WO2007009322A1 - Méthode optimisée de fonctionnement temps réel de procédure multi-entrée et multi-sortie de fabrication en continu - Google Patents

Méthode optimisée de fonctionnement temps réel de procédure multi-entrée et multi-sortie de fabrication en continu Download PDF

Info

Publication number
WO2007009322A1
WO2007009322A1 PCT/CN2005/002324 CN2005002324W WO2007009322A1 WO 2007009322 A1 WO2007009322 A1 WO 2007009322A1 CN 2005002324 W CN2005002324 W CN 2005002324W WO 2007009322 A1 WO2007009322 A1 WO 2007009322A1
Authority
WO
WIPO (PCT)
Prior art keywords
operating conditions
data
real
time
production process
Prior art date
Application number
PCT/CN2005/002324
Other languages
English (en)
French (fr)
Inventor
Jian Wang
Original Assignee
Jian Wang
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 Jian Wang filed Critical Jian Wang
Priority to US11/988,889 priority Critical patent/US7848831B2/en
Priority to CA2615727A priority patent/CA2615727C/en
Publication of WO2007009322A1 publication Critical patent/WO2007009322A1/zh

Links

Classifications

    • GPHYSICS
    • G05CONTROLLING; REGULATING
    • G05BCONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B11/00Automatic controllers
    • G05B11/01Automatic controllers electric
    • G05B11/36Automatic controllers electric with provision for obtaining particular characteristics, e.g. proportional, integral, differential
    • G05B11/42Automatic controllers electric with provision for obtaining particular characteristics, e.g. proportional, integral, differential for obtaining a characteristic which is both proportional and time-dependent, e.g. P.I., P.I.D.
    • GPHYSICS
    • G05CONTROLLING; REGULATING
    • G05BCONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B19/00Programme-control systems
    • G05B19/02Programme-control systems electric
    • G05B19/18Numerical control [NC], i.e. automatically operating machines, in particular machine tools, e.g. in a manufacturing environment, so as to execute positioning, movement or co-ordinated operations by means of programme data in numerical form
    • G05B19/41Numerical control [NC], i.e. automatically operating machines, in particular machine tools, e.g. in a manufacturing environment, so as to execute positioning, movement or co-ordinated operations by means of programme data in numerical form characterised by interpolation, e.g. the computation of intermediate points between programmed end points to define the path to be followed and the rate of travel along that path
    • GPHYSICS
    • G05CONTROLLING; REGULATING
    • G05BCONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B2219/00Program-control systems
    • G05B2219/30Nc systems
    • G05B2219/31From computer integrated manufacturing till monitoring
    • G05B2219/31265Control process by combining history and real time data
    • GPHYSICS
    • G05CONTROLLING; REGULATING
    • G05BCONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B2219/00Program-control systems
    • G05B2219/30Nc systems
    • G05B2219/32Operator till task planning
    • G05B2219/32015Optimize, process management, optimize production line

