WO2022121944A1 - 工业过程运行指标智能预报方法、装置、设备及存储介质 - Google Patents

工业过程运行指标智能预报方法、装置、设备及存储介质 Download PDF

Info

Publication number
WO2022121944A1
WO2022121944A1 PCT/CN2021/136453 CN2021136453W WO2022121944A1 WO 2022121944 A1 WO2022121944 A1 WO 2022121944A1 CN 2021136453 W CN2021136453 W CN 2021136453W WO 2022121944 A1 WO2022121944 A1 WO 2022121944A1
Authority
WO
WIPO (PCT)
Prior art keywords
model
deep learning
online
dynamic
prediction model
Prior art date
Application number
PCT/CN2021/136453
Other languages
English (en)
French (fr)
Inventor
柴天佑
张菁雯
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
Publication of WO2022121944A1 publication Critical patent/WO2022121944A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B19/00—Programme-control systems
    • G05B19/02—Programme-control systems electric
    • G05B19/418—Total factory control, i.e. centrally controlling a plurality of machines, e.g. direct or distributed numerical control [DNC], flexible manufacturing systems [FMS], integrated manufacturing systems [IMS] or computer integrated manufacturing [CIM]
    • G05B19/41885—Total factory control, i.e. centrally controlling a plurality of machines, e.g. direct or distributed numerical control [DNC], flexible manufacturing systems [FMS], integrated manufacturing systems [IMS] or computer integrated manufacturing [CIM] characterised by modeling, simulation of the manufacturing system
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B2219/00—Program-control systems
    • G05B2219/30—Nc systems
    • G05B2219/32—Operator till task planning
    • G05B2219/32339—Object oriented modeling, design, analysis, implementation, simulation language
    • Y—GENERAL 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
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
    • Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
    • Y02P90/02—Total factory control, e.g. smart factories, flexible manufacturing systems [FMS] or integrated manufacturing systems [IMS]

