WO2021208343A1 - 一种基于神经网络模型的喷水减温器喷水调整方法及装置 - Google Patents

一种基于神经网络模型的喷水减温器喷水调整方法及装置 Download PDF

Info

Publication number
WO2021208343A1
WO2021208343A1 PCT/CN2020/114593 CN2020114593W WO2021208343A1 WO 2021208343 A1 WO2021208343 A1 WO 2021208343A1 CN 2020114593 W CN2020114593 W CN 2020114593W WO 2021208343 A1 WO2021208343 A1 WO 2021208343A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
water spray
parameter data
neural network
network model
Prior art date
Application number
PCT/CN2020/114593
Other languages
English (en)
French (fr)
Inventor
文立斌
孙艳
李俊
张翌晖
吴健旭
丘浩
窦骞
Original Assignee
广西电网有限责任公司电力科学研究院
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 广西电网有限责任公司电力科学研究院 filed Critical 广西电网有限责任公司电力科学研究院
Publication of WO2021208343A1 publication Critical patent/WO2021208343A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • F—MECHANICAL ENGINEERING; LIGHTING; HEATING; WEAPONS; BLASTING
    • F22—STEAM GENERATION
    • F22G—SUPERHEATING OF STEAM
    • F22G5/00—Controlling superheat temperature
    • F22G5/12—Controlling superheat temperature by attemperating the superheated steam, e.g. by injected water sprays
    • F22G5/123—Water injection apparatus