Definitions

  • the invention belongs to the technical field of control systems and relates to a multiple input multiple output (Multiple
  • the continuous production process of continuous product output can be represented by Figure 1:
  • A is the upstream processing equipment group
  • its product is the raw material of processing unit B
  • the product of processing unit B is downstream.
  • A is called the device group
  • the device group is because there may be more than one upstream device to supply raw materials for B. Since there may be more than one product of B, its products may also be supplied to a plurality of downstream devices as raw materials, so the device downstream thereof is called device group C. It is worth noting that the serial flow of such device materials is continuous.
  • controllable key operating conditions U there are some production factors that can be adjusted, that is, controllable key operating conditions U, and these operating conditions are related to some technical indicators of processing device B, such as energy consumption, economic efficiency, product production.
  • the rate is related, where U can be regarded as the input of the production process, there can be one or more, J can be seen as the output of the production process, and there can be one or more.
  • the task of operational optimization is how to adjust these operating conditions in production so that the specified technical indicators (which can be one or more combinations) are optimized for a certain period of time (maximum, such as economic benefit; or minimum, such as Energy consumption), as shown in Figure 2. It is worth pointing out that the objective function may switch according to the needs of production. Operational optimization of a single objective function process is generally described in the following form:
  • J is the objective function
  • J* is the optimal value of the objective function
  • u is the operating condition, or the tuning variable.
  • is the optimal value for tuning variables. Since the function relationship J(U) of U and J is usually an unknown function, in order to obtain lT, there are two commonly used methods: mathematical model and real-time online search method.
  • the basic idea of the modeling method is to pre-establish a mathematical model between the objective function J and the tuning variable U, and then use the nonlinear and linear programming to find ⁇ according to the model and constraints.
  • mechanism modeling is to combine the operating mechanism equations of various parts of the whole system according to the process structure, using the principle of material balance and energy balance, so as to construct a mathematical equation that conforms to the specific production process.
  • the relationship between the objective function and the tuning variable u is determined according to the input, output and price of the system.
  • the empirical modeling method relies on a large number of experiments or daily running report data to construct an empirical relationship between the tuning variables of the system and the objective function.
  • the advantages of this modeling are: simple, universal. No matter how complex or different the process or system is, it can be modeled in the same simple way without the need for specialized process knowledge and priori equations.
  • the search method is a universal method and has nothing to do with the specific process.
  • the basic idea is to change the value of the tuning variable online and observe the change of the objective function to determine whether the direction of the tuning is correct.
  • many nonlinear programming methods such as the Golden Section method and the steepest descent method, can be applied online.
  • these methods are often very sensitive to interference.
  • the objective function is not only a function of tuning variables, but also a function of other uncontrollable variables (environment variables). Therefore, when the target function changes, it is difficult to judge whether it is due to the change of the tuning variable or the effect of the interference.
  • the tuning variable and the objective function are generally regarded as the sole causal relationship, so when there is an external disturbance, it is possible to make a wrong judgment or even reverse the action.
  • both the model method and the direct search method are generally based on the mathematical description of 1-1.
  • the relationship between the objective function J and U is defined as an algebraic relationship and does not contain external disturbances. Therefore, the algorithm derived from this principle is applicable only to static, non-interfering systems.
  • the tuning variable not only has a logical causal relationship with the objective function, but also has a dynamic process in time. That is to say, when the tuning variable changes, the objective function does not change immediately, but has a transition process.
  • the reality from many industrial processes In other words, the objective function is often in the undulating pulsation, it is difficult to find a static situation, which is usually caused by strong uncontrollable and uncontrollable disturbances. For example, changes in the composition of raw materials are often uncontrollable during production. These variables are often unmeasurable due to the difficulty of online component detection. On the other hand, many processes are very sensitive to changes in the composition of the raw materials.
  • the technical problem to be solved by the present invention is to provide a method for realizing the real-time operation optimization of the multi-input and multi-output continuous production process in the prior art, which has the defects of complicated process, high difficulty, many interference factors, narrow adaptability and poor effect, and provides a kind of defect.
  • Real-time operation optimization method for multi-input and multi-output continuous production process simplifying the optimization control process of the production process, simple operation, strong anti-interference, strong versatility, wide adaptability and good effect.
  • the technical solution adopted to solve the technical problem of the present invention is to adopt a real-time operation optimization method for a multi-input and multi-output continuous production process, which uses a plurality of key operating conditions in the production process as optimization variables to be associated with key operating conditions.
  • One or more technical indicators are objective functions.
  • the gradient vector between the key operating conditions and the technical indicators at the current time is calculated online, and then the operation is determined according to the gradient vector.
  • Condition adjustment direction when the gradient vector is negative or positive, the key operating conditions must be adjusted to change the gradient vector to zero, so that the technical indicators are optimal.
  • this gradient is continuous on the line, regardless of Whether the technical indicators are optimal, once the gradient is found to be non-zero, the key operating conditions are adjusted to change the gradient vector to zero direction to achieve the most advantageous tracking.
  • the process of adjusting the critical operating conditions to change the gradient vector to zero is the process of optimizing the critical operating conditions. Since the best of the production process may change over time, this optimization process is ongoing on-line. ⁇
  • the online method for calculating the gradient vector between the key operating conditions and the technical indicators at the current time is to use the relevant integral technique or other possible methods (such as dynamic model identification) to dynamically and historically analyze the key operational conditions and technical indicators of the production process. Perform the operation.
  • correlation integral is a stochastic process Related operations.
  • the objective function, interference, and optimization variables are treated as random processes, and the optimization variables are controllable by the mean.
  • the objective function (t) is determined first.
  • the objective function should be online or measurable. , then the objective function can be expressed as:
  • E ⁇ S(t) ⁇ is the mean value of the tuning variable, which can be the set value or valve position of the base controller.
  • is a constant. If the optimization goal is to find the maximum value, it is taken as a number greater than 0. If the optimization target is a minimum value, it is taken as a number less than 0.
  • the routine operation control is performed on the key operating conditions to be optimized, and the set value is calculated by the optimization control computer using the relevant integral technology. According to the specific conditions and requirements of the process, the set value is performed every certain time period. Adjustment
  • the routine fixed value control for the key operating conditions to be optimized is to perform the routine fixed value control on the operating conditions to be optimized by the distributed system computer, ie DCS or conventional instrument, and the set value is calculated by the optimization control computer using the relevant integral technology.
  • the set value is adjusted every certain time period (the time period is determined by the speed of the specific process).
  • the method is to establish real-time data with a certain data window width (the time width of the database should be greater than 3 times of the process optimization variable to the target function transition process time, generally 8-18 hours) according to specific process time characteristics.
  • a data acquisition system usually consisting of a distributed control system (ie DCS) to obtain historical data on key operating conditions and technical specifications of the production process.
  • the system collects key operating conditions and data for each objective function, ie, each objective function, at regular intervals (which are determined by the speed of the process, typically 30-90 seconds).
  • the data in the data window is stored in the database. Each time the data is sampled, the data window is moved forward by one sampling time. That is, the oldest data is discarded, and the latest data is added to the database.
  • Figure 3 shows an example of two operating conditions.
  • T, M is the integral constant.
  • the above M should be greater than the maximum time constant from the optimized variable to the objective function, and T is taken as 1-5 times M.
  • the gradient ⁇ of the integrated objective function J to the operating conditions is ⁇ + ⁇ 2 ⁇ , ⁇ , ⁇ Jn
  • a 2 A, a m m is a positive constant number (if required specifications maximum).
  • an operating condition to take the maximum value of the technical index as an example, and illustrate the method of adjustment. See Figure 4, Figure 5, Figure 6.
  • the technical index is the maximum value, and the operating condition does not need to be adjusted.
  • Fig. 5 when the gradient is calculated to be negative by the correlation integration method, the operating conditions should be reduced to improve the technical specifications.
  • Fig. 6 when the gradient is positive by the correlation integration method, the operating conditions should be increased to improve the technical specifications.
  • each adjustment step is ⁇ , as long as the appropriate value of ⁇ is taken, that is, if the maximum value of the technical index is obtained, ⁇ is taken as a positive value, otherwise a negative value is taken. , you can adjust the size and direction of the step.
  • the data is sampled again at a certain interval (30-90 seconds), and the steps are returned.
  • steps 3 to 6 The process of steps 3 to 6 is carried out online continuously, so that each operating point can finally reach the best advantage. It should be pointed out that even if the gradients have reached zero, steps 2 to 5 should be carried out continuously, because the functional relationship between the technical indicators and the operating conditions may change with time (such as changes in the properties of the raw materials, devices). The transformation, etc.), need to constantly observe whether the gradient is zero, if there is a change, adjust at any time. As shown in Figure :: When the relationship between technical indicators and operating conditions changes for some reason (such as raw material changes), so that the current operating point is no longer the most advantageous, it can be found that the gradient is not zero, according to the method of correlation The operating conditions are approached to the best.