Definitions

  • the invention belongs to the technical field of industrial artificial intelligence, and in particular relates to an intelligent forecasting method, device, equipment and storage medium of industrial process operation indicators.
  • the accurate forecast of the operational indicators that characterize the product quality, efficiency, and consumption of industrial process processing is crucial to realizing the operational optimization control of the industrial process. Due to the strong nonlinearity between the operation index and the input and output of the industrial process control system that affects the operation index, multivariable strong coupling, changes in operating conditions, fluctuations in raw materials, etc., the mechanism is unclear, and the dynamic model cannot be established.
  • Existing system identification methods and forecasting methods based on mechanism models establish forecasting models for operational indicators. Due to the dynamic change of the industrial process in the production process, the operation indicators, the input and output data of the industrial process control system are in a changing, open and uncertain information space, which cannot be established by the deep learning technology of the existing complete information space.
  • a forecast model for operating indicators Due to the dynamic change of the industrial process in the production process, the operation indicators, the input and output data of the industrial process control system are in a changing, open and uncertain information space, which cannot be established by the deep learning technology of the existing complete information space.
  • the present invention aims to solve one of the technical problems in the related art at least to a certain extent.
  • the technical scheme of the present invention is as follows:
  • An intelligent forecasting method for industrial process operation indicators comprising the following steps:
  • the dynamic model of operating indicators includes two parts: an identifiable model and an unmodeled dynamic;
  • the predicted value of the operation index is obtained from the output of the identifiable model in the dynamic model of the operation index and the output of the online intelligent prediction model of the unknown nonlinear dynamic system.
  • the online intelligent forecasting model includes an online deep learning forecasting model, a deep learning correction model and a self-correction mechanism; the online deep learning forecasting model is established by using the LSTM architecture; the same as the online deep learning forecasting model is adopted. structure to establish the deep learning correction model; when the error between the output of the online deep learning prediction model and the label data is greater than a set threshold, a self-correction mechanism is adopted, and the weight and offset of the deep learning correction model are used to correct the Weights and biases of the online deep learning forecasting model; wherein, the historical data used by the deep learning correction model is more than the historical data used by the online deep learning forecasting model.
  • both the online deep learning prediction model and the deep learning correction model include an input layer, a hidden layer, a fully connected layer and an output layer, wherein the number of hidden layers is L, and L is greater than or equal to 1 positive integers; fix the weight and bias of the hidden layer in the online deep learning forecast model, and correct the weight and bias of the fully connected layer in the online deep learning forecast model; online training The weights and biases of the hidden layer and the fully connected layer in the deep learning correction model; when the error between the output of the online deep learning prediction model and the label data is greater than the set threshold, a self-correction mechanism is used to use The weight and offset of the hidden layer of the deep learning correction model replace the weight and offset of the hidden layer of the online deep learning prediction model, and the weight of the fully connected layer of the deep learning correction model is used Replace the weights and biases of the fully connected layers of the online deep learning prediction model with biases.
  • the industrial process is the operation process of the fused magnesium group furnace, and the operation index is the power of the fused magnesium group furnace.
  • An industrial process operation index intelligent forecasting device comprising:
  • an operation index dynamic model modeling module which is used to establish an operation index dynamic model by utilizing the characteristics of the industrial process control system, and the operation index dynamic model includes two parts: an identifiable model and an unmodeled dynamic model;
  • a parameter identification module for estimating the parameters of the identifiable model in the dynamic model of the operating index
  • a nonlinear dynamic acquisition module configured to combine the identification error of the parameters of the identifiable model in the operating index dynamic model and the unmodeled dynamic in the operating index dynamic model into an unknown nonlinear dynamic system
  • an online intelligent prediction model modeling module used for establishing an online intelligent prediction model of the unknown nonlinear dynamic system by using adaptive deep learning
  • the forecasting module is configured to obtain the forecast value of the operation index from the output of the identifiable model in the dynamic model of the operation index and the output of the online intelligent forecasting model of the unknown nonlinear dynamic system.
  • the online intelligent prediction model modeling module includes a prediction model modeling module, a calibration model modeling module and a self-calibration module; the prediction model modeling module adopts an LSTM architecture to establish an online deep learning prediction model; the The correction model modeling module adopts the same structure as the online deep learning forecast model to establish a deep learning correction model; the self-correction module, when the error between the output of the online deep learning forecast model and the label data is greater than the set threshold, A self-correction mechanism is adopted to correct the weights and biases of the online deep learning prediction model with the weights and biases of the deep learning correction model; wherein, the historical data used by the deep learning correction model is higher than that of the online deep learning forecasting model.
  • the model uses a lot of historical data.
  • both the online deep learning prediction model and the deep learning correction model include an input layer, a hidden layer, a fully connected layer and an output layer, wherein the number of hidden layers is L, and L is greater than or equal to 1 positive integer; the forecast model modeling module fixes the weight and bias of the hidden layer in the online deep learning forecast model, and online corrects the weight of the fully connected layer in the online deep learning forecast model and bias; the correction model modeling module trains the weights and biases of the hidden layer and the fully connected layer in the deep learning correction model online; the self-correction module predicts in the online deep learning When the error between the output of the model and the label data is greater than the set threshold, a self-correction mechanism is used to replace the weight and bias of the hidden layer of the online deep learning prediction model with the weight and bias of the hidden layer of the deep learning correction model. Weight and bias, replacing the weight and bias of the fully connected layer of the online deep learning prediction model with the weight and bias of the fully connected layer of the deep learning correction model.
  • the industrial process is the operation process of the fused magnesium group furnace, and the operation index is the power of the fused magnesium group furnace.
  • An industrial process operation index intelligent forecasting device for realizing the above-mentioned industrial process operation index intelligent forecasting method, comprising: terminal-side sub-device, edge-side sub-device and cloud-side sub-device;
  • the end-side sub-equipment is used to collect input data and output data in the industrial process
  • the edge side sub-device uses the online deep learning prediction model to perform online prediction of the operating index
  • the cloud-side sub-device is used to train the deep learning correction model and implement the self-correction mechanism.
  • a computer-readable storage medium storing a computer program, when the program is executed by a processor, realizes the above-mentioned intelligent forecasting method for industrial process operation indicators.
  • the invention combines the system identification method based on the mechanism model with the deep learning method based on big data, and uses the characteristic that the change of the operation index depends on the dynamic characteristics of the industrial process control system, and proposes a
  • the intelligent forecasting method of industrial process operation indexes solves the problem of forecasting industrial process operation indexes.
  • Fig. 1 is the realization flow chart of the industrial process operation index intelligent forecast method of the embodiment of the present invention
  • Fig. 2 is the concrete realization flow chart of step S1 shown in Fig. 1;
  • Fig. 3 is the realization flow chart of the method for intelligently predicting power and demand of fused magnesium group furnaces according to an embodiment of the present invention
  • Fig. 4 is the concrete realization flow chart of step S1' shown in Fig. 3;
  • FIG. 5 is a structural diagram of an LSTM network according to an embodiment of the present invention.
  • Fig. 6 is the demand forecast result graph of one embodiment of the present invention.
  • FIG. 7 is a schematic structural diagram of an industrial process operation index intelligent forecasting device according to an embodiment of the present invention.
  • FIG. 8 is a schematic structural diagram of an industrial process operation index intelligent forecasting device according to an embodiment of the present invention.
  • Fig. 1 is the realization flow chart of the industrial process operation index intelligent forecast method of the embodiment of the present invention, the method comprises the following steps:
  • S1 Use the characteristics of the industrial process control system to establish a dynamic model of operating indicators, where the dynamic model of operating indicators includes two parts: an identifiable model and an unmodeled dynamic.
  • step S1 includes steps S11 and S12.
  • Step S11 is: establishing a closed-loop control system dynamic model of the industrial process; specifically, using the feature that the industrial process runs near the operating point, the industrial process is represented by a linear low-order model and an unknown high-order nonlinear term, and a PID control technology is used.
  • a dynamic model of an industrial process closed-loop control system consisting of a linear model and an unmodeled dynamic with unknown high-order nonlinear terms is established.
  • Step S12 is: establishing a dynamic model between the operation index and the input and output data of the industrial process closed-loop control system; specifically, using the dynamic model of the industrial process closed-loop control system, and using the industrial process closed-loop control system to determine the operation index of the industrial process
  • the characteristics of control within a certain range are described by the identifiable model and the unmodeled dynamic model of the operating index dynamic model.
  • the parameters of the identifiable model in the dynamic model of the operation index are estimated by the identification algorithm using the operation index and the input and output data of the industrial process control system.
  • S3 Combine the identification error of the parameters of the identifiable model in the dynamic model of operating indicators and the unmodeled dynamic in the dynamic model of operating indicators into an unknown nonlinear dynamic system.
  • the identification error of the parameters of the identifiable model in the dynamic model of operating indicators and the unmodeled dynamics of the operating indicators are combined into a nonlinear dynamic system with unknown model structure and order, and the output is Its input is the output y(k),...,y(kn) of the industrial process closed-loop control system and the input u(k-1),...,u(kn), and the output of the unknown nonlinear dynamic system is
  • the unknown constant n is used to represent the order of the dynamic system
  • the following formula is used to represent the unknown nonlinear dynamic system of the operating index:
  • f( ) is a nonlinear function of unknown variation, is the output of the unknown nonlinear dynamic system at (ki) time
  • r(k-i+1) is the running index at (k-i+1) time, is the identifiable model output of the operating index at (ki) time.
  • S4 Use adaptive deep learning to establish an online intelligent prediction model of the unknown nonlinear dynamic system.
  • the online intelligent prediction model consists of an online deep learning prediction model, a deep learning correction model and a self-correction mechanism.
  • the input variable in equation (1) is selected as the input of a single neuron, and the order n is used as the number of neurons.
