WO2020051795A1 - 一种钢铁煤气系统长期区间预测及其结构学习方法 - Google Patents

一种钢铁煤气系统长期区间预测及其结构学习方法 Download PDF

Info

Publication number
WO2020051795A1
WO2020051795A1 PCT/CN2018/105197 CN2018105197W WO2020051795A1 WO 2020051795 A1 WO2020051795 A1 WO 2020051795A1 CN 2018105197 W CN2018105197 W CN 2018105197W WO 2020051795 A1 WO2020051795 A1 WO 2020051795A1
Authority
WO
WIPO (PCT)
Prior art keywords
interval
prediction
data
layer
membership
Prior art date
Legal status (The legal status 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 status listed.)
Ceased
Application number
PCT/CN2018/105197
Other languages
English (en)
French (fr)
Inventor
韩中洋
赵珺
王伟
王霖青
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Dalian University of Technology
Original Assignee
Dalian University of Technology
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 Dalian University of Technology filed Critical Dalian University of Technology
Priority to PCT/CN2018/105197 priority Critical patent/WO2020051795A1/zh
Priority to US16/500,052 priority patent/US11526789B2/en
Publication of WO2020051795A1 publication Critical patent/WO2020051795A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06QINFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00Administration; Management
    • G06Q10/04Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/21Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/217Validation; Performance evaluation; Active pattern learning techniques
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/23Clustering techniques
    • G06F18/232Non-hierarchical techniques
    • G06F18/2321Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/23Clustering techniques
    • G06F18/232Non-hierarchical techniques
    • G06F18/2337Non-hierarchical techniques using fuzzy logic, i.e. fuzzy clustering
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/29Graphical models, e.g. Bayesian networks
    • G06F18/295Markov models or related models, e.g. semi-Markov models; Markov random fields; Networks embedding Markov models
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/004Artificial life, i.e. computing arrangements simulating life
    • G06N3/006Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/043Architecture, e.g. interconnection topology based on fuzzy logic, fuzzy membership or fuzzy inference, e.g. adaptive neuro-fuzzy inference systems [ANFIS]
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00Computing arrangements using knowledge-based models
    • G06N5/04Inference or reasoning models
    • G06N5/048Fuzzy inferencing
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N7/00Computing arrangements based on specific mathematical models
    • G06N7/01Probabilistic graphical models, e.g. probabilistic networks
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00Aspects of pattern recognition specially adapted for signal processing

