Nothing Special   »   [go: up one dir, main page]

WO2024060488A1 - 一种基于深度循环神经网络和进化计算的工业锅炉燃烧优化方法 - Google Patents

一种基于深度循环神经网络和进化计算的工业锅炉燃烧优化方法 Download PDF

Info

Publication number
WO2024060488A1
WO2024060488A1 PCT/CN2023/075002 CN2023075002W WO2024060488A1 WO 2024060488 A1 WO2024060488 A1 WO 2024060488A1 CN 2023075002 W CN2023075002 W CN 2023075002W WO 2024060488 A1 WO2024060488 A1 WO 2024060488A1
Authority
WO
WIPO (PCT)
Prior art keywords
neural network
boiler
parameters
target
control
Prior art date
Application number
PCT/CN2023/075002
Other languages
English (en)
French (fr)
Inventor
吴学成
陈玲红
吴迎春
陈华锋
张鑫
喻可
邱坤赞
郑成航
吴卫红
高翔
岑可法
Original Assignee
浙江大学
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 浙江大学 filed Critical 浙江大学
Publication of WO2024060488A1 publication Critical patent/WO2024060488A1/zh

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00Computer-aided design [CAD]
    • G06F30/20Design optimisation, verification or simulation
    • G06F30/27Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • G06N3/084Backpropagation, e.g. using gradient descent

