WO2022105294A1 - 一种热轧卷取温度的区间预测方法 - Google Patents

一种热轧卷取温度的区间预测方法 Download PDF

Info

Publication number
WO2022105294A1
WO2022105294A1 PCT/CN2021/109201 CN2021109201W WO2022105294A1 WO 2022105294 A1 WO2022105294 A1 WO 2022105294A1 CN 2021109201 W CN2021109201 W CN 2021109201W WO 2022105294 A1 WO2022105294 A1 WO 2022105294A1
Authority
WO
WIPO (PCT)
Prior art keywords
interval
neural network
prediction
picp
coiling temperature
Prior art date
Application number
PCT/CN2021/109201
Other languages
English (en)
French (fr)
Inventor
李旭
栾峰
李宗浩
吴艳
韩月娇
张殿华
Original Assignee
东北大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 东北大学 filed Critical 东北大学
Priority to JP2022530319A priority Critical patent/JP7393831B2/ja
Publication of WO2022105294A1 publication Critical patent/WO2022105294A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00Computer-aided design [CAD]
    • G06F30/20Design optimisation, verification or simulation
    • G06F30/27Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • BPERFORMING OPERATIONS; TRANSPORTING
    • B21MECHANICAL METAL-WORKING WITHOUT ESSENTIALLY REMOVING MATERIAL; PUNCHING METAL
    • B21BROLLING OF METAL
    • B21B37/00Control devices or methods specially adapted for metal-rolling mills or the work produced thereby
    • B21B37/74Temperature control, e.g. by cooling or heating the rolls or the product
    • BPERFORMING OPERATIONS; TRANSPORTING
    • B21MECHANICAL METAL-WORKING WITHOUT ESSENTIALLY REMOVING MATERIAL; PUNCHING METAL
    • B21BROLLING OF METAL
    • B21B2201/00Special rolling modes
    • B21B2201/06Thermomechanical rolling
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F2119/00Details relating to the type or aim of the analysis or the optimisation
    • G06F2119/08Thermal analysis or thermal optimisation
    • YGENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02PCLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
    • Y02P90/00Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
    • Y02P90/30Computing systems specially adapted for manufacturing

