Real-time automatic regulation of gene expression in living cells is a key technology for synthetic biology. Unlike traditional control engineering applications, cells grow and divide over time, and the natural variability arising in individual cell gene expression makes the population time-varying and heterogeneous. Therefore, the application of tried and tested control approaches can be often problematic. Here, using a microfluidics-based experimental platform, in which either glucose or galactose is provided to the cells, we measured expression from the galactose-inducible promoter in individual cells for thousands of minutes. We identified single cell linear dynamical models across hundreds of cells via a recently proposed linear mixed-effects identification strategy. We show that these models are able to capture the expression dynamics of single cells but also the mean and standard deviation of the population, thus making realistic simulation of gene expression possible. We then compared the performance of a Model-Predictive-Control strategy to solve regulation and tracking problem when based either on a deterministic model of the mean expression of the cell population, or on the individual models of the single cells.