Deep-learning modeling and control optimization framework for intelligent thermal power plants: A practice on superheated steam temperature

The operational flexibility requirement has brought great challenges to control systems of thermal power plants. Through the big data and deep-learning technology, intelligent thermal power plant can greatly improve the quality of deep peak-load regulation. Based on the framework of an intelligent thermal power plant, this paper proposes a control optimization framework by constructing a hybrid deep-learning simulation model adaptable for multiple disturbances and wide operational range. First, Gaussian naive Bayes is utilized to classify data for identification, in conjunction with prediction error method for fine data extraction. Second, deep long-short term memory is explored to fully learn extracted data attributes and identify the dynamic model. Third, based on the simulation model, two aspects are considered for control optimization: i) For a variety of immeasurable disturbances in thermal processes, the extended state observer is employed for disturbance rejection, and ii) as a widely used heuristic algorithm, particle swarm optimization is applied to optimize the parameters of controllers. Superheated steam temperature (SST) control system is the key system to maintain the safety and efficiency of a power plant; thus the proposed deep learning modeling and control optimization method is applied on the SST system of a 330 MW power plant in Nanjing, China. Simulation results compared with actual data and the index analysis demonstrated the effectiveness and superiority of the proposed method.