Optimal Control Of Networked Systems Under Communication Limitation And Its Application In Power Systems

This thesis focuses on the optimal control of discrete-time nonlinear systems in network environment and its application in load frequency control.Due to the limited bandwidth and the vulnerability of the network,various communication protocols(such as stochastic communication protocol and event-triggering protocol)are used to optimize the channel resources,and may suffer from some network induced phenomena(such as signal quantization and network attacks).For the optimal control problem of discrete-time nonlinear systems,several optimal controllers are designed based on adaptive dynamic programming(ADP)or goal-representation adaptive dynamic programming(Gr ADP),and sufficient conditions are obtained for the learning rate in the neural network parameter updating rules to meet the stability requirements.And the algorithms are applied to the load frequency control of power system.According to control method adopted,the research results of this thesis are mainly divided into three parts.Considering the unknown discrete-time nonlinear system with input saturation under stochastic communication protocol,a neural network-based observer is designed.By introducing auxiliary robust term,the influence of neural network approximation error is reduced.Then,considering the Markov characteristic of random communication protocol,an utility function reflecting the switching characteristic is constructed.Under the idea of optimal control and enhanced learning,a value iteration ADP algorithm for offline learning is designed and the convergence of the algorithm is proved.Furthermore,the ADP iterative algorithm and its weight updating rules are given with the help of the evaluation network-execution network structure.According to Lyapunov stability theory,sufficient conditions for system identification error,boundedness of neural network weight update error and stability of closed-loop system are obtained.On the other hand,for unknown discrete-time nonlinear systems under denial of service attack,an adaptive event-triggering rule is proposed to relieve the communication burden of the network.Then,the challenge of complex time series induced by attacks and event triggering is solved by using the rule of piecewise update of neural network weights.Under the influence of the input-to-state stability theory,sufficient conditions are obtained for the observer error to be bounded and stable depending on the attack frequency and duration.Furthermore,the ADP optimal control algorithm is designed,and the conditions for the learning rate of the neural network parameters and the stability of the closed-loop system satisfying the bounded approximation errors of the weights of the executive-evaluation network are obtained.The above ADP optimal control algorithm is applied to the load frequency control problem of power system successfully.The PID controller is designed based on the nominal system under the communication limitation,and then the ADP controller is used as the auxiliary control to obtain the optimal control method of the discrete-time nonlinear system.On the one hand,for the discrete-time nonlinear system under denial of service attack,a dynamic event triggering mechanism is constructed to schedule the transmitted information.By means of switching system theory and average dwell time method,the observer performance based on neural network is systematically analyzed,and a gain design algorithm which depends on the learning rate of neural network parameters is obtained,in which the weight of neural network is updated piecewise to solve the switching between protocol and attack sequence.In the framework of evaluation-execution network architecture,the ADP optimal control strategy and the learning rate condition in the neural network parameter updating rule are obtained.On the other hand,a neural network-based observer is designed to identify the unknown nonlinear dynamics of multi-controller discrete-time nonlinear systems under quantization conditions.In order to improve the control performance,based on the estimated state of zero-sum game and the ideal control input,a Gr HDP algorithm with enhanced terms is designed.The weight update rules are constructed by adding adjustable parameters to the algorithm to achieve the auxiliary control task.With the help of Lyapunov stability theory,the performance of closed-loop system and the learning rate condition of neural network parameter updating rule are analyzed.The above two kinds of ADP/Gr HDP controllers are applied to the load frequency control problem of power system,and good control performance is obtained.Model-free optimal control of discrete nonlinear systems with state constraints and dynamic quantization is studied in this thesis.Compared with the traditional uniform quantizer,the dynamic quantizer is used to encode the control signal and reduce the influence of the inherent quantization error.Using three groups of neural networks,a data-based parallel learning algorithm for goal heuristic dynamic programming is proposed to achieve model-free optimal control.In particular,historical data are used to establish new weight updating rules to avoid the dependence on continuous incentives.By introducing the barrier function into the cost function,the Gr HDP control strategy based on the barrier is obtained.When the state is forced to move to the constraint boundary,the repulsive force is generated,so as to meet the demand of the system state constraint.On this basis,a set of sufficient conditions are obtained to ensure that the weight updating error of neural networks is bounded.The optimal control strategy has been applied to the load frequency control problem of power system successfully.