ANN for Identification and Analysis of Continuous-Time Models in Energy Processing Systems

ECS-9820977 Stankovic本研究项目的主要目标是开发一个全面的框架，在能源处理系统的降阶动态建模。 展望了该方法在电力系统中面向控制的应用（用于推导负载模型，用于在瞬态稳定性研究中简化传统模型，以及用于控制器设计），电力驱动（带弹性轴的驱动器），和运动控制系统（通过机械子系统耦合的多个驱动器）我们的假设是，所研究的物理现象通常具有比分量模型起源于其中的环境坐标系的聚集显著更低的维度。 我们提出了一种方法，结合分析和物理为基础的方法与人工神经网络（ANN）的功能，旨在识别连续时间微分/代数（DAE）模型的能源处理systems.The建议的研究计划结合了数学分析和能源工程的实际见解。 我们的降阶模型（“动态等价物”）是在连续时间内制定的，这与从物理原理导出的组件模型是一致的。 在标准离散和连续动力系统之间存在许多重要的区别。 基于离散时间人工神经网络的模型可以预测虚假的瞬态，并具有连续时间系统不可能的吸引子。 我们的动态等效包括微分和代数关系，这是被广泛接受的能量处理系统的最准确的表示。 将在这个程序中开发的程序是唯一适合于识别的系统，表现出多个时间尺度（奇异摄动系统），这是非常常见的能量处理。 所得的ANN模型是基于测量，或（可能是多个）模拟的详细模型，要简化。 我们在电力系统和电机的两个示例中沿着这些方向描述了初步结果。该项目将开发专门用于能源处理系统建模任务的人工神经网络架构和训练方法。 特别是，新的连接和标准（基于物理的）模型和人工神经网络结构之间的比较将探讨在域中的“灰箱”模型，其中联合收割机的两个类。 新的发展将在三种建模方法中进行评估：（i）时域，（ii）标准相量，（iii）动态相量框架。

ECS-9820977StankovicThe main goal of this research project is to develop a comprehensive framework for reduced-order dynamical modeling in energy processing systems. Control-oriented applications of the proposed method are envisioned in power systems (for derivation of load models, for reduction of conventional models in transient stability studies, and for controller design), electric drives (drives with elastic shafts), and motion control systems (multiple drives coupled through the mechanical subsystem).Our hypothesis is that the physical phenomena under investigation often have significantly lower dimensions than the agglomeration of ambient coordinate systems in which component models originate. We propose a methodology that combines features of analytical and physics-based approaches with artificial neural networks (ANNs), and aims to identify continuous-time differential/algebraic (DAE) models of energy processing systems.The proposed research program combines mathematical analysis and practical insight from energy engineering. Our reduced-order models ("dynamic equivalents") are formulated in continuous time, which is consistent with component models derived from physical principles. There exists a number of important differences between standard discrete and continuous dynamical systems. Models based on discrete-time ANNs may predict spurious transients and have attractors that are impossible for continuous-time systems. Our dynamic equivalent comprises differential and algebraic relationships, which is widely accepted as the most accurate representation of energy processing systems. The procedures that will be developed in this program are uniquely suited for identification of systems that exhibit multiple time scales (singularly perturbed systems) which are very common in energy processing. Resulting ANN models are based on measurements, or on (possibly multiple) simulations of detailed models that are to be simplified. We describe initial results along these directions in two examples from power systems and electric machines.This project will develop ANN architectures and training methods that are specific for the modeling tasks in energy processing systems. In particular, new connections and comparisons between standard (physics-based) models and ANN structures will be explored in the domain of "gray-box" models which combine the two classes. New developments will be evaluated within three modeling approaches: (i) time-domain, (ii) standard phasor, and (iii) dynamic phasor framework.