Data driven hybrid model identification for control and optimisation of petrochemical and refining plants

通过确定最佳运行条件(优化)来实现加工厂的最佳运行 操作)，并控制它们的实时操作以保持这些最佳状态。这种控制和 优化需要准确的流程模型。 自20世纪70年代以来，出现了两种不同的模型开发流程：(I)源自 第一原理，以及(Ii)根据植物数据通过 自动控制社区。前者通常需要很大的努力来推导出第一原则 方程及其求解的有效算法的构造。后者已经沿着这条道路前进了 从工厂运行数据中识别经验模型。尽管已经取得了重大进展 并且有许多非常成功地使用这两种类型的模型进行优化和控制的实例， 分别而言，仍有许多改善的机会。例如，识别以下对象的模型 具有大时滞的过程 在过去的十年中，语音的人工智能方法取得了重大进展 识别、图像识别和分类、手写识别等。这些进步的基础 是由许多层组成的深层神经网络。人们已经发现，特定的神经网络结构 最擅长为特定类型的应用程序创建模型。这项研究提出了非常准确的识别 通过为特定类型的流程开发特定的模型结构，从操作数据中创建模型。会的 将一些第一原理方程(例如，质量和能量平衡)与深度神经网络或与 通过自动控制领域的识别方法开发的模型。预测两者的模型 将开发稳态和动态行为，并特别关注具有大型 时间延迟(例如乙烷/乙烯分离器)。已经为特定的项目开发了模型的“标准形式” 设备将使这些模型能够容易地进行调整，以代表特定的设备并可重复使用。

Excellence in operation of process plants is attained by determining the best operating conditions (optimisation of operation) and controlling their real-time operation to maintain these best conditions. Such control and optimisation require accurate process models. Since 1970s there have been two separate streams of model development: (i) rigorous models derived from the first principles, and (ii) empirical models identified from plant data via methods developed in the automatic control community. The former usually require a large effort in deriving the first principles equations and construction of efficient algorithms for solving them. The latter have proceeded along the path of identifying empirical models from the plant operating data. Even though significant advances have been made, and there are many instances of very successful use of both types of models for optimisation and control, respectively, there are still many opportunities for improvement. For instance, identification of models for processes with large time delays During the last decade there have been significant advances in artificial intelligence methods for speech recognition, image recognition and classification, handwriting recognition etc. Foundation for these advances are deep neural networks comprised on many layers. It has been found that specific neural network structures are best at creating models for specific types of applications. This research proposes to identify very accurate models from operating data by developing specific model structures for specific types of processes. It will combine some first principles equations (e.g. mass and energy balances) with deep neural networks or with models developed via identification methods from the automatic control field. Models predicting both steady-state and dynamic behaviour will be developed, with particular attention devoted to models with large time delays (e.g. ethane/ethylene splitter). Having developed a "standard form" of the model for specific equipment will enable such models to be readily adjusted to represent specific equipment and be re-used.