Simulation of compact heat exchangers using global regression and soft computing

This dissertation investigates enhancement in accuracy of heat rate predictions in compact fin-tube heat exchangers. The sources of error from a conventional approach based on correlating heat transfer coefficients, sometimes of 25–30%, are studied first. These include the idealized assumptions in the procedure by which correlations are found, the data compression that occurs through the correlation process, and the multiplicity of solutions for a proposed correlating function obtained using local regression.; To remove the non-uniqueness of regression results, a methodology based on global optimization techniques that include genetic algorithms, simulated annealing and interval analysis is introduced. Applications of globally-determined correlations to single-phase and condensing coils demonstrate improved accuracy with errors of only 3%. The issue of degradation due to the idealized assumptions in the procedure is addressed by correlating the error generated from these with flow rates and inlet temperatures. Heat rate estimations are improved by a factor of three. The enhancement in predictions by direct correlations of heat rates instead of heat transfer coefficients is also investigated. Using global regression, direct-heat-rate and heat-transfer-coefficient-based correlations are first determined and later applied to heat exchanger data. The direct heat-rate estimations are more accurate.; The effect of the functional form of the correlation is explored through a set of possible correlating functions. Two globally-determined correlations, a power-law and an inverse linear, are selected from this set to demonstrate that their accuracy in estimations are comparable, with errors of 2.7% and 3.4% respectively.; The compression of experimental information is addressed using artificial neural networks. The network is trained using information of operational and geometrical parameters. Comparisons against conventional correlations demonstrate enhancements in accuracy of sometimes 400%. Cluster analysis has also been applied to condensing heat exchangers with validation of its excellent characteristics for data classification. Finally, the problem of accuracy in estimations from neural networks using limited data is analyzed. A cross-validation methodology to estimate expected errors is developed. This approach produces an upper bound on the estimated error in heat rates as demonstrated from its application to evaporating coils.