Addressing the complex equipment and network challenges in new energy power plant industrial control systems, this study introduces a Markov time-varying machine learning algorithm, leveraging classification-constrained Boltzmann machines for real-time network security risk prediction. By employing a hybrid training mode for innovative feature extraction and classification, the algorithm forecasts future security risk states through an up-to-date state transition probability matrix. Integrating with the Markov time-varying model enhances the efficiency over traditional Boltzmann machines, facilitating nuanced network state analyses. The proposed model demonstrates high effectiveness against various network attacks, with average precision, recall, and F1 scores of 0.96, 0.93, and 0.94, respectively, and maintains over 80% accuracy under noise levels up to 40 dB. This research provides a solid foundation for proactive security defense mechanisms in industrial control systems.