Intelligent cloud-based RAS management: integration of DDPG reinforcement learning with AWS IoT for optimized aquaculture production

TL;DR

This paper introduces a cloud-edge system using AI to optimize aquaculture operations at commercial scales, solving issues like scalability and reliability.

Contribution

A novel cloud-edge hybrid architecture for deploying DDPG-based control systems in commercial aquaculture, addressing scalability and infrastructure challenges.

Findings

Edge optimization reduced DDPG model size by 74% while maintaining 98.5% performance retention during network disruptions.

Field validation showed 99.97% IoT message delivery rates and 98.7% reliability in parameter control across 108 tanks.

The system maintained robust performance with only 8.9% latency increase from small to large-scale operations.

Abstract

While Deep Deterministic Policy Gradient (DDPG) reinforcement learning has demonstrated significant potential for optimizing aquaculture operations in laboratory and controlled environments, its practical deployment in commercial-scale Recirculating Aquaculture Systems (RAS) faces critical scalability and infrastructure challenges. This paper presents a novel cloud-edge hybrid architecture that enables the deployment of DDPG-based control systems across diverse commercial aquaculture operations, from small research facilities to large-scale production systems. Building upon our previous work in DDPG-based feeding rate optimization and energy management, we develop a comprehensive framework that addresses the practical challenges of deploying AI-based control systems in real-world aquaculture environments. The proposed architecture integrates AWS IoT Core for sensor connectivity, AWS Greengrass for edge intelligence, and a suite of cloud services for scalable model deployment and management. Edge optimization techniques, including 16-bit quantization and architecture pruning, reduced the DDPG model size by 74% (32 MB to 8.3 MB) while maintaining accuracy within 1.5% of the full-precision version, enabling real-time inference with 47 ± 8 ms latency across all deployment scales. Field validation in a commercial facility with 108 tanks (3,132 m³ total volume) demonstrated exceptional scalability, with only 8.9% latency increase from small-scale (1,000 L) to large-scale (50,000 L) operations. The system achieved 99.97% IoT message delivery rates and maintained 98.7% reliability in critical parameter control, while comprehensive failsafe mechanisms ensured safe operation during network disruptions lasting up to 72 h. Network resilience testing validated robust performance under various connectivity challenges, maintaining 98.5% performance retention during minor network latency and 85.2% retention during 12-hour complete disconnections. This research establishes a practical blueprint for transitioning DDPG-based aquaculture management from research environments to commercial deployment, addressing critical gaps in scalability, reliability, and operational resilience that have previously limited the adoption of AI-based control systems in the aquaculture industry.

Figures