Prediction of daytime variations of HO2 radical concentrations in the marine boundary layer using BP network

A Back-Propagation Neural Network (BPNN) was established to predict the daytime variations of HO₂ radical concentration observed in the field campaign RISFEX 2003 (RIShiri Fall Experiment 2003) conducted in September 2003 at Rishiri Island (45.07 N, 141.12 E, and 35m asl) in the Sea of Japan. The initial weight matrices and bias vectors for the network were optimized by a bee evolutionary genetic algorithm (BEGA). It was found that the input variables sensitive to HO₂ variation were photolysis frequency of O₃ to O(¹D) (J(O¹D)), a composite parameter defined as the ratio of HC to NO _x reactivity towards OH radicals (Φ), and the total aerosol surface area (A). The predicted results are closely correlated with the experimental data with the coefficient of determination (R ²) close to 1. In addition, the means and ranges of the predicted HO₂ concentration agree with the observed data with the correlation coefficient (R), the index of agreement (IA) and the fractional bias (FB) in the range of 0.84–0.93, 0.88–0.95 and ?14%–7%, respectively. This study demonstrates that BPNN is a potential tool to predict the daytime variations of HO₂ radical concentrations in the marine boundary layer (MBL).