Definitions

  • the present invention relates to the technical field of boiler combustion, and specifically relates to an industrial boiler combustion optimization method based on deep circular neural network and evolutionary calculation.
  • boiler combustion optimization technology based on big data can provide optimized operation control strategies for boilers and learn better control experience from historical operation data, thereby replacing or assisting human decision-making and realizing boiler Further economical and clean operation.
  • main ones are to establish static combustion prediction models.
  • nonlinear prediction models such as feedforward neural networks and support vector machines are established. This type of method mainly has some shortcomings:
  • Static models have high requirements on data types because only the nonlinear characteristics between input variables and output variables are considered, but the characteristics of the variables themselves in the time dimension are not considered.
  • the difficulty lies in the design of the data structure and its combination with the optimization algorithm.
  • the purpose of the present invention is to provide an industrial boiler combustion optimization method based on deep circular neural network and evolutionary calculation, to achieve online combustion optimization of industrial boilers when there are few types of industrial boiler operating parameters, and to use the optimization decision results directly or indirectly. Adjust the boiler operation to make the boiler run cleaner and more efficiently.
  • An industrial boiler combustion optimization method based on deep recurrent neural network and evolutionary calculation includes:
  • Nonlinear offline modeling of boiler combustion A recurrent neural network structure is used to build a nonlinear model of boiler combustion BM.
  • the training parameters are trained using an offset data structure to obtain a nonlinear model BM_trd after training.
  • Nonlinear offline modeling of boiler combustion requires the use of recurrent neural network structures, such as ordinary recurrent neural network layers (RNN), long short-term memory neural network layers (LSTM) and gated recurrent neural network layers (GRU).
  • the built nonlinear boiler combustion model (BoilerModel, BM) is a neural network hierarchical stack structure. From top to bottom, it is a single-layer convolutional neural network layer (Conv), several recurrent neural network layers and several fully connected neural network layers ( Dense).
  • the built model has different numbers of cyclic network layers, fully connected network layers, number of network nodes at each layer, and node activation functions, so it is boiler customizable.
  • a batch normalization layer (Batch Normalization) is used between multiple fully connected layers to optimize the loss distribution of the underlying network.
  • Dropout regularization is used for the layer input units of the top-level recurrent neural network, and recurrent dropout regularization is used for the recurrent units to alleviate the over-fitting problem during neural network training.
  • the following defines the offset data structure: Y Target t+1
  • the offset data structure has three types of parameters: 1) Decision parameters to be optimized Control: Control t-n+1 , Control t and Control t+1 ; 2) Boiler operating status parameters Status: Status tn , Status t -1 and Status t ; 3) Optimize target parameters Target: Target tn , Target t-1 and Target t ;
  • the Control parameter in input The output Y is the Target parameter in the future, which is consistent with the time of the Control parameter; the time series length of Merit value; based on the offset data structure, the recurrent neural network can "learn" the nonlinear relationship between the feature x i and the target Y, as well as the nonlinear relationship between the past, present, and future time series distribution of the variable X and the target Y.
  • the boiler user determines the optimization target Target parameters, and selects the type of Control parameters and Status parameters based on different Target parameters. The determination of these parameters is based on the actual boiler control experience:
  • the Status parameters are: boiler load (thermal load or electrical load), main steam pressure (or water supply pressure), boiler combustion chamber temperature gradient (temperatures from the burner end to the furnace outlet Measured value), exhaust temperature (after the last stage heat exchanger);
  • Control parameters are: generally depending on the boiler type, mainly optimizing air volume, fuel volume and damper opening, generally primary air volume, furnace outlet oxygen concentration, supply Fuel quantity, secondary air door opening.
  • the Status parameter and Control parameter are based on 1), and the Status parameter should increase the properties of the fuel entering the furnace. For example, volatile matter parameters need to be added for NO x pollutants, sulfur parameters need to be added for SO 2 pollutants, and fuel properties are customized by the user for other pollutants.
  • System error position index artificially determine the target value range, for example, for thermal efficiency, when the thermal efficiency is greater than 1 or less than 0.7, it is considered invalid, and the position index of invalid data is extracted by traversing the target;
  • Random error processing filter the data, such as Kalman filtering
  • t is the current time
  • ⁇ C represents the continuous change of the Control parameter, and its specific optimal value is obtained by evolutionary calculation iterations. Define the range of the algorithm to search ⁇ C as [( ⁇ C) min , ( ⁇ C) max ], and add ⁇ C After X co , X coi is obtained, then the ⁇ C population is transformed into the X coi population:
  • Control t + ⁇ C Control t+1 ; ( ⁇ C) min and ( ⁇ C) max are defined by humans, and the maximum and minimum search ranges are generally set to be small to avoid excessive boiler control ranges that lead to operational fluctuations and accidents.
  • the optimal individual of the final ⁇ C population is the optimal decision change output.
  • the evolutionary computing methods used in the boiler combustion decision optimization method include swarm intelligence algorithms and evolutionary algorithms. For different problems, the evolutionary calculation method needs to establish an appropriate fitness function.
  • the fitness function of this method is defined as follows:
  • objV represents the fitness value
  • Norm(X c ) represents that X c adopts the same standardization and normalization process used by BM
  • X c [-1,-1] represents the Target value of the boiler operation at the current time t
  • Bm_trd ( Norm ( X c )) - to (0,+ ⁇ ) represents the selection operator of some evolutionary algorithms (objV>0).
  • the original ⁇ Target is usually small and difficult to distinguish.
  • the gap between ⁇ Target is widened, which is conducive to evolutionary calculations.
  • the combustion optimization process based on evolutionary calculation is to continuously iteratively evolve the ⁇ C population, and by finding the optimal combination of decision-making changes, the thermal efficiency at the next moment is finally optimized.
  • This process is consistent with the actual human control method of the boiler: first, through human observation of the current Operating data, such as load, steam pressure, etc., are then used to fine-tune some decision-making parameters, in the hope that a certain operating parameter of the boiler can be optimized in the future.
  • the combustion optimization target is specified by the user, such as boiler thermal efficiency, coal consumption, pollutant emission concentration, etc., and a recurrent neural network structure is mainly used to establish a nonlinear prediction model based on a small number of determined boiler operation monitoring parameters.
  • the optimization method provided by the present invention combines the actual boiler control experience and the principle of the recurrent neural network algorithm to build a special offset data structure, so that it can be combined with evolutionary calculation to meet the actual optimization needs of boiler production.
  • the boiler operation decision at the current moment is optimized through continuous evolutionary calculation.
  • the time consumption of evolutionary calculation is usually within an original working condition time interval, which meets the online optimization needs of the boiler.
  • Figure 1 is a structural block diagram of the industrial boiler combustion optimization method
  • Figure 2 shows the hierarchical stacking structure of the deep neural network
  • FIG3 is a detailed flow chart of an industrial boiler combustion optimization method
  • Figure 4 is a fitting diagram of the boiler combustion nonlinear model.
  • An industrial boiler combustion optimization method based on deep recurrent neural networks and evolutionary computation mainly consists of two parts: 1) nonlinear offline modeling of boiler combustion; 2) online optimization of boiler combustion decisions.
  • nonlinear offline modeling and online evolutionary calculations are performed to obtain real-time decision optimization results.
  • the nonlinear offline modeling of boiler combustion adopts a recurrent neural network structure.
  • This embodiment uses a long short-term memory neural network layer (LSTM_net).
  • LSTM_net long short-term memory neural network layer
  • the built nonlinear boiler combustion model (BoilerModel, BM) is a neural network hierarchical stack structure. After debugging, the final model consists of 1 layer of Conv_net network layer, 3 layers of LSTM_net network layer and 2 layers from top to bottom.
  • Layer Dense_net network layer in which the Conv_net extraction sequence length is 2, the output feature sequence length is 30, and the number of nodes in other network layers is 40, 40, 30, 30, 1 in order.
  • the Conv_net network layer activation function uses the ReLU function; the Dense_net network layer uses the Sigmoid function. Due to the large depth of the network, Batch Normalization is used between the two layers of Dense_net to optimize the loss distribution of the underlying network. Dropout regularization is used for the top LSTM_net network layer, and the dropout ratio is 0.1.
  • the offset data structure has three types of parameters: 1) Decision parameters to be optimized (Control); 2) Boiler operating status parameters (Status); 3) Optimization target parameters (Target).
  • the Control parameter in input
  • the output Y is the Target parameter in the future, which is consistent with the time of the Control parameter;
  • the recurrent neural network can "learn" the nonlinear relationship between the feature x i and the target Y, as well as the nonlinear relationship between the past, present, and future time series distribution of the variable X and the target Y.
  • the boiler user determines the optimization target Target parameters. According to different Target Select a certain type of Control parameter and Status parameter. The determination of these parameters is based on the actual boiler control experience.
  • the optimization object of this embodiment is a 130t/h small-scale circulating fluidized bed industrial boiler, and the optimization target Target is thermal efficiency. Then the Control parameter and Status parameter are determined as follows:
  • Control parameters left primary air volume, right primary air volume, total coal supply volume, and average oxygen concentration at the furnace outlet.
  • the final BM model Since the number of data features determined is only 10, and the number of neural network model nodes is also small after debugging, the final BM model has a total of 34,491 parameters to be trained. The fewer model parameters help reduce the time required for subsequent evolutionary calculation iterations.
  • Random error processing perform Kalman filtering on History(X,Y);
  • the sequence pool sequence is disrupted to facilitate model convergence.
  • the sequence data in the pool is divided into training set and test set in proportion.
  • the training set is imported into the BM model, and the weight parameters of the neural network are continuously updated by back propagation loss (the error between the true value of the training set and the model prediction value).
  • the final training and fitting results are shown in Figure 4.
  • the ideal fitting line and the model fitting line of the model on the test set overlap well, indicating that the model can better predict the boiler thermal efficiency value after training.
  • This step is offline training. After the training is completed, the optimal weight parameters are saved to obtain the trained nonlinear model BM_trd.
  • t is the current time.
  • ⁇ C represents the continuous change of the Control parameter, and its specific optimal value is obtained by differential evolution iteration.
  • the algorithm search range of ⁇ C is defined as [( ⁇ C) min , ( ⁇ C) max ].
  • ( ⁇ C) min [-3000, -3000, -0.5, -0.5]
  • ( ⁇ C) max [3000, 3000, 0.5, 0.5].
  • Control t + ⁇ C Control t + 1 .
  • objV represents the fitness value
  • Norm( Xc ) indicates that Xc adopts the same normalization process used by BM
  • Xc [-1,-1] represents the thermal efficiency value of the boiler at the current moment
  • BM_trd(Norm( Xc ))- Xc [-1,-1] represents ⁇ Target. If it is a negative value, it means that the adoption of the current ⁇ C strategy will lead to a decrease in future thermal efficiency, and vice versa
  • a is the exponential expansion coefficient, which is 200
  • ⁇ Target is standardized to (0,+ ⁇ ) to adapt to some evolutionary algorithm selection operators (objV>0).
  • the original ⁇ Target is usually small and difficult to distinguish.
  • the gap in ⁇ Target is widened, which is conducive to the evolutionary calculation to select the strategy ⁇ C with a larger ⁇ Target.
  • the optimal individual of the ⁇ C population is the optimal decision change amount of the output.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biophysics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Computer Hardware Design (AREA)
  • Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Biomedical Technology (AREA)
  • Geometry (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Control Of Steam Boilers And Waste-Gas Boilers (AREA)