  • the present invention uses a correlation integration method for gradient calculation, which is calculated based on fluctuation data of operating conditions and technical indicators over a period of time (ie, within the data window). Therefore, unlike the model method, the production process is not required in the optimization operation. Being static, can be in a wave state, and according to the relevant integral theory, the statistical characteristics of these fluctuations can be unknown.
  • the static and dynamic models of the process can be built without prior construction, greatly reducing the complexity of the operational optimization of the production process.
  • the traditional method whether it is mechanism modeling or statistical modeling, is to try to find the functional relationship between the technical index J and the operating condition U, and then find the optimal operating conditions based on this functional relationship. If the system is complex, it is difficult to get an accurate model and the cost is high. However, the present invention notifies that as long as there is information near the current operating point, the system can be optimized without establishing a large-scale model, and complicated process modeling can be omitted.
  • This method has strong adaptive performance. That is, when the production process changes for some reason, causing the operating conditions to deviate from the optimum, the present invention can find such deviations and automatically re-adjust the operating conditions to the optimum point.
  • This automatic tracking of the most advantageous performance has important practical value in actual production. Because of many operating conditions and technical indicators in the production process The relationship between them actually varies with the nature of the raw materials, the aging, replacement of the equipment, the improvement of the catalyst, and the like. Traditional modeling methods require re-modeling the production process or correcting the model.
  • This method is a versatile method. As long as the production process is continuous, the technical indicators to be optimized can be measured or calculated online, and the method can be used, which is independent of the specific production process and can be applied in a wide range. Traditional modeling methods must be modeled for a specific device, and the established optimization model has specific applicability.
  • the method has strong anti-interference characteristics, even under dynamic conditions, such as other factors such as the nature of the material causing the change of the technical index to be greater than the useful signal (the change of the objective function caused by the optimized operating conditions) Normal work, this strong anti-interference is of great significance in practical applications.
  • the core of the correlation real-time optimization method in the present invention is to calculate the gradient of the objective function to the operating conditions by using the principle of the correlation integral technique, and then adjust the operating conditions according to the gradient, and the calculation of the gradient and the adjustment of the operating conditions are continuously performed online. Therefore, no matter what calculation method is used, the device can be optimized online by using this principle as long as the operating conditions and the gradient of the objective function can be calculated on the line. Therefore, other possible methods such as dynamic model identification can also be used to calculate the gradient of the operating conditions and the objective function online, and then optimize the continuous production operation of the device in real time.
  • FIG. 1 Block diagram of a continuous production process for continuous feed and continuous product output
  • Figure 2 is a schematic diagram of the operation optimization process of the continuous production process.
  • FIG. 3 Schematic diagram of data window data processing and acquisition with 2 operating conditions
  • Figure 4 is a graph of the gradient of the curve of the operating conditions and technical indicators is zero.
  • Figure 5 is a graph when the gradient of the operating condition and the technical index is negative.
  • Figure 6 is a graph showing the gradient of the gradient between the operating conditions and the technical indicators.
  • Figure 7 is a graph showing the gradient of the gradient of the relationship between the operating conditions and the technical indicators.
  • Figure 8 is a control diagram of the present invention using a computer to optimize the continuous production process.
  • Figure 9 is a flow chart of the ARGG device reaction regeneration system.
  • Figure 10 is a process flow diagram of the ketone-benzene deoiling and dewaxing unit
  • the production process is required to be optimized by a computer, and the operating conditions to be optimized are firstly controlled by a distributed system computer (ie, DCS) or a conventional meter, and the setting is performed.
  • the fixed value is calculated by the optimization control computer by the method of the invention, and the set value is adjusted every certain time period, and the period of the adjustment is determined by the speed of the specific process. As shown in Figure 8. .
  • Embodiment 1 Online correlation integral optimization control of ARGG device
  • the ARGG unit is a unit used in petrochemical plants to crack low-value oil into high-value liquid hydrocarbons, gasoline, and diesel. It is a continuous production process.
  • the flow chart of the ARGG device reaction regeneration system is shown in Figure 9.
  • the feedstock oil from the tank zone is mixed with the refueled gasoline, the refinery oil and the refinery slurry, and is atomically sprayed into the riser reactor 1 in the lower portion of the riser reactor 1 to enter the reaction system.
  • the atomized reaction raw material and the injected steam are mixed with the high temperature catalyst from the catalyst regenerator 4 in the lower portion of the riser reactor 1, and rise up along the riser reactor 1 to carry out a catalytic cracking reaction, and the reacted oil and gas together with the catalyst at the riser outlet Entering the cyclone 2 rapidly separates the oil and gas from the solid particulate catalyst, and the separated oil and gas is sent to the fractionation system 6.
  • the separated catalyst enters the catalyst regenerator 4, and the surface of these catalysts has carbon deposits generated during the reaction, and these carbon deposits are burned and removed in the regenerator, a process called scorch.
  • the excess heat generated during the scorching process is taken away by the external heat extractor 5.
  • the charred regeneration catalyst is then passed to the bottom of the riser reactor 1 to be mixed with the feedstock for reaction.
  • a new catalyst is additionally supplied to the catalyst regenerator.
  • the oil and gas from the separator enters the fractionation system 6, and the separation of the system produces liquid hydrocarbons, gasoline, and light diesel oil.
  • the produced refining slurry and refining oil and part of the gasoline are returned to the riser reactor 1 .
  • the structure of the optimized control system is shown in Figure 8.
  • the setpoint control system consists of the HONEYWELL TPS3000's distributed control system, and the optimized control is carried out by the APP (Application Process Processor) computer of the HONEYWELL TPS3000.
  • APP Application Process Processor
  • one of the above five technical indicators is selected as the current technical indicator to be optimized. And follow the steps below.
  • An OPC-based data acquisition system was built using the HONEYWELL TPS3000 distributed system and APP computer.
  • a real-time database with a data window width of 8 hours is built in the APP computer.
  • the key operating conditions and data for each technical indicator (each objective function) are collected every 60 seconds.
  • the data in the data window is stored in the database. Each time the data is sampled, the data window moves forward one sample time, that is, the oldest data is discarded, and the latest data is added to the database.
  • the upper and lower limits 3600, -3600, 10800, -10800 in the above integration are determined based on the approximate response time of the process from the optimization variable to the target function channel.
  • the comprehensive objective function J has a gradient K of the operating conditions.
  • ⁇ , in, and 5 are five constants of 0 or 1, depending on the currently optimized objective function.
  • the principle is that if the calculated gradient is zero, the operating condition is currently in an optimal state; Zero, adjust the operating conditions according to the size and direction of the gradient.
  • the current operating condition value is known, and the adjusted operating condition value is obtained according to the following method, and the technical index is required to be maximized -
  • Example 2 Online correlation integral optimization control of ketone-benzene deoiling and dewaxing unit
  • the ketone-benzene dewaxing device is the main device for the production of lubricating oil in refineries.
  • the purpose is to separate the lubricating oil and paraffin in the raw material, which is also a continuous production process.
  • the process flow chart of the process is shown in Figure 10: After the feedstock oil enters the system, it is divided into 7 channels, each of which has 3 crystallizers. Each channel firstly adds filtrate to the feedstock oil for pre-dilution, and then enters the heat exchanger E101. When the feedstock oil enters E101, fresh solvent and cold-exchanged filtrate are added for one (1) dilution. After the feedstock oil is diluted once (1) and replaced by E101, the filtrate is added to the heat exchanger E101 for one time.
  • the raw material oil After dilution, the raw material oil enters the ammonia cold crystallizer E102, E103, and then fresh solvent is added for secondary dilution.
  • the crystallized raw material oil enters the buffer tank D101 and the dewaxing filter, and the cold washing solvent is added, and after filtration, The filtrate enters the solvent recovery system via filtrate tank D104.
  • the filtered wax enters the wax tanks D105 and D106, passes through a wax deoiling filter, and is added with a cold wash, and the filtrate enters the filtrate tank D110.
  • the filtered wax from a section of the wax deoiling filter enters the wax tank D112, is mixed with fresh solvent, and then enters the second-stage wax deoiling filter and is added to the second-stage cold washing.
  • the filtrate of the two-stage wax deoiling filter enters the filtrate tank Dili and then enters the dewaxing filter wax tank D105.
  • the wax filtered by the two-stage wax deoiling filter is mixed with the added two-stage dilution solvent in the wax liquid tank D113 to enter the solvent recovery system.
  • the technical indicator to be optimized for this system is the yield of dewaxed wax.
  • the key operating variables are the solvent ratios of the various flows, as follows -
  • the optimized controller design optimization variables are 23, respectively
  • the setpoint control system uses the YOKOGAWA Centem CS system, and the optimization control computer is the operating station of the system.
  • the data window width of the relevant integral optimization control is 13 hours. Every
  • u 7 seven-way pre-dilution ratio
  • u 8 one way - thin (1) dilution ratio
  • u 16 two-way - one diluted (2) dilution ratio
  • the data acquisition system was established on the operation station with the YOKOGAWA Centum CS distributed control system and the supporting data acquisition environment. And create a real-time database with a data window width of 13 hours.
  • the database collects key operating conditions and technical indicators through YOKOGAWA Centem CS every 60 seconds.
  • the data in the data window is stored in the database. Each time the data is sampled, the data window is moved forward by one sample time. That is, the oldest data is discarded and the latest data is added to the database.
  • the original m operating condition values are the adjusted new operating conditions.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Automation & Control Theory (AREA)
  • Computing Systems (AREA)
  • Theoretical Computer Science (AREA)
  • Human Computer Interaction (AREA)
  • Manufacturing & Machinery (AREA)
  • Feedback Control In General (AREA)

