Banca de QUALIFICAÇÃO: MARCO AURELIO ANDRADE CACHEADO

Uma banca de QUALIFICAÇÃO de MESTRADO foi cadastrada pelo programa.
STUDENT : MARCO AURELIO ANDRADE CACHEADO
DATE: 16/02/2024
TIME: 18:00
LOCAL: meet.google.com/epk-vkmk-igi
TITLE:

APPLICATION OF ARTIFICIAL NEURAL NETWORKS IN EVALUATING THE PERFORMANCE OF MICROBIAL ENERGY CELLS

KEY WORDS:

Computational intelligence;  Reactors; Microorganism.

PAGES: 60
BIG AREA: Engenharias
AREA: Engenharia Química
SUBÁREA: Processos Industriais de Engenharia Química
SUMMARY:

The responsibility of sustainable practices in society is undeniable, as evidenced by the growing human interest in economic and technological development while balancing environmental preservation for future generations. The production and consumption of energy, primarily derived from non-renewable sources until now, are frequently addressed due to the limitations of these sources and resources, emphasizing the need to explore alternative sources. Concerns about the increasing generation of effluents, both industrial and domestic, are heightened by population growth and various proposed public policies, particularly more stringent environmental regulations for wastewater treatment. An approach that aims to integrate these two concerns into a unified solution was initially proposed in the early twentieth century when the generation of electrical energy by microorganisms during the degradation of organic matter was documented. Since then, related studies have flourished in the field of Microbial Fuel Cells (MFCs). These cells are reactors containing microorganisms capable of converting chemical energy into electrical energy through their catalytic reactions. In addition to treating effluents, MFCs are considered self-sustainable technologies as they generate a potential difference that powers the system itself. This work proposes the use of artificial intelligence, specifically artificial neural networks (ANNs), to optimize the parameters of MFC utilization. ANNs are computational models trained based on provided information and continually updated to optimize pattern recognition and provide quick responses to a specific system. For the application and evaluation of ANNs in MFCs, a feedforward ANN model with the Levenberg-Marquardt training algorithm was used, programmed and implemented in MATLAB® 15 (MathWorks). Varying the number of hidden layers between 1 and 2, the activation function between logsig and tansig, and the number of neurons in each layer between 10 and 200, the input variables for ANNs were the anode area (cm2), external electrical resistance (Ω), and volume (mL), while the output variable was Power Density (mWm-2). Adjustments for validation, training, and overall fit of each network were evaluated, along with the associated error. For the configuration with 2 hidden layers, it was observed that the network with 120 neurons in the hidden layers, tansig activation function, 19 interactions, and R2 of 0.89562 in training adjustment, 0.74617 in validation adjustment, and 0.90633 in overall fit with all experimental data assessed, obtained an approximate error of 0.0845, proving to be the most effective. For the configuration with 2 hidden layers, with 25 neurons and logsig activation function, satisfactory results were also observed, with 113 interactions and R2 of 0.92 in training adjustment, 0.97872 in validation adjustment, and 0.86064 in overall fit with all experimental data assessed, resulting in an approximate error of 0.000593. In creating neural networks for configurations with only one hidden layer, different behaviors were observed. The most promising results were achieved for the configuration with a single hidden layer, 90 neurons, tansig activation function, 22 interactions, and R2 of 0.92628 in training adjustment, 0.82224 in validation adjustment, and 0.91009 in overall fit with all experimental data assessed, resulting in an approximate error of 0.0168. For the configuration with a single layer, 90 neurons, logsig activation function, 82 interactions, and R2 of 0.99361 in training adjustment, 0.89527 in validation adjustment, and 0.68667 in overall fit with all experimental data assessed, the approximate error was 0.0207. In conclusion, it is possible to use artificial neural networks to optimize microbial energy cells, obtaining satisfactory results, especially for Multilayer Perceptrons (MLPs).

BANKING MEMBERS:
Presidente - 1811284 - EDSON ROMANO NUCCI
Interno - 1742695 - JUAN CANELLAS BOSCH NETO
Externa ao Programa - 2029466 - ISABEL CRISTINA BRAGA RODRIGUES