Energy optimization in large-scale recirculating aquaculture systems: Implementation and performance analysis of a hybrid deep learning approach

Recirculating Aquaculture Systems (RAS) represent an increasingly important solution for sustainable fish production, yet their high energy consumption remains a significant operational challenge. This study extends our previous work on using Deep Deterministic Policy Gradient (DDPG) for optimizing feeding rates in Recirculating Aquaculture Systems (RAS) by developing a hybrid Long Short-Term Memory (LSTM)-DDPG approach for energy optimization in a large-scale commercial RAS facility. The system, comprising 108 tanks with a total water volume of 3,132 m³, was monitored over a complete annual cycle, collecting 8,760 hourly observations of environmental, biological, and operational parameters. The hybrid model achieved high predictive accuracy for energy consumption patterns, with R² values exceeding 0.91 for key components. Implementation resulted in a 15-20% reduction in daily energy consumption while maintaining optimal water quality. Economic analysis revealed a 17% decrease in energy costs per kilogram of fish production. The system's performance was validated under varying fish biomass densities (80-120 kg/m³) and seasonal temperature profiles. These findings demonstrate the effectiveness of integrating deep learning techniques for energy optimization in RAS, offering a scalable solution for enhancing the economic and environmental sustainability of intensive aquaculture operations.