Description

一种多输入多输出连续生产过程的实时操作优化方法 技术领域
本发明属于控制系统技术领域, 涉及一种多输入多输出 (Multiple
Input/Multiple Output, MIM0)连续生产过程的实时操作优化方法。 背景技术
对一个连续进料, 连续产品产出的连续的生产过程, 可以用图 1来表示: 图 中 A为上游加工装置群, 它的产品是加工装置 B的原料, 而加工装置 B的产品 为下游装置 C的原料。 A之所以称之为装置群, 是因为可能不止一套上游装置 为 B提供原料。 由于 B的产品可能不止一种, 其产品也可能提供给多个下游装 置作为原料, 因此其下游的装置称之为装置群 C。值得指出, 这种装置物料的串 接流动是连续的。 对于加工装置 B来说, 有一些生产中可以调整关键因数, 也 就是可控的关键操作条件 U,而这些操作条件又与加工装置 B的一些技术指标 J, 如能耗, 经济效益, 产品产率有关, 这里 U可看作生产过程的输入, 可以有一 个或多个, J看作为生产过程的输出, 也可以有一个或多个。 操作优化的任务是 如何在生产中调整这些操作条件, 使得在某一时间段内指定的技术指标 (可以是 一个, 或多个的组合)达到最优 (最大, 如经济效益; 或最小, 如能耗), 如图 2 所示。值得指出, 目标函数可能会根据生产的需要进行切换。对单一目标函数过 程的操作优化, 一般以下述形式描述:
J* = J(u*)=m-J(u) (1-1) 式中, J为目标函数, J*为目标函数最优值, u为操作条件, 或称调优变量。 ιΤ为调优变量的最佳值。 由于通常 U与 J的函数关系 J(U)为未知函数,为了得到 lT,有两种常用的 方法: 数学模型与实时在线搜索法。
1.建模法
建模法的基本思想是预先建立目标函数 J与调优变量 U间的数学模型, 然 后按照该模型和约束条件, 用非线性或线性规划求出 ιΤ。
在建立数学模型的过程中,依据建模原理的不同,可分为机理建模和经验建 模两类。
所谓机理建模,就是把整个系统中各部分设备的运转机理方程,按流程结构, 利用物料平衡、能量平衡原理组合起来,从而构造一套符合具体生产过程的数学 方程。 最后按系统的投入产出、 价格等确定出目标函数与调优变量 u间的关系。
但当被描述的过程过于复杂, 或者机理本身不清, 或者基本方程不准确, 往 往使机理建模工作难以进行。此外, 一个系统的机理模型一般不具有普遍性, 甚 至当加工产品变化, 或流程稍作改动, 就必须对模型做出修改或者推倒重来。
经验建模法就是依靠大量试验或日常运行报表数据为基础,构造系统的调优 变量与目标函数间的经验关系。这种建模的优点是: 简单、 具有普遍性。无论过 程或系统如何复杂或不同,都可用同样简单的手段建模,而不需要专门的工艺知 识和先验方程。
但是, 这种方法的可靠性较差。 当模型在线应用时的工作条件偏离或者超出 建模数据采集的工况时,模型将产生很大的误差而无法实用,工艺设备的微小改 动, 可能会导致模型结构很大的变化而使建模工作前功尽弃。
2.在线搜索法
搜索法是一种普遍性的方法,与具体过程无关。其基本思想是在线地变更调 优变量的数值, 观察目标函数的变化情况, 从而确定调优变化方向是否正确。从 原则上讲, 许多非线性规划方法, 如黄金分割法, 最速下降法等都能在线应用。 但是, 这些方法往往对干扰的敏感性很强。众所周知, 在通常情况下, 目标函数 不仅是调优变量的函数, 也是其他不可控变量(环境变量)的函数。 因而, 当目 标函数发生变化时,我们很难判断它是起因于调优变量的变化,还是干扰的作用。 在目前的在线搜索法中,一般都把调优变量与目标函数作为唯一的因果关系, 因 而当存在外界扰动时, 有可能做出错误的判断, 甚至使动作反向。
值得指出, 无论是模型法还是直接搜索法, 一般是建立在 1 - 1的数学描述 基础之上。 而反映目标函数 J与 U之间的关系被定义为代数关系, 也没有包含 外界干扰项. 因而, 由此导出的算法原则上只适用于静态的、 无干扰的系统。
在实际过程中, 情况要复杂得多。 首先, 调优变量不仅与目标函数在逻辑上 有因果关系, 而且在时间上也存在着动态过程, 也就是说当调优变量改变时, 目 标函数并不立刻变化, 而是有着一个过渡过程。其次, 从许多工业过程的实际情 况看, 目标函数常常处在上下起伏的脉动之中, 很难找到静态的情形, 这通常是 由于那些不可测、不可控的强扰动而引起的。例如生产过程中, 原料成份的变化 往往是不可控的, 由于在线成份检测的困难, 这些变量通常也是不可测的。而另 一方面,许多过程对于原料成份的变化却十分敏感,结果成份的变化对于目标函 数的影响常常要超出可控的温度、压力诸因素对目标函数的效应,有时可达数十 倍之多。这样一来, 调优变量变化所引起的目标函数变化成份往往要 "淹没 "于原 料成份对目标函数的干扰之中。在这种动态扰动的情况下,传统的方法就显得无 能为力。 发明内容
本发明所要解决的技术问题是针对现有技术中多输入多输出连续生产过程 的实时操作优化所存在的过程复杂、难度大、受干扰因素多、适应面窄、效果差 的缺陷,提供一种多输入多输出连续生产过程的实时操作优化方法,简化生产过 程的优化控制过程、 操作简单、 抗干扰性强、 通用性强、 适应面广、 效果好。
解决本发明技术问题所采用的技术方案是采用一种多输入多输出连续生产 过程的实时操作优化方法,该法将生产过程中多个关键操作条件为优化变量, 以 与关键操作条件相关联的一个或多个技术指标为目标函数,根据生产过程关键操 作条件和技术指标的动态历史数据,在线计算出关键操作条件与技术指标之间在 当前时间的梯度向量,然后根据这个梯度向量来确定操作条件的调整方向, 当梯 度向量为负值或正值时,都要调整关键操作条件使梯度向量向零的方向变化,使 技术指标达到最优,这种梯度的计算是在线不断进行的,无论技术指标是否达到 最优, 一旦发现梯度不为零, 则调整关键操作条件使梯度向量向零的方向变化, 以实现最优点的跟踪。调整关键操作条件使梯度向量向零的方向变化的过程就是 关键操作条件优化的过程, 由于生产过程的最优点可能会随时间变化, 因此这种 优化过程是在线不断进行的。 ·
在线计算出关键操作条件与技术指标之间在当前时间的梯度向量的方法是 利用相关积分技术或者其他可能的方法(如动态模型辩识方法)对生产过程关键 操作条件和技术指标的动态历史数据进行运算。
相关积分实时优化方法的依据是相关积分理论。相关积分是一种与随机过程 有关的运算。 在相关积分理论中, 目标函数、 干扰、 优化变量被视为随机过程, 而优化变量为均值可控, 一般地先确定目标函数 (t),该目标函数应当是在线 可计算的或在线可测量的, 那么目标函数可以表达为:
J(t) = f(u(t),p(t),t)
其中, S(t)为 m维均值可控优化变量
Figure imgf000006_0001
(t)为干扰, f为未知映射; 最优目标函数定义为:
J * ( = m x E{f[E{u(t)}, (t), t]}
E{u(()}
这里 E{S(t)}为调优变量的均值, 它可以是基层控制器的设定值或阀位等。 对于这个多变量的优化问题可以证明,在一定条件下目标函数均值对调优变量均 值的梯度 „ f、、满足下式 - dE{n(t)}
Figure imgf000006_0002
中, s(t)是均值为零的噪声项, 而 是调优变量与目标函数间互相关积分向 , 定义为 , i = 1,2,Λ m
Figure imgf000006_0003
k„„为调优变量自相关积分矩阵, 定义为
Figure imgf000006_0004
根据公式 kuiuJ 和优化变量的实 时测 量值 M; (t)
Figure imgf000007_0001
(/ = 1,2,Λ , m)计算优化变量的自相关积分矩阵 k„„;其中 T, Μ为大于 0的积 分常数;
根据公式
[ u(t - )J{t)dtdr, (i = 1,2,Λ , m)和目标函数的实时测量值 J(t)
Figure imgf000007_0002
计算优化变量与目标变量的互相关积分向量; 其中 Τ, Μ为大于 0的积分 常数;
上述公式中, «;. (t), ( = l,2,A , m) , J(t)分别为调优变量和目标函数的测量值。 可见 k,^ „能够通过调优变量和目标函数的观测值计算得到, 于是根据公式
(可用最小二乘法估
Figure imgf000007_0003
在算得目标函数的梯度后, 可以用直接迭代计算调优变量的新设定值 u,(/ + l) dEf
根据公式 I + 1) = + 计算优化变量的新设定值 (/ + 1);
dE{ (l)} 式中, α为常数, 如果优化目标是求极大值, 则取为大于 0的数, 如果优 化目标为极小值, 取为小于 0的数。
这种迭代过程在线持续进行, 直至梯度为零。
对于多目标的情况, 也有类似的结论。 由上, 可得出本发明的具体进一步的技术方案如下:
1. 根据被优化的过程的需要, 确定多个要优化的技术指标即目标函数 J15J23A 5J„, 这些目标函数必需是在线可以计算或测量的。 构造一个虚拟的综 合目标函数 J = + 2J2 + ,Λ ,+anJn, 这里 , σ2 ,Λ , σ„为各目标函数的加权 数, 根据工艺要求, 取值为 0到 1之间, 需要指出, 由于生产中目标函数可能切 换, 因此这里加权值可能是随时间变化的。
根据生产工艺的要求, 确定要优化的关键操作条件 ^,^,Λ 作为要优化 的变量。
对要优化的关键操作条件进行常规定值控制,而其设定值由优化控制计算机 釆用相关积分技术进行计算,根据工艺过程的具体情况和要求,每隔一定的时间 周期进行一次设定值的调整;