  • the input and output data of formula (1) are used to form large data samples, and a training algorithm is used to make the label data Output with online deep learning forecasting models forecast error of As small as possible, determine the number of neurons n, the number of unit nodes of LSTM
  • the number of layers L of the multi-layer neural network, the connection weight parameters and bias parameters of each layer, the time series window length of the input data of the online deep learning forecast model is N, and the online deep learning forecast model of the unknown nonlinear dynamic system is established.
  • the input data of the sequence window length N online corrects the connection weight parameters and bias parameters of the fully connected layer of the model.
  • a deep learning correction model is established, and the input data of formula (1) at the current moment and all previous times are used as the input data of the deep learning correction model, and each layer of the deep learning correction model is trained.
  • the self-correction mechanism sets the upper bound of the forecast error interval as ⁇ .
  • the connection weight parameters and bias parameters of each layer of the deep learning correction model are used to replace the connection weight parameters and bias parameters of the corresponding layers of the online deep learning prediction model.
  • the intelligent forecasting method for industrial process operation indexes can be used for intelligent forecasting of the operation indexes—demands of electric fused magnesia group furnaces in an electric fused magnesia plant.
  • the fused magnesia furnace is a large-scale energy-consuming equipment, and its product, fused magnesia, is an important refractory material widely used in metallurgy, chemical industry, aerospace and other fields.
  • the fused magnesia furnace uses the submerged arc method to add magnesite ore while melting, and uses the PID current control system to control the melting current to melt the magnesite ore to produce fused magnesia. Since the fused magnesia furnace is a major energy-consuming equipment, its demand for Monitoring and forecasting are of great significance to energy conservation.
  • Group furnace demand at time k is the average value of the group furnace power at time k and the previous m-1 time, namely
  • p(k) is the group furnace power at time k
  • Fig. 3 is the realization flow chart of the method for intelligent prediction of power and demand of fused magnesium group furnaces according to an embodiment of the present invention, and the method comprises the following steps:
  • S1' Use the characteristics of the industrial process control system to establish a dynamic model of the group furnace power, and the dynamic model of the group furnace power includes two parts: an identifiable model and an unmodeled dynamic part.
  • step S1' includes steps S11' and S12'.
  • Step S11' is: establishing a dynamic model of the melting current closed-loop control system of the i-th fused magnesia furnace.
  • a i (z -1 ) 1+a i1 z -1 , a i1 is a variable constant, z -1 is a backward shift operator; y i (k) is the i-th fused magnesia furnace at time k Melting current; b i0 is a variable constant; u i (k-1) is the rotation direction and frequency of the variable frequency motor of the i-th electric fused magnesia furnace at time (k-1), its positive and negative indicate the direction, and the size indicates the frequency; v i ( ⁇ ) is the unknown higher-order nonlinear term.
  • the PID controller is
  • Step S12' is: establishing a dynamic model between the power of the group furnace and the input and output data of the fused magnesia furnace melting current closed-loop control system.
  • the power of the i-th fused magnesium furnace at time k is
  • the power of m electric fused magnesia furnaces at time k is:
  • the least squares estimation algorithm can be used to obtain
  • S3' Combine the identification error of the parameters of the identifiable model in the dynamic model of the group furnace power and the unmodeled dynamics in the dynamic model of the group furnace power into an unknown nonlinear dynamic system.
  • formula (10) can be expressed as:
  • the operating index r(k) p(k)
  • the output of the identifiable model of the operating index is f( ⁇ ) is a nonlinear function of unknown variation
  • n is the unknown order of the dynamic system.
  • S4' Adopt adaptive deep learning to establish an online intelligent prediction model of the unknown nonlinear dynamic system.
  • the online intelligent prediction model consists of an online deep learning prediction model, a deep learning correction model and a self-correction mechanism.
  • the establishment method of the online deep learning forecast model is as follows:
  • the online deep learning forecasting model of is selected to be 1, and the input of the jth neuron at (k-1) time is h(k+jn-2) is the output of the (j-1)th neuron, the number of neurons is n, and the number of nodes of a single neuron of LSTM is The length of the input data time series window is N, and the M input and output data in formula (15) are used Constitute a large data sample, and use the following training algorithm to make the labeled data Output with online deep learning forecasting models error as small as possible. determine n, and N, the objective function of the training algorithm is:
  • connection weight of the fully connected layer for the row vector of is the bias of the fully connected layer
  • h(k-1) is the output of the nth neuron
  • W f , W in , W c are The matrix of , b f , b in , b c is column vector of .
  • Gradient descent algorithm is used to train network weights W o , W f , W in , W c and biases b o , b f , bin , b c .
  • ⁇ is the learning rate.
  • the same training algorithm is used to determine other connection weights and biases.
  • n 1, 2, 3, 4, ... and calculate the forecast error
  • the root mean squared error (Root Mean Squared Error, RMSE)
  • RMSE Root Mean Squared Error
  • the RMSE is the smallest, so the number of network layers L is 3, and the weights and biases of the fully connected layers are determined by the gradient descent method.
  • the online deep learning prediction model shown in Figure 5 is established.
  • the online deep learning prediction model of is:
  • connection weight of the fully connected layer in the formula for row vector is the bias of the fully connected layer
  • h 3 (k) is the output of the 30th neuron in the third layer.
  • the connection weights and biases of the first, second and third layers are fixed, and only the weights of the fully connected layers in Eq. (26) are corrected online. and bias
  • the updated dataset is for
  • the establishment method of the deep learning correction model is as follows:
  • the data set updated with all the data in the past time is: All weights and biases in online training of LSTM networks.
  • the self-correction mechanism is as follows:
  • the upper bound of the interval is ⁇ , at time k+1, if Then, the weights and biases of each layer of the deep learning correction model are used to correct the weights and biases of the corresponding layers of the online deep learning prediction model.
  • S5' Obtain the predicted value of the group furnace power from the output of the identifiable model in the dynamic model of the group furnace power and the output of the online intelligent prediction model of the unknown nonlinear dynamic system.
  • the prediction model of the group furnace power p(k+1) is:
  • S6' obtain the predicted value of the demand of the group furnace from the predicted value of the group furnace power.
  • the forecast model for is:
  • the forecasting accuracy of demand is 99.96%
  • the forecasting accuracy of demand rising trend is 96.46%
  • the forecasting accuracy of demand falling trend is 92.78%.
  • the accuracy requirements of energy-saving control for demand forecasting are met.
  • an intelligent forecasting device for industrial process operation indicators including: a dynamic model modeling module for operation indicators, a parameter identification module, a nonlinear dynamic acquisition module, and an online intelligent forecasting model modeling module. modules and forecast modules, where:
  • the operation index dynamic model modeling module is used to establish an operation index dynamic model by utilizing the characteristics of the industrial process control system, and the operation index dynamic model includes two parts: an identifiable model and an unmodeled dynamic model;
  • the parameter identification module is used for estimating the parameters of the identifiable model in the dynamic model of the operating index
  • the nonlinear dynamic acquisition module is configured to combine the identification error of the parameters of the identifiable model in the operating index dynamic model and the unmodeled dynamic in the operating index dynamic model into an unknown nonlinear dynamic system;
  • the online intelligent prediction model modeling module is used to establish an online intelligent prediction model of the unknown nonlinear dynamic system by using adaptive deep learning
  • the forecast module is configured to obtain the forecast value of the operation index from the output of the identifiable model in the dynamic model of the operation index and the output of the online intelligent prediction model of the unknown nonlinear dynamic system.
  • the online intelligent prediction model modeling module includes a prediction model modeling module, a calibration model modeling module and a self-calibration module.
  • the prediction model modeling module adopts the LSTM architecture to establish an online deep learning prediction model;
  • the correction model modeling module adopts the same structure as the online deep learning prediction model to establish a deep learning calibration model;
  • the self-calibration module in the When the error between the output of the online deep learning prediction model and the label data is greater than the set threshold, a self-correction mechanism is adopted to correct the weight and bias of the online deep learning prediction model with the weight and bias of the deep learning correction model;
  • the historical data used by the deep learning correction model is more than the historical data used by the online deep learning prediction model.
  • both the online deep learning prediction model and the deep learning correction model include an input layer, a hidden layer, a fully connected layer and an output layer, wherein the number of hidden layers is L, and L is greater than or A positive integer equal to 1; the prediction model modeling module fixes the weight and bias of the hidden layer in the online deep learning prediction model, and online corrects the fully connected layer in the online deep learning prediction model
  • the weights and biases of the correction model modeling module train the weights and biases of the hidden layer and the fully connected layer in the deep learning correction model online;
  • the self-correction module in the online depth
  • a self-correction mechanism is used to replace the hidden layer of the online deep learning prediction model with the weight and bias of the hidden layer of the deep learning correction model.
  • the weights and biases of the layers are replaced by the weights and biases of the fully connected layers of the deep learning correction model to replace the weights and biases of the fully connected layers of the online
  • the industrial process operation index intelligent forecasting device is used for power forecasting of fused magnesium group furnaces.
  • Each module in the above-mentioned intelligent forecasting device for industrial process operation indicators can be implemented in whole or in part by software, hardware and combinations thereof.
  • the above modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.
  • an industrial process operation index intelligent prediction device for implementing the industrial process operation index intelligent prediction method in the above embodiments, including: terminal side sub-device, edge side Sub-devices and cloud-side sub-devices; the terminal-side sub-devices are used to collect input data and output data in the industrial process; the edge-side sub-devices use the online deep learning forecast model to perform online analysis of the operating indicators Forecast; the cloud side sub-device is used to train the deep learning correction model and implement the self-correction mechanism.
  • a computer-readable storage medium which stores a computer program, and when the program is executed by a processor, implements the methods for intelligently predicting industrial process operation indicators in the foregoing embodiments.
  • the method, device and device for intelligent forecasting of industrial process operation indicators proposed in the embodiments of the present invention cannot be used for the existing model-based forecasting methods and deep learning methods to characterize the operation of product quality, efficiency and consumption of industrial process processing.
  • the problem of predicting the indicators is to represent the dynamic model between the operation indicators and the input and output of the process control system with identifiable models and unmodeled dynamics, using the operation indicators, the input and output data of the industrial process control system, using The identification algorithm estimates the parameters of the identifiable model, and uses the identification error and unmodeled dynamics to form a nonlinear dynamic system with unknown model structure and order.
  • an online intelligent prediction model for the dynamic system is established.
  • the online intelligent prediction model of nonlinear dynamic system realizes the accurate prediction of industrial process operation indicators.

