ADAPTIVE CRITIC DESIGN NEUROCONTROLLERS FOR NONLINEAR LARGE SCALE SYSTEMS AND ADVANCED NETWORK MANAGEMENT

电网、电信网、金融、交通等大系统是难以控制和管理的非线性非平稳大系统。这类系统的数学模型通常是基于线性技术得出的，广泛的安全裕度确保了稳定的运行。先进的辨识和控制方案是改善这类系统的吞吐量与可靠性之间权衡的必要手段。这些PI的初步工作表明，先进的、基于值的控制方法可以显著提高单个汽轮发电机对意外系统扰动的容忍能力。该项目将把这项工作扩展到发电机和电网的分布式网络。基本技术是一种高级形式的基于学习的近似动态规划或“强化学习”，它输出单个变量的价值衡量标准，或动态“影子价格”，而不是全球战略效用的传统衡量标准。这一领域的成功也可能给需要分布式控制和分布式智能的其他领域带来好处。

0080764WunschNonlinear non-stationary large-scale systems such as electric power networks, telecommunication networks, financial and transportation systems are difficult to control and manage. Mathematical models of such systems are typically derived based on linear techniques, and wide margins of safety ensure stable operation. Advanced identification and control schemes are necessary in order to improve the throughput-versus-reliability tradeoff of such systems.Preliminary work by these PIs has shown that advanced, value-based control methods can substantially improve the ability of individual turbogenerators to tolerate unexpected system disturbances. This project will extend that work to distributed networks of power generators and power grids in general. The underlying technology is an advanced form of learning-based Approximate Dynamic Programming or "Reinforcement Learning" which outputs measures of value for individual variables, or dynamic "shadow prices," rather than the conventional measures of global strategic utility. Success in this area may also have benefits for other domains where distributed control and distributed intelligence are needed.