对要优化的关键操作条件进行常规定值控制是先由集散系统计算机即 DCS 或常规仪表对要优化的操作条件进行常规定值控制,而其设定值由优化控制计算 机采用相关积分技术进行计算,每隔一定的时间周期(时间周期由具体的工艺过 程的快慢来确定)进行一次设定值的调整。
2. 采集关键操作条件和各技术指标 (即各目标函数) 的数据。 其方法是根 据具体的过程时间特性,建立具有一定数据窗宽度(该数据库的时间宽度应当大 于该工艺过程优化变量到目标函数过渡过程时间的 3倍以上,一般为 8— 18小时) 的实时数据采集数据系统, 通常该系统由集散控制系统 (即 DCS)构成, 以获 得生产过程关键操作条件和技术指标的历史数据。该系统每隔一定的采样时间间 隔 (该时间间隔由工艺过程的快慢来决定, 一般为 30-90秒)采集关键操作条 件和各技术指标即各目标函数的数据。数据窗里的数据存储在数据库之中,每采 一次样, 数据窗就向前移动一个采样时间, 也就是说, 最老的数据被抛弃, 而最 新的数据被加入到数据库之中, 如图 3所示为一个 2个操作条件的例子。
3.数据釆样完毕后,对各操作条件进行自相关积分矩阵 kTO的计算:设有 m 个操作条件
Figure imgf000009_0001
式中: = H ίΓ λ― τ)άλάτ i, j = \,2,K m Τ, Μ为积分常数。
4.计算各操作变量与技术指标之间的互相关积分矩阵 KUJ: 设有 n个目标函数
Figure imgf000009_0002
K UJ1
Figure imgf000009_0005
Figure imgf000009_0003
( = 1,2,Λ ,m;s = \,2, ,ή)
T, M为积分常数。
根据相关积分理论,以上的 M应当大于从优化变量到目标函数最大的时间常数, 而 T取为 1-5倍的 M。
5. 根据以上操作条件的自相关积分矩阵和操作变量与各技术指标间的互相 关积分向量, 计算出操作条件对综合技术指标间的梯度向量 K 。 先通过计算以 下的线性方程得到 KD :
Figure imgf000009_0004
k uu 0 Λ 0
0 kuu 0 0
M 0 0 M
0 Λ 0 kuu 而
Figure imgf000010_0001
综合目标函数 J对操作条件的梯度 κ 为 κχ + σ2Κ^,Λ ,Κ Jn
6.根据所得到的操作条件对综合技术指标的梯度向量 K^,计算出操作条件的 变化方向,其原则是如果计算出来的梯度为零,则该操作条件目前已经在最优状 态; 如果不为零, 按照梯度的大小与方向进行对操作条件调整。例如: 现在的操 作条件值是已知的,按照下面的方法求出调整后的操作条件值,要求综合技术指 标最大化:
式 为调整后的
Figure imgf000010_0002
新操作条件 (第 +1步) 的值。
Figure imgf000010_0003
a2,A ,am为 m个正的常数 (如果求技术指标最大值)。 我们以一个操作条件的 生产过程以求技术指标最大值为例, 图示调整的方法。见图 4、 图 5、 图 6所示。 图 4中示出了当用相关积分方法计算出梯度为零时,技术指标为最大值,操作条 件不用调整这种情况。图 5中示出了当用相关积分方法计算出梯度为负时,操作 条件应当减小,才能提高技术指标。图 6中示出了当用相关积分方法计算出梯度 为正时, 操作条件应当增大, 才能提高技术指标。 显然,对第 i个操作条件 ui来说,每次调整步长为 αΛ^,只要取适当的 ^的 值, 即如果求技术指标的最大值, 取 α,.为正值, 否则取负值, 就可以调整步长 的大小和方向。
调整完成后, 间隔一定的时间 (30— 90秒), 再次进行数据采样, 返回步骤
3。
步骤 3至 6的过程是在线不断进行的,可以使各操作点最终达到最优点。需 要指出的是即便是各梯度已经到了零,但是步骤 2到 5还应当不断地进行,这是 因为技术指标与操作条件之间的函数关系可能随着时间而变化(如原料性质的变 化, 装置的改造, 等), 需要不断地观测梯度是否为零, 如果发生变化, 随时进 行调整。如图 Ί所示:当技术指标与操作条件的关系由于某种原因(如原料变化) 发生变化, 使得当前的操作点不再是最优点时, 可以发现梯度不为零, 按照相关 积分的方法将操作条件向最优点迫近。
这就使得本方法能够随时发现生产过程是否偏离了最优点,并进行最优点的 追踪。
相关积分实时优化方法有以下不同以往优化技术的特点:
• 本发明采用了相关积分方法进行梯度计算, 它是根据操作条件与技术指 标在一段时间内 (即数据窗内) 的波动数据进行计算, 因此与模型方法 不同, 在优化操作时不要求生产过程处于静态, 可以处于波动状态, 而 且根据相关积分理论, 这些波动的统计特征可以是未知的。
• 可以不需要事先建立过程的静态和动态模型, 大大减少生产过程操作优 化的复杂性。传统的方法, 无论是机理建模还是统计建模都是事先设法 求出技术指标 J与操作条件 U之间的函数关系,然后在这个函数关系的 基础上去求出最佳的操作条件。如果系统很复杂, 就很难得到准确的模 型,所需的代价也很高。而本发明注意到只要有当前操作点附近的信息, 无须建立大范围的模型即可以对系统进行优化操作, 可以省去复杂的过 程建模。
• 本法具有很强的自适应性能。 也就是说, 当生产过程由于某种原因发生 变化, 使得操作条件偏离最优点时, 本发明可以发现这种偏离, 并且自 动地把操作条件重新调整到最佳点上。这种自动追踪最优点的性能在实 际生产中具有重要实用价值。因为许多生产过程中操作条件与技术指标 之间的关系实际上是随着原料性质的变化, 设备的老化、 更替, 催化剂 的改良等而变化。 而传统的建模法就需要对生产过程重新建模, 或对模 型进行修正。
• 本方法是一种通用性的方法。 只要生产过程是连续的, 要优化的技术指 标是可以在线测量或计算的, 都可以使用本方法, 与具体的生产过程无 关, 可以应用的范围很广。而传统的建模法必须针对某一具体的装置进 行建模, 建立的优化模型具有特定适用性。
• 利用过程正常运行的自然波动进行工作, 无需另对过程加入测试信号, 因而对过程操作的干扰很小。
• 该法具有很强的抗干扰特性, 甚至在动态强干扰, 即其它因素如原料性 质引起技术指标的变化大于有用信号 (由被优化的操作条件引起目标函 数的变化)的恶劣条件下仍能正常工作,这种强抗干扰性在实际应用中具 有重要的意义。
本发明中采用相关积分实时优化方法的核心是利用相关积分技术的原理计 算目标函数对操作条件的梯度,然后根据梯度进行操作条件的调整, 并且梯度的 计算与操作条件的调整不断地在线进行。 因此, 无论用什么计算方法, 只要能在 线计算操作条件与目标函数的梯度的都可以用这一原理对装置进行在线优化。因 此采用其他可能的方法如动态模型辩识方法同样能在线计算操作条件与目标函 数的梯度, 进而对装置的连续生产操作进行实时优化。 附图说明
图 1 一个连续进料、 连续产品产出的连续的生产过程方框图
图 2 为连续生产过程的操作优化流程示意图
图 3 具有 2个操作条件的数据窗数据处理采集示意图
图 4 为操作条件与技术指标的曲线的梯度为零时的曲线图
图 5 为操作条件与技术指标的曲线的梯度为负时的曲线图
图 6 为操作条件与技术指标的曲线的梯度为正时的曲线图
图 7 为操作条件与技术指标的关系发生变化时期曲线的梯度随之变化的曲 线图 图 8 为本发明采用计算机对连续生产过程进行优化操作的控制图 图 9 为 ARGG装置反应再生系统流程图
图 10 为酮苯脱油脱蜡联合装置工艺流程图
图中: 1-提升管反应器 2-旋风分离器 3-沉降器 4-催化剂再生器 5-外取热 器, 6-分馏系统 具体实施方式
以下是本发明的非限定实施例,在实施时,要求生产过程采用计算机进行优 化控制, 先由集散系统计算机(即 DCS) 或常规仪表对要优化的操作条件进行 常规定值控制,而其设定值由优化控制计算机采用本发明的方法进行计算,每隔 一定的时间周期进行一次设定值的调整,这个调整的周期与具体的工艺过程的快 慢来决定。 如图 8所示。 .
实施例 1: ARGG装置的在线相关积分优化控制
ARGG装置是石油化工厂用于将低价值的油料裂解成高价值的液态烃、 汽 油、 柴油的装置, 属于连续生产过程。 ARGG装置反应再生系统流程图如图 9 所示。来自罐区的原料油与回炼汽油、 回炼油和回炼油浆混和, 在提升管反应器 1下部雾化喷入提升管反应器 1, 进入反应系统。
雾化的反应原料和喷入的蒸汽与来自催化剂再生器 4的高温催化剂在提升管 反应器 1下部混和,沿提升管反应器 1上升进行催化裂化反应, 反应后的油气连 同催化剂在提升管出口处进入旋风分离器 2迅速将油气与固体颗粒状的催化剂 分离, 分离出的油气送入分馏系统 6。
分离出的催化剂进入催化剂再生器 4, 这些催化剂表面有在反应过程中产生 的积碳, 这些积碳在再生器中燃烧除去, 这个过程称为烧焦。烧焦过程中产生的 过剩热量由外取热器 5取走。经烧焦再生后的催化剂再进入提升管反应器 1底部, 与原料混合再进行反应。 为了保持催化剂的活性和补充消耗的催化剂, 外部还 向催化剂再生器补充新催化剂。
来自分离器的油气进入分馏系统 6, 经该系统的分离, 产出液态烃, 汽油, 轻柴油。 而产出的回炼油浆和回炼油和部分汽油返回到提升管反应器 1。
在该装置的应用实例中, 以下几个关键操作变量选为优化变量-
• 提升管反应器出口温度
• 预提升干气流量
• 进料温度
• 终止剂流量
• 回炼汽油比
• 回炼油流量
• 回炼油浆流量 • 新催化剂加料量
而被优化的技术指标 (目标函数)有 5个:
• 液态烃收率
• 汽油收率
• 柴油收率
• 总液收收率
