Hybrid Multivariate Statistical and Neural Network Model to Predict Greenhouse Gas Emissions

Abstract

This paper discusses a set of hybrid predictive and descriptive model to estimate the emission of $$\hbox {CO}_{2}$$ in tropical reservoirs based on field measurements of physical and chemical water quality parameters. With this model, it is possible to evaluate the influence of the water quality on the greenhouse gases (GHG) emission at the reservoir and predict the GHG emissions. The quality of the proposed method is evaluated at the reservoir Hidrosogamoso in Colombia. Results are compared with traditional statistical methods. This model combines statistical and artificial neural network (ANN) models. The statistical component improves the characterization of the observations and avoids the multiple collinearity, while the ANN improves the optimization and provides the model with prediction ability. Factor analysis and multiple linear regression have been used to reduce the number of input variables required to train the ANN. The reduction in the number of inputs for the ANN in this case is fundamental since the amount of available samples is fundamental to avoid problems with a high number of degrees of freedom that can cause a memorization problem. Also to reduce the impact of a reduced number of samples, the output of the ANN consists in an ensemble of several networks trained with different datasets obtained from a bootstrap procedure. All the results let us conclude the best model is the combination of statistical and neural network models, hybrid model, with ME = − 0.40, MAE = 0.48, MSE = 0.34 and correlation coefficient R = 0.89. The results showed that combining multivariate statistical analysis and neural network improves the efficiency of both models, having the ability of detailed descriptions, numerical estimations of the effects from the inputs over the outputs from statistical methods and the final prediction power from the neural network.