NeuralRiverOps: An Operational Framework for Implementing MLOps and Agentic AI in LSTM-based Flood Forecasting for Large-scale River Basins

Since the early 2020s, artificial intelligence (AI) has gained substantial attention across both industry and academia. In the field of water resources, AI-based approaches for improving the prediction of floods, water supply, droughts, and related hydrological phenomena have been actively explored. More recently, the emergence of agentic AI, in which large language models (LLMs) orchestrate multiple AI tools for analysis, prediction, and operational services, has attracted increasing attention.

Despite these advances, most research efforts remain focused on model development, while the establishment of sustainable operational systems, such as those enabled by machine learning operations (MLOps), remains limited. This gap is particularly evident in water resources applications, where continuous retraining, performance evaluation, and system-level reproducibility are critical for real-world deployment.

In this study, we propose NeuralRiverOps, an operational framework that integrates MLOps and agentic AI for multi-point flood prediction in large-scale river basins using long short-term memory (LSTM) networks. First, we design a workflow that supports sequential model development and prediction from upstream to downstream and from tributaries to main streams, leveraging the neuralhydrology Python library as the core modeling engine. Second, to enable systematic model retraining, storage, inference, and performance evaluation, we construct an MLOps pipeline based on MLflow. PostgreSQL is employed for structured time-series data management (e.g., rainfall, dam releases, and river water levels), while MinIO is used for scalable object storage, such as trained LSTM models. Furthermore, we develop an agentic AI system that allows users to interactively invoke the MLOps pipeline through a chat-based interface. This system is implemented using Ollama as an open-source LLM platform and OpenWebUI as the conversational interface. All components - including AI models, MLflow, PostgreSQL, MinIO, Ollama, and OpenWebUI - are containerized and orchestrated using Docker Compose to enhance computational reproducibility, scalability, and maintainability.

The proposed framework demonstrates a practical architecture for integrating agentic AI into analytical systems and highlights the essential role of MLOps in the sustainable operation of AI models for disaster preparedness, such as flood and drought forecasting. This study provides a pathway for future research to move beyond isolated model development toward robust, operational AI systems supported by MLOps and agentic AI.