Abstract To maintain optimum economic process performance, a good process model is the cornerstone of an optimal scheduling strategy and controller design. Up to now, approaches to dynamic modeling have already been studied, but the models they constructed are only valid in their corresponding operating conditions. As operating conditions switch fast during the production, the constructed model may lack the extrapolating capability and may not describe the process behaviors in the new operating condition properly. Especially in the case that only a small number of data can be collected from the new operating condition for the construction of the model; the performance of the model may not be guaranteed for online new data. In this paper, a dynamic transfer modeling approach based on the Gaussian process model (GPM) is proposed. It can quickly model the target process and get correct predictions, by transferring source model knowledge trained with a sufficient number of historical data to a target model with a small number of available target data. This can significantly reduce the amount of time waiting for getting the target process data and quickly achieve a good process model. The statistical approach leverages GPM to transfer the knowledge. GPM is introduced to capture the uncertainty that propagates from the source process to the target process. Thus, the multi-step ahead prediction of the target model can provide the mean prediction as well as probabilistic information for its prediction in the form of a predictive variance. Finally, CSTR and the real furnace system are used to demonstrate the features of the proposed method and the applicability to a real plant process.

中文翻译：

---

基于合并源知识的迁移学习使用高斯过程模型快速建模动态目标过程

摘要 为了保持最佳的经济过程性能，良好的过程模型是优化调度策略和控制器设计的基石。到目前为止，已经研究了动态建模的方法，但是他们构建的模型仅在其相应的操作条件下有效。由于生产过程中工况变化很快，构建的模型可能缺乏外推能力，可能无法正确描述新工况下的工艺行为。尤其是在新工况下只能采集少量数据用于模型构建的情况下；对于在线新数据，可能无法保证模型的性能。本文提出了一种基于高斯过程模型（GPM）的动态传递建模方法。它可以通过将使用足够数量的历史数据训练的源模型知识转移到具有少量可用目标数据的目标模型来快速建模目标过程并获得正确的预测。这可以显着减少等待获取目标流程数据的时间，并快速实现良好的流程模型。统计方法利用 GPM 来转移知识。引入 GPM 来捕获从源进程传播到目标进程的不确定性。因此，目标模型的多步提前预测可以以预测方差的形式为其预测提供平均预测和概率信息。最后，