Comparative Analysis of Statistical and Machine Learning Models for Carbon Dioxide Fugacity in Vridi Canal

Abstract

This study compared the performance of Multiple Linear Regression (MLR), Multilayer Perceptron (MLP), and a hybrid MLR/MLP model for estimating carbon dioxide fugacity (fCO₂) in the surface waters of Vridi Canal. Measurements and supporting hydroclimatic data collected between May and September 2023 included redox potential, conductivity, dissolved oxygen content, cumulative rainfall, ambient temperature, and tidal coefficient. In the MLR framework, fCO₂ was treated as the dependent variable, while in the MLP model it was treated as the output parameter. For the hybrid model, the MLP was applied to the residuals obtained from the MLR predictions, using the relevant variable selected from the regression analysis. The MLR results showed weak linear association between fCO₂ and the selected variables, with redox potential being the only variable showing pseudo-linearity during the study period. The standalone MLP model also showed limited ability to reproduce the experimental fCO₂ values, with the best-performing 1-1-1 configuration remaining below the required validation threshold. By contrast, the hybrid MLR/MLP approach, particularly the MLR/1-7-1 configuration, produced substantially stronger predictive performance and accounted for 99.84% of the variance in the testing phase. These findings indicate that the hybrid framework can represent complex, partly nonlinear ecological behaviour in Vridi Canal more effectively than the individual statistical or neural-network models.