Definitions

  • the invention belongs to the field of information technology, and relates to technologies such as fuzzy modeling, reinforcement learning, and parallel computing, and is a long-term interval prediction and structural learning method for a steel industry gas system combining granular calculation and reinforcement learning.
  • the invention uses industrial real data, first constructs a multi-level information granularity non-equal length distribution structure, and establishes a corresponding optimization model; further, considering the importance of the model structure to prediction accuracy, the present invention uses Monte Carlo method to multi-level models Based on the optimal multi-layer granularity calculation structure, a parallel calculation strategy is used to obtain the long-term interval prediction results of gas production and consumption. The accuracy of the results obtained by this method is high, and the calculation efficiency meets the requirements of practical applications. It can also be popularized and applied in other energy media systems in the steel industry.
  • the by-product gas of the iron and steel industry mainly includes three types of blast furnace gas, coke oven gas and converter gas. Due to changes in production demand and equipment switching operations, the production and production balance often appears on the scene. At this time, a scheduling plan needs to be formulated to achieve the pipe network The new balance of production and consumption ensures smooth production and avoids waste of resources. In the above process, the production and consumption change trend is an important basis for energy dispatchers to make decisions, so the prediction of by-product gas has important practical significance. (Xiong Chao. Discussion on Energy Conservation of Gas System in Iron and Steel Enterprises [C]. (2011) .Metallurgical Technology Economics Academic Conference of Chinese Metal Society.)
  • the invention mainly solves the problems of long-term interval prediction and model structure learning of the production and consumption of the by-product gas system of the iron and steel industry.
  • Methods Using real industrial data collected from the field, first, a multi-level information granularity non-equal length distribution structure and a corresponding optimization model are established.
  • the present invention uses Monte Carlo method to perform reinforcement learning on structural parameters; finally Hierarchical and efficient solution with parallel computing to get long-term interval prediction results.
  • a method for long-term interval prediction and structural learning of a steel gas system is as follows:
  • FCM Fuzzy C-Means
  • the beneficial effect of the present invention is that the long-term interval prediction model of the present invention optimizes information granularity hierarchically, and overcomes the problems that the traditional single-layer method requires too many solution parameters and has low average accuracy.
  • the established granularity distribution optimization model describes the coverage of information as a constraint condition, and the objective function is only a specific objective, thereby avoiding the tediousness of solving multi-objective problems.
  • the application of Monte Carlo method provides a reinforcement learning mechanism for the structural learning of long-term interval prediction models, so that the multi-level granular computing structure can be determined adaptively.
  • the use of parallel computing in the above optimization model solution and reinforcement learning process ensures that the calculation efficiency of the method can meet the requirements of practical applications.
  • Figure 1 is a schematic diagram of the by-product gas system of the iron and steel industry.
  • FIG. 2 is an application flowchart of the present invention.
  • Figure 3 is a schematic diagram of the multi-level information granularity allocation and optimization structure.
  • Figure 4 (a) is a long-term prediction result of the # 2 blast furnace gas generation amount by the MVE method.
  • Figure 4 (b) is a long-term prediction result of the single-layer particle size calculation method for # 2 blast furnace gas generation.
  • Fig. 4 (c) is a long-term interval prediction result of the method of the present invention for the # 2 blast furnace gas generation amount.
  • Fig. 5 (a) is a long-term prediction result of the # 1 coke oven gas consumption by the MVE method.
  • Figure 5 (b) is a long-term interval prediction result of the single-layer particle size calculation method for # 1 coke oven gas usage.
  • FIG. 5 (c) is a long-term prediction result of the method of the present invention with respect to # 1 coke oven gas usage.
  • MVE in the figure refers to Mean-Variance Estimate
  • the by-product gas system of Shanghai Baoshan Iron and Steel Plant which has a high level of automation in the domestic steel industry, will be further described below.
  • the schematic diagram of Baosteel's gas system shown in Figure 1 it can be seen that four blast furnaces, six coke ovens and six converters constitute three main by-product gas generating units, while the consumption units include cold / hot rolling, sintering, etc.
  • low-pressure boilers and power plants are often used as adjustable units.
  • the pipeline network also contains multiple gas cabinets, which play the role of temporary storage and buffer.
  • the gas mixing station and the pressurizing station are used as the transmission and distribution system, which is responsible for the gas pressure delivery to each consumption unit.
  • the present invention aims at this problem, and carries out research and application work on the prediction method of the by-product gas production and consumption.
  • Step 1 Data preprocessing
  • Step 3 Establish a multi-level granular computing model
  • i 1, 2, ..., M
  • j 1,2,..., n i , and n 1 ⁇ n 2 ⁇ ... ⁇ n m .
  • two measures of coverage cov and specific spec are defined as follows:
  • T represents the number of data points contained in the sample
  • ⁇ i is an identification variable, which is equal to 1 when the interval covers the data points of the sample, otherwise equal to 0
  • range refers to the difference between the maximum and minimum values of the sample data
  • z i respectively represent the upper and lower limits of the prediction interval.
  • the optimization goal of the information granularity model is to make (1) and (2) take maximum values, where cov should be at least equal to the target confidence (1- ⁇ ) ⁇ 100%, and ⁇ ⁇ [0,1] is a significant level.
  • the present invention considers (1) as a constraint condition, that is, cov must be greater than or equal to the target confidence interval.
  • the order of optimization of information granularity is opposite to the order of allocation.
  • the optimization model of each layer is established as follows:
  • range (2) refers to the difference between the maximum value and the minimum value of the corresponding data samples in the second layer
  • z i (2) represents the upper and lower limits of the interval results obtained at the second layer
  • is a hyperparameter that controls the granularity of the overall information; with Used to control ⁇ i so that it does not deviate too much from ⁇ ; It is an identification variable similar to ⁇ i , that is, when the interval obtained by the second layer covers the data points of the sample, otherwise it is 0.
  • the first layer needs to calculate a series of optimization problems, which can be expressed as follows:
  • the present invention uses a differential evolution (DE) algorithm to solve the above optimization problem. It should be particularly pointed out that the optimization problems of the first layer are independent of each other. Therefore, the present invention adopts a parallel strategy to process, so that the calculation time can be greatly reduced to meet the real-time requirements of the site.
  • DE differential evolution
  • Step 4 Long-term interval prediction
  • the probability of the data segment z k can be estimated as Corresponding predicted value It can be obtained by the central method, namely:
  • Step 5 Reinforcement learning of model structure parameters
  • state S, action A and reward R as follows:
  • ⁇ ⁇ (s, a) be a multilayer perceptron neural network:
  • ⁇ ⁇ (s, a) f ( ⁇ T ⁇ ⁇ (s, a) + b) (9)
  • represents the changing step size
  • is the discount factor
  • r t is the reward obtained at time t, that is:
  • this step (12) Computing can be accelerated with the help of parallel strategies;
  • the present invention divides information granularity into layers and optimizes it in parallel to improve the computing efficiency while ensuring prediction accuracy; on the other hand, it also adaptively determines the structure of multi-layer granularity calculation through reinforcement learning parameter.
  • Figures 4 and 5 are the long-term interval prediction results for # 2 blast furnace gas generation and # 1 coke oven gas generation, respectively.
  • the predicted duration is 480 points, that is, 8 hours, where (a) is the statistical mean -Mean-Variance Estimate (MVE) method, (b) is a general single-layer granularity calculation long-term interval prediction model, and (c) is the method of the present invention.
  • MVE statistical mean -Mean-Variance Estimate
  • b is a general single-layer granularity calculation long-term interval prediction model
  • (c) is the method of the present invention.
  • the dashed line is the true value
  • the gray band-shaped area is the predicted interval of the structure. Table 1 shows the prediction interval accuracy and calculation efficiency comparison.
  • the measurement indicators include prediction interval coverage (Prediction Intervals Coverage Probability (PICP)), regularized average interval width (Prediction Intervals Normalized Average Width (PINAW)), interval score (Interval score, IS) ) And computing time (ComputingTime, CT), where the definition of PIP, PINAW and IS are as follows:
  • T test is the total number of data points contained in the test set, and ⁇ i is an identification variable.
  • z i are the upper and lower limits of the prediction interval; the maximum and minimum values in the test set are d max and d min respectively ;
  • e i is a piecewise defined variable:

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Mathematical Physics (AREA)
  • Business, Economics & Management (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computing Systems (AREA)
  • Fuzzy Systems (AREA)
  • Automation & Control Theory (AREA)
  • Human Resources & Organizations (AREA)
  • Economics (AREA)
  • Strategic Management (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computational Linguistics (AREA)
  • Probability & Statistics with Applications (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Computational Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Quality & Reliability (AREA)
  • Development Economics (AREA)
  • Tourism & Hospitality (AREA)
  • General Business, Economics & Management (AREA)
  • Game Theory and Decision Science (AREA)
  • Marketing (AREA)
  • Operations Research (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Biophysics (AREA)

Abstract

一种钢铁煤气系统长期区间预测及其结构学习方法,属于信息技术领域。该方法采用工业真实数据,首先构造多层次的信息粒度非等长分配结构,建立相应优化模型;进而,考虑到模型结构对预测精度的重要性,借助蒙特卡洛方法,对多层次模型的结构参数进行强化学习;最终基于最优的多层粒度计算结构,运用并行计算策略,求得煤气产消量的长期区间预测结果。该方法所得到结果精度较高,且计算效率符合实际应用要求,在钢铁工业其它能源介质系统中亦可推广应用。

Description

一种钢铁煤气系统长期区间预测及其结构学习方法 技术领域
本发明属于信息技术领域,涉及到模糊建模、强化学习、并行计算等技术,是一种粒度计算与强化学习相结合的钢铁工业煤气系统长期区间预测及其结构学习方法。本发明采用工业真实数据,首先构造多层次的信息粒度非等长分配结构,建立相应优化模型;进而,考虑到模型结构对预测精度的重要性,本发明借助蒙特卡洛方法,对多层次模型的结构参数进行强化学习;最终基于最优的多层粒度计算结构,运用并行计算策略,求得煤气产消量的长期区间预测结果。此方法所得到结果精度较高,且计算效率符合实际应用要求,在钢铁工业其它能源介质系统中亦可推广应用。
背景技术
节能减排始终是钢铁工业日常生产的重要目标之一。作为随生产过程产生的副产煤气这一二次能源,能否实现对其合理利用将直接关系到企业经济利益和降耗效果。钢铁工业的副产煤气主要包括高炉煤气、焦炉煤气和转炉煤气三种,由于生产需求变化、设备切换操作等原因,现场常出现产消失衡情况,此时需制定调度方案以使管网达到新的产消平衡,从而保障生产顺畅、避免资源浪费。在上述过程中,产消量变化趋势是能源调度人员做出决策的重要依据,因此对副产煤气的预测工作具有重要的实际应用意义。(熊超.钢铁企业煤气系统节能探讨[C].(2011).中国金属学会冶金技术经济学会学术年会.)
随着研究应用工作的广泛开展,以粒度计算为代表的预测模型已可实现超过4小时的长期趋势估计(J.Zhao,Z.Y.Han,W.Pedrycz,W.Wang.(2016).Granular model of long-term prediction for energy system in steel industry[J].IEEE transactions on cybernetics,46(2),388-400)(Z.Y.Han,J.Zhao,Q.Liu,W.Wang.(2016).Granular-computing based hybrid collaborative fuzzy clustering for long-term prediction of multiple gas holders levels[J].Information Sciences,330,175-185)。然而,这些方法的预测结果均以点为呈现形式,无法满足现场对结果可靠性衡量的需求。鉴于此,除了支持向量机(Support Vector Machine,SVM)、基于统计学的一些模型均可被用来实现区间预测以外(C.Y.Sheng,J.Zhao,W.Wang,H.Leung.(2013).Prediction intervals for a noisy nonlinear time series based on a bootstrapping reservoir computing network ensemble[J].IEEE Transactions on neural networks and learning systems,24(7),1036-1048)(A.Khosravi,S.Nahavandi,D.(2010).Creighton.Construction of optimal prediction intervals for load forecasting problems[J].IEEE Transactions on Power Systems,25(3),1496-1503.),将粒度计算中的聚类中心纵向拓展为区间值、进而求解相应优化模型,也是一个可行的解决方案(Z.Y.Han,J.Zhao,H.Leung,W.Wang.(2018).Construction of prediction intervals for gas flow systems in steel industry based on granular computing[J].Control Engineering Practice,78,79-88)。
然而,这些方法存在明显不足:首先,由于以迭代机制构造样本,SVM、基于统计学等模型在长期预测方面精度难以令人满意,一般仅能给出60个点以内的良好预测结果;其次,虽然单层的信息粒度分配方式可计算出较长时间的区间估计结果,但大量的待优化参数导致其运算时间和精度均存在不确定性,若能将这些参数分层次处理,则有望在计算效率和平均精度上获得明显提升;此外,若应用多层次粒度计算模型,其结构将直接左右长期区间预测结果的精度,因此需要设计方法以高效、合理地获取最优模型结构参数。
