Abstract:

Sintering is an important production unit in the blast furnace ironmaking system, and its production level is directly related to the production efficiency of ironmaking enterprises. As the modernization process of iron-making enterprises continues to accelerate, higher demands are placed on the stability of the sintering production process. At present, the existing manual adjustment method in sintering production is not enough to achieve stable production. Constructing an advance prediction model for the state of the sintering system can enhance the immediateness and accuracy of the operation and improve the sintering production level. Sintering process has two prominent characteristics: time-delay and non-linear. To precisely predict the sintering production state, this paper builds a causal-based sintering state prediction model by integrating autocorrelation analysis, convergent cross-mapping and error back-propagation neural network. At a given threshold level, the autocorrelation length of the six state variables (the negative pressure and exhaust gas temperature of No. 14 and No. 22 bellows and exhaust gas temperature, sintering end position and temperature) is 27 min, and the influence window of the operating variable on the state variable is 40 min. The selected 9 sets of operating variables can produce a significant causal relationship with 6 state variables, and have a good explanation for the changes of 6 state variables. An error back-propagation neural network model was constructed, and the combined window input of the operating variable and the state variable was used to predict the average relative error of the six state variables. The sintering state prediction model proposed in this paper combines causal mechanism and black-box model, and achieves an accurate prediction of six key state variables in the sintering unit. The average relative error is within 3.1% according to industrial data test, which shows its capability to aid decision making in iron-making production.

Key words: convergent cross-mapping, causality, neural networks, sinter, model-predictive control, process systems

摘要：

烧结是高炉炼铁系统中的重要生产单元，其生产水平高低直接关系炼铁企业的生产效益。烧结过程具有时滞性和非线性特征，为了实现对烧结生产状态的准确预测，本文综合自相关分析、收敛交叉映射和误差反向传播神经网络等方法，融合因果性机理和黑箱模型，建立了基于因果分析的烧结生产状态预测模型。该模型通过因果分析层选取解释变量集、自相关窗口和因果性窗口，并通过神经网络层实现对6个烧结生产状态关键变量的准确预测。经过工业数据测试，该模型预测平均误差控制在0.5%～3.1%之间，能够有效辅助工厂进行烧结状态调整。

关键词: 收敛交叉映射, 因果性, 神经网络, 烧结, 模型预测控制, 过程系统