Definitions

  • the invention relates to the technical field of temperature interval prediction, in particular, to a method for interval prediction of hot-rolled coiling temperature.
  • Controlling the coiling temperature is an important step in the production of hot-rolled strip, and its prediction is beneficial to improve the properties of the strip.
  • most predictions for coiling temperature are point predictions.
  • Scholars usually use BP neural network for point prediction of coiling temperature.
  • wavelet neural network for point prediction of strip coiling temperature and have achieved good results.
  • Point forecasting provides a forecast point for a target value and can only provide forecast error, not the probability of a correct forecast. And since point prediction cannot handle contingent data, the prediction accuracy is reduced.
  • This patent performs interval prediction on the crimping temperature, on the one hand, a larger amount of prediction information can be obtained, and on the other hand, the validity of the prediction is enhanced.
  • Delta, Bayesian, and Bootstrap are three traditional methods commonly used to construct neural network-based interval forecasting.
  • the Delta method mainly applies the nonlinear regression technique of the neural network, and firstly obtains a set of parameters of the linear neural network by minimizing the sum of squared errors. Then, standard asymptotic theory is applied to construct linearized models for interval prediction. Prediction intervals are constructed assuming that the noise is uniformly normally distributed. Since the noise is not uniform in most real case studies, the constructed intervals are not accurate.
  • the Bayesian method is another method for constructing prediction intervals based on neural networks. Training a neural network using Bayesian techniques allows error bars to be assigned to the predicted values of the network.
  • an interval prediction method of hot-rolled coiling temperature is provided.
  • the present invention mainly utilizes a kind of interval prediction method of hot rolling coiling temperature, which is characterized in that it comprises the following steps:
  • Step 1 According to the original data of the rolled piece and the measured sample data of the finishing rolling outlet, the input data and output data can be known; wherein the input data include: the final rolling temperature, the strip speed, that is, the F6 speed, the average strip thickness and the target coiling temperature. ; Described output data includes: the relative deviation of actual measured coiling temperature and target coiling temperature;
  • Step 2 data preprocessing; divide the input data and output data into a training set, a verification set and a test set according to a certain proportion; and normalize all the data;
  • Step 3 set the artificial neural network; set the number of hidden layers of the artificial neural network and the number of nodes in each layer of the hidden layer; use the sigmoid function as the activation function, set the learning rate to 1, and scan the data.
  • the number of times is set to 40, and the average gradient step size is set to 10 for neural network training;
  • Step 4 optimize the artificial neural network through the whale optimization algorithm, and perform optimization by minimizing the cost function to obtain the optimal weight and bias of the artificial neural network;
  • Step 5 According to the input, the upper and lower prediction limits of the relative deviation between the measured coiling temperature and the target coiling temperature are obtained, and point prediction and interval prediction can be performed simultaneously;
  • Step 6 Perform interval prediction performance analysis according to the proposed evaluation index NCWC.
  • a single-layer artificial neural network, a double-layer artificial neural network and a three-layer artificial neural network are respectively used for interval prediction; wherein, the number of hidden layer nodes of the single-layer artificial neural network is 15; The number of nodes in the hidden layer of the two-layer artificial neural network is 16 and 18 respectively; the number of nodes in the hidden layer of the three-layer artificial neural network is 18, 10 and 20 respectively.
  • step 4 first randomly initializes the weights and biases of the neural network, and trains the neural network through the whale optimization algorithm; it specifically includes the following steps:
  • Step 4.2 Determine the optimal search agent; the target prey is the optimal position of the search space, set the best search agent, and other search agents will update their positions with the position of the best search agent; wherein: calculate each search agent The fitness function of , represents the best search agent with the smallest fitness;
  • Step 4.3 Search agent position update; when the current number of iterations is less than the maximum number of iterations, for each search agent, update the parameters: l, p,
  • t represents the current iteration number
  • b represents a constant that defines the shape of the logarithmic spiral
  • Step 4.5 Return the weight and bias value of the artificial neural network; add 1 to the current number of iterations, and return to step 4.3 until the current number of iterations is equal to the maximum number of iterations; return the optimal position vector That is, the weights and biases of the artificial neural network.
  • the fitness function in the step 4.2 includes: a comprehensive evaluation index of coverage probability and coverage width; specifically, the following steps are included:
  • Step 4.2.1 Calculate the interval prediction coverage probability;
  • PICP represents the probability that the target value is covered by the upper and lower limits, and is defined as follows:
  • N represents the total number of samples
  • Step 4.2.2 Calculate the average width of the interval forecast; the quantitative measure of the interval forecast width is defined as the normalized average width of the forecast interval, PINAW, which is expressed mathematically as follows:
  • R represents the difference between the maximum value and the minimum value of the objective function value
  • Step 4.2.3 Calculate interval forecast mean square error; introduce interval forecast mean square error PIMSE, and obtain a symmetrical interval closer to the true confidence interval by minimizing the PIMSE index; expressed in mathematical formulas as follows:
  • Step 4.2.4 Redefine the standard based on coverage and width; use the standard NCWC based on coverage and width as the evaluation index for the final interval prediction;
  • NCWC PINAW+ ⁇ (PICP)e ⁇ (PICP ⁇ ) +PIMSE;
  • the pre-specified PICP is less than ⁇ , it is the balance between PINAW and PICP when PICP reaches 95%;
  • ⁇ (PICP) is a step function, and ⁇ (PICP) is determined by the satisfaction of PICP:
  • the present invention has the following advantages:
  • the invention realizes interval prediction of hot-rolled coiling temperature. And by changing the structure of artificial neural network, comparing single-layer ANN, double-layer ANN and three-layer ANN, it is found that using three-layer artificial neural network for prediction can significantly improve the prediction accuracy of the model.
  • the interval prediction method of the invention adopts a group of intelligent optimization algorithms such as whale optimization algorithm, which has strong global optimization ability, avoids easily falling into the local optimal problem, strengthens the optimization breadth and precision, and significantly improves the convergence speed.
  • whale optimization algorithm which has strong global optimization ability, avoids easily falling into the local optimal problem, strengthens the optimization breadth and precision, and significantly improves the convergence speed.
  • the invention proposes a new cost function which is also an evaluation index for interval prediction.
  • the contradictory multi-objective optimization problem of interval prediction coverage and interval prediction width is transformed into a single-objective optimization problem, and the concept of interval prediction mean square error is proposed, and the overall consideration is more comprehensive.
  • the method can realize point prediction and interval prediction at the same time.
  • FIG. 1 is a flow chart of a method for predicting an interval of hot-rolled coiling temperature according to the present invention.
  • FIG. 2 is a schematic diagram of a network structure of a neural network of a method for interval prediction of hot-rolled coiling temperature according to an embodiment of the present invention.
  • FIG. 3 is a schematic diagram of an interval prediction result of an interval prediction method for hot-rolled coiling temperature according to an embodiment of the present invention.
  • FIG. 4 is a schematic diagram of a point prediction result of an interval prediction method for hot-rolled coiling temperature according to an embodiment of the present invention.
  • the present invention provides an interval prediction method for hot-rolled coiling temperature, comprising the following steps:
  • Step 1 According to the original data of the rolled piece and the measured sample data of the finishing rolling outlet, the input data and output data can be known; wherein the input data include: the final rolling temperature, the strip speed, that is, the F6 speed, the average strip thickness and the target coiling temperature. ; Described output data includes: the relative deviation of actual measured coiling temperature and target coiling temperature;
  • Step 2 data preprocessing; divide the input data and output data into a training set, a verification set and a test set according to a certain proportion; and normalize all the data.
  • it is generally necessary to ensure that the training set is larger than the validation set and the test set.
  • the ratio of training set:validation set:test set is 4:3:3 or 6:2:2.
  • Step 3 set the artificial neural network; set the number of hidden layers of the artificial neural network and the number of nodes in each layer of the hidden layer; use the sigmoid function as the activation function, set the learning rate to 1, and scan the data.
  • the number of times is set to 40, and the average gradient step size is set to 10 for neural network training.
  • a single-layer artificial neural network, a double-layer artificial neural network and a three-layer artificial neural network are respectively used for interval prediction; wherein, the number of hidden layer nodes of the single-layer artificial neural network is 15; The number of nodes in the hidden layer of the neural network is 16 and 18 respectively; the number of nodes in the hidden layer of the three-layer artificial neural network is 18, 10 and 20 respectively.
  • Step 4 The artificial neural network is optimized by the whale optimization algorithm, and the optimization is performed by minimizing the cost function to obtain the optimal weight and bias of the artificial neural network.
  • the step 4 first randomly initializes the weights and biases of the neural network, and trains the neural network through the whale optimization algorithm; it specifically includes the following steps:
  • Step 4.2 Determine the optimal search agent; the target prey is the optimal position of the search space, set the best search agent, and other search agents will update their positions with the position of the best search agent; wherein: calculate each search agent The fitness function of , represents the best search agent with the smallest fitness.
  • the fitness function in the step 4.2 includes: a comprehensive evaluation index of coverage probability and coverage width; specifically, the following steps are included:
  • Step 4.2.1 Calculate the interval prediction coverage probability;
  • PICP represents the probability that the target value is covered by the upper and lower limits, and is defined as follows:
  • N represents the total number of samples
  • Step 4.2.2 Calculate the average width of the interval predictions; the quantitative measure of the interval prediction width is defined as the prediction interval normalized average width PINAW, which is expressed mathematically as follows:
  • R represents the difference between the maximum value and the minimum value of the objective function value
  • Step 4.2.3 Calculate interval forecast mean square error; introduce interval forecast mean square error PIMSE, and obtain a symmetrical interval closer to the true confidence interval by minimizing the PIMSE index; expressed in mathematical formulas as follows:
  • Step 4.2.4 Redefine the standard based on coverage and width; use the standard NCWC based on coverage and width as the evaluation index for the final interval prediction;
  • NCWC PINAW+ ⁇ (PICP)e ⁇ (PICP ⁇ ) +PIMSE;
  • the pre-specified PICP is less than ⁇ , it is the balance between PINAW and PICP when PICP reaches 95%;
  • ⁇ (PICP) is a step function, and ⁇ (PICP) is determined by the satisfaction of PICP:
  • Step 4.3 Search agent position update; when the current number of iterations is less than the maximum number of iterations, for each search agent, update the parameters: l, p,
  • t represents the current iteration number
  • b represents a constant that defines the shape of the logarithmic spiral
  • Step 4.4 Update the search agents; determine if any search agents are out of the search space and correct them, calculate the fitness function of each search agent, and replace it with the search agent with the smallest fitness value
  • Step 4.5 Return the weight and bias value of the artificial neural network; add 1 to the current number of iterations, and return to step 4.3 until the current number of iterations is equal to the maximum number of iterations; return the optimal position vector That is, the weights and biases of the artificial neural network.
  • Step 5 According to the input, the upper and lower prediction limits of the relative deviation between the measured coiling temperature and the target coiling temperature are obtained, and point prediction and interval prediction can be performed simultaneously;
  • Step 6 Perform interval prediction performance analysis according to the proposed evaluation index NCWC.
  • FIG. 1 An interval prediction method for hot-rolled coiling temperature, the process of the prediction steps is shown in the above-mentioned FIG. 1 .
  • the present invention will be further described below in conjunction with examples.
  • the method selects the relevant data of the coiler in the laminar cooling system of a hot rolling plant, and the data used are 1600 groups in total.
  • Each set of data includes 4 input data and 1 output data.
  • the input data include: rolling temperature, strip speed (F6 speed), average strip thickness, and target coiling temperature.
  • the output data is: the relative deviation between the measured coiling temperature and the target coiling temperature.
  • the method first preprocesses the data, and normalizes all the data to data between 0 and 1.
  • the data is divided into training set, validation set and test set according to a certain proportion.
  • the artificial neural network is parameterized.
  • the method adopts the sigmoid function as the activation function, the learning rate is set to 1, the number of data scans is set to 40, and the average gradient step size is set to 10 for training.
  • a single-layer artificial neural network, a two-layer artificial neural network and a three-layer artificial neural network are used for interval prediction, respectively.
  • the number of hidden layer nodes of a single-layer ANN is 15.
  • the number of nodes in the hidden layer of the two-layer ANN is 16 and 18, respectively.
  • the number of hidden layer nodes of the three-layer ANN is 18, 10 and 20, respectively.
  • the relevant parameters of the whale optimization algorithm used in the method are set, the number of search agents is set to 50, and the maximum number of iterations T is set to 2000.
  • the upper and lower sidebands of the search space are set to 1 and -1, respectively.
  • the two control parameters ⁇ and ⁇ distribution in the evaluation index NCWC are set to 0.95 and 50.
  • the method uses the whale optimization algorithm to optimize the artificial neural network to obtain the optimal weights and biases of the artificial neural network.
  • the specific steps are as follows: First, initialize the algorithm, calculate the fitness function value of all search agents after given parameters, and determine the search agent with the smallest fitness function value as the optimal search agent. The search agent is then updated with the location based on the new formula for the location. After the update, the fitness function value of each search agent is calculated again, and the search agent corresponding to the minimum function value is selected to replace the original optimal search agent. Determine whether the number of iterations is equal to the maximum number of iterations, and if they are equal, output the optimal position vector, that is, the weight and bias of the artificial neural network.
  • the cost function of the optimization algorithm involves three indicators, the interval prediction coverage rate, the interval prediction average width, and the interval prediction mean square error.
  • the specific formula is as follows:
  • NCWC PINAW+ ⁇ (PICP)e ⁇ (PICP ⁇ ) +PIMSE;
  • the proposal of this new evaluation index takes into account both information and effectiveness, which can strengthen the optimization effect and conduct a more comprehensive evaluation of interval prediction methods.
  • the interval prediction method finally provides the upper and lower prediction limits of prediction and the output result of point prediction, which can realize point prediction and interval prediction at the same time.
  • Table 1 is the comparison between the three-layer ANN-based hot-rolling coiling temperature interval prediction method proposed in the example and the single-layer artificial neural network and double-layer artificial neural network hot-rolling coiling temperature interval prediction methods.
  • the comparison results show that the interval prediction method adopted by the patent has significantly reduced the value of the comprehensive evaluation index NCWC and has a better prediction effect.

