面向知识学习与资源高效利用的复杂非线性系统自适应评判控制

With the rapid development of information technology, such as artificial intelligence, big data, and deep learning, adaptive critic and reinforcement learning are regarded as promising schemes involving evaluation component to accomplish intelligent optimization. Nowadays, increasing communication burdens are often brought in by network-based techniques. In addition, more and more dynamical systems are encountered with complex uncertain factors and the difficulty of building accurately mathematical model. In this project, the adaptive critic control theory and methods of complex nonlinear systems are investigated toward knowledge learning and efficient resource utilization. First, the adaptive optimal control approaches of complex nonlinear systems are developed based on an improved learning mechanism, followed by sufficient discussions on their robustness performance. Then, the event-triggered self-learning robust control and adaptive H∞ control of nonlinear systems are studied respectively, thereby extending the framework of traditional adaptive critic designs. Moreover, the novel adaptive critic control policies of nonlinear systems are established by using a mixed data and event driven scheme. Finally, the neural network implementation and closed-loop stability of the present methods are performed in detail. In addition, by carrying out experimental simulation and application study on practical plants, such as power systems and overhead crane systems, the effectiveness and superiority of the present approaches are verified. Through improving the adaptive critic learning mechanism and efficiently utilizing data and communication resources, it is important to note that the given strategies can not only strengthen the results of control theory with data-based and event-triggered formulations, but also provide new avenues to robust control and H∞ control designs of complex nonlinear systems. Therefore, it will be beneficial for the construction of advanced automation techniques and intelligent systems as well as of great significance both in theory and application.

随着人工智能、大数据、深度学习等信息技术的快速发展，自适应评判与强化学习成为融合评价机制、实现智能优化的重要方法。鉴于网络化技术带来通信负担加剧，越来越多的动态系统存在复杂不确定因素、难以精确建模等，本项目面向知识学习与资源高效利用，研究复杂非线性系统的自适应评判控制理论与方法。首先，提出基于改进学习机制的自适应优化控制方法并讨论鲁棒性能。其次，研究非线性系统事件驱动自学习鲁棒控制和自适应H∞控制，推广传统的自适应评判体系。再次，建立数据、事件混合驱动的新型自适应评判控制策略。最后，探讨神经网络实现，分析控制系统稳定性，并针对电力、吊车等实际系统开展仿真实验与应用研究。通过改善自适应评判学习机制，高效利用数据、通信资源，本项目能够深化基于数据和事件驱动的控制理论成果，为复杂非线性系统的鲁棒控制和H∞控制设计开辟新途径，也将有助于构建先进自动化技术与智能系统，具有重要的理论意义和应用价值。

复杂动态系统的优化控制设计是控制领域的研究热点之一。由于实际非线性系统往往存在着不确定性、关联、建模困难等复杂因素，而且网络化技术的普及加剧了控制系统的通信负担，在智能科学与技术的有力推动下，获取人工智能驱动的先进优化控制方案正在成为相关领域的研究难点。本项目面向知识学习与资源高效利用，研究复杂非线性系统的自适应评判控制理论与方法，旨在逐步完善自适应评判控制的研究框架。首先，提出采用改进学习机制的自适应优化控制方法并研究其鲁棒性能，针对一般连续时间非线性系统的自适应优化控制、不确定系统的自学习鲁棒镇定与跟踪控制、关联系统的自学习分散控制设计、含有输入约束非线性系统的自学习干扰抑制等问题，分别获得了新颖的研究成果。其次，探讨非线性系统的事件驱动自学习鲁棒控制和自适应H∞控制，得到实现鲁棒镇定和保成本控制的先进设计策略，并在考虑外部扰动的情况下获得事件驱动干扰抑制方案，由此显著推广了自适应评判方法。再次，充分利用知识学习与提高资源利用效率的要求，初步建立了针对连续时间系统的数据、事件混合驱动自适应评判控制架构，在数据驱动控制和事件驱动控制方面分别深化了自适应评判控制理论。最后，开展神经网络实现与控制系统稳定性分析，并针对电力、吊车等实际系统进行仿真应用研究，验证了相关技术的有效性。此外，项目组成员也在强化学习与智能控制、污水处理过程智能优化控制等方面进行了延伸，为在未来开展更加深入的研究工作打下了基础。通过开展本项工作，深化了不确定环境下非线性系统自适应优化控制的研究成果，显著扩大了自适应评判方法的使用范围，由此推动了复杂非线性系统智能优化控制的研究进展。本项目提出的一系列方法能够应用于不确定环境下几类电力系统的优化控制设计，也为开展其他复杂系统尤其是污水处理过程的智能评判控制设计提供了一定的技术支撑，从而有助于构建先进自动化与智能系统，具有潜在可观的经济效益和社会效益。