• 装置总经济效益 优化控制系统的结构如图 8所示, 给定值控制系统由 HONEYWELL TPS3000的 集散控制系统组成, 而优化控制由 HONEYWELL TPS3000的 APP (应用过程处 理器)计算机承担。根据当前工艺的要求从以上五个技术指标中选一个作为当前 要优化的技术指标。 并按照以下步骤进行 .
1.命名各关键操作变量
u尸提升管反应器出口温度
U2=预提升干气流量
u3=进料温度
U4=终止剂流量
U5=回炼汽油比
U6=回炼油流量
u7=回炼油浆流量
u8=新催化剂加料量 命名各目标函数:
Jl=液态烃收率
J2=汽油收率
J3=柴油收率
J4=总液收收率
J5=装置总经济效益
2. 利用 HONEYWELL TPS3000集散系统和 APP计算机建立了一个基于 OPC的 数据采集系统。 在 APP计算机中建立具有数据窗宽度为 8小时的实时数据库。 每隔 60秒釆集关键操作条件和各技术指标 (各目标函数) 的数据。 数据窗里的 数据存储在数据库之中, 每采一次样, 数据窗就向前移动一个采样时间, 也就是 说, 最老的数据被抛弃, 而最新的数据被加入到数据库之中。
3.数据采样完毕后,对各操作条件进行自相关积分矩阵 kw的计算:现有 8个操 作条件
Figure imgf000015_0001
(3600 1 0800
式中: k'-i = —— u} (X)u , (X - τ)άλάτ
y J-3600 29TT7 -10800 1 J
上述积分中的上下限值 3600、 -3600、 10800、 -10800 是根据该过程的从优 化变量到目标函数通道的大致的响应时间来确定的。
4. 计算各操作变量与技术指标之间的互相关积分矩阵 KUJ: 现有 5个目标函数
Figure imgf000015_0002
式中:
Figure imgf000015_0003
Figure imgf000015_0004
f棚 1 (Ϊ0800
—— u, ( ) J'. (λ― τ)άλάτ
3600 2T 上画 ο
(i = 1,2,Λ ,8;^ = 1,2,Λ ,5)
5. 根据以上的操作条件的自相关积分矩阵与操作变量与各技术指标间的互 相关积分向量, 计算出操作条件对综合技术指标间的梯度向量 Κ 。 先通过 计算以下的线性方程得到 Kd:
KUJ = KUUK
Figure imgf000016_0001
而 d
K J2
d =
M
K 综合目标函数 J对操作条件的梯度 K 为
Figure imgf000016_0002
式中, ^,入, 5为5个0或1的常数, 根据当前优化的目标函数而定。例如, 当前要求优化的是液态烃收率, 则取 σ1 = 1, 其余取为 0, 其他情况依次类推。 6.根据所得到的操作条件对综合技术指标的梯度向量 Κ ,计算出操作条件的 变化方向,其原则是如果计算出来的梯度为零,则该操作条件目前已经在最优状 态; 如果不为零, 按照梯度的大小与方向进行对操作条件调整。例如: 现在的操 作条件值是已知的,按照下面的方法求出调整后的操作条件值,要求技术指标最 大化-
Figure imgf000016_0003
式中: + 1)
为原来的 8个操作条件值, 而 "2 (7
为调整后的新操作条件
M M
"8 (7+ 1) M σ
.
2,入,《8为 8个正的常数, 这些常数的大小与优化控制的收敛速度相关, 需 要在现场进行调整。 根据现场的调整结果, 均取 0.001。
7. 回到第 3步。 根据实际系统的测试, 对液态烃收率、总液收收率、和装置经济效益这三种技术 指标得到如下的结果- 优化前后的技术指标的对比
Figure imgf000017_0001
从以上的测试结果看, 该系统有着较好的应用效果。 实施例 2: 酮苯脱油脱蜡联合装置的在线相关积分优化控制
酮苯脱蜡装置是炼油厂润滑油生产的主要装置,目的是把原料中的润滑油和石蜡 分离, 也是属于连续生产过程。 该过程的工艺流程图如图 10所示: 原料油进入系统后分成 7路, 每路设有 3个结晶器, 各路首先向原料油添 滤液以进行预稀释, 之后进入换热器 E101 , 在原料油进入 E101时, 加入新鲜溶 剂和换冷后的滤液以进行一次 (1 )稀释, 原料油经一次 (1 )稀释并经 E101换 冷后, 在换热器 E101后加入滤液以进行一次(2)稀释, 之后原料油进入氨冷结 晶器 E102、 E103, 然后加入新鲜溶剂以进行二次稀释, 结晶后的原料油进入缓 冲罐 D101、 脱蜡滤机, 期间加入冷洗溶剂, 过滤后的滤液经滤液罐 D104进入 溶剂回收系统。过滤出来的蜡进入蜡液罐 D105和 D106,经过一段蜡脱油滤机, 期间加入一段冷洗, 其滤液进入滤液罐 D110。 一段蜡脱油滤机的过滤出的蜡进 入蜡液罐 D112, 加入新鲜溶剂与之混合后, 进入二段蜡脱油滤机并加入二段冷 洗。 二段蜡脱油滤机的滤液进入滤液罐 Dili , 然后进入脱蜡滤机蜡液罐 D105。 二段蜡脱油滤机过滤出来的蜡在蜡液罐 D113中与加入的二段稀释溶剂混合后形 成的脱油蜡进入溶剂回收系统。
该系统要优化的技术指标是去油蜡的收率。 而关键的操作变量 (优化变量) 是各流各路的溶剂比, 具体如下- 优化控制器的设计优化变量是 23个, 分别是
1、 一路预稀释比
2、 二路预稀释比
3、 三路预稀释比
4、 四路预稀释比
5、 五路预稀释比
6、 六路预稀释比
7、 七路预稀释比
8、 一路一稀 (1 )稀释比
9、 二路一稀 (1 ) 稀释比
10、 三路- -稀 ( 1 ) 稀释比
11、 四路一 -稀 ( 1 ) 稀释比
12、 五路一稀 ( 1 ) 稀释比
13、 六路- -稀 ( 1 ) 稀释比
14、 七路一 -稀 ( 1 ) 稀释比
15、 一路— -稀 (2) 稀释比
16、 二路— (2) 稀释比
17、 三路一 -稀 (2) 稀释比
18、 四路一 -稀 (2) 稀释比
19、 五路一 -稀 (2) 稀释比
20、 六路一 -稀 (2) 稀释比
21、 七路一 -稀 (2) 稀释比
22、 二次比
23、 冷洗比 在这个系统中,给定值控制系统用的是 YOKOGAWA Centem CS系统,优化控制 计算机是该系统的操作站。 相关积分优化控制的数据窗宽度是 1 3小时。 每隔
60秒采集关键操作条件和技术指标 (目标函数: 去油蜡的收率) 的数据。 按照 以下步骤进行控制计算:
1.命名各关键操作变量
ui=—路预稀释比
u2=二路预稀释比
U3=三路预稀释比
U4= 四路预稀释比
U5=五路预稀释比
U6=六路预稀释比
u7=七路预稀释比 u8=一路—稀 (1) 稀释比
u9=二路— -稀 (1) 稀释比
Ul。=三路-一稀 (1) 稀释比
Un= 四路-一稀 (1) 稀释比
12=五路-一稀 (1) 稀释比
u13=六路-一稀 (1) 稀释比
u14=七路-一稀 (1) 稀释比
u15=一路-一稀 (2) 稀释比
u16=二路-一稀 (2) 稀释比
u17=三路-一稀 (2) 稀释比
u18= 四路-一稀 (2) 稀释比
u19=五路- —稀 (2) 稀释比
u20=六路- —稀 (2)稀释比
21=七路-一稀 (2) 稀释比
22=二次比
u23=冷洗比 命名目标函数:
Jl=去油蜡的收率
2.用 YOKOGAWA Centum CS集散控制系统和配套的数据采集环境在操作 站上建立了数据采集系统。 并建立具有数据窗宽度为 13小时的实时数据库。 该 数据库每隔 60秒通过 YOKOGAWA Centem CS釆集关键操作条件和各技术指标
(各目标函数)的数据。数据窗里的数据存储在数据库之中, 每采一次样, 数据 窗就向前移动一个采样时间, 也就是说, 最老的数据被抛弃, 而最新的数据被加 入到数据库之中。
3.数据釆样完毕后, 对各操作条件进行自相关积分矩阵 kOT的计算: 现有 m=23个操作条件
Figure imgf000019_0001
i,j = \,2, m T=7200, M=14000为积分常数, 该常数是由脱蜡过 程大致的时间常数决定的。 4.计算各操作变量与技术指标之间的互相关积分矩阵 K κ UJ
式中:
uUl
k, u2J\
K UJ1
M
umj\
Figure imgf000020_0001
(i = 1,2,Λ , m)
T, M为积分常数, T=7200, M=14000。 5. 根据以上的操作条件的自相关积分矩阵与操作变量与各技术指标间的互 相关积分向量, 计算出操作条件对综合技术指标间的梯度向量 K 。 先通过计算 以下的线性方程得到 Kd:
KUJ = KuuKd kuu 0 Λ 0
0 0 0
M 0 0 M
0 Λ 0 kuu
综合目标函数: r对操作条件的梯度 Kd ff
式中:
σ, = 1
6.根据所得到的操作条件对综合技术指标的梯度向量 Κ^,计算出操作条件 的变化方向,其原则是如果计算出来的梯度为零,则该操作条件目前已经在最优 状态; 如果不为零, 按照梯度的大小与方向进行对操作条件调整。例如: 现在的 操作条件值是已知的,按照下面的方法求出调整后的操作条件值,要求技术指标 最大化:
Figure imgf000021_0001
式中: u2 {l)
为原来的 m个操作条件值, 而 为调整后的新操作条件 。 M M
",,, ( + 1)
Figure imgf000021_0002
«12,八,0;„1111个正的常数, 这些常数的大小与优化控制的收敛速度相关, 需 要在现场进行调整。 根据现场的调整结果, 均取值为 0.001。 .
7. 回到第 3步。 根据测试我们得到了下表: 优化前后产品收率
Figure imgf000021_0003
从以上结果中可以看出, 要优化的技术指标, 去油蜡的收率增加了 1.21%, 有一 定的优化效果。

