fine-tuned Chronos achieved cross-station generalization superior to other benchmark models, and we found this transferability is dictated by water-quality-data complexity. This study first evaluated the large time series foundation models as a powerful and data-efficient tool for environmental science, poised to remarkably improve pollution prediction and water resource management, especially in data-scarce regions.