Comparison model learning methods for methane emission prediction of reservoirs on a regional field scale: Performance and adaptation of methods with different experimental datasets

In this study, we explored different methods to build methane emissions prediction models of temperate reservoirs on a regional field scale, and then we examined the performances and adaptation of these prediction models. First, four statistical model learning methods and two machine learning methods were used to develop methane emissions prediction models based on environmental factors (i.e., temperature and atmospheric pressure) and methane fluxes at three reservoirs (Miyun, Yudushan, and Baihepu) in Beijing, China, from 2009 to 2012. In general, decision trees (DT) exhibited better performance with higher r2 (the coefficient of determination) and lower root mean squared error, mean deviation, mean squared error, and mean absolute error. Second, in order to examine model adaptation, two experimental datasets were used to build methane emissions prediction models separately: D1 (samples only from Miyun) and D2 (samples from Miyun, Yudushan, and Baihepu). Then, three test data groups which used samples from the three reservoirs separately were used. In general, decision trees (DT) exhibited better performance and adaptation compared to those of other model learning methods. Moreover, our study indicated that it could be necessary to compare greenhouse gas prediction models when compiling greenhouse gas inventories according to IPCC. •Methane emission prediction models were developed using different learning methods.•Four statistical model learning and two machine learning methods were compared.•Decision tree models performed better than others either from the model performance or from the model adaptation.