Orthogonalization-based manipulated-variable ranking for identifying and addressing gain-conditioning problems in multivariable control systems

其目的是开发和测试一种新的软件工具，以减少高级控制工程师在炼油厂和化工厂实施线性模型预测控制应用时所需的时间和精力。目前，高技能的控制工程师进行耗时的增益调节调查，然后对在线线性规划（LP）和动态矩阵控制（DMC）计算中使用的过程增益矩阵进行手动调整。为了实现安全和经济的过程操作，这些调查用于检测在调整多个操纵变量（如流量和加热器设置）时可能发生的潜在不良相互作用。所提出的工具将帮助帝国石油公司和其他公司的工程师在改进条件下生成增益矩阵，从而提高过程自动化。所提出的增益调节工具将能够有效地结合来自两类模型的信息：i)从动态过程数据中识别的经验传递函数或阶跃响应模型，ii)用于实时优化（RTO）计算的稳态基本模型。这些模型的预测包含误差（由于参数不确定性，干扰和结构缺陷），但RTO模型确保遵守物质和能量平衡，为增益调节问题提供关键见解。该算法的另一种版本也将被开发出来，以帮助在RTO模型不可用时进行增益调节。将开发具有代表性的案例研究，并用于测试当模型预测与真实过程行为之间存在不匹配时，所提出的方法是否有助于防止不希望的控制行为。

The aim is to develop and test a new software tool that will reduce time and effort required by advanced control engineers when they implement linear model-predictive control applications at refineries and chemical plants. Currently, highly-skilled control engineers undertake time-consuming gain-conditioning investigations and then make manual adjustments to process gain matrices used in online Linear Programming (LP) and Dynamic Matrix Control (DMC) calculations. These investigations are used to detect potential undesirable interactions that may occur when multiple manipulated variables (MVs), such as flow rates and heater settings, are adjusted to achieve safe and economical process operation. The proposed tool will help engineers at Imperial Oil and other companies to generate gain matrices with improved conditioning that will lead to improved process automation. The proposed gain-conditioning tool will be able to effectively combine information from two types of models: i) empirical transfer-function or step-response models identified from dynamic process data and ii) steady-state fundamental models used in real-time optimization (RTO) calculations. Predictions from these models contain errors (due to parameter uncertainties, disturbances and structural imperfections), but RTO models ensure that material and energy balances are obeyed, providing key insights into the gain-conditioning problem. An alternative version of the algorithm will also be developed to aid gain conditioning when RTO models are not available. Representative case studies will be developed and used to test whether the proposed methods help to prevent undesirable control actions when there is mismatch between model predictions and the true process behaviour.