A New CCN Number Concentration Prediction Method Based on Multiple Linear Regression and Non-Negative Matrix Factorization: 1. Development, Validation, and Comparison Using the Measurement Data Over the Korean Peninsula

Abstract

To reduce uncertainty in climate change prediction, a large amount of cloud condensation nuclei number concentration (N-CCN) data must be obtained. This study aimed to develop a new N-CCN prediction method, hereafter MLRNMF method, that applies multiple linear regression (MLR) and non-negative matrix factorization (NMF) to aerosol number size distribution data measured in Seoul and over the Yellow Sea. To verify the reliability of the MLRNMF method, a data set separated from the training data set was used, and sufficient time differences of several years were given between the two data sets to make them as independent as possible from each other. The predicted N-CCN was in acceptable agreement with the measured N-CCN. The coefficient of determination (R-2) values between measured and predicted N-CCN for the Yellow Sea and Seoul were 0.81 and 0.71, respectively. Mean fractional bias (MFB) and mean fractional error (MFE) also met the performance goals (<+/- 30% and <+50%, respectively). The MLRNMF method had similar accuracy to the backward integration method but showed strength in robustness and the capability to take into account external mixing. The N-CCN prediction methods trained using data of a specific season tended to underestimate/overestimate somewhat, but MFB and MFE for all four seasons met the performance goals except for MFB in using June-August (JJA) data, implying that the MLRNMF method trained using only the data of a specific season can predict N-CCN for all four seasons to some extent. It is expected that abundant N-CCN data can be obtained through the MLRNMF method in future studies.