Agent AI Empowers California Water Resource Prediction: A New Paradigm for Large Model-Driven Seasonal Runoff Forecasting

This article introduces a collaborative framework combining an agent AI assistant and an automatic code mutation system for developing seasonal runoff prediction models. The model reduces errors by 29% in forecasts for 23 watersheds in the Sierra Nevada Mountains of California, providing more accurate decision support for water resource management.

California's water resources are highly dependent on winter snowmelt in the Sierra Nevada Mountains. Accurate seasonal runoff prediction is crucial for reservoir operation, agricultural irrigation, urban water supply, and ecological protection. However, climate change has altered snow patterns, reducing the accuracy of traditional prediction models based on historical statistics, creating an urgent need for systems that adapt to new hydrological dynamics.

The study proposes a dual-component architecture: the agent AI assistant (Explorer) is responsible for data discovery, knowledge synthesis, and architecture exploration; the automatic code mutation system (Refiner) optimizes solutions through Monte Carlo Tree Search, including automated feature engineering and hyperparameter tuning, achieving a combination of broad exploration and deep optimization.

The prediction system adopts an adaptive integration strategy of three XGBoost quantile regression sub-models to provide prediction intervals for quantifying uncertainty; feature engineering incorporates physical laws such as snow energy balance, soil moisture memory, and topographic features; for 23 watersheds in the Sierra Nevada Mountains, it predicts monthly natural total runoff with a time horizon of 1-6 months.

Compared to California's current operational forecasting system (Bulletin 120), the 2021-2025 evaluation shows that the agent AI system reduces quantile prediction errors by 29%, with significant improvements especially in early-season cumulative runoff prediction and extreme event capture; among different time horizons, the mid-term (3-4 months) shows the most obvious advantages.

This study provides a new paradigm for AI applications in Earth sciences: human-machine collaboration (humans define problems + AI explores and optimizes), code as an optimization space (discovering new strategies), and the framework can be transferred to other Earth science problems such as flood forecasting and meteorological prediction.

The study's limitations include data dependency (high uncertainty in sparse areas), weak interpretability of the XGBoost model, and unvalidated generalization to extreme events. Future directions: hybrid physics and machine learning models, real-time data assimilation, and uncertainty quantification for multi-model integration.