Landscapes

  • Engineering & Computer Science (AREA)
  • Manufacturing & Machinery (AREA)
  • General Engineering & Computer Science (AREA)
  • Quality & Reliability (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Automation & Control Theory (AREA)
  • Feedback Control In General (AREA)

Abstract

提供了一种工业过程运行指标智能预报方法、装置、设备及存储介质。工业过程运行指标智能预报方法包括：利用运行指标的变化取决于工业过程控制系统的动态特性的特点，建立运行指标动态模型，运行指标动态模型包括可辨识模型和未建模动态两部分（S1）；估计运行指标动态模型中的可辨识模型的参数（S2）；将运行指标动态模型中的可辨识模型的参数的辨识误差与运行指标动态模型中的未建模动态合并为未知非线性动态系统（S3）；建立未知非线性动态系统的在线智能预报模型（S4）；由运行指标动态模型中的可辨识模型的输出与未知非线性动态系统的在线智能预报模型的输出得到运行指标的预报值（S5）。针对工业过程运行指标难以预报的难题，将基于机理模型的系统辨识方法与基于大数据的深度学习方法相结合，提出了工业过程运行指标的智能预报方法，解决了工业过程运行指标的预报难题。

Description

工业过程运行指标智能预报方法、装置、设备及存储介质 技术领域
本发明属于工业人工智能技术领域，尤其涉及一种工业过程运行指标智能预报方法、装置、设备及存储介质。
背景技术
表征工业过程加工的产品质量、效率、消耗的运行指标的准确预报对实现该工业过程的运行优化控制至关重要。由于运行指标与影响运行指标的工业过程控制系统输入与输出之间具有强非线性、多变量强耦合、运行工况变化、原料波动等造成机理不清难以建立动态模型等综合复杂性，无法采用已有的基于机理模型的系统辨识方法和预报方法建立运行指标的预报模型。由于生产过程中工业过程处于动态变化中，导致运行指标、工业过程控制系统的输入与输出数据处于变化的、开放的、不确定的信息空间，无法采用已有的完备信息空间的深度学习技术建立运行指标的预报模型。
发明内容
本发明旨在至少在一定程度上解决相关技术中的技术问题之一。本发明的技术方案如下：
一种工业过程运行指标智能预报方法，包括如下步骤：
利用工业过程控制系统特性建立运行指标动态模型，所述运行指标动态模型包括可辨识模型和未建模动态两部分；
估计所述运行指标动态模型中的所述可辨识模型的参数；
将所述运行指标动态模型中的所述可辨识模型的参数的辨识误差与所述运行指标动态模型中的所述未建模动态合并为未知非线性动态系统；
采用自适应深度学习建立所述未知非线性动态系统的在线智能预报模型；
由所述运行指标动态模型中的所述可辨识模型的输出与所述未知非线性动态系统的所述在线智能预报模型的输出得到所述运行指标的预报值。
进一步，作为优选，所述在线智能预报模型包括在线深度学习预报模型、深度学习校正模型和自校正机制；采用LSTM架构建立所述在线深度学习预报模型；采用与所述在线深度学习预报模型相同的结构建立所述深度学习校正模型； 当所述在线深度学习预报模型的输出与标签数据的误差大于设定阈值时，采用自校正机制，用所述深度学习校正模型的权重与偏置校正所述在线深度学习预报模型的权重与偏置；其中，所述深度学习校正模型所用的历史数据比所述在线深度学习预报模型所用的历史数据多。
进一步，作为优选，所述在线深度学习预报模型和所述深度学习校正模型均包括输入层、隐藏层、全连接层和输出层，其中，隐藏层的层数为L，L为大于或等于1的正整数；固定所述在线深度学习预报模型中的所述隐藏层的权重与偏置，并在线校正所述在线深度学习预报模型中的所述全连接层的权重与偏置；在线训练所述深度学习校正模型中的所述隐藏层和所述全连接层的权重与偏置；当所述在线深度学习预报模型的输出与标签数据的误差大于设定阈值时，采用自校正机制，用所述深度学习校正模型的所述隐藏层的权重与偏置替换所述在线深度学习预报模型的所述隐藏层的权重与偏置，用所述深度学习校正模型的所述全连接层的权重与偏置替换所述在线深度学习预报模型的所述全连接层的权重与偏置。
进一步，作为优选，所述工业过程为电熔镁群炉运行过程，所述运行指标为电熔镁群炉功率。
一种工业过程运行指标智能预报装置，包括：
运行指标动态模型建模模块，用于利用工业过程控制系统特性建立运行指标动态模型，所述运行指标动态模型包括可辨识模型和未建模动态两部分；
参数辨识模块，用于估计所述运行指标动态模型中的所述可辨识模型的参数；
非线性动态获取模块，用于将所述运行指标动态模型中的所述可辨识模型的参数的辨识误差与所述运行指标动态模型中的所述未建模动态合并为未知非线性动态系统；
在线智能预报模型建模模块，用于采用自适应深度学习建立所述未知非线性动态系统的在线智能预报模型；
预报模块，用于由所述运行指标动态模型中的所述可辨识模型的输出与所述未知非线性动态系统的所述在线智能预报模型的输出得到所述运行指标的预报值。
进一步，作为优选，所述在线智能预报模型建模模块包括预报模型建模模块、 校正模型建模模块和自校正模块；所述预报模型建模模块采用LSTM架构建立在线深度学习预报模型；所述校正模型建模模块采用与所述在线深度学习预报模型相同的结构建立深度学习校正模型；所述自校正模块，在所述在线深度学习预报模型的输出与标签数据的误差大于设定阈值时，采用自校正机制，用所述深度学习校正模型的权重与偏置校正所述在线深度学习预报模型的权重与偏置；其中，所述深度学习校正模型所用的历史数据比所述在线深度学习预报模型所用的历史数据多。
进一步，作为优选，所述在线深度学习预报模型和所述深度学习校正模型均包括输入层、隐藏层、全连接层和输出层，其中，隐藏层的层数为L，L为大于或等于1的正整数；所述预报模型建模模块固定所述在线深度学习预报模型中的所述隐藏层的权重与偏置，并在线校正所述在线深度学习预报模型中的所述全连接层的权重与偏置；所述校正模型建模模块在线训练所述深度学习校正模型中的所述隐藏层和所述全连接层的权重与偏置；所述自校正模块，在所述在线深度学习预报模型的输出与标签数据的误差大于设定阈值时，采用自校正机制，用所述深度学习校正模型的所述隐藏层的权重与偏置替换所述在线深度学习预报模型的所述隐藏层的权重与偏置，用所述深度学习校正模型的所述全连接层的权重与偏置替换所述在线深度学习预报模型的所述全连接层的权重与偏置。
进一步，作为优选，所述工业过程为电熔镁群炉运行过程，所述运行指标为电熔镁群炉功率。
一种用于实现上述工业过程运行指标智能预报方法的工业过程运行指标智能预报设备，包括：端侧子设备、边缘侧子设备和云侧子设备；
所述端侧子设备用于采集所述工业过程中的输入数据和输出数据；
所述边缘侧子设备利用所述在线深度学习预报模型进行所述运行指标的在线预报；
所述云侧子设备用于训练所述深度学习校正模型，并实现所述自校正机制。
一种计算机可读存储介质，其存储有计算机程序，所述程序被处理器执行时实现上述工业过程运行指标智能预报方法。
本发明针对工业过程运行指标难以预报的难题，将基于机理模型的系统辨识方法与基于大数据的深度学习方法相结合，利用运行指标的变化取决于工业过程 控制系统的动态特性的特点，提出了工业过程运行指标的智能预报方法，解决了工业过程运行指标的预报难题。
附图说明
图1为本发明实施例的工业过程运行指标智能预报方法实现流程图；
图2为图1所示的步骤S1的具体实现流程图；
图3为本发明一个实施例的电熔镁群炉功率及需量智能预报方法实现流程图；
图4为图3所示的步骤S1’的具体实现流程图；
图5为本发明一个实施例的LSTM网络结构图；
图6为本发明一个实施例的需量预报结果图；
图7为本发明一个实施例的工业过程运行指标智能预报装置的结构示意图；
图8为本发明一个实施例的工业过程运行指标智能预报设备的结构示意图；
具体实施方式
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作出创造性劳动前提下所获得的所有其他实施例，都属于本发明保护的范围。
图1为本发明实施例的工业过程运行指标智能预报方法实现流程图，该方法包括以下步骤：
S1：利用工业过程控制系统特性建立运行指标动态模型，所述运行指标动态模型包括可辨识模型和未建模动态两部分。
如图2所示，步骤S1包括：步骤S11和S12。
步骤S11为：建立工业过程的闭环控制系统动态模型；具体的，利用工业过程运行在工作点附近的特点，将工业过程用线性低阶模型和未知高阶非线性项来表示，采用PID控制技术控制该工业过程，建立由线性模型和含未知高阶非线性项的未建模动态组成的工业过程闭环控制系统动态模型。
步骤S12为：建立运行指标与工业过程闭环控制系统的输入和输出数据之间的动态模型；具体的，采用工业过程闭环控制系统的动态模型，利用工业过程闭环控制系统将该工业过程的运行指标控制在一定的区间范围内的特点，将运行指 标动态模型用可辨识模型和未建模动态来描述。
S2:估计所述运行指标动态模型中的所述可辨识模型的参数。
具体的，利用运行指标与工业过程控制系统的输入与输出数据采用辨识算法估计运行指标动态模型中的可辨识模型的参数。
S3:将所述运行指标动态模型中的所述可辨识模型的参数的辨识误差与所述运行指标动态模型中的所述未建模动态合并为未知非线性动态系统。
具体的，将运行指标动态模型中的可辨识模型的参数的辨识误差与运行指标的未建模动态合并为模型结构与阶次未知的非线性动态系统，其输出为
Figure PCTCN2021136453-appb-000001
其输入为工业过程闭环控制系统的输出y(k),…,y(k-n)与输入u(k-1),…,u(k-n)，并将该未知非线性动态系统的输出
Figure PCTCN2021136453-appb-000002
作为输入变量，采用未知常数n表示该动态系统的阶次，采用下式表示运行指标的未知非线性动态系统：
Figure PCTCN2021136453-appb-000003
其中，f(·)是未知变化的非线性函数，
Figure PCTCN2021136453-appb-000004
为(k-i)时刻未知非线性动态系统的输出，y(k-j)(j＝0,1,…n)为(k-j)时刻工业过程闭环控制系统的输出，u(k-g)(g＝1,2,…,n)为(k-g)时刻工业过程闭环控制系统的输入；
Figure PCTCN2021136453-appb-000005
r(k-i+1)为(k-i+1)时刻的运行指标，
Figure PCTCN2021136453-appb-000006
为(k-i)时刻的运行指标的可辨识模型输出。
S4:采用自适应深度学习建立所述未知非线性动态系统的在线智能预报模型。
