How Effective are Large Time Series Models in Hydrology? A Study on Water Level Forecasting in Everglades (2505.01415v1)

Abstract: The Everglades play a crucial role in flood and drought regulation, water resource planning, and ecosystem management in the surrounding regions. However, traditional physics-based and statistical methods for predicting water levels often face significant challenges, including high computational costs and limited adaptability to diverse or unforeseen conditions. Recent advancements in large time series models have demonstrated the potential to address these limitations, with state-of-the-art deep learning and foundation models achieving remarkable success in time series forecasting across various domains. Despite this progress, their application to critical environmental systems, such as the Everglades, remains underexplored. In this study, we fill the gap by investigating twelve task-specific models and five time series foundation models across six categories for a real-world application focused on water level prediction in the Everglades. Our primary results show that the foundation model, Chronos, significantly outperforms all other models while the remaining foundation models exhibit relatively poor performance. Moreover, the performance of task-specific models varies with the model architectures. Lastly, we discuss the possible reasons for the varying performance of models.

Large Time Series Models in Hydrology: A Study of Water Level Forecasting in Everglades

This paper evaluates the performance of large time series models within the domain of hydrology, specifically focusing on the application to water level forecasting in the Everglades. Recognizing the critical role of accurate water level predictions in regulating floods, droughts, and ecosystem management, the paper investigates the efficacy of twelve task-specific models and five foundation models. These models are examined across multiple forecasting horizons and measuring stations, providing comprehensive insights into their strengths and limitations.

Key Findings

Foundation Models vs. Task-Specific Models

The paper finds that the foundation model Chronos significantly excels in zero-shot inference across various locations and forecasting horizons. Contrarily, other foundation models, including Moirai, Timer, and TimesFM, demonstrate relatively poor performance. This emphasizes the selective success of certain pre-trained models, such as Chronos, which might be attributed to the diverse datasets used during its training, potentially including climate datasets that better correlate with Everglades data.

Performance Across Stations

Model performance inherently varies across measurement stations, with NP205 exhibiting weaker correlations with other stations resulting in higher prediction errors. Chronos maintains superior accuracy across all station types, showcasing its robustness. Most task-specific models such as NBEATS, PatchTST, and RMoK, offer strong performance but require retraining for each specific dataset or experimental condition, highlighting a key limitation in scalability and adaptability compared to foundation models.

Extreme Value Prediction

Sediments of extreme values, both high and low, are particularly challenging to forecast accurately. Chronos again outperforms other models in terms of capturing these extremes, as measured by the Symmetric Extremal Dependence Index (SEDI). The analysis highlights the difficulty all models face in predicting abrupt changes, a common occurrence in hydrological datasets.

Implications for Hydrology and Water Management

This research conveys significant practical and theoretical implications. On the practical aspect, it highlights the potential for scalable deployment of foundation models in real-world water management tasks, such as flood control and ecological conservation efforts. Theoretical implications involve the opportunities for advancing these models further in hydrology, particularly by integrating more contextual hydrological datasets during pre-training.

Speculations for Future Research

The paper suggests that future advances could be made by improving the architectural paradigms utilized by foundation models. This involves not just scaling parameters but also optimizing the depth and complexity of correlation-capturing mechanisms. Moreover, expanding datasets for pre-training across diverse environmental and hydrological conditions could drastically improve generalizability and accuracy in zero-shot settings. Addressing extreme value predictions remains an area ripe for innovation, potentially through hybrid models unifying data-driven approaches with traditional hydrological theory.

Overall, while the task-specific models offer targeted accuracy, foundation models such as Chronos provide promising solutions for broad, adaptive applications within environmental hydrology. Further exploration and refinement in these areas hold the potential for significant advancements in predictive accuracy and operational efficiency.