Abstract

本发明公开了一种基于深度循环神经网络和进化计算的工业锅炉燃烧优化方法:(1)锅炉燃烧非线性离线建模:采用循环神经网络结构搭建锅炉燃烧非线性模型BM,训练参数采用偏移数据结构对BM进行训练,得到训练完成后的非线性模型BM_trd;锅炉燃烧决策在线优化:根据非线性模型BM_trd和偏移数据结构建立在线决策优化方法,寻找最优的决策变化量组合,使得下一时刻的热效率得到优化。该方法在工业锅炉运行参数种类较少的情况下,实现工业锅炉在线燃烧优化,将优化决策结果直接或间接用于锅炉运行调整,使锅炉运行更加清洁、高效。

Description

一种基于深度循环神经网络和进化计算的工业锅炉燃烧优化方法 技术领域
本发明涉及锅炉燃烧技术领域,具体涉及一种基于深度循环神经网络和进化计算的工业锅炉燃烧优化方法。
背景技术
当前,我国工业锅炉在市场中仍然占据较大的比例。然而,工业锅炉并不像电站锅炉那样,有着完善的自动化控制系统和丰富的运行监测数据类型。随着国内地方标准对工业锅炉污染物排放标准愈发严格、锅炉运行成本不断提高,当前工业锅炉面临降低污染物排放和实现经济运行的困难。在污染物控制方面,为每台工业锅炉配备昂贵的污染物处理装置往往不切实际;在经济运行方面,工业锅炉仍然以人为调控锅炉运行为主,往往只关注锅炉是否能顺利启动、是否能稳定运行,并不关注锅炉是否经济运行。
随着人工智能的快速发展,基于大数据的锅炉燃烧优化技术能为锅炉提供优化的运行控制策略,能从历史运行数据中学习较优的控制经验,从而代替人为决策或辅助人为决策,实现锅炉进一步的经济、清洁运行。目前,已有较多基于大数据的大型锅炉燃烧优化方法,其中以建立静态燃烧预测模型为主,例如基于稳态工况历史数据,建立前馈神经网络、支持向量机等非线性预测模型,该类方法主要存在一些缺陷:
1)模型数据所需工况要求高,需要筛选稳态运行工况下的数据,而实际锅炉运行过程中,通常难以保持稳态运行,多随外界负荷需求的变动不断调节工况。
2)静态模型只适应于锅炉稳态工况下的优化,而当锅炉负荷波动时,静态的燃烧模型对当前锅炉运行的预测较差。
3)静态模型对数据类型要求高,因为仅考虑输入变量与输出变量之间的非线性特征,而未考虑变量本身在时间维度上的特征。
目前基于动态模型采用进化计算方法相对较少,一些采用循环神经网络建模的方法中,通常依赖较多的锅炉监测数据类型,一方面模型较为复杂,计算缓慢,另一方面不适用于工业锅炉数据缺乏的情况。如公开号为CN109670629A的中国专利公开的基于长短期记忆神经网络的燃煤锅炉热效率预测方法;公开号为CN110705881A的中国专利公开的一种基于人工神经网络的过滤效率在线计算方法及系统。
因此,基于动态建模的大数据工业锅炉燃烧优化方法的建立较为困难,其难点在于数据结构的设计和与优化算法的联合。
发明内容
本发明的目的在于提供一种基于深度循环神经网络和进化计算的工业锅炉燃烧优化方法,在工业锅炉运行参数种类较少的情况下,实现工业锅炉在线燃烧优化,将优化决策结果直接或间接用于锅炉运行调整,使锅炉运行更加清洁、高效。
本发明提供如下技术方案:
一种基于深度循环神经网络和进化计算的工业锅炉燃烧优化方法,所述方法包括:
(1)锅炉燃烧非线性离线建模:采用循环神经网络结构搭建锅炉燃烧非线性模型BM,训练参数采用偏移数据结构对BM进行训练,得到训练完成后的非线性模型BM_trd;
(2)锅炉燃烧决策在线优化:根据非线性模型BM_trd和偏移数据结构建立在线决策优化方法,寻找最优的决策变化量组合,使得下一时刻的热效率得到优化。
锅炉燃烧非线性离线建模需采用循环神经网络结构,例如普通循环神经网络层(RNN)、长短期记忆神经网络层(LSTM)和门控循环神经网络层(GRU)。所搭建的锅炉燃烧非线性模型(BoilerModel,BM)为神经网络层级堆叠结构,从上往下依次为单层卷积神经网络层(Conv)、若干循环神经网络层和若干全连接神经网络层(Dense)。
其中,根据不同的锅炉对象,所搭建模型具有不同的循环网络层数、全连接网络层数、各层网络节点数和节点激活函数,因而具有锅炉自定义性。当网络深度较大时,在多层全连接层之间采用批标准化层(Batch Normalization),优化底层网络的损失分布。对顶层循环神经网络的层输入单元采用dropout正则化,对循环单元采用recurrent dropout正则化,缓解神经网络训练时的过拟合问题。
BM模型具有特殊的训练数据结构:偏移数据结构。由于循环神经网络具有时间序列记忆特性,因而输入数据需包含时间维度,例如{(Xt-2,Xt-1,Xt),X=(x1,x2,…,xn)}表示时间序列长度为3的n维输入X。以下对偏移数据结构进行定义:
Y=Targett+1
所述偏移数据结构,具有三类参数分别为:1)待优化决策参数Control:Controlt-n+1、Controlt和Controlt+1;2)锅炉运行状态参数Status:Statust-n、Statust-1和Statust;3)优化目标参数Target:Targett-n、Targett-1和Targett
在时间维度上,输入X中的Control参数需比Status参数和Target参数多前进一个时刻,即所谓的“偏移”为Control参数在时间维度上的前向偏移(向未来方向偏移);输出Y为将来时刻的Target参数,与Control参数所处时刻一致;X的时间序列长度定义为无偏移参数的时间间隔(t-(t-n)+1=n+1),需根据调试获得较优值;基于偏移数据结构,循环神经网络可“学习”特征xi与目标Y的非线性关系,以及变量X的过去、现在、将来的时间序列分布与目标Y的非线性关系。
偏移数据结构中的3类参数,由锅炉用户确定优化目标Target参数,根据不同的Target参数选择确定类型的Control参数和Status参数,这些参数的确定依据了锅炉实际调控经验:
1)Target参数为热效率、煤耗量时,Status参数为:锅炉负荷(热负荷或电负荷)、主蒸汽压力(或供水压力)、锅炉燃烧室温度梯度(从燃烧器端至炉膛出口处各温度测量值)、排烟温度(最后一级换热器之后);Control参数为:一般视锅炉类型而定,主要优化风量、燃料量以及风门开度,一般为一次风量、炉膛出口氧气浓度、给燃料量、二次风门开度。
2)Target参数为污染物浓度时,Status参数和Control参数在1)的基础上,Status参数应增加入炉燃料性质。例如针对NOx污染物需增加挥发分参数,针对SO2污染物需增加硫分参数,针对其他污染物由用户自定义选择燃料性质。
由于确定的数据特征数较少,通常为10-15,满足工业锅炉数据类型较少的现状,同时可减小神经网络模型节点数,搭建“轻量化”模型,减少后续优化时的计算量。
最终,锅炉非线性离线建模流程如下:
1)从锅炉原始运行数据中提取数据集,包括Target参数、Control参数和Status参数;
2)系统误差位置索引:人为确定目标值范围,例如针对热效率,当热效率大于1或小于0.7时认定无效,通过遍历目标提取无效数据的位置索引;
3)数据标准化或归一化;
4)随机误差处理:对数据进行滤波处理,例如卡尔曼滤波;
5)系统误差处理:基于2)的位置索引,提取相邻位置索引间隔大于100的中间数据段,该操作既可分离误差,又可保证数据的连续性;
6)建立序列池;
7)将5)提取得到的所有数据段按偏移数据结构进行变换,将每一条序列(X,Y)存放至序列池中。
8)将序列池内的序列顺序打乱,利于模型收敛。按比例将数据分为模型训练集和测试集,将训练集导入BM模型中,通过反向传播损失(训练集真值与模型预测值的误差),不断迭代更新神经网络的权重参数。通过不断调整模型超参数(神经网络节点数、神经网络层数、偏移数据结构的时间序列长 度)得到最优化模型。该步骤为离线训练,训练完成后保存最优的权重参数,得到训练后的非线性模型BM_trd。
基于BM_trd,建立在线决策优化方法,其过程依赖偏移数据结构,具体方法定义如下:
1)提取当前锅炉运行数据,建立一个序列长度的输入Xc(X_current):
t为当前时刻;
2)将Control参数向前偏移,得到偏移后的输入Xco(X_current_offset):
3)建立优化对象ΔC种群:ΔC表示Control参数的连续变化量,其具体最优值由进化计算迭代所得,定义算法搜索ΔC的范围为[(ΔC)min,(ΔC)max],将ΔC加入Xco后得到Xcoi,则ΔC种群转化为Xcoi种群:
其中,Controlt+ΔC=Controlt+1;(ΔC)min和(ΔC)max由人为定义,一般设定搜索的最大最小范围较小,避免锅炉调控幅度过大导致运行波动、发生事故等。最终ΔC种群的最优个体即为输出的最优决策变化量。
锅炉燃烧决策优化方法采用的进化计算方法包括群智能算法和进化算法。针对不同问题,进化计算方法需要建立适宜的适应性函数(fitness function),以下定义本方法的适应性函数:
s.t.(ΔC)min≤ΔC≤(ΔC)max
其中,objV表示适应度值;Norm(Xc)表示Xc采用BM所用的相同标准化、归一化过程;Xc[-1,-1]表示当前时刻t锅炉运行的Target值,则Bm_trd(Norm(Xc))-Xc[-1,-1]表示ΔTarget,若为负值,表示采用当前ΔC策略时将导致未来Target降低,反之亦然;a为指数扩大系数,主要将ΔTarget规范至(0,+∞),以适应部分进化算法的选择算子(objV>0),同时原始ΔTarget通常较小,难以区分,而通过指数扩大后,ΔTarget的差距被拉大,有利于进化计算对拥有较大ΔTarget的策略ΔC进行选择,例如利于“轮盘赌选择”等;最终根据调试获得适宜大小的a。
基于进化计算的燃烧优化过程,就是不断迭代进化ΔC种群,通过寻找最优的决策变化量组合,最终使得下一时刻的热效率得到优化,该过程与锅炉实际人为调控方法一致:先通过人为观察当前运行数据,例如负荷、蒸汽压力等,再去微调部分决策参数,以此期待未来锅炉某个运行参数能得到优化。
本发明在锅炉燃烧非线性离线建模时,燃烧优化目标由用户指定,例如锅炉热效率、煤耗量、污染物排放浓度等,主要采用循环神经网络结构,基于少量确定的锅炉运行监测参数建立非线性预测模型。
本发明提供的优化方法中由于锅炉燃烧非线性离线建模中所需数据类别较少,适应于工业锅炉运行参数种类偏少的状况。工业锅炉种类包括小型生 物质锅炉、燃气锅炉、燃油锅炉、循环流化床锅炉等,不包含小型电加热锅炉。然而锅炉大同小异,本燃烧优化方法也同样适用于大型电站锅炉。
本发明提供的优化方法结合锅炉实际调控经验和循环神经网络算法原理,搭建特殊的偏移数据结构,从而可联合进化计算,符合锅炉生产实际优化需求。基于离线模型,通过连续的进化计算,对当前时刻锅炉运行决策进行优化,进化计算耗时通常在一个原始工况时间间隔以内,满足锅炉在线优化需求。
附图说明
图1为工业锅炉燃烧优化方法的结构框图;
图2为深度神经网络层级堆叠结构;
图3为工业锅炉燃烧优化方法详细流程图;
图4为锅炉燃烧非线性模型拟合图。
具体实施方式
下面结合附图对本发明具体实施方式做进一步说明,但不限于这些实施例。
一种基于深度循环神经网络和进化计算的工业锅炉燃烧优化方法,如图1所示,主要包含两个部分:1)锅炉燃烧非线性离线建模;2)锅炉燃烧决策在线优化。针对不同的工业锅炉对象,通过采集历史数据和在线数据,分别进行非线性离线建模和在线进化计算得到实时的决策优化结果。