具体的，该在线智能预报模型由在线深度学习预报模型、深度学习校正模型和自校正机制组成。采用长短周期记忆多层神经网络LSTM，选择(1)式中的输入变量作为单个神经元的输入，阶次n作为神经元的个数，
Figure PCTCN2021136453-appb-000007
作为标签数据，采用(1)式的输入、输出数据组成大数据样本，采用训练算法，使标签数据
Figure PCTCN2021136453-appb-000008
与在线深度学习预报模型输出
Figure PCTCN2021136453-appb-000009
的预报误差
Figure PCTCN2021136453-appb-000010
尽可能小，确 定神经元个数n、LSTM的单元节点数
Figure PCTCN2021136453-appb-000011
多层神经网络层数L、各层的连接权参数和偏置参数，在线深度学习预报模型的输入数据的时间序列窗口长度取N，建立未知非线性动态系统的在线深度学习预报模型，采用时间序列窗口长度N的输入数据在线校正该模型的全连接层的连接权参数和偏置参数。采用该在线深度学习预报模型的相同结构，建立深度学习校正模型，采用当前时刻以及以前所有时刻的公式(1)的输入数据作为深度学习校正模型的输入数据，训练该深度学习校正模型的各层的连接权参数和偏置参数。自校正机制设定预报误差的区间上界为δ，当在线深度学习预报模型的预报误差
Figure PCTCN2021136453-appb-000012
则采用深度学习校正模型的各层连接权参数和偏置参数替换在线深度学习预报模型的相应层的连接权参数和偏置参数。
S5:由所述运行指标动态模型中的所述可辨识模型的输出与所述未知非线性动态系统的所述在线智能预报模型的输出得到所述运行指标的预报值。
进一步的，在一个实施例中，工业过程运行指标智能预报方法可用于电熔镁砂厂的电熔镁群炉的运行指标—需量的智能预报。
电熔镁炉是一种大型耗能设备，其产品电熔镁砂是一种广泛应用于冶金、化工、航天等领域的重要耐火材料。电熔镁炉采用埋弧方式边熔化边加入菱镁矿石，采用PID电流控制系统控制熔化电流，熔化菱镁矿石生产电熔镁砂，由于电熔镁炉是重大耗能设备，对其需量进行监控与预报对节能具有重要意义。
k时刻的群炉需量
Figure PCTCN2021136453-appb-000013
为k时刻与前m-1个时刻的群炉功率的平均值，即
Figure PCTCN2021136453-appb-000014
其中p(k)为k时刻的群炉功率，熔炼过程中定义m＝30。根据群炉需量的定义式(2)可知(k+1)时刻的需量为
Figure PCTCN2021136453-appb-000015
由(3)式可以看出(k+1)时刻的需量
Figure PCTCN2021136453-appb-000016
预报的关键在预报(k+1)时刻的功率 p(k+1)。
图3为本发明一个实施例的电熔镁群炉功率及需量智能预报方法实现流程图，该方法包括以下步骤：
S1’：利用工业过程控制系统特性建立群炉功率的动态模型，所述群炉功率的动态模型包括可辨识模型和未建模动态两部分。
如图4所示，步骤S1’包括：步骤S11’和S12’。
步骤S11’为：建立第i个电熔镁炉熔化电流闭环控制系统动态模型。
具体的：
第i个电熔镁炉的电流的动态模型为
A i(z -1)y i(k)＝b i0u i(k-1)+v i(k)     (4)
其中，A i(z -1)＝1+a i1z -1，a i1为可变常数，z -1为后移算子；y i(k)为k时刻第i个电熔镁炉的熔化电流；b i0为可变常数；u i(k-1)为(k-1)时刻第i个电熔镁炉的变频电机转动方向与频率，其正负表示方向，大小表示频率；v i(·)为未知高阶非线性项。
PID控制器为
(1-z -1)u i(k)＝G i(z -1)e i(k)      (5)
其中，G i(z -1)＝g i0+g i1z -1+g i2z -2，g i0,g i1,g i2为可变常数；e i(k)＝y i(k)-y *，y *是熔化电流设定值。
由式(4)，式(5)和z -1u i(k)＝u i(k-1)可得第i个电熔镁炉熔化电流闭环控制系统动态模型为：
T i(z -1)y i(k)＝-b i0G i(z -1)y *+(1-z -1)v i(k)    (6)
其中z -1y *＝y *，T i(z -1)为设计PID控制器参数选择的闭环系统的理想特征多项式，
T i(z -1)＝(1-z -1)A i(z -1)-z -1b i0G i(z -1)
＝(1-z -1)(1+a i1z -1)-z -1b i0(g i0+g i1z -1+g i2z -2)
＝1+(a i1-b i0g i0-1)z -1-(b i0g i1+a i1)z -2-b i0g i2z -3
＝1+t i1z -1+t i2z -2+t i3z -3
步骤S12’为：建立群炉功率与电熔镁炉熔化电流闭环控制系统的输入和输 出数据之间的动态模型。
具体的：
k时刻第i个电熔镁炉的功率为
Figure PCTCN2021136453-appb-000017
其中，U为电压，
Figure PCTCN2021136453-appb-000018
为功率因数。
由式(6)和式(7)可得第i个电熔镁炉的功率动态模型为
Figure PCTCN2021136453-appb-000019
其中，b i0G i(z -1)p *＝b i0(g i0+g i1+g i2)p *＝d i0p *，p *是熔化电流设定值y *对应的功率，
Figure PCTCN2021136453-appb-000020
k时刻m台电熔镁炉的功率为：
Figure PCTCN2021136453-appb-000021
采用t 1,t 2,t 3分别代替t i1,t i2,t i3，d 0代替d i0，根据式(8)和式(9)可得电熔镁群炉功率动态模型为：
Figure PCTCN2021136453-appb-000022
其中，ψ(k-1)＝[p(k-1),p(k-2),p(k-3),p *]；θ＝(θ 0,θ 1.θ 2,θ 3) T；θ j＝-t j+1,j＝0,1,2；θ 3＝-d 0，ψ(k-1)θ为可辨识模型，v(k-1)由未知非线性项
Figure PCTCN2021136453-appb-000023
和引入t 1,t 2,t 3,d 0产生的模型误差组成。
S2’:估计所述群炉功率的动态模型中的所述可辨识模型的参数θ。
具体的，模型(10)的参数辨识方程为：
p(k)-v(k-1)＝ψ(k-1)θ     (11)
使用实际功率数据，采用最小二乘估计算法可得
Figure PCTCN2021136453-appb-000024
S3’:将所述群炉功率的动态模型中的所述可辨识模型的参数的辨识误差与所述群炉功率的动态模型中的所述未建模动态合并为未知非线性动态系统。
具体的，式(10)可以表示为：
Figure PCTCN2021136453-appb-000025
其中未知非线性动态系统
Figure PCTCN2021136453-appb-000026
为：
Figure PCTCN2021136453-appb-000027
由式(12)，(k+1)时刻的群炉功率p(k+1)为：
Figure PCTCN2021136453-appb-000028
根据式(13)，p(k+1)的未知非线性动态系统为：
Figure PCTCN2021136453-appb-000029
其中，运行指标r(k)＝p(k)，运行指标可辨识模型输出为
Figure PCTCN2021136453-appb-000030
Figure PCTCN2021136453-appb-000031
f(·)为未知变化的非线性函数，n为该动态系统的未知阶次。
S4’:采用自适应深度学习建立所述未知非线性动态系统的在线智能预报模型。
具体的，该在线智能预报模型由在线深度学习预报模型、深度学习校正模型和自校正机制组成。
在线深度学习预报模型的建立方法如下：
采用长短记忆多层神经网络LSTM架构建立
Figure PCTCN2021136453-appb-000032
的在线深度学习预报模型。 选择网络层数为1，(k-1)时刻第j个神经元的输入为
Figure PCTCN2021136453-appb-000033
h(k+j-n-2)为第(j-1)个神经元的输出，神经元个数为n，LSTM单个神经元的节点数为
Figure PCTCN2021136453-appb-000034
输入数据时间序列窗口长度为N，采用式(15)中的M个输入输出数据
Figure PCTCN2021136453-appb-000035
构成大数据样本，采用下列训练算法使标签数据
Figure PCTCN2021136453-appb-000036
与在线深度学习预报模型输出
Figure PCTCN2021136453-appb-000037
的误差
Figure PCTCN2021136453-appb-000038
尽可能小。确定n，
Figure PCTCN2021136453-appb-000039
和N，训练算法的目标函数为：
Figure PCTCN2021136453-appb-000040
其中，
Figure PCTCN2021136453-appb-000041
为
Figure PCTCN2021136453-appb-000042
其中全连接层的连接权
Figure PCTCN2021136453-appb-000043
为
Figure PCTCN2021136453-appb-000044
的行向量，
Figure PCTCN2021136453-appb-000045
为全连接层的偏置，h(k-1)为第n个神经元的输出
h(k-1) T＝o n⊙tanh(C n)      (18)
其中tanh(x)＝(1-e -2x)/(1+e -2x)，输出门o n和状态门C n分别为
o n＝σ(W o·x n(k-1) T+b o)       (19)
Figure PCTCN2021136453-appb-000046
式(19)中σ(x)＝1/(1+e -x)，
Figure PCTCN2021136453-appb-000047
为第n个神经元的输入，W o为
Figure PCTCN2021136453-appb-000048
的矩阵，b o为
Figure PCTCN2021136453-appb-000049
的列向量。式(20)中遗忘门f n，输入门i n和状态候选值
Figure PCTCN2021136453-appb-000050
分别为
f n＝σ(W f·x n(k-1) T+b f)       (21)
i n＝σ(W in·x n(k-1) T+b in)     (22)
Figure PCTCN2021136453-appb-000051
式(21)-(23)中W f,W in,W c为
