Otimização do coeficiente de performance em sistemas de refrigeração utilizando o algoritmo simulated annealing assistido por redes neurais artificiais

Abstract

The growing demand for energy efficiency in refrigeration systems, combined with the urgent need to replace synthetic refrigerants with lower-environmental-impact natural alternatives, drives the need to develop new optimization methods. Propane (R290), a hydrocarbon with a unitary global warming potential and excellent thermodynamic properties, represents a promising alternative to conventional refrigerants. However, its operation under maximumefficiency conditions requires tools capable of capturing and interpreting, even if approximately, the complexity and nonlinearity present in real refrigeration cycles. In this context, the present work proposes and validates an original hybrid methodology based on the Simulated Annealing metaheuristic algorithm assisted by Artificial Neural Networks to maximize the Coefficient of Performance in an experimental refrigeration bench operating with R290. The experimental data, comprising more than 12,000 records obtained under multiple operating conditions of the instrumented bench, cover suction and discharge pressures, temperatures at the four thermodynamic points of the cycle, and the opening percentage of the electronic expansion valve. The Coefficient of Performance for each operating point was rigorously calculated from enthalpy balances across the compression, condensation, and isenthalpic expansion processes, using precise thermodynamic properties of R290 obtained through the CoolProp library. These data were used to train a Multilayer Perceptron (MLP) neural network architecture capable of simultaneously mapping the input variables to four outputs — Coefficient of Performance, condensation temperature, evaporation temperature, and valve opening — achieving a mean squared error below 0.003 on the test set, which demonstrates the model's high predictive capability. The Simulated Annealing algorithm was then applied over the search space delimited by the real boundaries of the experimental data, using the trained neural network as a computational surrogate for the physical system. This eliminated the need for analytical models during optimization, drastically reducing computational cost without compromising fidelity to the nonlinear behavior, since the machine learning model implicitly incorporates the irreversibilities and nonlinearities of the real system. The results demonstrated that the proposed hybrid approach is capable of efficiently, robustly, and reproducibly identifying locally optimal operating conditions, establishing itself as a relevant contribution to the field of energy optimization in refrigeration systems.