Abstract

本发明提供一种热轧卷取温度的区间预测方法,包括以下步骤:根据轧件原始数据以及精轧出口的实测样本数据可知输入数据和输出数据;对数据进行预处理;对人工神经网络进行设置;通过鲸鱼优化算法优化所述人工神经网络,进而通过最小化代价函数进行寻优,获取人工神经网络最优的权重和偏置量;根据所述输入得到实测卷取温度与目标卷取温度的相对偏差的预测上限和预测下限,可同时进行点预测与区间预测;根据所提出的评价指标NCWC进行区间预测性能分析。本发明在卷取温度预测领域,相比于采用传统数学模型进行点预测,所述发明实现了热轧卷取温度的区间预测。并通过改变人工神经网络结构,对比单层ANN、双层ANN与三层ANN,发现采用三层人工神经网络进行预测能够显著提高模型预测精度。

Description

一种热轧卷取温度的区间预测方法 技术领域
本发明涉及温度区间预测的技术领域,具体而言,尤其涉及一种热轧卷取温度的区间预测方法。
背景技术
控制卷取温度是生产热轧带钢的重要步骤,对其进行预测有利于改善带钢的性能。影响卷取温度的因素多且复杂,很难用传统数学模型对卷曲温度进行预测。目前,对于卷取温度的预测大多是点预测。学者通常对卷取温度采用BP神经网络进行点预测,另外有学者采用小波神经网络对带钢卷取温度进行点预测取得了不错效果。点预测为一个目标值提供一个预测点,只能提供预测误差,没有说明正确预测的概率。并且由于点预测不能处理偶然性数据,降低了预测准确度。本专利对卷曲温度进行区间预测,一方面可以获得更大的预测信息量,另一方面增强了预测的有效性。
Delta、Bayesian和Bootstrap是三种常用于构建基于神经网络的区间预测传统方法。Delta方法主要应用了神经网络的非线性回归技术,首先通过最小化平方误差和而获得的一组线性神经网络的参数。然后,将标准渐近理论应用于构造区间预测的线性化模型。假设噪声是均匀正态分布的以此来构建预测区间。由于噪声在大多真实的案例研究中并不均匀,构建的区间并不准确。Bayesian方法是另一种基于神经网络的预测区间构造方法。使用贝叶斯技术训练神经网络允许误差条被分配给网络的预测值。尽管有较强的理论支持,但该方法计算负担比较大,需要计算成本函数的Hessian矩阵来构建概率神经网络。Bootstrap方法也是利用神经网络进行区间预测的方法之一,不需要复杂的导数计算,但其主要缺点是需要对大量数据进行计算,计算代价较大。由此可见,以上传统的区间预测方法实施较为困难,其缺点阻碍了传统区间预测方法的广泛应用。A.Khosravi提出一种新的区间预测方法,称之为上下界区间估计法。此方法没有对数据分布做任何假设,并且避免了诸如Jacobian矩阵和Hessian矩阵的计算。
发明内容
根据上述提出的技术问题,而提供一种热轧卷取温度的区间预测方法。本发明主要利用一种热轧卷取温度的区间预测方法,其特征在于,包括以下步骤:
步骤1:根据轧件原始数据以及精轧出口的实测样本数据可知输入数据和输出数据;其中所述输入数据包括:终轧温度、带钢速度即F6速度、带钢平均厚度以及目标卷取温度;所述输出数据包括:实测卷取温度与目标卷取温度的相对偏差;
步骤2:数据预处理;将所述输入数据和输出数据按一定比例分为训练集、验证集和测试集;并将所有数据进行归一化处理;
步骤3:对人工神经网络进行设置;设置所述人工神经网络的隐藏层层数以及所述隐藏层每一层节点个数;通过sigmoid函数作为激活函数,将学习率设置为1,数据的扫描次数设置为40,平均梯度步长设置为10以进行神经网络训练;
步骤4:通过鲸鱼优化算法优化所述人工神经网络,通过最小化代价函数进行寻优,获取人工神经网络最优的权重和偏置量;
步骤5:根据所述输入得到实测卷取温度与目标卷取温度的相对偏差的预测上限和预测下限,可同时进行点预测与区间预测;
步骤6:根据所提出的评价指标NCWC进行区间预测性能分析。
进一步地,所述步骤3分别采用单层人工神经网络、双层人工神经网络和三层人工神经网络进行区间预测;其中,所述单层人工神经网络的隐藏层节点个数为15;所述双层人工神经网络隐藏层的节点个数分别为16和18;所述三层人工神经网络的隐藏层节点个数分别是18、10和20。
更进一步地,所述步骤4首先随机初始化神经网络的权重和偏置,并通过鲸鱼优化算法训练神经网络;具体包括以下步骤:
步骤4.1:初始化算法;初始化时,包括座头鲸数量Xi,当前的迭代次数为t,t=0,最大迭代次数T并对猎物的位置进行随机初始化;
步骤4.2:确定最优搜索代理;目标猎物即搜索空间的最优位置,设置最 佳的搜索代理,其他搜索代理将以所述最佳搜索代理的位置更新自身位置;其中:计算每一个搜索代理的适应度函数,
Figure PCTCN2021109201-appb-000001
表示最佳的搜索代理,所述最佳的搜索代理的适应度最小;
步骤4.3:搜索代理位置更新;当当前迭代次数小于最大迭代次数时,对于每一个搜索代理,进行参数的更新:
Figure PCTCN2021109201-appb-000002
l,p,
Figure PCTCN2021109201-appb-000003
其中,
Figure PCTCN2021109201-appb-000004
在迭代过程中从2到0线性递减,
Figure PCTCN2021109201-appb-000005
为0到1的随机数,l为介于[-1,1]之间的随机数,P为[0,1]之间的随机数,
Figure PCTCN2021109201-appb-000006
Figure PCTCN2021109201-appb-000007
是两个向量系数;当p<0.5,
Figure PCTCN2021109201-appb-000008
时按照下式更新当前搜索代理的位置公式:
Figure PCTCN2021109201-appb-000009
Figure PCTCN2021109201-appb-000010
其中,t表示当前迭代次数,
Figure PCTCN2021109201-appb-000011
表示当前最优的位置向量;向量系数
Figure PCTCN2021109201-appb-000012
的计算如下:
Figure PCTCN2021109201-appb-000013
Figure PCTCN2021109201-appb-000014
当p<0.5,
Figure PCTCN2021109201-appb-000015
时随机选择一个搜索代理
Figure PCTCN2021109201-appb-000016
即随机位置向量,按照下式更新当前搜索代理的位置公式:
Figure PCTCN2021109201-appb-000017
Figure PCTCN2021109201-appb-000018
当p>0.5时按照下式更新当前搜索代理的位置公式:
Figure PCTCN2021109201-appb-000019
Figure PCTCN2021109201-appb-000020
其中,b表示定义对数螺线形状的常数;
步骤4.4:更新搜索代理;判断是否有任何搜索代理超出了搜索空间并对其进行修正,计算每个搜索代理的适合度函数,用适应度值最小的搜索代理替换
Figure PCTCN2021109201-appb-000021
步骤4.5:返回人工神经网络的权重和偏置值;当前迭代次数加1,返回 步骤4.3直到当前迭代次数等于最大迭代次数;返回最优位置向量
Figure PCTCN2021109201-appb-000022
也就是人工神经网络的权重与偏置。