Figure PCTCN2021136453-appb-000052
的矩阵，b f,b in,b c为
Figure PCTCN2021136453-appb-000053
的列向量。采用梯度下降算法训练网络权重W o,W f,W in,W c与偏置b o,b f,b in,b c。首先计算输出门权重W o的梯度为：
Figure PCTCN2021136453-appb-000054
根据下式更新输出门权重W o
Figure PCTCN2021136453-appb-000055
其中α为学习率。采用相同的训练算法确定其他连接权和偏置。
随机初始化节点数
Figure PCTCN2021136453-appb-000056
首先确定神经元的个数n。令n＝1,2,3,4,…并计算预报误差
Figure PCTCN2021136453-appb-000057
的均方根误差(Root Mean Squared Error,RMSE)，当n＝30时，
Figure PCTCN2021136453-appb-000058
的RMSE最小。因此该动态系统的阶次n为30，即神经元的个数为30。
确定节点数
Figure PCTCN2021136453-appb-000059
固定神经元的个数为30，令
Figure PCTCN2021136453-appb-000060
分别计算
Figure PCTCN2021136453-appb-000061
的RMSE，当
Figure PCTCN2021136453-appb-000062
时，
Figure PCTCN2021136453-appb-000063
的RMSE最小，因此
Figure PCTCN2021136453-appb-000064
取100。
确定输入数据时间序列窗口长度N。固定深度学习模型的n和
Figure PCTCN2021136453-appb-000065
令窗口长度N＝1,…，2500，并计算
Figure PCTCN2021136453-appb-000066
的RMSE，当N＝2000时，
Figure PCTCN2021136453-appb-000067
的RMSE最小，因此窗口长度N＝2000。
固定n，
Figure PCTCN2021136453-appb-000068
和N，增加网络层数L＝1,2,3,4,…并分别计算
Figure PCTCN2021136453-appb-000069
的RMSE，当L＝3时，
Figure PCTCN2021136453-appb-000070
的RMSE最小，因此网络层数L为3，并通过梯度下降法确定全连接层的权重与偏置。
建立如图5所示的在线深度学习预报模型。该在线深度学习预报模型的第1层第j个神经元的输入为x j(k)，输出为h 1(k+j-31)(j＝1,…,30)，第i层第j个神经元的输入为[h i(k+j-32),h i-1(k+j-31)] T，输出为h i(k+j-31)，(i＝2,3,j＝1,…,30)。
Figure PCTCN2021136453-appb-000071
的在线深度学习预报模型为：
Figure PCTCN2021136453-appb-000072
式中全连接层的连接权
Figure PCTCN2021136453-appb-000073
为
Figure PCTCN2021136453-appb-000074
行向量，
Figure PCTCN2021136453-appb-000075
为全连接层的偏置，h 3(k)为第3层第30个神经元的输出。第一层、第二层和第三层的连接权和偏置固定不变，只在线校正式(26)中的全连接层权重
Figure PCTCN2021136453-appb-000076
和偏置
Figure PCTCN2021136453-appb-000077
在k+1时刻，更新的数据集为
Figure PCTCN2021136453-appb-000078
Figure PCTCN2021136453-appb-000079
为
Figure PCTCN2021136453-appb-000080
在线校正
Figure PCTCN2021136453-appb-000081
和
Figure PCTCN2021136453-appb-000082
的目标函数和校正算法如下：
Figure PCTCN2021136453-appb-000083
Figure PCTCN2021136453-appb-000084
Figure PCTCN2021136453-appb-000085
其中
Figure PCTCN2021136453-appb-000086
和
Figure PCTCN2021136453-appb-000087
为：
Figure PCTCN2021136453-appb-000088
深度学习校正模型的建立方法如下：
采用与在线深度学习预报模型完全相同的结构，即n＝30，
Figure PCTCN2021136453-appb-000089
和L＝3建立深度学习校正模型。在(k+1)时刻采用过去时刻全部数据更新的数据集为
Figure PCTCN2021136453-appb-000090
在线训练LSTM网络中的所有权重与偏置。
自校正机制如下：
设定在线深度学习预报模型误差
Figure PCTCN2021136453-appb-000091
的区间上界为δ，k+1时刻，若
Figure PCTCN2021136453-appb-000092
则采用深度学习校正模型的各层的权重与偏置校正在线深度学习 预报模型的对应层的权重与偏置。
S5’:由所述群炉功率的动态模型中的所述可辨识模型的输出与所述未知非线性动态系统的所述在线智能预报模型的输出得到所述群炉功率的预报值。
具体的，群炉功率p(k+1)的预报模型为：
Figure PCTCN2021136453-appb-000093
S6’:由所述群炉功率的预报值得到群炉需量的预报值。
具体的，群炉需量
Figure PCTCN2021136453-appb-000094
的预报模型为：
Figure PCTCN2021136453-appb-000095
如图6所示，采用上述电熔镁群炉需量预报方法，需量的预报精度为99.96％，需量上升趋势预报准确率为96.46％，需量下降趋势预报准确率为92.78％，满足了节能控制对需量预测的精度要求。
在一个实施例中，如图7所示，提供了一种工业过程运行指标智能预报装置，包括：运行指标动态模型建模模块、参数辨识模块、非线性动态获取模块、在线智能预报模型建模模块和预报模块，其中：
运行指标动态模型建模模块用于利用工业过程控制系统特性建立运行指标动态模型，所述运行指标动态模型包括可辨识模型和未建模动态两部分；
参数辨识模块用于估计所述运行指标动态模型中的所述可辨识模型的参数；
非线性动态获取模块用于将所述运行指标动态模型中的所述可辨识模型的参数的辨识误差与所述运行指标动态模型中的所述未建模动态合并为未知非线性动态系统；
在线智能预报模型建模模块用于采用自适应深度学习建立所述未知非线性动态系统的在线智能预报模型；
预报模块用于由所述运行指标动态模型中的所述可辨识模型的输出与所述未知非线性动态系统的所述在线智能预报模型的输出得到所述运行指标的预报值。
在其中一个实施例中，在线智能预报模型建模模块包括预报模型建模模块、校正模型建模模块和自校正模块。所述预报模型建模模块采用LSTM架构建立在线深度学习预报模型；所述校正模型建模模块采用与所述在线深度学习预报模 型相同的结构建立深度学习校正模型；所述自校正模块，在所述在线深度学习预报模型的输出与标签数据的误差大于设定阈值时，采用自校正机制，用所述深度学习校正模型的权重与偏置校正所述在线深度学习预报模型的权重与偏置；其中，所述深度学习校正模型所用的历史数据比所述在线深度学习预报模型所用的历史数据多。
在其中一个实施例中，所述在线深度学习预报模型和所述深度学习校正模型均包括输入层、隐藏层、全连接层和输出层，其中，隐藏层的层数为L，L为大于或等于1的正整数；所述预报模型建模模块固定所述在线深度学习预报模型中的所述隐藏层的权重与偏置，并在线校正所述在线深度学习预报模型中的所述全连接层的权重与偏置；所述校正模型建模模块在线训练所述深度学习校正模型中的所述隐藏层和所述全连接层的权重与偏置；所述自校正模块，在所述在线深度学习预报模型的输出与标签数据的误差大于设定阈值时，采用自校正机制，用所述深度学习校正模型的所述隐藏层的权重与偏置替换所述在线深度学习预报模型的所述隐藏层的权重与偏置，用所述深度学习校正模型的所述全连接层的权重与偏置替换所述在线深度学习预报模型的所述全连接层的权重与偏置。
在其中一个实施例中，所述工业过程运行指标智能预报装置用于电熔镁群炉功率预报。
关于工业过程运行指标智能预报装置的具体限定可以参见上文中对于工业过程运行指标智能预报方法的限定，在此不再赘述。上述工业过程运行指标智能预报装置中的各个模块可全部或部分通过软件、硬件及其组合来实现。上述各模块可以以硬件形式内嵌于或独立于计算机设备中的处理器中，也可以以软件形式存储于计算机设备的存储器中，以便于处理器调用执行以上各个模块对应的操作。
在一个实施例中，如图8所示，提供了一种用于实现上述各实施例中的工业过程运行指标智能预报方法的工业过程运行指标智能预报设备，包括：端侧子设备、边缘侧子设备和云侧子设备；所述端侧子设备用于采集所述工业过程中的输入数据和输出数据；所述边缘侧子设备利用所述在线深度学习预报模型进行所述运行指标的在线预报；所述云侧子设备用于训练所述深度学习校正模型，并实现所述自校正机制。
在一个实施例中，提供了一种计算机可读存储介质，其存储有计算机程序，所述程序被处理器执行时实现上述各实施例中的工业过程运行指标智能预报方法。
在不相互矛盾的情况下，本领域的技术人员可以将本说明书中描述的不同实施例以及不同实施例的特征进行结合和组合。
综上所述，本发明实施例提出的工业过程运行指标智能预报方法、装置和设备针对已有的基于模型的预报方法和深度学习方法无法对表征工业过程加工的产品质量、效率与消耗的运行指标进行预报的难题，将运行指标与过程控制系统的输入与输出之间的动态模型用可辨识的模型与未建模动态来表示，利用运行指标、工业过程控制系统的输入与输出数据，采用辨识算法估计可辨识模型的参数，将辨识误差与未建模动态构成模型结构与阶次未知的非线性动态系统，采用工业大数据，建立该动态系统的在线智能预报模型，利用辨识模型与未知非线性动态系统的在线智能预报模型，实现了工业过程运行指标的精准预报。
以上所述，仅为本发明的具体实施方式，但本发明的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本发明揭露的技术范围内，可轻易想到其各种变化或替换，这些都应涵盖在本发明的保护范围之内。因此，本发明的保护范围应以所述权利要求的保护范围为准。

