基于神经动力学优化和有监督学习的非线性鲁棒预测控制方法及其应用

Model predictive control (MPC) is an optimization-based advanced control strategy which has been widely recognized in academia and industries. MPC generates control actions by means of real-time optimization of a performance index over a ?nite moving horizon of predicted future, subject to system constraints. MPC has several desirable features from theoretical and practical viewpoints; e.g., it handles multivariable control problems naturally, it optimizes its dynamic performance over a prediction horizon, it takes account of input and output constraints, and it takes account of structural changes. A major challenge of the MPC research and development lies in the realization of nonlinear and robust MPC methods. This project aims at addressing several critical issues for analysis and design of nonlinear and robust MPC based on neural networks. First, in the cases of unknown dynamic systems or partially known dynamic systems with unmodeled dynamics, new types of neural networks with fast learning efficiency (e.g., echo state network, extreme learning machine) can be applied to build accurate prediction models. Secondly, with nonlinear models to predict dynamic behaviors, real-time optimization of nonconvex optimization problems have to be performed for nonlinear MPC, which is extreme demanding in terms of solution optimality and computational efficiency. For a class of nonlinear systems with weak nonlinearity, convexity or generalized convexity could be achieved by means of reformulating or transforming MPC performance indices. As a promising approach to real-time optimization, recurrent neural networks can be designed and employed for nonlinear MPC. Thirdly, for nonconvex optimization problems resulted from strong nonlinearity, they could be convexified by means of decompositions and approximations via affine transformation and supervised learning.The resulting convexified problems then could be solved by means of real-time neurodynamic optimization. In the presence of various system uncertainties, robust MPC could be synthesized by minimizing minimax optimization problems or solving linear matrix inequalitiesbased on neural networks. In addition, the research will develop computationally intelligent multiple-objective MPC based on goal programming, value function, and other scalarization methods. Finally, the research will explore the developed MPC schemes for the automatic control of robot manipulators and surface/underwater marine vehicles. It is expected that the project will make significant contributions in advancing the research of neural networks and real-time optimization as well as model predictive control.

模型预测控制是一种基于时域模型和实时优化的拥有广阔应用前景的先进控制方法。该领域当前的主要挑战在于非线性系统建模和优化的实现。本项目旨在开发多种基于神经网络的自主学习和实时优化等固有功能的非线性和鲁棒模型预测控制方法。针对系统未知或含有未建模动态的情况，应用学习速度快和泛化能力强的新型神经网络，建立高精度预测模型。针对非线性模型预测控制中普遍存在的非凸优化难题，通过构造凸或泛凸性能指标，确定一类弱非线性系统的凸化范围，进而设计收敛快和结构简单的递归神经网络作为并行计算模型实现实时优化。针对强非线性模型导致的非凸优化问题，通过线性化分解和监督学习等逼近方法凸化原问题，设计递归神经网络实时求解。针对多种系统不确定性，建立基于神经网络的鲁棒模型预测控制方法。并采用目标规划价值函数等标量化方法，建立多目标神经网络模型预测控制方法。进而探索基于神经网络的机器人和水面及水下载体的模型预测控制方案。

模型预测控制在石油、化工、电力、航空、航海等领域有着广泛的应用。本项目针对非线性系统建模和优化目前面临的挑战性问题，提出了多种基于神经网络的自主学习和实时优化等固有功能的非线性和鲁棒模型预测控制方法。针对系统未知或含有未建模动态的情况，应用学习速度快和泛化能力强的神经网络，建立了高精度预测模型。针对非线性模型预测控制中普遍存在的非凸优化问题，通过构造凸或泛凸性能指标，设计了收敛快和结构简单的递归神经网络作为并行计算模型用于实时优化。针对强非线性模型导致的非凸优化问题，通过线性化分解和监督学习等逼近方法凸化原问题，设计了递归神经网络实时求解。针对不确定性系统，建立了基于神经网络的鲁棒模型预测控制方法。针对自主水面船舶和水下机器人，提出了基于递归神经网络的在线建模和模型预测控制方法。该项目的完成有力地推动了神经动力学优化和模型预测控制的进步和发展。.研究小组在Neural Networks，IEEE Transactions on Fuzzy Systems，IEEE Transactions on Automatic Control，IEEE Transactions on Neural Networks and Learning Systems, IEEE Transactions on Mechatronics, IEEE Transactions on Industrial Electronics等国际学术期刊和会议中发表基金标注SCI/EI检索论文34篇，其中SCI检索期刊论文20篇（含录用），EI检索论文14篇。项目主持人多次担任国际学术会议大会主席或程序委员会主席，应邀在多个国际学术会议上作大会报告和国内外众多院校及研究机构作学术讲座。担任IEEE Transactions on Cybernetics主编，IEEE杰出讲师 (Distinguished Lecturer), Neural Networks编委和International Journal of Neural Systems顾问编委。项目主持人于2014年获得IEEE计算智能协会神经网络先驱奖，2016年获得“吴文俊人工智能科技技术奖”终身成就奖。