发明内容
本发明主要解决钢铁工业副产煤气系统的产消量长期区间预测及模型结构学习问题。方法使用采集自现场的真实工业数据,首先建立多层次的信息粒度非等长分配结构及相应优化模型;为自适应地确定模型结构,本发明借助蒙特卡洛方法对结构参数进行强化学习;最终借助并行计算进行分层次的高效求解,得到长期区间预测结果。
本发明的技术方案:
一种钢铁煤气系统长期区间预测及其结构学习方法,步骤如下:
(1)由现场实时数据库采集副产煤气系统的产消量数据,经过除噪、滤波及填补等预处理后作为基础数据样本;
(2)应用模糊C均值(FuzzyC-Means,FCM)算法进行聚类,得到聚类中心及隶属度;
(3)逐层次的对聚类中心分配信息粒度,将其由点值拓展为区间值,并建立各层次的信息粒度优化模型;
(4)利用并行计算策略,逐层求解最优结构参数下的信息粒度分配优化模型,并结合概率、模糊建模等手段,得到初步的长期区间预测结果;
(5)将长期区间预测模型的当前结构定义为“状态”、参数变化定义为“动作”,借助蒙特卡洛方法,以强化学习的方式求取近似的策略值函数,从而确定模型的最优结构参数,并再次运用步骤4,得到最终的长期区间预测结果;
本发明的有益效果:本发明的长期区间预测模型将信息粒度分层优化,克服了传统单层方法需求解参数过多、平均精度较低的问题。所建立的粒度分配优化模型将信息的覆盖性描述为约束条件,而目标函数仅为专一性一个目标,从而避免了求解多目标问题的繁琐。此外,蒙特卡洛方法的运用,为长期区间预测模型的结构学习提供了强化学习机制,从而可自适应地确定多层次粒度计算结构。加之并行计算在上述优化模型求解、强化学习过程中的运用,保障了方法的计算效率可符合实际应用之要求。
附图说明
图1为钢铁工业副产煤气系统示意图。
图2为本发明应用流程图。
图3为多层次信息粒度分配及优化结构示意图。
图4(a)为MVE方法对#2高炉煤气发生量的长期区间预测结果图。
图4(b)为单层粒度计算方法对#2高炉煤气发生量的长期区间预测结果图。
图4(c)为本发明方法对#2高炉煤气发生量的长期区间预测结果图。
图5(a)为MVE方法对#1焦炉煤气使用量的长期区间预测结果图。
图5(b)为单层粒度计算方法对#1焦炉煤气使用量的长期区间预测结果图。
图5(c)为本发明方法对#1焦炉煤气使用量的长期区间预测结果图。
图中的MVE指均值-方差估计(Mean-Variance Estimation)
具体实施方式
为了更好地理解本发明的技术路线和实施方案,下面以国内钢铁工业自动化水平较高的上海宝山钢铁厂副产煤气系统做进一步说明。由附图1所示的宝钢煤气系统示意图可看出,四座高炉、六座焦炉和六座转炉构成了三种主要副产煤气的发生单元,而消耗单元则包括冷/热轧、烧结等,其中低压锅炉和电厂常作为可调节单元。管网内还包含多座煤气柜,起到暂存和缓冲的作用。此外,煤气混合站和加压站作为输配系统,负责将煤气压送至各消耗单元。在日常生产中,维持产消平衡既可以保障生产的顺利进行,也有助于达成节能减排的目标,因此是能源调度人员的工作重点。由于副产煤气网络复杂,遍布炼铁、炼钢、轧钢等多个生产区域,具有非线性、大时滞等显著特点,能源调度人员难以准确判断产消量的未来趋势。本发明即针对此问题,开展副产煤气产消量的预测方法研究与应用工作。
本发明的具体实施步骤如下:
步骤1:数据预处理
从工业现场实时关系数据库,读取副产煤气系统发生和消耗单元数据,并做基本的除噪、滤波和填补等预处理工作。
步骤2:FCM
将数据分割为数段等长度的片段,即Z={z 1,z 2,…,z N},其中
Figure PCTCN2018105197-appb-000001
n代表每个数据段所含数据点个数,N则为数据段个数。应用FCM聚类算法,获得聚类中心V={v 1,v 2,…,v c}和对应隶属度U={u 1,u 2,…,u N},其中
Figure PCTCN2018105197-appb-000002
c是聚类中心维度。
步骤3:建立多层次粒度计算模型
如附图1所示,由下至上逐层次地对聚类中心V={v 1,v 2,…,v c}分配信息粒度α i,j和β i,其中i=1,2,…,m,j=1,2,…,n i,且n 1≠n 2≠…≠n m。如此即可将聚类中心由点值延展为 区间值。为优化求解上述信息粒度参数,首先定义覆盖性cov和专一性spec两个衡量指标如下:
Figure PCTCN2018105197-appb-000003
Figure PCTCN2018105197-appb-000004
其中,T表示样本所含数据点个数;λ i是标识变量,即当区间覆盖样本的数据点时等于1,否则等于0;range指样本数据的最大值与最小值之差,
Figure PCTCN2018105197-appb-000005
z i分别表示预测区间上下限。
信息粒度模型的优化目标是使得(1)和(2)均取极大值,其中cov至少应等于目标置信度(1-ρ)×100%,ρ∈[0,1]是显著水平。为避免求解多目标问题的困难与繁琐,本发明将(1)作为约束条件考虑,即cov必须大于或等于目标置信区间。信息粒度的优化顺序与分配顺序相反,各层优化模型建立如下:
①第二层
Figure PCTCN2018105197-appb-000006
其中,range (2)指第二层对应数据样本的最大值与最小值之差,
Figure PCTCN2018105197-appb-000007
z i (2)表示第二层所得区间结果的上下限;ε是控制总体信息粒度的超参数;
Figure PCTCN2018105197-appb-000008
Figure PCTCN2018105197-appb-000009
用来控制β i,以使其不至于过分偏离ε;
Figure PCTCN2018105197-appb-000010
是类似λ i的标识变量,即当第二层所得区间覆盖样本的数据点时为1,否则为0。
②第一层
与第二层仅处理一个优化问题不同,第一层要计算合计为m个的一系列优化问题,其中任意一个可表述如下:
Figure PCTCN2018105197-appb-000011
其中,
Figure PCTCN2018105197-appb-000012
指第一层各优化问题对应数据样本的最大值与最小值之差,i=1,2,…,m;
Figure PCTCN2018105197-appb-000013
Figure PCTCN2018105197-appb-000014
分别为第一层各优化问题求解所得区间上下限;
Figure PCTCN2018105197-appb-000015
Figure PCTCN2018105197-appb-000016
则用来控制α i,j不过分偏离β i
考虑到收敛性和求解速度,本发明运用差分进化(Differential Evolution,DE)算法求解上述优化问题。需要特别指出的是,第一层的各个优化问题之间相互独立,因此本发明采用并 行策略处理,如此便可较大程度地缩短计算时间,以符合现场对于实时性的要求。
