Ciprian George Piuleac

Modeling Methodology Based on Stacked Neural Networks Applied to the Photocatalytic Degradation of Triclopyr

In the present paper, we propose a modeling methodology based on stacked neural networks by combining several individual networks in parallel, whose outputs are weighted to provide the output of the stack. Also, a procedure was included for finding the optimal set of weights that leads to the best performance of modeling, on both training and validation data. As a case study, we consider the photocatalytic oxidation of triclopyr where the final concentration was evaluated depending on the reaction conditions, irradiation time and amounts of reactants. We show that the performance of the stack is better than those of individual networks, especially for the validation phase.

Optimization methodology based on neural networks and genetic algorithms applied to electro-coagulation processes

AbstractAn optimization methodology based on neural networks and genetic algorithms was developed and used to optimize a real world process — an electro-coagulation process involving three pollutants at different concentrations: kaolin (250–1000 mg L−1), Eriochrome Black T solutions (50–200 mg L−1), and oil/water emulsion (1500–4500 mg L−1). Feed-forward neural networks using heterogeneous combination of transfer functions were developed, leading to good results in the validation stage (relative error about 8%). The parameters of the process (concentration of pollutant, time, pH0, conductivity and current density) were optimized handling the genetic algorithm parameters, in order to obtain a maximum removal efficiency for each pollutant. Therefore, the optimization methodology combines neural networks as modeling tools with genetic algorithms as solving method. Validation of the optimization results using supplementary experimental data reveals errors under 11%.

Instance-based regression with missing data applied to a photocatalytic oxidation process

AbstractIn this paper, a modified nearest-neighbor regression method (kNN) is proposed to model a process with incomplete information of the measurements. This technique is based on the variation of the coefficients used to weight the distances of the instances. The case study selected for testing this algorithm was the photocatalytic degradation of Reactive Red 184 (RR184), a dye belonging to the group of azo compounds, which is widely used in manufacturing paint paper, leather and fabrics. The process is conducted with TiO2 as catalyst (an inexpensive semiconductor material, completely inert chemically and biologically), in the presence of H2O2 (with the role of increasing the rate of photo-oxidation), at different pH values. The final concentration of RR184 is predicted accurately with the modified kNN regression method developed in this article. A comparison with other machine learning methods (sequential minimal optimization regression, decision table, reduced error pruning tree, M5 pruned model tree) proves the superiority and efficiency of the proposed algorithm, not only for its results, but for its simplicity and flexibility in manipulating incomplete experimental data.