This paper continues work on creating digital twins of nuclear power plant (NPP) units based on flexible modeling systems (FMS). The paper examines integrating machine learning methods into the FMS architecture to improve the accuracy of process parameter prediction and optimize control of NPP units. An analysis of existing approaches to the use of artificial intelligence algorithms in modeling complex energy systems was conducted, identifying their advantages and limitations as they apply to nuclear power. A hybrid architecture is proposed that combines physical and mathematical models of NPP with neural network correction algorithms, significantly improving the system’s adaptability to changing operating conditions. A methodology for training neural network modules using historical operating data and deterministic model results has been developed. The results of numerical experiments demonstrate a 15–25% reduction in forecasting errors for key power unit parameters compared to classical modeling methods. The prospects for the practical application of the developed solutions for information support for NPP operators, predictive maintenance, and equipment lifespan assessment are discussed.