锅炉燃烧非线性离线建模采用循环神经网络结构,本实施例采用长短期记忆神经网络层(LSTM_net)。如图2所示,所搭建的锅炉燃烧非线性模型(BoilerModel,BM)为神经网络层级堆叠结构,经调试,最终模型从上往下依次为1层Conv_net网络层、3层LSTM_net网络层和2层Dense_net网络层,其中,Conv_net提取序列长度尺寸为2、输出特征序列长度为30,其他网络层节点数依次为40,40,30,30,1。Conv_net网络层激活函数采用ReLU函数;Dense_net网络层均采用Sigmoid函数。由于网络深度较大,在两层Dense_net之间采用Batch Normalization,优化底层网络的损失分布。对顶层LSTM_net网络层采用dropout正则化,dropout的比例为0.1。
BM模型训练数据采用“偏移数据结构”,其定义如下:
Y=Targett+1
偏移数据结构,具有三类参数分别为:1)待优化决策参数(Control);2)锅炉运行状态参数(Status);3)优化目标参数(Target)。在时间维度上,输入X中的Control参数需比Status参数和Target参数多前进一个时刻,即所谓的“偏移”为Control参数在时间维度上的前向偏移(向未来方向偏移);输出Y为将来时刻的Target参数,与Control参数所处时刻一致;X的时间序列长度定义为无偏移参数的时间间隔(t-(t-n)+1=n+1,本案例中n=2);基于偏移数据结构,循环神经网络可“学习”特征xi与目标Y的非线性关系,以及变量X的过去、现在、将来的时间序列分布与目标Y的非线性关系。偏移数据结构中的3类参数,由锅炉用户确定优化目标Target参数,根据不同的Target 选择确定类型的Control参数和Status参数,这些参数的确定依据了锅炉实际调控经验。
本实施例的优化对象为130t/h小型循环流化床工业锅炉,优化目标Target为热效率,则确定Control参数和Status参数如下:
1)Status参数:锅炉负荷、主蒸汽压力、平均床温、炉膛出口平均温度、平均排烟温度;
2)Control参数:左侧一次风量、右侧一次风量、总给煤量、炉膛出口平均氧气浓度。
由于确定的数据特征数只有10个,神经网络模型节点数经过调试也较少,最终BM模型待训练总参数共34491个,模型参数较少,利于减少后续进化计算迭代所需的时间。
基于搭建的模型结构和数据结构,完成锅炉燃烧非线性离线建模,其流程如图3中的(a)所示:
1)提取锅炉历史运行数据若干条,建立总数据集History(X,Y);
2)系统误差位置索引:当热效率大于1或小于0.7时认定无效,通过遍历History(Y)提取无效数据的位置索引;
3)对总数据集History(X)进行标准化;
4)随机误差处理:对History(X,Y)进行卡尔曼滤波处理;
5)系统误差处理:基于2)的位置索引,提取相邻位置索引间隔大于100的中间数据段,该操作既可分离误差,又可保证数据的连续性;
6)建立序列池;
7)将5)提取到的中间数据段按偏移数据结构进行变换,将每一条序列(X,Y)存放至序列池中。
8)将序列池序列打乱,利于模型收敛。按比例将池内序列数据为训练集和测试集。将训练集导入BM模型,通过反向传播损失(训练集真值与模型预测值的误差),采用RMSProp梯度下降算法不断迭代更新神经网络的权重参数,最终训练拟合结果如图4所示,图中模型在测试集上理想拟合线和模型拟合线重合情况较好,表明模型训练后能较好的预测锅炉热效率值。该步骤为离线训练,训练完成后保存最优的权重参数,得到训练后的非线性模型BM_trd。
依靠偏移数据结构,结合实际调控人员的锅炉控制经验,基于差分进化算法建立决策在线优化方法,其流程如图3中的(b)所示:
1)提取当前锅炉运行数据,建立一个序列长度的输入Xc(X_current):
t为当前时刻。
2)将Control参数向前偏移,得到偏移后的输入Xco(X_current_offset):
3)建立优化对象ΔC种群:ΔC表示Control参数的连续变化量,其具体最优值由差分进化迭代所得,定义算法搜索ΔC的范围为[(ΔC)min,(ΔC)max],其 中(ΔC)min=[-3000,-3000,-0.5,-0.5],(ΔC)max=[3000,3000,0.5,0.5]。将ΔC加入Xco后得到Xcoi,则ΔC种群转化为Xcoi种群:
其中,Controlt+ΔC=Controlt+1
4)计算个体适应度:将Xcoi输入BM_trd模型后,得到下一时刻热效率的预测值,则定义适应性函数:
s.t.(ΔC)min≤ΔC≤(ΔC)max
其中,objV表示适应度值;Norm(Xc)表示Xc采用BM所用的相同标准化过程;Xc[-1,-1]表示当前时刻锅炉运行的热效率值,则BM_trd(Norm(Xc))-Xc[-1,-1]表示ΔTarget,若为负值,表示采用当前ΔC策略时导致未来热效率降低,反之亦然;a为指数扩大系数,为200,将ΔTarget规范至(0,+∞),以适应部分进化算法选择算子(objV>0),同时原始ΔTarget通常较小,难以区分,而通过指数扩大后,ΔTarget的差距被拉大,有利于进化计算对拥有较大ΔTarget的策略ΔC进行选择。
5)差分进化:若不满足算法收敛条件,则进行种群的差分变异、重组和选择,其优化过程,就是不断迭代进化ΔC种群,通过寻找最优的决策变化量组合,最终使得下一时刻的热效率得到优化,该过程与锅炉实际人为调控方法一致:先通过人为观察当前运行数据,例如负荷、蒸汽压力等,再去微调部分决策参数,以此期待未来锅炉某个运行参数能得到优化。
6)ΔC种群的最优个体即为输出的最优决策变化量。
最终,该实施例决策优化结果如表1所示,针对20个不同的原始工况进行优化,平均计算时长为130s,远小于原始工况时间间隔(10min),满足在线优化要求。热效率变化量基本为正值,表明通过燃烧决策优化实现热效率提升具有较大的空间。该实施例的优化结果表明,本工业锅炉燃烧优化方法具有可行性。
表1实施例决策优化结果

