Qualitative Modeling and Machine Learning Applied to the Real-Time Estimation of Chemical Process Dynamics

动态变化过程的控制，例如那些以批或半批模式操作的过程，是复杂的，因为用于实现控制策略的数学模型没有解析解。有时会使用递归最小二乘（RLS）估计等数值解决方案，但对于可能是高阶，时变，非线性或三者兼而有之的化学处理系统，这些解决方案并不容易实现。已经成功地使用了结合特定系统特性修改RLS的方法，但解决方案的适用性有限。在这项工作中，PI计划使用RLS生成自动控制算法，该算法基于将干扰分类为三类之一并相应地规划控制动作。所得到的方法有望具有普遍的适用性。干扰有三种基本类型：1)系统状态缓慢漂移——对于这样的系统，存在一个允许使用线性化模型的可接受的操作条件范围，但过程最终将超出该范围（实际例子是由催化剂衰变、结垢和结渣引起的）。等等)。2)短周期稳态运行-这导致线性动态模型不可操作。3)设定值的突然变化——紧随此类干扰之后，线性模型将非常不准确（例如，当饲料批次发生变化时，实际例子就会发生）。PI计划设计他的算法，以便将所有干扰分解成一个或多个干扰的组合。然后，该方法包括对过程行为的连续诊断，以更新控制动作。

The control of dynamically changing processes, for example those operated in a batch or semi-batch mode, are complex because the mathematical models used to implement the control strategies do not have analytical solutions. Numerical solutions such as recursive least-squares (RLS) estimation are sometimes used, but these are not easy to implement for chemical processing systems which may be high order, time-varying, non-linear or all three. Modifying RLS incorporating specific system characteristics has been used successfully resulting in solutions that have limited applicability. In this work, the PI plans to generate automatic control algorithms using RLS that are based on classifying disturbances into one of three categories and planning control action accordingly. The resulting method is expected to have general applicability. There are three basic types of disturbances: 1) Slow drift in system state - for such systems there is an acceptable band of operating conditions which will permit the use of linearized models, but the process will ultimately move out of this band (real examples are caused by catalyst decay, fouling and slagging, etc.). 2) Short periods of steady state operation - this causes linear dynamic models to be inoperable. 3) Sudden change in setpoint - immediately following such disturbances linear models will be grossly inaccurate (real examples occur when feed lots are changed, for example). The PI plans to design his algorithm so that it will decompose all disturbances into combinations of one or more of these. The method then involves continuous diagnosis of process behavior to update control action.