The leakage of stored and transported CO

2

is a risk for geological sequestration technology. One of the most challenging problems is to recognize and determine CO

2

leakage signal in the complex atmosphere background. In this work

a time series model was proposed to forecast the atmospheric CO

2

variation and the approximation error of the model was utilized to recognize the leakage. First

the fitting neural network trained with recently past CO

2

data was applied to predict the daily atmospheric CO

2

. Further

the recurrent nonlinear autoregressive with exogenous input (NARX) model was adopted to get more accurate prediction. Compared with fitting neural network

the approximation errors of NARX have a clearer baseline

and the abnormal leakage signal can be seized more easily even in small release cases. Hence

the fitting approximation of time series prediction model is a potential excellent method to capture atmospheric abnormal signal for CO

2

storage and transportation technologies.