步骤4:长期区间预测
基于粒度计算的长期区间预测本质上是预测模糊隶属度,即
Figure PCTCN2018105197-appb-000017
其中
Figure PCTCN2018105197-appb-000018
为隶属度预测值,
Figure PCTCN2018105197-appb-000019
是隶属度矩阵U的一部分,n I表示隶属度关系的输入个数。本发明将通过概率估计实现隶属度的预测。为便于理解,以下定义及叙述以点形式给出,。首先定义粒度计算框架下的聚类中心概率
Figure PCTCN2018105197-appb-000020
数据段概率
Figure PCTCN2018105197-appb-000021
和共生矩阵
Figure PCTCN2018105197-appb-000022
如下:
Figure PCTCN2018105197-appb-000023
Figure PCTCN2018105197-appb-000024
Figure PCTCN2018105197-appb-000025
其中,
Figure PCTCN2018105197-appb-000026
是聚类中心矩阵V中的一部分,
Figure PCTCN2018105197-appb-000027
是数据片段集合Z中的一部分;
Figure PCTCN2018105197-appb-000028
是标识型变量,即考虑
Figure PCTCN2018105197-appb-000029
Figure PCTCN2018105197-appb-000030
分别是隶属度
Figure PCTCN2018105197-appb-000031
中的元素,且{h1,h2,…,hn I}∈[1,c]),当满足h1=i1,…,hn I=in I时为1,否则为0;p(v i|v j)是条件概率,即当
Figure PCTCN2018105197-appb-000032
的最大隶属度出现在聚类中心集合
Figure PCTCN2018105197-appb-000033
时,z k的最大隶属度出现在v i的概率。
基于上述定义,数据段z k的概率可估计为
Figure PCTCN2018105197-appb-000034
对应预测值
Figure PCTCN2018105197-appb-000035
可由中心法得到,即:
Figure PCTCN2018105197-appb-000036
其中,
Figure PCTCN2018105197-appb-000037
Figure PCTCN2018105197-appb-000038
中元素。
步骤5:模型结构参数的强化学习
本发明将多层次粒度计算的模型结构确定问题看做单步马尔科夫决策过程,进而采用蒙特卡洛方法,强化学习包括m和n i(i=1,2,…,m)在内的结构参数。首先定义状态S、动作A和奖赏R如下:
S–确定的多层次粒度计算模型结构
A–变更参数m和n i(i=1,2,…,m)
R–所得预测区间的spec
由于待确定结构参数数量较庞大,本发明采用梯度下降近似函数策略,学习策略值函数π ω(s,a)。设π ω(s,a)为多层感知机神经网络:
π ω(s,a)=f(ω T·φ(s,a)+b)   (9)
其中,φ(s,a)是表征状态-动作对的特征向量,本发明定义为φ(s,a)=(m,n 1,n 2,…,n m) T;b表示多层感知机的偏置量,f表示激活函数,本发明采用的是sigmoid函数。
定义可导的性能函数如下:
Figure PCTCN2018105197-appb-000039
其中,
Figure PCTCN2018105197-appb-000040
是π ω(s,a)的真值函数,s 0是起始状态。对J(ω)求关于ω的梯度并应用策略梯度定理,可最终得出ω的更新公式为:
Figure PCTCN2018105197-appb-000041
其中,τ表示变化步长,γ是贴现因子,r t是t时刻获得的奖赏,即:
Figure PCTCN2018105197-appb-000042
设结构参数的搜索空间维度为L,模型参数确定过程可总结如下:
①初始化τ>0,γ>0,
Figure PCTCN2018105197-appb-000043
在L中取l个结构参数样本作为训练子集;
②取t=1到l,利用步骤四表述过程循环先后计算(12)和(11),获得学习到的策略值函数参数ω opt,考虑到问题的相互独立性,此步中(12)的计算可借助并行策略加速;
③取t=1到L,计算(9),选取其中策略值函数最大值对应的特征向量φ opt(s,a)为最优结构参数,即:
Figure PCTCN2018105197-appb-000044
④利用获得最优结构参数,再次计算(3)-(8),即可获得最终的长期区间预测结果。
由以上过程可以看出,本发明一方面将信息粒度分层次分配和并行优化,在提高运算效率的同时保证预测精度;另一方面还通过强化学习方式,自适应地确定多层粒度计算的结构参数。
图4、图5分别是针对#2高炉煤气发生量、#1焦炉煤气发生量的长期区间预测结果,预测时长为480个点,即8个小时,其中(a)为统计学中的均值-方差估计(Mean-VarianceEstimation,MVE)方法,(b)为一般的单层粒度计算长期区间预测模型,(c)则是本发明方法。虚线为真实值,灰色带状区域为构造的预测区间。表1给出预测区间精度及运算效率比较,衡量指标包括预测区间覆盖率(Prediction Intervals Coverage Probability,PICP)、正则化平均区间宽度(Prediction Intervals Normalized Average Width,PINAW)、区间得分(Interval Score,IS)以及计算耗时(ComputingTime,CT),其中PICP、PINAW和IS的定义如下:
Figure PCTCN2018105197-appb-000045
Figure PCTCN2018105197-appb-000046
Figure PCTCN2018105197-appb-000047
其中,T test是测试集所含数据点总数,λ i为标识变量,当测试集中的数据点在预测区间内时λ i=1,否则λ i=0;
Figure PCTCN2018105197-appb-000048
z i分别是预测区间上下限;测试集中的极大值、极小值分别是d max和d min;e i是个分段定义变量:
Figure PCTCN2018105197-appb-000049
其中,d i是测试集中的数据点,实验中取显著水平ρ=0.1。综合图表结果可以明显看出,本发明在精度表现和运算效率上均优于其它工业常用的区间预测方法。
表1 三种方法在单次长期区间预测的精度及耗时结果比较
Figure PCTCN2018105197-appb-000050