Claims (10)

  1. 一种工业过程运行指标智能预报方法，其特征在于，所述方法包括：
    利用工业过程控制系统特性建立运行指标动态模型，所述运行指标动态模型包括可辨识模型和未建模动态两部分；
    估计所述运行指标动态模型中的所述可辨识模型的参数；
    将所述运行指标动态模型中的所述可辨识模型的参数的辨识误差与所述运行指标动态模型中的所述未建模动态合并为未知非线性动态系统；
    采用自适应深度学习建立所述未知非线性动态系统的在线智能预报模型；
    由所述运行指标动态模型中的所述可辨识模型的输出与所述未知非线性动态系统的所述在线智能预报模型的输出得到所述运行指标的预报值。
  2. 根据权利要求1所述的方法，其特征在于，所述在线智能预报模型包括在线深度学习预报模型、深度学习校正模型和自校正机制；
    采用LSTM架构建立所述在线深度学习预报模型；
    采用与所述在线深度学习预报模型相同的结构建立所述深度学习校正模型；
    当所述在线深度学习预报模型的输出与标签数据的误差大于设定阈值时，采用自校正机制，用所述深度学习校正模型的权重与偏置校正所述在线深度学习预报模型的权重与偏置；
    其中，所述深度学习校正模型所用的历史数据比所述在线深度学习预报模型所用的历史数据多。
  3. 根据权利要求2所述的方法，其特征在于，所述在线深度学习预报模型和所述深度学习校正模型均包括输入层、隐藏层、全连接层和输出层，其中，隐藏层的层数为L，L为大于或等于1的正整数；
    固定所述在线深度学习预报模型中的所述隐藏层的权重与偏置，并在线校正所述在线深度学习预报模型中的所述全连接层的权重与偏置；
    在线训练所述深度学习校正模型中的所述隐藏层和所述全连接层的权重与偏置；
    当所述在线深度学习预报模型的输出与标签数据的误差大于设定阈值时，采用自校正机制，用所述深度学习校正模型的所述隐藏层的权重与偏置替换所述在线深度学习预报模型的所述隐藏层的权重与偏置，用所述深度学习校正模型的所述全连接层的权重与偏置替换所述在线深度学习预报模型的所述全连接层的权 重与偏置。
  4. 根据权利要求1-3任一所述的方法，其特征在于，所述工业过程为电熔镁群炉运行过程，所述运行指标为电熔镁群炉功率。
  5. 一种工业过程运行指标智能预报装置，其特征在于，所述装置包括：
    运行指标动态模型建模模块，用于利用工业过程控制系统特性建立运行指标动态模型，所述运行指标动态模型包括可辨识模型和未建模动态两部分；
    参数辨识模块，用于估计所述运行指标动态模型中的所述可辨识模型的参数；
    非线性动态获取模块，用于将所述运行指标动态模型中的所述可辨识模型的参数的辨识误差与所述运行指标动态模型中的所述未建模动态合并为未知非线性动态系统；
    在线智能预报模型建模模块，用于采用自适应深度学习建立所述未知非线性动态系统的在线智能预报模型；
    预报模块，用于由所述运行指标动态模型中的所述可辨识模型的输出与所述未知非线性动态系统的所述在线智能预报模型的输出得到所述运行指标的预报值。
  6. 根据权利要求5所述的装置，其特征在于，所述在线智能预报模型建模模块包括预报模型建模模块、校正模型建模模块和自校正模块；
    所述预报模型建模模块采用LSTM架构建立在线深度学习预报模型；
    所述校正模型建模模块采用与所述在线深度学习预报模型相同的结构建立深度学习校正模型；
    所述自校正模块，在所述在线深度学习预报模型的输出与标签数据的误差大于设定阈值时，采用自校正机制，用所述深度学习校正模型的权重与偏置校正所述在线深度学习预报模型的权重与偏置；
    其中，所述深度学习校正模型所用的历史数据比所述在线深度学习预报模型所用的历史数据多。
  7. 根据权利要求6所述的装置，其特征在于，所述在线深度学习预报模型和所述深度学习校正模型均包括输入层、隐藏层、全连接层和输出层，其中，隐藏层的层数为L，L为大于或等于1的正整数；
    所述预报模型建模模块固定所述在线深度学习预报模型中的所述隐藏层的 权重与偏置，并在线校正所述在线深度学习预报模型中的所述全连接层的权重与偏置；
    所述校正模型建模模块在线训练所述深度学习校正模型中的所述隐藏层和所述全连接层的权重与偏置；
    所述自校正模块，在所述在线深度学习预报模型的输出与标签数据的误差大于设定阈值时，采用自校正机制，用所述深度学习校正模型的所述隐藏层的权重与偏置替换所述在线深度学习预报模型的所述隐藏层的权重与偏置，用所述深度学习校正模型的所述全连接层的权重与偏置替换所述在线深度学习预报模型的所述全连接层的权重与偏置。
  8. 根据权利要求5-7任一所述的装置，其特征在于，所述工业过程为电熔镁群炉运行过程，所述运行指标为电熔镁群炉功率。
  9. 一种用于实现权利要求2或3所述的方法的工业过程运行指标智能预报设备，其特征在于，所述设备包括：端侧子设备、边缘侧子设备和云侧子设备；
    所述端侧子设备用于采集所述工业过程中的输入数据和输出数据；
    所述边缘侧子设备利用所述在线深度学习预报模型进行所述运行指标的在线预报；
    所述云侧子设备用于训练所述深度学习校正模型，并实现所述自校正机制。
  10. 一种计算机可读存储介质，其存储有计算机程序，其特征在于，所述程序被处理器执行时实现如权利要求1-4中任一所述的方法。
PCT/CN2021/136453 2020-12-10 2021-12-08 工业过程运行指标智能预报方法、装置、设备及存储介质 WO2022121944A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202011435311.2A CN112631215B (zh) 2020-12-10 2020-12-10 工业过程运行指标智能预报方法、装置、设备及存储介质
CN202011435311.2 2020-12-10

