| Abstract | Receptor models are commonly classified into two main branches, Chemical Mass Balances (CMB) and multivariate approaches. They have different advantages, which normally compensate each other. Moreover, all receptor models are generally applied for only measurement data collected at a single receptor site. When the multivariate models are applied for multiple sites (in the same region), there is a strong pulling effect introduced by certain species in the process of calculation, which may result in poor source profiles. Separating calculation of each receptor site. however, produces different composition profile of a similar source in a region. Receptor modeling results normally contain large uncertainty but in fact there is no established procedure for evaluation of the receptor model results. This research is designed to test the following hypotheses: (1) A new receptor model with combined advantages of both CMB and PMF can be developed, (2) A new receptor model for multiple sites in the same area can be developed, and (3) Evaluation procedure for receptor model results may he developed and verified. Two new algorithms of Evolutionary Receptor Model (EVORM) and Dual-site Receptor Model (DUALM) have been developed. EVORM model is based on the combination of advantages of both CMB and multivariate approaches. It uses a few source composition profiles, which are known, i.e. similar to the CMB approach, as the input. The remaining source profiles are assumed to be unknown and will be identified and produced by the model, i.e. similar to the multivariate approach. A system of linear equations is solved where both the uncertainty of source compositions and the uncertainty of ambient composition data are also accounted to control the fitting of source composition factors. The model can provide global optimal solutions with the application procedure of Evolutionary Programming. Non-negative constrain is also integrated to provide reasonable solutions. DUALM model has been developed for simultaneous source apportionment of ambient air particulate matter (PM) at 2 different monitoring sites, where some similar sources exist. The model can then identify other local/specific sources at each site. The model has been developed based on the expanded Multilinear Engine (ME-2) model. General meteorological information of each receptor site (wind speed, wind direction) and other information (seasons, and date) are assigned into the model as auxiliary information to prevent the rotational ambiguity and to add necessary constraints into the calculation. The 2 newly-developed receptor models has been tested for source apportionment of PM2.5 based on the data collected by the Interagency Monitoring of Protected Visual Environments (IMPROVE) program in 2004 at 2 different monitoring sites, Great Smoky Mountain National Park (GRSM) station, Sevier County, Tennessee, and-Cohutta (COI IU) station, Whitfield County, Georgia. Both of them are located in the area where contribution of secondary particle is significant. Hourly meteorological data from the nearby US's national meteorological stations were used as auxiliary information for DUALM model. To apply EVORM for the 2 datasets, 3 sources were assumed to be known including secondary sulfate I, diesel exhaust, and industry. These 3 sources were also assumed to be the similar sources at both sites. A framework of evaluation procedure of receptor model results was established and used to evaluate the performance of the 2 newly developed models at the 2 sites. The performance of the EVORM and DUALM shows satisfactory source apportionment results. Both EVORM and DUALM models produced almost similar and comparable sources at both stations. DUALM model can indicate highest number of pptential sources in both stations and provides more interpretable profile output due to less colinearity effects as compared to EVORM. CPF plots revealed that both models captured the effects of local sources in the source apportionment results. The source apportionment results of the 2 newly developed models and their statistical performance parameters (i.e. R2, Qobj, Fractional Bias, Normalized Mean Square Error, maximum of' residuals. minimum of' residuals, standard deviation of residuals, and % o mass explain.) were also compared with the results of' the other currently used receptor models i.e. PMF2, PMF EPA version (ME engine). ME, Expanded ME, COPREM, and PCA for model performance evaluation. Advantages and disadvantages of the 2 newly developed models were also analyzed. EVORM has the limitations due to multilinear equation used in the model. Strong pulling in one side of matrix (G- or F-Factor) may introduce transformations of all components in the matrix from another side. Although EVORM model replaces the negative values with a number closed to zero, high chi square is produced. Due to the colinearity problem inherited in the niultilinear equation. the model can not separate some sources distinctly. Hence less number of sources was recognized by the model. Time consumption and requirement of high CPU performance are also limitations of EVORM when apply for the large input datasets. DUALM model can handle limitations of EVORM using the Conjugate Gradient algorithm. The model also includes meteorological information in the algorithm to constrain the results and reduce colinearity problem. However, different initial pseudo-random number in the DUALM model always produces different results. To avoid the problems. trial and errors need to be performed with different initial pseudorandom number until the best explainable and interpretable outputs are produced. Further development of EVORM model should be performed by incorporating the built-in equations in the MI model to produce better solutions. Improvement in the programming should be done to reduce computation time. Multi-site receptor model in Windows-based should be further developed which allow flexibility for multiple site to be included. More data applications for different sites should be performed to test the models |