Claims

权利要求书
1、 一种多输入多输出连续生产过程的实时操作优化方法, 其特征在于将生 产过程中多个关键操作条件为优化变量,以与关键操作条件相关联的一个或多个 技术指标为目标函数,根据生产过程关键操作条件和技术指标的历史数据,在线 计算出关键操作条件与技术指标之间在当前时间的梯度向量,然后根据这个梯度 向量来确定操作条件的调整方向, 当梯度向量为负值或正值时,都要调整关键操 作条件使梯度向量向零的方向变化,使技术指标达到最优,这种梯度的计算是在 线不断进行的, 无论技术指标是否达到最优,一旦发现梯度不为零, 则调整关键 操作条件使梯度向量向零的方向变化, 以实现最优点的跟踪。
2、 根据权利要求 1 所述的多输入多输出连续生产过程的实时操作优化方 法, 其特征在于在线计算出关键操作条件与技术指标之间在当前时间的梯度向 量的方法是利用相关积分技术对生产过程关键操作条件和技术指标的历史数据 进行积分运算。
3、 根据权利要求 2所述的多输入多输出连续生产过程的实时操作优化方 法, 其特征在于所述相关积分技术具体包括- 构造目标函数 J(t),该目标函数应当是在线可计算的或在线可测量的, 那么: (t) = wt , /), 其中, (t)为 m维均值可控优化变量, (t)为干 扰, /为未知映射;
根据公式
A
A Jr
M M M M
A ^um m
kuiuJ = ^ ^t - T^j i dtdT和优化变量的实 时测 量值 u人 t
1,2,Λ ,m)计算优化变量的自相关积分矩阵 k„,,;其中 T, M为大于 0的积 分常数;
根据公式
Figure imgf000023_0001
hu = £ £ M0 - r)J{t)dtdr, · = 1,2,Λ , )和目标函数的实时测量值 J(t) 计算优化变量与目标变量的互相关积分向量; 其中 T, Μ为大于 0的积分 常数;
dEf , ,、 ... 、,,, ,, ,、, _ Mi
根据公式 =fc ·ε( 求出目标函数的梯度
dE{u(t)} dE{u(t)} dEf
根据公式 1 + 1) = + α 计算优化变量的新设定值 Ui(/ + 1);
dE{u(l)} 式中, α为常数,
4、根据权利要求 3所述的多输入多输出连续生产过程的实时操作优化方法, 其特征在于包括以下步骤:
(1) 根据被优化的过程的需要, 确定多个要优化的技术指标即目标函数 J15J2,A,J„, 这些目标函数必需是在线可以计算或测量的,构造一个综合的目标 函数
Figure imgf000023_0002
TnJ,,, 这里 σι2,Λ ,σ„为各目标函数的加权数, 根 据工艺要求, 取值为 0到 1之间, 这里加权值可能是随时间变化的,
根据生产工艺的要求, 确定要优化的关键操作条件 ^,^,Λ 作为要优化 的多个变量;
对要优化的关键操作条件进行常规定值控制,而其设定值由优化控制计算机 采用相关积分技术进行计算,根据工艺要求,每隔一定的时间周期进行一次设定 值的调整;
(2) 建立具有数据窗的实时数据采集系统, 采集关键操作条件和技术指标 的数据, 以获得生产过程关键操作条件和技术指标的历史数据; (3 )数据采样完毕后, 对各关键操作条件进行自相关积分矩阵 kutJ的计算 设有 Π1个操作条件, 则
Figure imgf000024_0001
式中: · = J IT £ ( 一 τ)άΜτ
i, j = 1,2,A m T, M为积分常数,
在自相关积分矩阵中包含了各操作条件之间相关性的信息,
(4) 计算各操作变量与技术指标之间的互相关积分矩阵 kUi : 设有 n个目 标函数, 则
Figure imgf000024_0002
Figure imgf000024_0003
(X)J. (λ - τ)άλάτ
= 1,2,Λ ,m;s = 1,2,Λ,"
Τ, Μ为积分常数;
(5 ) 根据以上的操作条件的自相关积分矩阵与操作变量与各技术指标间的 互相关积分向量, 计算出操作条件对综合技术指标间的梯度向量 κ 先通 过计算以下的线性方程得到 Kd:
K UJ
0 Λ 0
0 kuu 0 0
K UU
M 0 0 Μ
0 Λ 0 kuu
Figure imgf000025_0001
综合目标函数 J对操作条件的梯度 K 为
= σχ Α η + σ2Κ^2,Λ ,Κ Jn
(6) 根据所得到的操作条件对综合技术指标的梯度向量 Κ , 计算出操作条 件的变化方向, 其原则是如果计算出来的梯度为零, 则该操作条件目前已经在最 优状态; 如果不为零, 按照梯度的大小与方向进行对操作条件调整;
(7) 调整完成后, 间隔一定的时间, 再次进行数据采样, 返回步骤 3, 步骤 3至 6的过程是在线不断进行的, 可以使各操作点最终达到最优点。
5、根据权利要求 4所述的多输入多输出连续生产过程的实时操作优化方法, 其特征在于步骤 (1 ) 中对要优化的关键操作条件进行常规定值控制是先由集散 系统计算机即 DCS或常规仪表对要优化的操作条件进行常规定值控制, 而其设 定值由优化控制计算机采用相关积分技术进行计算,每隔一定的时间周期进行一 次设定值的调整。
6、根据权利要求 4所述的多输入多输出连续生产过程的实时操作优化方法, 其特征在于步骤 (2) 中, 是根据具体的过程时间特性, 建立具有数据窗的实时 数据采集系统,采集关键操作条件和技术指标数据的方法是建立具有一定数据窗 宽度的实时数据釆集数据系统, 一般该系统由集散控制系统即 DCS构成, 以获 得生产过程关键操作条件和技术指标的历史数据,该系统每隔一定的采样间隔采 集关键操作条件和各技术指标即各目标函数的数据,数据窗里的数据存储在数据 库之中, 每采一次样, 数据窗就向前移动一个采样时间, 最老的数据被抛弃, 而 最新的数据被加入到数据库之中。
7、根据权利要求 6所述的多输入多输出连续生产过程的实时操作优化方法, 其特征在于步骤 (2) 中, 建立的数据窗宽度的时间宽度大于该工艺过程优化变 量到目标函数过渡过程时间的 3倍以上。
8、根据权利要求 6所述的多输入多输出连续生产过程的实时操作优化方法, 其特征在于步骤 (2) 中, 实时数据采集数据系统采集关键操作条件和各技术指 标即各目标函数的数据是根据工艺过程的快慢每隔 30-90秒采集一次。
9、根据权利要求 4所述的多输入多输出连续生产过程的实时操作优化方法, 其特征在于步骤 (6) 中, 计算出来的梯度为零, 则该操作条件目前已经在最优 状态; 计算出来的梯度不为零, 按照梯度的大小与方向进行对操作条件调整, 在 操作条件值是已知的,按照下面的方法求出调整后的操作条件值,要求综合技术 指标最大化:
Λ 0
w2(/ + l) 0 0 M u2 {l)
M M 0 0 M
",,,( + 1) 0 0 r,
式中: 为调整后的新操作条件
Figure imgf000026_0001
如果是求技术指标最大值, 则 Α, ,Λ ,am为 m个正的常数。
10、 根据权利要求 7所述的多输入多输出连续生产过程的实时操作优化方 法, 其特征在于步骤(6)中, 对第 i个操作条件 Ui来说, 每次调整步长为《, , 取适当的 α;的值即如果求技术指标的最大值, 取 为正值, 否则取负值, 以调 整步长的大小。
PCT/CN2005/002324 2005-07-20 2005-12-27 Méthode optimisée de fonctionnement temps réel de procédure multi-entrée et multi-sortie de fabrication en continu WO2007009322A1 (fr)