Claims (1)

  1. 一种钢铁煤气系统长期区间预测及其结构学习方法,其特征在于,步骤如下:
    步骤1:数据预处理
    从工业现场实时关系数据库,读取副产煤气系统发生和消耗单元数据,并做基本的除噪、滤波和填补;
    步骤2:FCM
    将步骤1读取的数据分割为数段等长度的片段,即Z={z 1,z 2,…,z N},其中
    Figure PCTCN2018105197-appb-100001
    n代表每个数段所含数据点个数,N则为数段个数;应用FCM聚类算法,获得聚类中心V={v 1,v 2,…,v c}和对应隶属度U={u 1,u 2,…,u N},其中
    Figure PCTCN2018105197-appb-100002
    c是聚类中心维度;
    步骤3:建立多层次粒度计算模型
    由下至上逐层次地对聚类中心矩阵V={v 1,v 2,…,v c}分配信息粒度α i,j和β i,其中i=1,2,…,m;j=1,2,…,n i;且n 1≠n 2≠…≠n m;如此即将聚类中心由点值延展为区间值;为优化求解上述信息粒度参数,首先定义覆盖性cov和专一性spec两个衡量指标如下:
    Figure PCTCN2018105197-appb-100003
    Figure PCTCN2018105197-appb-100004
    其中,T表示数据样本所含数据点个数;λ i是标识变量,即当区间覆盖样本的数据点时等于1,否则等于0;range指样本数据的最大值与最小值之差,
    Figure PCTCN2018105197-appb-100005
    Figure PCTCN2018105197-appb-100006
    分别表示预测区间上限和下限;
    信息粒度模型的优化目标是使得覆盖性cov和专一性spec均取极大值,其中cov至少等于目标置信度(1-ρ)×100%,ρ∈[0,1]是显著水平;将式(1)作为约束条件考虑,即cov必须大于或等于目标置信区间;信息粒度的优化顺序与分配顺序相反,各层优化模型建立如下:
    (1)第二层
    Figure PCTCN2018105197-appb-100007
    其中,range (2)指第二层对应数据样本的最大值与最小值之差,
    Figure PCTCN2018105197-appb-100008
    表示第二层所得区间结果的上限和下限;ε是控制总体信息粒度的超参数;
    Figure PCTCN2018105197-appb-100009
    Figure PCTCN2018105197-appb-100010
    用来控制β i,以使其不至于过分偏离ε;
    Figure PCTCN2018105197-appb-100011
    是类似λ i的标识变量,即当第二层所得区间覆盖样本的数据点时为1,否则为0;
    (2)第一层
    与第二层仅处理一个优化问题不同,第一层要计算合计为m个的一系列优化问题,其中任意一个表述如下:
    Figure PCTCN2018105197-appb-100012
    其中,
    Figure PCTCN2018105197-appb-100013
    指第一层各优化问题对应数据样本的最大值与最小值之差,i=1,2,…,m;
    Figure PCTCN2018105197-appb-100014
    Figure PCTCN2018105197-appb-100015
    分别为第一层各优化问题求解所得区间的上限和下限;
    Figure PCTCN2018105197-appb-100016
    Figure PCTCN2018105197-appb-100017
    则用来控制α i,j不过分偏离β i
    步骤4:长期区间预测
    基于粒度计算的长期区间预测本质上是预测模糊隶属度,即
    Figure PCTCN2018105197-appb-100018
    Figure PCTCN2018105197-appb-100019
    其中
    Figure PCTCN2018105197-appb-100020
    为隶属度预测值,
    Figure PCTCN2018105197-appb-100021
    是隶属度矩阵U的一部分,n I表示隶属度关系的输入个数;将通过概率估计实现隶属度的预测;为便于理解,以下定义及叙述以点形式给出,首先定义粒度计算框架下的聚类中心概率
    Figure PCTCN2018105197-appb-100022
    数据段概率
    Figure PCTCN2018105197-appb-100023
    和共生矩阵
    Figure PCTCN2018105197-appb-100024
    如下:
    Figure PCTCN2018105197-appb-100025
    Figure PCTCN2018105197-appb-100026
    Figure PCTCN2018105197-appb-100027
    其中,
    Figure PCTCN2018105197-appb-100028
    {i1,i2,…,in I}∈[1,c],是聚类中心矩阵V中的一部分;
    Figure PCTCN2018105197-appb-100029
    是数据片段集合Z中的一部分;
    Figure PCTCN2018105197-appb-100030
    是标识型变量,即考虑
    Figure PCTCN2018105197-appb-100031
    分别是隶属度
    Figure PCTCN2018105197-appb-100032
    中的元素,且{h1,h2,…,hn I}∈[1,c];当满足h1=i1,…,hn I=in I时为1,否则为0;p(v i|v j)是条件概率,即当
    Figure PCTCN2018105197-appb-100033
    的最大隶属度出现在聚类中心集合
    Figure PCTCN2018105197-appb-100034
    时,z k的最大隶属度出现在v i的概率;
    基于上述定义,数据段z k的概率估计为
    Figure PCTCN2018105197-appb-100035
    Figure PCTCN2018105197-appb-100036
    对应预测值
    Figure PCTCN2018105197-appb-100037
    由中心法得到,即:
    Figure PCTCN2018105197-appb-100038
    其中,
    Figure PCTCN2018105197-appb-100039
    Figure PCTCN2018105197-appb-100040
    中元素;
    步骤5:模型结构参数的强化学习
    将多层次粒度计算的模型结构确定问题看做单步马尔科夫决策过程,进而采用蒙特卡洛方法,强化学习包括m和n i,i=1,2,…,m在内的结构参数;首先定义状态S、动作A和奖赏R如下:
    S为确定的多层次粒度计算模型结构;
    A为变更参数m和n i,i=1,2,…,m;
    R为所得预测区间的spec;
    由于待确定结构参数数量较庞大,采用梯度下降近似函数策略,学习策略值函数π ω(s,a);设π ω(s,a)为多层感知机神经网络:
    π ω(s,a)=f(ω T·φ(s,a)+b)  (9)
    其中,φ(s,a)是表征状态-动作对的特征向量,定义为φ(s,a)=(m,n 1,n 2,…,n m) T;b表示多层感知机的偏置量,f表示激活函数,本发明采用的是sigmoid函数;
    定义可导的性能函数如下:
    Figure PCTCN2018105197-appb-100041
    其中,
    Figure PCTCN2018105197-appb-100042
    是π ω(s,a)的真值函数,s 0是起始状态;对J(ω)求关于ω的梯度并应用策略梯度定理,最终得出ω的更新公式为:
    Figure PCTCN2018105197-appb-100043
    其中,τ表示变化步长,γ是贴现因子,r t是t时刻获得的奖赏,即:
    Figure PCTCN2018105197-appb-100044
    设结构参数的搜索空间维度为L,模型参数确定过程总结如下:
    (1)初始化τ>0,γ>0,
    Figure PCTCN2018105197-appb-100045
    在L中取l个结构参数样本作为训练子集;
    (2)取t=1~l,利用步骤四表述过程循环先后计算(12)和(11),获得学习到的策略值函数参数ω opt,考虑到问题的相互独立性,此步中(12)的计算借助并行策略加速;
    (3)取t=1~l,计算(9),选取其中策略值函数最大值对应的特征向量φ opt(s,a)为最优结构参数,即:
    Figure PCTCN2018105197-appb-100046
    (4)利用获得最优结构参数,再次计算(3)-(8),即获得最终的长期区间预测结果。
PCT/CN2018/105197 2018-09-12 2018-09-12 一种钢铁煤气系统长期区间预测及其结构学习方法 Ceased WO2020051795A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
PCT/CN2018/105197 WO2020051795A1 (zh) 2018-09-12 2018-09-12 一种钢铁煤气系统长期区间预测及其结构学习方法
US16/500,052 US11526789B2 (en) 2018-09-12 2018-09-12 Method for construction of long-term prediction intervals and its structural learning for gaseous system in steel industry