Claims (8)

  1. 一种基于深度循环神经网络和进化计算的工业锅炉燃烧优化方法,其特征在于,所述方法包括:
    (1)锅炉燃烧非线性离线建模:采用循环神经网络结构搭建锅炉燃烧非线性模型BM,训练参数采用偏移数据结构对BM进行训练,得到训练完成后的非线性模型BM_trd;
    (2)锅炉燃烧决策在线优化:根据非线性模型BM_trd和偏移数据结构建立在线决策优化方法,寻找最优的决策变化量组合,使得下一时刻的热效率得到优化。
  2. 根据权利要求1所述的基于深度循环神经网络和进化计算的工业锅炉燃烧优化方法,其特征在于,所述循环神经网络结构为普通循环神经网络RNN、长短期记忆神经网络层LSTM或门控循环神经网络层GRU;搭建的锅炉燃烧非线性模型BM为神经网络层级堆叠结构,从上往下依次为单层卷积神经网络层Conv、若干循环神经网络层和若干全连接神经网络层Dense。
  3. 根据权利要求2所述的基于深度循环神经网络和进化计算的工业锅炉燃烧优化方法,其特征在于,在步骤(1)中,所述偏移数据结构为:
    Y=Targett+1
    所述偏移数据结构,具有三类参数分别为:1)待优化决策参数Control:Controlt-n+1、Controlt和Controlt+1;2)锅炉运行状态参数Status:Statust-n、Statust-1和Statust;3)优化目标参数Target:Targett-n、Targett-1和Targett
    在时间维度上,输入X中的Control参数需比Status参数和Target参数多前进一个时刻,即所谓的“偏移”为Control参数在时间维度上的前向偏移,即,向未来方向偏移;输出Y为将来时刻的Target参数,与Control参数所处时刻一致;X的时间序列长度定义为无偏移参数的时间间隔t-(t-n)+1=n+1。
  4. 根据权利要求3所述的基于深度循环神经网络和进化计算的工业锅炉燃烧优化方法,其特征在于,由锅炉用户确定优化目标Target参数,根据不同的Target参数选择确定类型的Control参数和Status参数,具体为:
    1)当Target参数为热效率、煤耗量时,Status参数为:锅炉负荷、主蒸汽压力或供水压力、锅炉燃烧室温度梯度、排烟温度;Control参数为:风量、燃料量以及风门开度;
    2)当Target参数为污染物浓度时,Status参数和Control参数在1)的基础上,Status参数增加入炉燃料性质。
  5. 根据权利要求4所述的基于深度循环神经网络和进化计算的工业锅炉燃烧优化方法,其特征在于,所述锅炉负荷为热负荷或电负荷,锅炉燃烧室温度梯度为从燃烧器端至炉膛出口处各温度测量值,排烟温度为最后一级换热器之后的排烟温度;Control参数为:一次风量、炉膛出口氧气浓度、给燃料量、二次风门开度;入炉燃料性质:当针对NOx污染物时为挥发分参数,当针对SO2污染物时为硫分参数。
  6. 根据权利要求5所述的基于深度循环神经网络和进化计算的工业锅炉燃烧优化方法,其特征在于,锅炉燃烧非线性离线建模的过程为:
    1)从锅炉原始运行数据中提取数据集,包括Target参数、Control参数和Status参数;
    2)系统误差位置索引;
    3)数据标准化或归一化;
    4)随机误差处理;
    5)系统误差处理:基于步骤2)的位置索引,提取相邻位置索引间隔大于100的中间数据段进行系统误差处理;
    6)建立序列池;
    7)将5)提取得到的所有数据段按偏移数据结构进行变换,将每一条序列(X,Y)存放至序列池中;
    8)将序列池内的序列顺序打乱;按比例将数据分为模型训练集和测试集,将训练集导入BM模型中,通过反向传播损失,不断迭代更新神经网络的权重参数;通过不断调整模型超参数得到最优化模型;该步骤为离线训练,训练完成后保存最优的权重参数,得到训练后的非线性模型BM_trd。
  7. 根据权利要求6所述的基于深度循环神经网络和进化计算的工业锅炉燃烧优化方法,其特征在于,在步骤2)中,系统误差位置索引的方法为人为确定目标值范围,当针对热效率:当热效率大于1或小于0.7时认定无效,通过遍历目标提取无效数据的位置索引;在步骤4)中随机误差处理为对数据进行滤波处理;在步骤8)中,模型超参数包括神经网络节点数、神经网络层数、偏移数据结构的时间序列长度。
  8. 根据权利要求1所述的基于深度循环神经网络和进化计算的工业锅炉燃烧优化方法,其特征在于,在步骤(2)中,所述决策在线优化方法为:
    1)提取当前锅炉运行数据,建立一个序列长度的输入Xc(X_current):
    t为当前时刻;
    2)将Control参数向前偏移,得到偏移后的输入Xco(X_current_offset):
    3)建立优化对象ΔC种群:ΔC表示Control参数的连续变化量,其具体最优值由进化计算迭代所得,定义算法搜索ΔC的范围为[(ΔC)min,(ΔC)max],将ΔC加入Xco后得到Xcoi,则ΔC种群转化为Xcoi种群:
    其中,Controlt+ΔC=Controlt+1
    4)计算个体适应度:将Xcoi输入BM_trd模型后,得到下一时刻热效率的预测值,则定义适应性函数:

    s.t.(ΔC)min≤ΔC≤(ΔC)max
    其中,objV表示适应度值;Norm(Xc)表示Xc采用标准化、归一化过程;Xc[-1,-1]表示当前时刻t锅炉运行的Target值,则BM_trd(Norm(Xc))-Xc[-1,-1]表示ΔTarget,若为负值,表示采用当前ΔC策略时将导致未来Target 降低,反之亦然;a为指数扩大系数,用于将适应度值放大,最终根据实际调试获得适宜值;
    5)差分进化:若不满足算法收敛条件,则进行种群的差分变异、重组和选择,其优化过程,就是不断迭代进化ΔC种群,通过寻找最优的决策变化量组合,最终使得下一时刻的热效率得到优化,该过程与锅炉实际人为调控方法一致:先通过人为观察当前运行数据,再去微调部分决策参数,以此期待未来锅炉某个运行参数能得到优化;
    6)ΔC种群的最优个体即为输出的最优决策变化量。
PCT/CN2023/075002 2022-09-21 2023-02-08 一种基于深度循环神经网络和进化计算的工业锅炉燃烧优化方法 WO2024060488A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202211153466.6A CN115495984A (zh) 2022-09-21 2022-09-21 一种基于深度循环神经网络和进化计算的工业锅炉燃烧优化方法
CN202211153466.6 2022-09-21

Publications (1)

Publication Number Publication Date
WO2024060488A1 true WO2024060488A1 (zh) 2024-03-28

Family

ID=84469794

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/075002 WO2024060488A1 (zh) 2022-09-21 2023-02-08 一种基于深度循环神经网络和进化计算的工业锅炉燃烧优化方法

Country Status (2)

Country Link
CN (1) CN115495984A (zh)
WO (1) WO2024060488A1 (zh)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114418358A (zh) * 2021-12-31 2022-04-29 新奥数能科技有限公司 锅炉调度方法及装置
CN118070713A (zh) * 2024-04-18 2024-05-24 苏州工业园区蒙纳士科学技术研究院 数值模拟与人工神经网络相耦合的焦化过程实时模拟方法
CN118194607A (zh) * 2024-05-16 2024-06-14 江苏太湖锅炉股份有限公司 一种用于锅炉机组数据动态建模方法及系统
CN118286838A (zh) * 2024-04-01 2024-07-05 浙江浩普智能科技有限公司 燃烧-污染物治理全流程智能调控减污降碳方法及系统
CN118429137A (zh) * 2024-05-27 2024-08-02 淮阴工学院 一种林场建筑物运行系统及方法
CN118499816A (zh) * 2024-07-18 2024-08-16 大连富氧燃烧新技术开发有限责任公司 一种煤粉锅炉燃烧优化及增氧助燃节能降碳方法

