Abstract: The real-time optimization control technology for reservoir development is a focus of the research of intelligent reservoirs This paper proposes a new method for real-time reservoir production optimization based on data space inversion. First, only a number of priori production dynamics from reservoir models are required to construct the proxy model based on Bayesian theory and then to perform model training by fitting of historical observation data based on the random maximum likelihood principle. Finally, the posteriori estimation of production dynamics conforming to the actual observation data can be obtained by inversion. Using a simultaneous perturbation stochastic approximation (SPSA), a mathematical model of production optimization has been established to realize the real time optimization of injection-production parameters and improve economic development benefits. On the basis of proxy model, this method avoids repeated numerical simulation calculations in the history matching stage, and obtains the posterior estimation of real production dynamics while considering the actual geological characteristics of reservoir model. The optimized scheme takes into consideration the uncertainties of reservoirs and reduces the risks of reservoir development. The actual application in oil reservoirs showed that the production forecast results of this method are consistent with those obtained by the conventional multiple model data assimilation method. The calculation efficiency of the history matching process has been improved by five times and good waterflooding effect has also been achieved using the injection production optimization scheme. The proposed production optimization method of reservoir development has provided new ideas for the real-time production optimization of oilfields.