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2018/105197 WO2020051795A1 (zh) 2018-09-12 2018-09-12 一种钢铁煤气系统长期区间预测及其结构学习方法

Publications (1)

Publication Number Publication Date
WO2020051795A1 true WO2020051795A1 (zh) 2020-03-19

Family

ID=69777375

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/105197 Ceased WO2020051795A1 (zh) 2018-09-12 2018-09-12 一种钢铁煤气系统长期区间预测及其结构学习方法

Country Status (2)

Country Link
US (1) US11526789B2 (zh)
WO (1) WO2020051795A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112215464A (zh) * 2020-09-04 2021-01-12 北京天泽智云科技有限公司 一种多工况下高炉煤气的预测平衡调度系统
CN118377777A (zh) * 2024-06-20 2024-07-23 无锡东雄重型电炉有限公司 一种获取炼钢过程中工序数据的方法及系统

Families Citing this family (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111914479A (zh) * 2020-07-03 2020-11-10 天津大学 一种用于大坝基岩的灌浆功率混合区间预测方法
CN112195302B (zh) * 2020-10-16 2023-05-12 中冶赛迪技术研究中心有限公司 一种预测转炉一次烟气电除尘爆炸风险的方法及装置
CN112948125A (zh) * 2021-03-29 2021-06-11 北京深睿科技有限责任公司 基于gpu并行强化学习的建筑节能方法
CN115099510A (zh) * 2022-07-06 2022-09-23 山东钢铁集团永锋临港有限公司 一种转炉煤气预测平衡调度方法
CN115097737B (zh) * 2022-08-24 2022-11-08 北京航空航天大学 一种可重入制造系统的多层级调控方法
CN117930687B (zh) * 2024-01-24 2024-08-27 日照港集装箱发展有限公司动力分公司 一种用于港口的智慧能源优化控制系统
CN118550262B (zh) * 2024-07-29 2024-10-25 江苏协航能源科技有限公司 基于模糊控制的光伏组件生产控制方法及系统
CN119815469B (zh) * 2024-12-30 2025-10-17 西安电子科技大学 一种大规模接入节能优化方法、装置及电子设备

