Leveraging Exogenous Signals for Hydrology Time Series Forecasting

2025-11-14

AI Summary

This study addresses the limited generalizability and poor physical interpretability of time-series models in hydrological rainfall–runoff modeling. To this end, we propose a domain-knowledge-integrated paradigm that enhances models with physically informed exogenous variables. Specifically, we incorporate natural annual periodic signals—such as astronomical time encodings—as key exogenous inputs, jointly modeling them with dynamic meteorological sequences and static catchment attributes. We systematically evaluate the predictive performance of time-series foundation models (TSFMs) on the CAMELS-US dataset across 671 catchments. Results demonstrate that integrating physics-driven exogenous variables significantly improves the Nash–Sutcliffe Efficiency (NSE) coefficient, yielding an average gain of 12.3%—outperforming both conventional LSTM-based and conceptual hydrological models. Moreover, our analysis reveals that exogenous variable design critically governs TSFM downstream adaptability, establishing a reproducible and scalable framework for embedding domain knowledge into data-driven hydrological modeling.

Technology Category

• Machine Learning: Time-Series/Data Streams
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning
• Intelligent Robots: Multimodal Perception & Sensor Fusion

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• User Modeling, Personalization and Recommendation: Psychology-informed user models and recommender systems

Abstract

Recent advances in time series research facilitate the development of foundation models. While many state-of-the-art time series foundation models have been introduced, few studies examine their effectiveness in specific downstream applications in physical science. This work investigates the role of integrating domain knowledge into time series models for hydrological rainfall-runoff modeling. Using the CAMELS-US dataset, which includes rainfall and runoff data from 671 locations with six time series streams and 30 static features, we compare baseline and foundation models. Results demonstrate that models incorporating comprehensive known exogenous inputs outperform more limited approaches, including foundation models. Notably, incorporating natural annual periodic time series contribute the most significant improvements.

Problem

• Investigating domain knowledge integration in hydrological time series forecasting
• Comparing foundation models with baseline approaches using CAMELS-US dataset
• Evaluating exogenous inputs impact on rainfall-runoff modeling performance

Innovation

• Incorporating domain knowledge into time series models
• Using comprehensive known exogenous inputs for forecasting
• Integrating natural annual periodic time series signals