进一步地,所述步骤4.2中适应度函数为包括:覆盖概率和覆盖宽度的综合评价指标;具体地包括以下步骤:
步骤4.2.1:计算区间预测覆盖概率;PICP表示目标值被上限和下限覆盖的概率,定义如下:
Figure PCTCN2021109201-appb-000023
其中,N表示样本总数,∈ i表示布尔变量,表示预测区间的覆盖率;如果目标值yi的数值大于等于下限Li且小于等于上限Ui时,则∈ i=1;如果目标值yi的数值小于下限Li或大于上限Ui时,则∈ i=0;用数学公式表示为:
Figure PCTCN2021109201-appb-000024
步骤4.2.2:计算区间预测平均宽度;区间预测宽度的定量测量被定义为预测区间归一化平均宽度PINAW,用数学公式表示如下:
Figure PCTCN2021109201-appb-000025
其中,R表示目标函数值的最大值与最小值之差;
步骤4.2.3:计算区间预测均方误差;引入区间预测均方误差PIMSE,通过最小化PIMSE指数,获取更接近真实置信区间的对称区间;用数学公式表示如下:
Figure PCTCN2021109201-appb-000026
步骤4.2.4:重新定义基于覆盖率和宽度的标准;通过基于覆盖率和宽度的标准NCWC作为最终区间预测的评价指标;
NCWC=PINAW+γ(PICP)e -η(PICP-μ)+PIMSE;
其中,对于训练集来说,γ(PICP)=1;μ和η表示两个控制参数;名义置信水平[(1-α)%]为选择μ的参考,其中μ为95%;μ表示预先指定必须满足的区间预测覆盖率;η表示超参数,η放大PICP和μ之间的差异;
如果预先指定的PICP小于μ,当PICP达到95%时,它就是PINAW和 PICP之间的平衡;
因此,对于测试样本来说,γ(PICP)是阶跃函数,γ(PICP)由PICP的满意度确定:
Figure PCTCN2021109201-appb-000027
即对于评价区间预测的结果来说,如果PICP不小于指定的μ,γ(PICP)=0,同样PICP的测量值也为0;否则,γ(PICP)=1,相应的处罚通过NCWC计算得出。
较现有技术相比,本发明具有以下优点:
(1)在卷取温度预测领域,相比于采用传统数学模型进行点预测,所述发明实现了热轧卷取温度的区间预测。并通过改变人工神经网络结构,对比单层ANN、双层ANN与三层ANN,发现采用三层人工神经网络进行预测能够显著提高模型预测精度。
(2)所述发明的区间预测方法采用鲸鱼优化算法这一群智能优化算法,全局寻优能力强,避免了容易陷入局部最优问题,加强寻优广度与精度,明显提高收敛速度。
(3)所述发明提出了新的代价函数同时也是区间预测的评价指标。将区间预测覆盖率与区间预测宽度这一矛盾的多目标优化问题转化成单目标优化问题,并提出区间预测均方误差这一概念,整体考虑更加综合全面。
(4)所述方法能够同时实现点预测与区间预测。
附图说明
为了更清楚地说明本发明实施例或现有技术中的技术方案,下面将对实施例或现有技术描述中所需要使用的附图做以简单地介绍,显而易见地,下面描述中的附图是本发明的一些实施例,对于本领域普通技术人员来讲,在不付出创造性劳动性的前提下,还可以根据这些附图获得其他的附图。
图1为本发明一种热轧卷取温度的区间预测方法的流程框图。
图2为本发明实施例所述的一种热轧卷取温度的区间预测方法的神经网络的网络结构示意图。
图3为本发明实施例所述的一种热轧卷取温度的区间预测方法的区间预测结果示意图。
图4为本发明实施例所述的一种热轧卷取温度的区间预测方法的点预测结果示意图。
具体实施方式
为了使本技术领域的人员更好地理解本发明方案,下面将结合本发明实施例中的附图,对本发明实施例中的技术方案进行清楚、完整地描述,显然,所描述的实施例仅仅是本发明一部分的实施例,而不是全部的实施例。基于本发明中的实施例,本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例,都应当属于本发明保护的范围。
需要说明的是,本发明的说明书和权利要求书及上述附图中的术语“第一”、“第二”等是用于区别类似的对象,而不必用于描述特定的顺序或先后次序。应该理解这样使用的数据在适当情况下可以互换,以便这里描述的本发明的实施例能够以除了在这里图示或描述的那些以外的顺序实施。此外,术语“包括”和“具有”以及他们的任何变形,意图在于覆盖不排他的包含,例如,包含了一系列步骤或单元的过程、方法、系统、产品或设备不必限于清楚地列出的那些步骤或单元,而是可包括没有清楚地列出的或对于这些过程、方法、产品或设备固有的其它步骤或单元。
如图1-4所示,本发明提供了一种热轧卷取温度的区间预测方法,包括以下步骤:
步骤1:根据轧件原始数据以及精轧出口的实测样本数据可知输入数据和输出数据;其中所述输入数据包括:终轧温度、带钢速度即F6速度、带钢平均厚度以及目标卷取温度;所述输出数据包括:实测卷取温度与目标卷取温度的相对偏差;
步骤2:数据预处理;将所述输入数据和输出数据按一定比例分为训练集、验证集和测试集;并将所有数据进行归一化处理。作为一种优选的实施方式,一般需保证训练集大于验证集和测试集在本申请中采用训练集:验证集:测试集的比例为4:3:3或6:2:2。
步骤3:对人工神经网络进行设置;设置所述人工神经网络的隐藏层层数以及所述隐藏层每一层节点个数;通过sigmoid函数作为激活函数,将学习率设置为1,数据的扫描次数设置为40,平均梯度步长设置为10以进行神经网络训练。所述步骤3分别采用单层人工神经网络、双层人工神经网络和三层人工神经网络进行区间预测;其中,所述单层人工神经网络的隐藏层节点个数为15;所述双层人工神经网络隐藏层的节点个数分别为16和18;所述三层人工神经网络的隐藏层节点个数分别是18、10和20。
步骤4:通过鲸鱼优化算法优化所述人工神经网络,通过最小化代价函数进行寻优,获取人工神经网络最优的权重和偏置量。
所述步骤4首先随机初始化神经网络的权重和偏置,并通过鲸鱼优化算法训练神经网络;具体包括以下步骤:
步骤4.1:初始化算法;初始化时,包括座头鲸数量Xi,当前的迭代次数为t,t=0,最大迭代次数T并对猎物的位置进行随机初始化;
步骤4.2:确定最优搜索代理;目标猎物即搜索空间的最优位置,设置最佳的搜索代理,其他搜索代理将以所述最佳搜索代理的位置更新自身位置;其中:计算每一个搜索代理的适应度函数,
Figure PCTCN2021109201-appb-000028
表示最佳的搜索代理,所述最佳的搜索代理的适应度最小。
所述步骤4.2中适应度函数为包括:覆盖概率和覆盖宽度的综合评价指标;具体地包括以下步骤:
步骤4.2.1:计算区间预测覆盖概率;PICP表示目标值被上限和下限覆盖的概率,定义如下:
Figure PCTCN2021109201-appb-000029
其中,N表示样本总数,∈ i表示布尔变量,表示预测区间的覆盖率;如果目标值yi的数值大于等于下限Li且小于等于上限Ui时,则∈ i=1;如果目标值yi的数值小于下限Li或大于上限Ui时,则∈ i=0;用数学公式表示为:
Figure PCTCN2021109201-appb-000030
步骤4.2.2:计算区间预测平均宽度;区间预测宽度的定量测量被定义为 预测区间归一化平均宽度PINAW,用数学公式表示如下:
Figure PCTCN2021109201-appb-000031
其中,R表示目标函数值的最大值与最小值之差;
步骤4.2.3:计算区间预测均方误差;引入区间预测均方误差PIMSE,通过最小化PIMSE指数,获取更接近真实置信区间的对称区间;用数学公式表示如下:
Figure PCTCN2021109201-appb-000032
步骤4.2.4:重新定义基于覆盖率和宽度的标准;通过基于覆盖率和宽度的标准NCWC作为最终区间预测的评价指标;
NCWC=PINAW+γ(PICP)e -η(PICP-μ)+PIMSE;
其中,对于训练集来说,γ(PICP)=1;μ和η表示两个控制参数;名义置信水平[(1-α)%]为选择μ的参考,其中μ为95%;μ表示预先指定必须满足的区间预测覆盖率;η表示超参数,η放大PICP和μ之间的差异;
如果预先指定的PICP小于μ,当PICP达到95%时,它就是PINAW和PICP之间的平衡;
因此,对于测试样本来说,γ(PICP)是阶跃函数,γ(PICP)由PICP的满意度确定:
Figure PCTCN2021109201-appb-000033
即对于评价区间预测的结果来说,如果PICP不小于指定的μ,γ(PICP)=0,同样PICP的测量值也为0;否则,γ(PICP)=1,相应的处罚通过NCWC计算得出。NCWC越小越好,γ(PICP)=1显著增大了NCWC,故称为处罚,可将处罚改为代价。
步骤4.3:搜索代理位置更新;当当前迭代次数小于最大迭代次数时,对于每一个搜索代理,进行参数的更新:
Figure PCTCN2021109201-appb-000034
l,p,
Figure PCTCN2021109201-appb-000035
其中,
Figure PCTCN2021109201-appb-000036
在迭代过程中从2到0线性递减,
Figure PCTCN2021109201-appb-000037
为0到1的随机数,l为介于[-1,1]之间的随机数,P为[0,1]之间的随机数,
Figure PCTCN2021109201-appb-000038
Figure PCTCN2021109201-appb-000039
是两个向量系数;当p<0.5,
Figure PCTCN2021109201-appb-000040
时按照下式更新当前搜索代理的位置公式:
Figure PCTCN2021109201-appb-000041
Figure PCTCN2021109201-appb-000042
其中,t表示当前迭代次数,
Figure PCTCN2021109201-appb-000043
表示当前最优的位置向量;向量系数
Figure PCTCN2021109201-appb-000044
的计算如下:
Figure PCTCN2021109201-appb-000045
Figure PCTCN2021109201-appb-000046
当p<0.5,
Figure PCTCN2021109201-appb-000047
时随机选择一个搜索代理
Figure PCTCN2021109201-appb-000048
即随机位置向量,按照下式更新当前搜索代理的位置公式:
Figure PCTCN2021109201-appb-000049
Figure PCTCN2021109201-appb-000050
当p>0.5时按照下式更新当前搜索代理的位置公式:
Figure PCTCN2021109201-appb-000051
Figure PCTCN2021109201-appb-000052
其中,b表示定义对数螺线形状的常数;
步骤4.4:更新搜索代理;判断是否有任何搜索代理超出了搜索空间并对其进行修正,计算每个搜索代理的适合度函数,用适应度值最小的搜索代理替换
Figure PCTCN2021109201-appb-000053
步骤4.5:返回人工神经网络的权重和偏置值;当前迭代次数加1,返回步骤4.3直到当前迭代次数等于最大迭代次数;返回最优位置向量
Figure PCTCN2021109201-appb-000054
也就是人工神经网络的权重与偏置。
步骤5:根据所述输入得到实测卷取温度与目标卷取温度的相对偏差的预测上限和预测下限,可同时进行点预测与区间预测;
步骤6:根据所提出的评价指标NCWC进行区间预测性能分析。
实施例1
一种热轧卷取温度的区间预测方法,其预测步骤流程如上述图1所示。 下面结合实例对本发明进一步说明。所述方法选择某热轧厂的层流冷却系统中卷取机的相关数据,所采用的数据共1600组。每组数据包括4个输入数据和1个输出数据。其中输入数据包括:轧温度、带钢速度(F6速度)、带钢平均厚度、目标卷取温度。输出数据是:实测卷取温度与目标卷取温度的相对偏差。
为了更好的进行预测并说明效果,所述方法首先对数据进行预处理,将全部数据归一化为0到1之间的数据。将数据按照一定比例分为训练集、验证集和测试集。对人工神经网络进行参数设置,所述方法采用sigmoid函数作为激活函数,将学习率设置为1,数据的扫描次数设置为40,平均梯度步长设置为10以进行训练。分别使用单层人工神经网络、双层人工神经网络和三层人工神经网络进行区间预测。其中单层ANN的隐藏层节点个数为15。双层ANN隐藏层的节点个数分别为16和18。三层ANN的隐藏层节点个数分别是18、10和20。设置所述方法中采用的鲸鱼优化算法的相关参数,搜索代理数量设为50,最大迭代次数T设置为2000。搜索空间的上下边带分别设置为1和-1。所述评价指标NCWC中的两个控制参数μ和η分布设置为0.95和50。
所述方法使用鲸鱼优化算法优化人工神经网络,以获得人工神经网络最优的权重和偏置量。具体的步骤如下:首先对算法进行初始化,给定参数后计算所有搜索代理的适应度函数值,将适应度函数值最小的搜索代理确定为最优搜索代理。随后根据位置新公式对搜索代理进行位置更新。更新后再次计算每个搜索代理的适应度函数值,选择最小函数值所对应的搜索代理替换掉原最优搜索代理。判断迭代次数是否与最大迭代次数相等,相等则输出最优位置向量,即人工神经网络的权重与偏置量。所述优化算法的代价函数分别涉及到区间预测覆盖率、区间预测平均宽度和区间预测均方误差三个指标。具体公式如下:
NCWC=PINAW+γ(PICP)e -η(PICP-μ)+PIMSE;
这一新评价指标的提出兼顾信息性与有效性,能够加强优化效果,更加全面的进行区间预测方法评估。通过对输入数据的训练,最终所述区间预测方法给出预测上限与预测下限以及点预测输出结果,能够同时实现点预测与 区间预测。
实验结果表明,所述发明提出的热轧卷取温度区间预测方法能够提高预测精度,符合生产应用的实际需求,有助与实现高质量的卷取温度控制。表1为所述实例提出的基于三层ANN的热轧卷取温度区间预测方法与单层人工神经网络、双层人工神经网络的热轧卷取温度区间预测方法的对比。对比结果表明所述专利采用的区间预测方法显著缩小了综合评价指标NCWC值,拥有更好的预测效果。
表1预测性能比较
Figure PCTCN2021109201-appb-000055
上述本发明实施例序号仅仅为了描述,不代表实施例的优劣。
在本发明的上述实施例中,对各个实施例的描述都各有侧重,某个实施例中没有详述的部分,可以参见其他实施例的相关描述。在本申请所提供的几个实施例中,应该理解到,所揭露的技术内容,可通过其它的方式实现。
最后应说明的是:以上各实施例仅用以说明本发明的技术方案,而非对其限制;尽管参照前述各实施例对本发明进行了详细的说明,本领域的普通技术人员应当理解:其依然可以对前述各实施例所记载的技术方案进行修改,或者对其中部分或者全部技术特征进行等同替换;而这些修改或者替换,并不使相应技术方案的本质脱离本发明各实施例技术方案的范围。

