METHANE EBULLITION PREDICTORS FOR RESERVOIR EMISSIONS COMBINING MODELING AND MACHINE LEARNING

Código

XXIV-SBRH0413

Autores

Lediane Marcon, Mayra Matsume Ishikawa, Michael Mannich, Tobias Bernward Bleninger, Andreas Lorke

Tema

SE04.C - Engenharia Limnológica e Mecânica dos Fluidos Ambiental

Resumo

Freshwater reservoirs have been reported as contributors to greenhouse gas emissions to the atmosphere, but their contribution is uncertain. The produced methane in such systems can be transported to the atmosphere by ebullition and diffusion. Ebullition is highly variable in space and time, as the formation of bubbles depends on physical, chemical and biological factors. Therefore, there are still gaps in the processes underlying ebullition for accurate emissions estimation. Data scarcity is one of the problems faced to comprehend the process, hence the application of lake models can assist to overcome this issue. In addition to the physical based model, machine learning methods can be applied to provide predictions while classical approaches are still not resolved. Thus, in this study simulation results of a calibrated 3D hydrodynamic-water quality model (temperature, dissolved oxygen, chlorophyll-a and water depth) were combined with observed ebullition data to bring insights about methane release from a drinking water reservoir in Paraná - Brazil, then machine learning techniques were explored to predict ebullition (MatLab toolbox). A supervised machine learning regression model confirmed that the magnitude of ebullition cannot be well predicted by considering only water quality conditions (RMSE = 0.17 and R2 = 0.34). On the other hand a classification model predicted the occurrence of ebullition events independent of its magnitude with an accuracy of 77%, indicating a potential use of the combined techniques of data analysis. This study provided preliminary results on a new approach for gas ebullition assessment, and improvements are still required.