Priority Applications (2)

Application Number Priority Date Filing Date Title
US11/988,889 US7848831B2 (en) 2005-07-20 2005-12-27 Real-time operating optimized method of multi-input and multi-output continuous manufacturing procedure
CA2615727A CA2615727C (en) 2005-07-20 2005-12-27 Real-time operating optimized method of multi-input and multi-output continuous manufacture procedure

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN200510085063 2005-07-20
CN200510085063.2 2005-07-20

Publications (1)

Publication Number Publication Date
WO2007009322A1 true WO2007009322A1 (fr) 2007-01-25

Family

ID=37668433

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2005/002324 WO2007009322A1 (fr) 2005-07-20 2005-12-27 Méthode optimisée de fonctionnement temps réel de procédure multi-entrée et multi-sortie de fabrication en continu

Country Status (5)

Country Link
US (1) US7848831B2 (zh)
KR (1) KR100979363B1 (zh)
CA (1) CA2615727C (zh)
SG (1) SG164373A1 (zh)
WO (1) WO2007009322A1 (zh)

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20120239169A1 (en) * 2011-03-18 2012-09-20 Rockwell Automation Technologies, Inc. Transparent models for large scale optimization and control
US8874242B2 (en) * 2011-03-18 2014-10-28 Rockwell Automation Technologies, Inc. Graphical language for optimization and use
WO2013142556A2 (en) * 2012-03-21 2013-09-26 Salt Water Solutions, Llc Fluid treatment systems, methods and applications
US11663679B2 (en) 2019-10-11 2023-05-30 International Business Machines Corporation Generating mode change alerts with automatic detection from sensor data
CN111177854B (zh) * 2019-12-31 2023-05-30 西北工业大学 基于直接搜索法的起落架收放机构的优化设计方法
US11868932B2 (en) 2020-09-30 2024-01-09 International Business Machines Corporation Real-time opportunity discovery for productivity enhancement
CN112486111B (zh) * 2020-11-17 2021-12-14 湖南师范大学 一种基于数据分析的食用油碱炼工艺智能调节方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH07135474A (ja) * 1993-11-09 1995-05-23 Nec Corp スペクトル拡散受信装置
JP2001129044A (ja) * 1999-11-04 2001-05-15 Toshio Fukuda 体調調整装置
CN1399176A (zh) * 2002-08-30 2003-02-26 王建 相关积分法催化裂化在线优化控制系统
CN1414070A (zh) * 2002-09-26 2003-04-30 王建 相关积分法酮苯脱蜡装置蜡收率在线优化控制系统

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6701223B1 (en) * 2000-09-11 2004-03-02 Advantica, Inc. Method and apparatus for determining optimal control settings of a pipeline
WO2002088995A1 (en) * 2001-04-30 2002-11-07 Nagabhushana Prabhu Optimization on nonlinear surfaces
US20050187643A1 (en) * 2004-02-19 2005-08-25 Pavilion Technologies, Inc. Parametric universal nonlinear dynamics approximator and use
US7272454B2 (en) * 2003-06-05 2007-09-18 Fisher-Rosemount Systems, Inc. Multiple-input/multiple-output control blocks with non-linear predictive capabilities
US7117046B2 (en) * 2004-08-27 2006-10-03 Alstom Technology Ltd. Cascaded control of an average value of a process parameter to a desired value
GB2446343B (en) * 2005-12-05 2011-06-08 Fisher Rosemount Systems Inc Multi-objective predictive process optimization with concurrent process simulation

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH07135474A (ja) * 1993-11-09 1995-05-23 Nec Corp スペクトル拡散受信装置
JP2001129044A (ja) * 1999-11-04 2001-05-15 Toshio Fukuda 体調調整装置
CN1399176A (zh) * 2002-08-30 2003-02-26 王建 相关积分法催化裂化在线优化控制系统
CN1414070A (zh) * 2002-09-26 2003-04-30 王建 相关积分法酮苯脱蜡装置蜡收率在线优化控制系统

Also Published As

Publication number Publication date
KR20080033419A (ko) 2008-04-16
US20090037003A1 (en) 2009-02-05
KR100979363B1 (ko) 2010-08-31
CA2615727A1 (en) 2007-01-25
SG164373A1 (en) 2010-09-29
CA2615727C (en) 2013-03-19
US7848831B2 (en) 2010-12-07

Similar Documents

Publication Publication Date Title
WO2007009322A1 (fr) Méthode optimisée de fonctionnement temps réel de procédure multi-entrée et multi-sortie de fabrication en continu
CN104463327A (zh) 一种预测催化裂化焦炭产率的方法
CN109814513A (zh) 一种基于数据模型的催化裂化装置优化方法
CN106845796B (zh) 一种加氢裂化流程产品质量在线预测方法
CN104392098A (zh) 一种预测催化裂化汽油产率的方法
CN104463343A (zh) 一种预测催化裂化轻质油产率的方法
CN102213949B (zh) 一种乙烯装置价值优化方法
CN104789256A (zh) 一种催化裂化装置的收率实时预测方法
CN105278520A (zh) 基于t-kprm复杂工业过程运行状态评价方法及应用
CN105205327B (zh) 一种基于工况的乙烯生产能效动态评估方法
CN102663235B (zh) 一种变种群规模dna遗传算法的催化裂化主分馏塔建模方法
CN103323484B (zh) 甘蔗煮糖过程结晶状态预测方法
CN110376987A (zh) 基于案例库和稳态检测的炼油过程在线优化方法及装置
CA2339966C (en) Component mapper for use in connection with real-time optimization process
CN104657586A (zh) 基于异常工况检测的高含硫天然气净化工艺建模优化方法
US20200108327A1 (en) High Purity Distillation Process Control
CN106647285B (zh) 一种基于软测量技术的催化剂活性检测方法
CN1327376C (zh) 基于支持向量机的软测量仪表建模方法
CN103087753B (zh) 一种重质原料工业裂解炉价值最大化模型的构建方法
CN113703314A (zh) 碳三液相加氢反应器在线预测自动控制方法及系统
US10953377B2 (en) Delta temperature control of catalytic dehydrogenation process reactors
CN113536601B (zh) 一种提高连续重整装置催化剂碳含量预估精度的方法
CN101859103A (zh) 催化裂化反应深度在线计算和自适应非线性预测控制方法
CN114492988A (zh) 一种催化裂化过程的产品收率的预测方法及装置
Farsang et al. Role of steady state data reconciliation in process model development

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application
WWE Wipo information: entry into national phase

Ref document number: 11988889

Country of ref document: US

WWE Wipo information: entry into national phase

Ref document number: 2615727

Country of ref document: CA

NENP Non-entry into the national phase

Ref country code: DE

WWW Wipo information: withdrawn in national office

Ref document number: DE

WWE Wipo information: entry into national phase

Ref document number: 1020087003815

Country of ref document: KR

122 Ep: pct application non-entry in european phase

Ref document number: 05847867

Country of ref document: EP

Kind code of ref document: A1