Otimização em tempo real de uma planta de tratamento de efluentes utilizando técnicas de aprendizado de máquina.

Abstract

Wastewater treatment using the activated sludge process is a highly efficient technique for the removal of organic compounds and nitrogenous products, becoming widely used and enabling the reuse of treated waters. Computational modeling has been an essential tool to enhance the performance of treatment systems, allowing for the planning and analysis of effluent treatment plants. Several predictive models, such as ASM for activated sludge systems and BSM for biological treatment in activated sludge reactors, have been developed to evaluate control strategies in treatment plants. To improve these processes, the use of metamodels has been explored, offering a simplified and optimized representation of the original model and resulting in time and computational resources savings across various fields, including engineering and science. This research proposes real-time optimization of an effluent treatment plant through machine learning techniques, utilizing the BSM1 model and kriging metamodels. The study aimed to understand the performance of optimizers in simulations of wastewater treatment processes, evaluating different metrics to identify trends and efficiency. The results demonstrated the reliability of kriging in generating metamodels, with all combinations yielding satisfactory outcomes. The optimizers "matlab," "filtersd," and "ipopt" proved effective in the objective function and compliance with constraints, while "nomad" and "nlopt" exhibited lower performance. The RTO optimization approach yielded satisfactory results, enabling a better understanding of the involved processes. The combination of these techniques holds promise to enhance the operational efficiency of wastewater treatment plants, with the potential to significantly contribute to sustainable water resource management and environmental impact reduction.