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7013249B1 (en) * 2001-07-16 2006-03-14 Kinder Morgan, Inc. Method for detecting near neutral/low pH stress corrosion cracking in steel gas pipeline systems
CN103426035A (zh) * 2013-08-12 2013-12-04 浙江大学 钢铁行业副产高炉煤气自平衡调度系统及产消量预测方法
CN103473469A (zh) * 2013-09-25 2013-12-25 南京航空航天大学 一种基于客观指标的扇区交通态势多层次模糊评价方法
CN103514486A (zh) * 2012-06-15 2014-01-15 上海宝信软件股份有限公司 基于因素分析的高炉煤气受入流量预测方法
CN106779384A (zh) * 2016-12-07 2017-05-31 大连理工大学 一种基于信息粒度最优分配的钢铁工业高炉煤气长期区间预测方法

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9569804B2 (en) * 2012-08-27 2017-02-14 Gridium, Inc. Systems and methods for energy consumption and energy demand management
US9461876B2 (en) * 2012-08-29 2016-10-04 Loci System and method for fuzzy concept mapping, voting ontology crowd sourcing, and technology prediction
US9418337B1 (en) * 2015-07-21 2016-08-16 Palantir Technologies Inc. Systems and models for data analytics

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7013249B1 (en) * 2001-07-16 2006-03-14 Kinder Morgan, Inc. Method for detecting near neutral/low pH stress corrosion cracking in steel gas pipeline systems
CN103514486A (zh) * 2012-06-15 2014-01-15 上海宝信软件股份有限公司 基于因素分析的高炉煤气受入流量预测方法
CN103426035A (zh) * 2013-08-12 2013-12-04 浙江大学 钢铁行业副产高炉煤气自平衡调度系统及产消量预测方法
CN103473469A (zh) * 2013-09-25 2013-12-25 南京航空航天大学 一种基于客观指标的扇区交通态势多层次模糊评价方法
CN106779384A (zh) * 2016-12-07 2017-05-31 大连理工大学 一种基于信息粒度最优分配的钢铁工业高炉煤气长期区间预测方法

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112215464A (zh) * 2020-09-04 2021-01-12 北京天泽智云科技有限公司 一种多工况下高炉煤气的预测平衡调度系统
CN112215464B (zh) * 2020-09-04 2024-05-07 北京天泽智云科技有限公司 一种多工况下高炉煤气的预测平衡调度系统
CN118377777A (zh) * 2024-06-20 2024-07-23 无锡东雄重型电炉有限公司 一种获取炼钢过程中工序数据的方法及系统

Also Published As

Publication number Publication date
US11526789B2 (en) 2022-12-13
US20200285982A1 (en) 2020-09-10

Similar Documents

Publication Publication Date Title
WO2020051795A1 (zh) 一种钢铁煤气系统长期区间预测及其结构学习方法
CN109242188B (zh) 一种钢铁煤气系统长期区间预测及其结构学习方法
CN112186743B (zh) 一种基于深度强化学习的动态电力系统经济调度方法
Wang et al. Low-carbon development quality of cities in China: Evaluation and obstacle analysis
US11126765B2 (en) Method for optimal scheduling decision of air compressor group based on simulation technology
WO2021189739A1 (zh) 一种基于生产计划的钢铁企业氧气负荷预测方法
CN112036633B (zh) 一种基于水库生态发电多目标中长期随机调度模型的优化调度方法
WO2019237316A1 (zh) 一种基于知识迁移的高炉煤气调度系统建模方法
CN107918368B (zh) 钢铁企业煤气产生量与消耗量的动态预测方法及设备
CN103942422B (zh) 一种基于粒度计算的冶金企业转炉煤气柜位长期预测方法
CN105631528B (zh) 一种基于nsga-ii和近似动态规划的多目标动态最优潮流求解方法
CN106355285B (zh) 一种基于参数修正的用电负荷预测方法
CN110443418A (zh) 基于ga-bp神经网络的城市用水量预测方法
CN113902213B (zh) 面向电力现货市场基于碳排放和源网荷储互动的电网规划方法
CN109146121A (zh) 基于pso-bp模型的停限产情况下的电量预测方法
CN110400009A (zh) 基于自适应遗传算法的高炉炼铁多目标智能优化方法
CN106779384B (zh) 一种基于信息粒度最优分配的钢铁工业高炉煤气长期区间预测方法
CN111461404A (zh) 一种基于神经网络预测区间的短期负荷和水电预测方法
CN114202086A (zh) 一种矿山开采方案多目标优化方法
CN104298214A (zh) 一种高炉铁水生产过程综合优化控制方法
CN104134102B (zh) 基于leap模型的电网中长期电力需求分布预测方法
CN108734419A (zh) 一种基于知识迁移的高炉煤气调度系统建模方法
Timplalexis et al. A comprehensive review on industrial demand response strategies and applications
Miao et al. Short-term Load Forecasting Based on Echo State Network and LightGBM
Luo et al. Analysis of influencing factors of green building energy consumption based on genetic algorithm

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: 18933163

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: 18933163

Country of ref document: EP

Kind code of ref document: A1