Research On Modeling And Control Algorithm Of Denitrification System Of Thermal Power Units Based On Deep Reinforcement Learning

With the increase of new energy penetration,the role of traditional coal-fired generating units in the power system has also changed.In order to absorb the volatility brought by new energy and ensure the safe and stable operation of the power system,coal-fired generating units are more frequently in variable load conditions and low load conditions,weakening their profitability.At the same time,with the implementation of stricter environmental protection regulations,stricter requirements for NOx emissions from coal-fired generating units have been put forward,increasing operating costs.These factors have made the operating environment of traditional thermal power generation enterprises more difficult,and there is an urgent need to introduce new technologies to achieve cost reduction and efficiency increase,improve the competitiveness of enterprises,save the consumption of reductant and reduce the operating cost of denitrification systems while ensuring that NOx emissions meet the standards.The rapid penetration and integration of artificial intelligence technology with various industries has guided the new development direction and become the core driving force for a new round of technological innovation and industrial upgrading.Some research has been conducted on the application of artificial intelligence technologies such as supervised learning and soft measurement for ammonia injection control in denitrification systems.However,due to the nature of supervised learning relying on labeled data,these methods are mostly used as measurement signals and rarely as control quantities,and there is still room for more in-depth application of AI techniques to ammonia injection control in denitrification systems.In order to achieve the optimal control goal of saving the consumption of reductant while ensuring the standard emission of NOx,this paper uses deep reinforcement learning algorithms to design the ammonia injection controller for denitrification systems.Compared with other artificial intelligence techniques,the reinforcement learning algorithm’s unique ability of autonomous exploration allows it to surpass the shortcomings of other artificial intelligence techniques and enable the ammonia injection controller to independently explore better control strategies.This paper presents a systematic study of deep reinforcement learning algorithms for ammonia injection control in denitrification systems,which includes the following three aspects:(1)extracting high-quality data of denitrification systems for reinforcement learning virtual environment construction and intelligentsia training;(2)constructing a virtual environment with high simulation accuracy and speed to support reinforcement learning ammonia injection controller training;(3)designing a reinforcement learning intelligentsia ammonia injection controller and training it to autonomously explore the nitrogen and nitrogen balance.The main research contents and innovations of this paper are as follows:1.This paper proposes an all-in-one data anomaly detection and independent homogeneous distribution verification method that incorporates two different principles of self-encoder and Gaussian mixture model.Compared with the separated method,this method optimizes the degree of adaptation of feature extraction and Gaussian mixture model,and the two different principles of self-encoder and Gaussian mixture model verify each other to further improve the reliability of historical data set acquisition.As an unsupervised learning method,the method has been experimentally validated to have wide applicability.High-quality data is the basis for training deep reinforcement learning intelligences.In order to obtain high-quality datasets,the novel method is applied to coal-fired generating units historical data,and high-quality generating units historical datasets are obtained.2.In this paper,a neural network modeling method based on recurrent neural network suitable for modeling multiple-input single-output nonlinear systems through attention mechanism to improve dynamic simulation accuracy is proposed.After experimental verification,the method can better simulate the dynamic characteristics of denitrification system.The neural network modeling method is based on a long and short term memory neural network with a "sequence to sequence" network structure,which can be well applied to model nonlinear systems with large delay and large inertia.By introducing the attention mechanism,this neural network modeling method improves the simulation accuracy of dynamic processes.Based on the historical data of the denitrification system,this neural network modeling method is used to establish the dynamic model of the denitrification system.3.This paper proposes a virtual environment modeling method forMulti-input and multi-output nonlinear systems using parallel network structure to optimize the operation speed,which improves the simulation speed based on recurrent neural network and makes it more suitable for providing training virtual environment as a reinforcement learning agent.The new virtual environment modeling method uses an extended causal convolutional neural network with a parallel computational structure and a multi-headed attention mechanism,which has a higher computational speed compared with the dynamic model of denitrification system built using recurrent neural network proposed in the previous point 2.Moreover,this new denitrification system virtual environment uses a multi-decoder structure and alternate training of each decoder to overcome the influence of mutual coupling among simulation variables,which makes it possible to simulate multiple output variables simultaneously with high accuracy.Using this new method,a virtual environment model of denitrification system with high simulation accuracy and simulation speed is established,and its role is to provide a training environment for the reinforcement learning agent denitrification controller in the subsequent research.4.In this paper,a novel deep reinforcement learning algorithm for process control object control tasks is proposed,according to which a deep reinforcement learning agent controller is designed and trained.The algorithm innovatively employs a Bayesian neural network supervisor to plan the actions of the intelligentsia exploring the environment,overcoming the shortcomings of traditional model-free reinforcement learning algorithms that cannot utilize historical data.Based on the high-quality data set and the virtual environment of denitrification system,this new deep reinforcement learning algorithm is used to design the reinforcement learning agent denitrification controller.After simulation verification,the new reinforcement learning agent denitrification controller is able to achieve the reduction of reductant usage and reduce the unit operation cost while ensuring the NOx emission standard.