Claims (4)

  1. 一种热轧卷取温度的区间预测方法,其特征在于,包括以下步骤:
    步骤1:根据轧件原始数据以及精轧出口的实测样本数据可知输入数据和输出数据;其中所述输入数据包括:终轧温度、带钢速度即F6速度、带钢平均厚度以及目标卷取温度;所述输出数据包括:实测卷取温度与目标卷取温度的相对偏差;
    步骤2:数据预处理;将所述输入数据和输出数据按一定比例分为训练集、验证集和测试集;并将所有数据进行归一化处理;
    步骤3:对人工神经网络进行设置;设置所述人工神经网络的隐藏层层数以及所述隐藏层每一层节点个数;通过sigmoid函数作为激活函数,将学习率设置为1,数据的扫描次数设置为40,平均梯度步长设置为10以进行神经网络训练;
    步骤4:通过鲸鱼优化算法优化所述人工神经网络,进而通过最小化代价函数进行寻优,获取人工神经网络最优的权重和偏置量;
    步骤5:根据所述输入得到实测卷取温度与目标卷取温度的相对偏差的预测上限和预测下限,可同时进行点预测与区间预测;
    步骤6:根据所提出的评价指标NCWC进行区间预测性能分析。
  2. 根据权利要求1所述的一种热轧卷取温度的区间预测方法,其特征在于:所述步骤3分别采用单层人工神经网络、双层人工神经网络和三层人工神经网络进行区间预测;其中,所述单层人工神经网络的隐藏层节点个数为15;所述双层人工神经网络隐藏层的节点个数分别为16和18;所述三层人工神经网络的隐藏层节点个数分别是18、10和20。
  3. 根据权利要求1所述的一种热轧卷取温度的区间预测方法,其特征在于:所述步骤4首先随机初始化神经网络的权重和偏置,并通过鲸鱼优化算法训练神经网络;具体包括以下步骤:
    步骤4.1:初始化算法;初始化时,包括座头鲸数量Xi,当前的迭代次数为t,t=0,最大迭代次数T并对猎物的位置进行随机初始化;
    步骤4.2:确定最优搜索代理;目标猎物即搜索空间的最优位置,设置最佳的搜索代理,其他搜索代理将以所述最佳搜索代理的位置更新自身位置;其中:计算每一个搜索代理的适应度函数,
    Figure PCTCN2021109201-appb-100001
    表示最佳的搜索代理,所述最佳的搜索代理的适应度最小;
    步骤4.3:搜索代理位置更新;当当前迭代次数小于最大迭代次数时,对于每一个搜索代理,进行参数的更新:
    Figure PCTCN2021109201-appb-100002
    l,p,
    Figure PCTCN2021109201-appb-100003
    其中,
    Figure PCTCN2021109201-appb-100004
    在迭代过程中从2到0线性递减,
    Figure PCTCN2021109201-appb-100005
    为0到1的随机数,l为介于[-1,1]之间的随机数,P为[0,1]之间的随机数,
    Figure PCTCN2021109201-appb-100006
    Figure PCTCN2021109201-appb-100007
    是两个向量系数;当p<0.5,
    Figure PCTCN2021109201-appb-100008
    时按照下式更新当前搜索代理的位置公式:
    Figure PCTCN2021109201-appb-100009
    Figure PCTCN2021109201-appb-100010
    其中,t表示当前迭代次数,
    Figure PCTCN2021109201-appb-100011
    表示当前最优的位置向量;向量系数
    Figure PCTCN2021109201-appb-100012
    的计算如下:
    Figure PCTCN2021109201-appb-100013
    Figure PCTCN2021109201-appb-100014
    当p<0.5,
    Figure PCTCN2021109201-appb-100015
    时随机选择一个搜索代理
    Figure PCTCN2021109201-appb-100016
    即随机位置向量,按照下式更新当前搜索代理的位置公式:
    Figure PCTCN2021109201-appb-100017
    Figure PCTCN2021109201-appb-100018
    当p>0.5时按照下式更新当前搜索代理的位置公式:
    Figure PCTCN2021109201-appb-100019
    Figure PCTCN2021109201-appb-100020
    其中,b表示定义对数螺线形状的常数;
    步骤4.4:更新搜索代理;判断是否有任何搜索代理超出了搜索空间并对其进行修正,计算每个搜索代理的适合度函数,用适应度值最小的搜索代理替换
    Figure PCTCN2021109201-appb-100021
    步骤4.5:返回人工神经网络的权重和偏置值;当前迭代次数加1,返回步骤4.3直到当前迭代次数等于最大迭代次数;返回最优位置向量
    Figure PCTCN2021109201-appb-100022
    也就是人工神经网络的权重与偏置。
  4. 根据权利要求1所述的一种热轧卷取温度的区间预测方法,其特征在于:所述步骤4.2中适应度函数为包括:覆盖概率和覆盖宽度的综合评价指标;具体地包括以下步骤:
    步骤4.2.1:计算区间预测覆盖概率;PICP表示目标值被上限和下限覆盖的概率,定义如下:
    Figure PCTCN2021109201-appb-100023
    其中,N表示样本总数,ε i表示布尔变量,表示预测区间的覆盖率;如果目标值yi的数值大于等于下限Li且小于等于上限Ui时,则ε i=1;如果目标值yi的数值小于下限Li或大于上限Ui时,则ε i=0;用数学公式表示为:
    Figure PCTCN2021109201-appb-100024
    步骤4.2.2:计算区间预测平均宽度;区间预测宽度的定量测量被定义为预测区间归一化平均宽度PINAW,用数学公式表示如下:
    Figure PCTCN2021109201-appb-100025
    其中,R表示目标函数值的最大值与最小值之差;
    步骤4.2.3:计算区间预测均方误差;引入区间预测均方误差PIMSE,通过最小化PIMSE指数,获取更接近真实置信区间的对称区间;用数学公式表示如下:
    Figure PCTCN2021109201-appb-100026
    步骤4.2.4:重新定义基于覆盖率和宽度的标准;通过基于覆盖率和宽度的标准NCWC作为最终区间预测的评价指标;
    NCWC=PINAW+γ(PICP)e -η(PICP-μ)+PIMSE;
    其中,对于训练集来说,γ(PICP)=1;μ和η表示两个控制参数;名义置信水平[(1-α)%]为选择μ的参考,其中μ为95%;μ表示预先指定必 须满足的区间预测覆盖率;η表示超参数,η放大PICP和μ之间的差异;
    如果预先指定的PICP小于μ,当PICP达到95%时,它就是PINAW和PICP之间的平衡;
    因此,对于测试样本来说,γ(PICP)是阶跃函数,γ(PICP)由PICP的满意度确定:
    Figure PCTCN2021109201-appb-100027
    即对于评价区间预测的结果来说,如果PICP不小于指定的μ,γ(PICP)=0,同样PICP的测量值也为0;否则,γ(PICP)=1,相应的处罚通过NCWC计算得出。
PCT/CN2021/109201 2020-11-20 2021-07-29 一种热轧卷取温度的区间预测方法 WO2022105294A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
JP2022530319A JP7393831B2 (ja) 2020-11-20 2021-07-29 熱間圧延巻取り温度の区間予測法

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202011311991.7A CN112507610B (zh) 2020-11-20 2020-11-20 一种热轧卷取温度的区间预测方法
CN202011311991.7 2020-11-20

Publications (1)

Publication Number Publication Date
WO2022105294A1 true WO2022105294A1 (zh) 2022-05-27

Family

ID=74959168

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/109201 WO2022105294A1 (zh) 2020-11-20 2021-07-29 一种热轧卷取温度的区间预测方法

Country Status (3)

Country Link
JP (1) JP7393831B2 (zh)
CN (1) CN112507610B (zh)
WO (1) WO2022105294A1 (zh)

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115453868A (zh) * 2022-08-31 2022-12-09 中国农业大学 基于番茄光响应差分特征的全生长期光强调控方法
CN115458167A (zh) * 2022-09-13 2022-12-09 郑州市中心医院 基于正余弦机制改进飞蛾扑火算法的肺癌预测方法
CN115608793A (zh) * 2022-12-20 2023-01-17 太原科技大学 一种机理融合数据的精轧温度调控方法
CN116341372A (zh) * 2023-03-02 2023-06-27 中国石油大学(华东) 一种基于人工神经网络的换热器性能预测及优化方法
CN117066301A (zh) * 2023-10-13 2023-11-17 江苏甬金金属科技有限公司 结合生产质量监测的钛带卷加工调控方法及系统
CN117309195A (zh) * 2023-11-30 2023-12-29 中控技术股份有限公司 测温仪表的检测方法、装置及电子设备
CN117724433A (zh) * 2024-02-07 2024-03-19 东北大学 基于多通道分布式深度集成预测的冷轧生产前馈控制方法

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112507610B (zh) * 2020-11-20 2023-09-26 东北大学 一种热轧卷取温度的区间预测方法