Publications (1)

Publication Number Publication Date
WO2022121944A1 true WO2022121944A1 (zh) 2022-06-16

Family

ID=75309024

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/136453 WO2022121944A1 (zh) 2020-12-10 2021-12-08 工业过程运行指标智能预报方法、装置、设备及存储介质

Country Status (2)

Country Link
CN (1) CN112631215B (zh)
WO (1) WO2022121944A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115438271A (zh) * 2022-11-08 2022-12-06 商飞软件有限公司 一种工业机理模型及app的管理系统
CN115857360A (zh) * 2023-02-14 2023-03-28 青岛海大新星软件咨询有限公司 一种工业系统机理建模仿真平台
CN117055487A (zh) * 2023-08-24 2023-11-14 北京科技大学 一种基于机理数据混合驱动的二辊斜轧穿孔参数优化方法

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112631215B (zh) * 2020-12-10 2022-06-24 东北大学 工业过程运行指标智能预报方法、装置、设备及存储介质

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2004059400A1 (en) * 2002-12-27 2004-07-15 Inea D.O.O. Self-tuning controller for non-linear processes described by set of local linear models
CN101813916A (zh) * 2009-02-19 2010-08-25 中国石油化工股份有限公司 非线性生产过程的自适应预测函数控制方法
CN103064282A (zh) * 2012-05-04 2013-04-24 浙江大学 非线性参数变化模型辨识方法（npv）
CN104122878A (zh) * 2014-08-13 2014-10-29 黑龙江省科学院自动化研究所 工业节能减排控制装置及控制方法
CN106033555A (zh) * 2015-03-13 2016-10-19 中国科学院声学研究所 基于满足k度稀疏约束的深度学习模型的大数据处理方法
CN110703718A (zh) * 2019-11-13 2020-01-17 东北大学 一种基于信号补偿的工业过程控制方法
CN111768000A (zh) * 2020-06-23 2020-10-13 中南大学 在线自适应微调深度学习的工业过程数据建模方法
CN112631215A (zh) * 2020-12-10 2021-04-09 东北大学 工业过程运行指标智能预报方法、装置、设备及存储介质

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20030009240A1 (en) * 2001-04-20 2003-01-09 Honda Giken Kogyo Kabushiki Kaisha Control system for plant
CN103092072B (zh) * 2012-12-28 2015-06-17 东北大学 磨矿过程运行控制实验系统及方法
CN109088406A (zh) * 2018-06-26 2018-12-25 河海大学常州校区 一种基于lstm神经网络的微电网等效建模方法
CN109272146B (zh) * 2018-08-23 2021-10-19 河海大学 一种基于深度学习模型和bp神经网络校正的洪水预测方法
CN111930010A (zh) * 2020-06-29 2020-11-13 华东理工大学 一种基于lstm网络的通用mfa控制器设计方法

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2004059400A1 (en) * 2002-12-27 2004-07-15 Inea D.O.O. Self-tuning controller for non-linear processes described by set of local linear models
CN101813916A (zh) * 2009-02-19 2010-08-25 中国石油化工股份有限公司 非线性生产过程的自适应预测函数控制方法
CN103064282A (zh) * 2012-05-04 2013-04-24 浙江大学 非线性参数变化模型辨识方法（npv）
CN104122878A (zh) * 2014-08-13 2014-10-29 黑龙江省科学院自动化研究所 工业节能减排控制装置及控制方法
CN106033555A (zh) * 2015-03-13 2016-10-19 中国科学院声学研究所 基于满足k度稀疏约束的深度学习模型的大数据处理方法
CN110703718A (zh) * 2019-11-13 2020-01-17 东北大学 一种基于信号补偿的工业过程控制方法
CN111768000A (zh) * 2020-06-23 2020-10-13 中南大学 在线自适应微调深度学习的工业过程数据建模方法
CN112631215A (zh) * 2020-12-10 2021-04-09 东北大学 工业过程运行指标智能预报方法、装置、设备及存储介质

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
JANUARY SINICA, ZHANG YA-JUN, TIAN-YOU CHAI, JIE YANG: "Alternating Identification Algorithm and Its Application to a Class of Nonlinear Discrete-time Dynamical Systems", ACTA AUTOMATICA SINICA, 15 January 2017 (2017-01-15), pages 101 - 113, XP055942242, [retrieved on 20220713], DOI: 10.16383/j.aas.2017.c150759 *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115438271A (zh) * 2022-11-08 2022-12-06 商飞软件有限公司 一种工业机理模型及app的管理系统
CN115438271B (zh) * 2022-11-08 2023-03-24 商飞软件有限公司 一种工业机理模型及app的管理系统
CN115857360A (zh) * 2023-02-14 2023-03-28 青岛海大新星软件咨询有限公司 一种工业系统机理建模仿真平台
CN115857360B (zh) * 2023-02-14 2023-07-25 青岛海大新星软件咨询有限公司 一种工业系统机理建模仿真平台
CN117055487A (zh) * 2023-08-24 2023-11-14 北京科技大学 一种基于机理数据混合驱动的二辊斜轧穿孔参数优化方法
CN117055487B (zh) * 2023-08-24 2024-04-16 北京科技大学 一种基于机理数据混合驱动的二辊斜轧穿孔参数优化方法

Also Published As

Publication number Publication date
CN112631215B (zh) 2022-06-24
CN112631215A (zh) 2021-04-09

Similar Documents

Publication Publication Date Title
WO2022121944A1 (zh) 工业过程运行指标智能预报方法、装置、设备及存储介质
WO2022121923A1 (zh) 复杂工业过程数字孪生系统智能建模方法、装置、设备及存储介质
CN109685283B (zh) 一种基于置信规则库推理的矿热炉工况预测方法
CN104846306B (zh) 一种镀锌厚度控制系统及方法
Yi et al. Event-triggered globalized dual heuristic programming and its application to networked control systems
CN107464023A (zh) 生命周期预测模型的在线学习方法和设备
CN110347192B (zh) 基于注意力机制和自编码器的玻璃炉温智能预测控制方法
CN112560338B (zh) 基于自适应深度学习的复杂工业系统智能预报方法、装置、设备及存储介质
CN113325721B (zh) 一种工业系统无模型自适应控制方法及系统
CN110097929A (zh) 一种高炉铁水硅含量在线预测方法
CN113919545A (zh) 一种多数据模型融合的光伏发电功率预测方法和系统
CN101639902A (zh) 生物发酵过程中基于支持向量机的软测量仪表的建模方法
CN112083694A (zh) 氧气底吹铜吹炼过程反馈控制方法、装置及电子设备
CN114429248A (zh) 一种变压器视在功率预测方法
CN116880191A (zh) 一种基于时序预测的过程工业生产系统的智能控制方法
CN114384931B (zh) 一种基于策略梯度的无人机多目标最优控制方法和设备
Tian et al. A new incremental learning modeling method based on multiple models for temperature prediction of molten steel in LF
CN115718883A (zh) 电池储能系统热预警方法及装置
Zhang et al. Demand forecasting of a fused magnesia smelting process based on LSTM and FRA
CN102486632A (zh) 一种对二甲苯氧化过程中对苯二甲酸结晶粒径在线分析方法
Liu et al. A PSO-RBF neural network for BOD multi-step prediction in wastewater treatment process
CN117784852B (zh) 一种基于鱼鳞仿生优化算法的多模态传感器温度控制方法
CN103135444A (zh) 一种钢铁生产能耗免疫预测控制模型
Jiang et al. Reinforcement Learning for Blast Furnace Ironmaking Operation With Safety and Partial Observation Considerations
Wang et al. A new model based on improved ACA and BP to predict Silicon content in hot metal

Legal Events

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

Ref document number: 21902649

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: 21902649

Country of ref document: EP

Kind code of ref document: A1