Abstract:

PID feedback control has been considered as the primary control strategy for chemical production processes ever since. However, due to the large time-delay and nonlinearity of complex chemical processes, PID control usually suffers poor performances in controlling key process variables. For this reason, in actual engineering, it is usually the on-site operators who implement manual prediction and control based on their own experience. In order to effectively learn manual predictive regulation strategies from historical operating data, the paper proposes a multi-task learning cascade network (LSTM multi-task network cascades, LSTM-MNC). Therein, process prediction models that predict short-term changes and long-term trends are established based on different trends of process variables in the long and short term, and the causal relationship between estimated information of the process prediction model and manipulated variable sequences is extracted. Supported by predictions of controlled variable deviations, manipulated variable sequences are generated, achieving intelligent manipulations of production processes. Experiments were conducted on an industrial heat exchanger process simulation platform, leading to satisfactory results, which verify the effectiveness of the proposed method.

Key words: predictive manipulating strategy, multi-task learning, cascade structure, chemical process, process control

摘要：

当前，PID反馈控制依然是化工生产过程的主要控制策略。然而，由于复杂化工过程通常具有大时滞和非线性等特性，使得PID控制对于一些关键过程参数控制的性能不佳。为此，在实际工程中通常是现场操作人员依据自身经验对其实施人工预测调控。为了能够从历史调控数据中学习人工预测调控策略，提出了一种多任务学习级联网络（LSTM multi-task network cascades，LSTM-MNC）。根据过程变量长短期不同趋势建立预测短期变化和长期趋势的过程预测模型，并学习过程预测模型估计信息与操纵变量序列的因果关系，由被控变量偏差预测支持操纵变量序列生成，实现生产过程的智能化调控。在工业换热器过程仿真平台上进行实验，获得了满意的结果，验证了所提方法的有效性。

关键词: 预测调控策略, 多任务学习, 级联结构, 化工过程, 过程控制