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR100643373B1 (ko) * 2005-12-21 2006-11-10 주식회사 포스코 열간압연 후물재 길이방향 온도 제어방법
CN101428301A (zh) * 2007-11-09 2009-05-13 株式会社日立制作所 卷取温度控制装置及控制方法
CN103464475A (zh) * 2013-09-06 2013-12-25 鞍钢股份有限公司 一种基于关联神经网络的热轧卷取温度预报方法
CN110110930A (zh) * 2019-05-08 2019-08-09 西南交通大学 一种改进鲸鱼算法的循环神经网络短期电力负荷预测方法
CN112507610A (zh) * 2020-11-20 2021-03-16 东北大学 一种热轧卷取温度的区间预测方法

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2004331992A (ja) * 2003-04-30 2004-11-25 Jfe Steel Kk 熱間圧延における金属板の温度予測方法および冷却方法
JP2007301603A (ja) 2006-05-11 2007-11-22 Kobe Steel Ltd 圧延材の巻き取り温度制御方法及び圧延装置
JP6642289B2 (ja) 2016-06-14 2020-02-05 東芝三菱電機産業システム株式会社 圧延ラインの数学モデル算出装置および圧延材の温度制御装置
CN108205698A (zh) * 2017-06-15 2018-06-26 广东工业大学 一种基于正余混沌双弦鲸鱼优化算法的云资源负载预测方法
CN108022025B (zh) * 2017-12-28 2020-08-18 华中科技大学 一种基于人工神经网络的风速区间预测方法与系统
CN111062533A (zh) * 2019-12-16 2020-04-24 国家能源集团谏壁发电厂 一种基于鲸鱼优化算法优化加权最小二乘支持向量机的风机故障预测方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR100643373B1 (ko) * 2005-12-21 2006-11-10 주식회사 포스코 열간압연 후물재 길이방향 온도 제어방법
CN101428301A (zh) * 2007-11-09 2009-05-13 株式会社日立制作所 卷取温度控制装置及控制方法
CN103464475A (zh) * 2013-09-06 2013-12-25 鞍钢股份有限公司 一种基于关联神经网络的热轧卷取温度预报方法
CN110110930A (zh) * 2019-05-08 2019-08-09 西南交通大学 一种改进鲸鱼算法的循环神经网络短期电力负荷预测方法
CN112507610A (zh) * 2020-11-20 2021-03-16 东北大学 一种热轧卷取温度的区间预测方法

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115453868B (zh) * 2022-08-31 2024-04-12 中国农业大学 基于番茄光响应差分特征的全生长期光强调控方法
CN115453868A (zh) * 2022-08-31 2022-12-09 中国农业大学 基于番茄光响应差分特征的全生长期光强调控方法
CN115458167A (zh) * 2022-09-13 2022-12-09 郑州市中心医院 基于正余弦机制改进飞蛾扑火算法的肺癌预测方法
CN115608793A (zh) * 2022-12-20 2023-01-17 太原科技大学 一种机理融合数据的精轧温度调控方法
CN115608793B (zh) * 2022-12-20 2023-04-07 太原科技大学 一种机理融合数据的精轧温度调控方法
CN116341372B (zh) * 2023-03-02 2024-01-23 中国石油大学(华东) 一种基于人工神经网络的换热器性能预测及优化方法
CN116341372A (zh) * 2023-03-02 2023-06-27 中国石油大学(华东) 一种基于人工神经网络的换热器性能预测及优化方法
CN117066301B (zh) * 2023-10-13 2023-12-22 江苏甬金金属科技有限公司 结合生产质量监测的钛带卷加工调控方法及系统
CN117066301A (zh) * 2023-10-13 2023-11-17 江苏甬金金属科技有限公司 结合生产质量监测的钛带卷加工调控方法及系统
CN117309195A (zh) * 2023-11-30 2023-12-29 中控技术股份有限公司 测温仪表的检测方法、装置及电子设备
CN117309195B (zh) * 2023-11-30 2024-04-12 中控技术股份有限公司 测温仪表的检测方法、装置及电子设备
CN117724433A (zh) * 2024-02-07 2024-03-19 东北大学 基于多通道分布式深度集成预测的冷轧生产前馈控制方法
CN117724433B (zh) * 2024-02-07 2024-04-19 东北大学 基于多通道分布式深度集成预测的冷轧生产前馈控制方法

Also Published As

Publication number Publication date
JP7393831B2 (ja) 2023-12-07
CN112507610B (zh) 2023-09-26
CN112507610A (zh) 2021-03-16
JP2023505958A (ja) 2023-02-14

Similar Documents

Publication Publication Date Title
WO2022105294A1 (zh) 一种热轧卷取温度的区间预测方法
CN111222698B (zh) 面向物联网的基于长短时记忆网络的积水水位预测方法
CN109858647B (zh) 一种耦合gis和gbdt算法的区域洪涝灾害风险评价与预估方法
Zhang et al. An integrated model of rolling force for extra-thick plate by combining theoretical model and neural network model
Celikoglu Application of radial basis function and generalized regression neural networks in non-linear utility function specification for travel mode choice modelling
Lu et al. Generalized radial basis function neural network based on an improved dynamic particle swarm optimization and AdaBoost algorithm
CN111461453B (zh) 一种基于多模型组合的中长期径流集合预报方法
CN103489039B (zh) 具有在线自整定优化能力的高速公路交通流量融合预测方法
El Hamzaoui et al. Application of artificial neural networks to predict the selling price in the real estate valuation process
CN109409614A (zh) 一种基于贝叶斯正则化神经网络的电力负荷预测方法
CN108280998A (zh) 基于历史数据动态选择的短时交通流预测方法
CN112085254A (zh) 基于多重分形协同度量门控循环单元的预测方法及模型
CN108229750A (zh) 一种股票收益率预测方法
CN116689503A (zh) 一种基于记忆性功能网络的带钢全长厚度预测方法
CN110264007A (zh) 股票交易控制方法及装置
Hansun et al. Brown’s Weighted Exponential Moving Average Implementation in Forex Forecasting
Purnawansyah et al. An inflation rate prediction based on backpropagation neural network algorithm
CN109493921A (zh) 一种基于多代理模型的常压精馏过程建模方法
CN116689502A (zh) 基于sa-scso-scn神经网络的热连轧厂精轧机组工艺的弯辊力参数设定方法
Bashir et al. Short-term load forecasting using artificial neural network based on particle swarm optimization algorithm
Wang et al. Rolling forecast nature gas spot price with back propagation neural network
CN116189794A (zh) 一种夯土水盐含量测量方法
Purbasari et al. Time-series Modeling for Consumer Price Index Forecasting using Comparison Analysis of Autoregressive Integrated Moving Average and Artificial Neural Network
CN115099469A (zh) 基于优选气候因子和精度权重系数的中长期径流预测方法
CN108053266A (zh) 一种专利价值预估方法以及装置

Legal Events

Date Code Title Description
ENP Entry into the national phase

Ref document number: 2022530319

Country of ref document: JP

Kind code of ref document: A

121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21893446

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21893446

Country of ref document: EP

Kind code of ref document: A1