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115495984A (zh) * 2022-09-21 2022-12-20 浙江大学 一种基于深度循环神经网络和进化计算的工业锅炉燃烧优化方法
CN117490092B (zh) * 2024-01-02 2024-03-19 中电行唐生物质能热电有限公司 基于人工智能的生物质电厂燃烧控制决策优化方法及系统

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20060178762A1 (en) * 2005-02-08 2006-08-10 Pegasus Technologies, Inc. Method and apparatus for optimizing operation of a power generating plant using artificial intelligence techniques
CN104763999A (zh) * 2015-03-04 2015-07-08 内蒙古瑞特优化科技股份有限公司 电厂煤粉锅炉燃烧性能在线优化方法和系统
CN109670629A (zh) * 2018-11-16 2019-04-23 浙江蓝卓工业互联网信息技术有限公司 基于长短期记忆神经网络的燃煤锅炉热效率预测方法
CN115495984A (zh) * 2022-09-21 2022-12-20 浙江大学 一种基于深度循环神经网络和进化计算的工业锅炉燃烧优化方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20060178762A1 (en) * 2005-02-08 2006-08-10 Pegasus Technologies, Inc. Method and apparatus for optimizing operation of a power generating plant using artificial intelligence techniques
CN104763999A (zh) * 2015-03-04 2015-07-08 内蒙古瑞特优化科技股份有限公司 电厂煤粉锅炉燃烧性能在线优化方法和系统
CN109670629A (zh) * 2018-11-16 2019-04-23 浙江蓝卓工业互联网信息技术有限公司 基于长短期记忆神经网络的燃煤锅炉热效率预测方法
CN115495984A (zh) * 2022-09-21 2022-12-20 浙江大学 一种基于深度循环神经网络和进化计算的工业锅炉燃烧优化方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
SHAO-YOU NI, YAO CHE-PENG, CHEN LING-HONG, WU XUE-CHENG, CEN KE-FU: "Prediction Analysis of SO2 Emissions from Circulating Fluidized Bed Boiler Based on BP Neural Network", ENERGY ENGINEERING, vol. 6, 1 December 2021 (2021-12-01), pages 31 - 37, XP093149268 *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114418358A (zh) * 2021-12-31 2022-04-29 新奥数能科技有限公司 锅炉调度方法及装置
CN118286838A (zh) * 2024-04-01 2024-07-05 浙江浩普智能科技有限公司 燃烧-污染物治理全流程智能调控减污降碳方法及系统
CN118070713A (zh) * 2024-04-18 2024-05-24 苏州工业园区蒙纳士科学技术研究院 数值模拟与人工神经网络相耦合的焦化过程实时模拟方法
CN118194607A (zh) * 2024-05-16 2024-06-14 江苏太湖锅炉股份有限公司 一种用于锅炉机组数据动态建模方法及系统
CN118429137A (zh) * 2024-05-27 2024-08-02 淮阴工学院 一种林场建筑物运行系统及方法
CN118499816A (zh) * 2024-07-18 2024-08-16 大连富氧燃烧新技术开发有限责任公司 一种煤粉锅炉燃烧优化及增氧助燃节能降碳方法

Also Published As

Publication number Publication date
CN115495984A (zh) 2022-12-20

Similar Documents

Publication Publication Date Title
WO2024060488A1 (zh) 一种基于深度循环神经网络和进化计算的工业锅炉燃烧优化方法
CN105020705B (zh) 循环流化床锅炉燃烧性能实时优化控制方法及系统
CN104534507B (zh) 一种锅炉燃烧优化控制方法
CN111144609A (zh) 一种锅炉废气排放预测模型建立方法、预测方法及装置
CN104763999A (zh) 电厂煤粉锅炉燃烧性能在线优化方法和系统
CN106019935B (zh) 基于约束模糊关联规则的多目标锅炉燃烧优化
CN104613468A (zh) 基于模糊自适应推理的循环流化床锅炉燃烧优化控制方法
CN115111601B (zh) 多变负荷下内嵌算法融合的多目标锅炉燃烧优化控制方法
CN113359425A (zh) 一种基于lstm神经网络pid优化的火电厂锅炉主汽温智能控制系统
CN113742997B (zh) 城市固废焚烧过程风量智能优化设定方法
CN114881207A (zh) 基于lstm深度循环神经网络的蒸汽压力预测方法及装置
WO2020052413A1 (zh) 火力发电机组燃烧控制优化方法、装置及可读存储介质
CN111881120A (zh) 一种锅炉智能运行优化方法
Zheng et al. Data-driven based multi-objective combustion optimization covering static and dynamic states
CN114958395B (zh) 一种生物质炭化炉余热利用方法及其利用系统
Tian et al. Furnace Temperature Control based on Adaptive TS-FNN for Municipal Solid Waste Incineration Process
CN113887130B (zh) 一种基于集成学习的工业锅炉运行优化方法
CN117031938A (zh) 一种城市固废焚烧炉膛温度的模糊神经网络控制方法
CN111428906B (zh) 一种基于图像变换的工业锅炉蒸汽量预测方法
Wang et al. Data-driven multi-objective intelligent optimal control of municipal solid waste incineration process
Wenjie et al. The optimizing for boiler combustion based on fuzzy association rules
CN113408185A (zh) 一种城市生活垃圾焚烧过程炉排空气流量设定方法
Pan et al. Feature Selection and Hyperparameter Optimization for Combustion Status Recognition Model of MSWI Process Based on Parallel Differential Evolution
Manke A Performance Comparison of Different Back Propagation Neural Networks for Nitrogen Oxides Emission Prediction in Thermal Power Plant
Arafat et al. The Development of a Matlab-Based Fuzzy PID Controller and The Simulation

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: 23866792

Country of ref document: EP

Kind code of ref document: A1