Definitions

  • the invention relates to the technical field of water spray adjustment of a water spray desuperheater for thermal power generation, in particular to a water spray adjustment method and device of a water spray desuperheater based on a neural network model.
  • Boiler water spray desuperheating system is indispensable in large thermal power units. Its function is to regulate, reduce temperature or pressure of the over-temperature and large-flow steam flowing through the boiler; it is an important guarantee for the safe operation of boilers and steam turbines, and makes the steam entering the steam turbine temperature Within the range of normal operation requirements of the unit; when the operating conditions of the unit change suddenly, especially after the fast load shedding (FCB) of the steam turbine unit, the temperature control of the steam produced by the boiler has an important connection with the coordinated operation of the steam turbine; the existing boiler In the water spray desuperheating system, the water spray adjustment method of the water spray desuperheater still stays after the problem occurs, and the adjustment is carried out by manual control. It is more difficult to achieve real-time adjustment, which may bring safety hazards to the National Chiao Tung University.
  • the purpose of the present invention is to overcome the shortcomings of the prior art.
  • the present invention provides a method and device for adjusting the water spray of a water spray desuperheater based on a neural network model.
  • the real-time correction and prediction of the amount of heat makes the steam heat and pressure output by the superheater within the preset range, ensuring the operating efficiency and operating safety of the thermal power boiler.
  • an embodiment of the present invention also provides a method for adjusting the spray water of a boiler spray desuperheater based on a neural network model, the method comprising:
  • the normalized current superheat parameter data and the normalized current water spray parameter data are used to construct a data characteristic matrix, wherein the data characteristic matrix is a matrix of M rows and N columns, and one row consists of the normalized current superheat parameter data The other line is composed of the current water spray parameter data after normalization;
  • the water spray desuperheater of the thermal power boiler performs water spray adjustment according to the corrected prediction result.
  • the current superheat parameter data includes steam volume data, steam outlet density data, steam inlet mass flow data, steam outlet mass flow data, pipe working fluid inlet specific enthalpy data, pipe working fluid outlet specific enthalpy data, and pipe
  • the current water spray parameter data includes: inlet steam mass flow data, inlet desuperheating water mass flow data, outlet steam mass flow data, inlet desuperheating water specific enthalpy data, inlet steam specific enthalpy data, outlet steam specific enthalpy data, inlet desuperheating water Temperature data and inlet steam temperature data.
  • the normalization processing is performed on the current superheat parameter data and the current water spray parameter data respectively to obtain normalized current superheat parameter data and normalized current water spray parameter data, include:
  • the current water spray parameter data is normalized based on the maximum and minimum value normalization method to obtain the current water spray parameter data after unitization.
  • the training step of training the convergent neural network model includes:
  • the training data feature matrix is input into the neural network model to be trained after the loss function is updated, and feature learning training is performed until convergence, so as to obtain a neural network model with training convergence.
  • the neural network model to be trained is any one of a deep neural network model, an incomplete neural network model, or a convolutional neural network model.
  • the updating the initial loss function based on a regularization term includes:
  • the initial loss function is a cross-entropy loss function or a negative log-likelihood loss function.
  • the input of the training data feature matrix into the neural network model to be trained after the updated loss function performs feature learning training until convergence;
  • the back propagation algorithm is used to reset the coefficients of all layer parameter vectors of the trained neural network model, and the training sample data feature matrix is used for retraining.
  • normalized current superheat parameter data and normalized current water spray parameter data to construct a data characteristic matrix includes:
  • the normalized current superheat parameter data and the normalized current water spray parameter data are used to construct a data characteristic matrix according to the preset data sorting position.
  • an embodiment of the present invention also provides a water spray adjusting device for a water spray desuperheater based on a neural network model, the device comprising:
  • Data acquisition module used to obtain the current superheat parameter data of the superheater in the thermal power boiler and the current water spray parameter data of the water spray desuperheater in the thermal power boiler;
  • Normalization module used to normalize the current overheating parameter data and the current spraying parameter data, respectively, to obtain normalized current overheating parameter data and normalized current spraying parameter data ;
  • Matrix building module used to construct a data feature matrix using normalized current overheating parameter data and normalized current water spray parameter data, where the data feature matrix is a matrix of M rows and N columns, and one row is normalized After the current overheating parameter data, the other line is composed of the normalized current water spray parameter data;
  • Correction prediction module used to input the data feature matrix into the neural network model for training convergence to correct and predict the water injection volume of the thermal power boiler water spray desuperheater, and obtain a correction prediction result;
  • Adjustment module used for the thermal power boiler water spray desuperheater to adjust the water spray according to the corrected prediction result.
  • real-time correction and prediction of the water injection volume of the water injection desuperheater in the thermal power boiler can be realized, so that the steam heat and pressure output by the superheater are within a preset range, and the operating efficiency of the thermal power boiler is guaranteed And operational safety.
  • Fig. 1 is a schematic flowchart of a method for adjusting spray water of a spray desuperheater based on a neural network model in an embodiment of the present invention
  • Fig. 2 is a schematic diagram of the structural composition of a water spray adjusting device of a water spray desuperheater based on a neural network model in an embodiment of the present invention.
  • FIG. 1 is a schematic flow chart of a method for adjusting the spray water of a water spray desuperheater based on a neural network model in an embodiment of the present invention.
  • a method for adjusting spray water of a boiler spray water desuperheater based on a neural network model includes:
  • the current superheat parameter data includes steam volume data, steam outlet density data, steam inlet mass flow data, steam outlet mass flow data, pipe working fluid inlet specific enthalpy data, pipe working fluid outlet ratio Enthalpy data, heat transfer data of the working fluid in the tube and the metal tube wall, the inlet density data of the working fluid in the tube, the inlet pressure data of the working fluid in the tube, the outlet pressure data of the working fluid in the tube, the flue gas inlet temperature data outside the tube, the flue gas outlet outside the tube Temperature, pipe wall temperature data, heat transfer data of flue gas to metal, heat exchange area data and pipe wall quality data;
  • the current water spray parameter data includes: inlet steam mass flow data, inlet desuperheating water mass flow data, Outlet steam mass flow data, inlet desuperheating water specific enthalpy data, inlet steam specific enthalpy data, outlet steam specific enthalpy data, inlet desuperheating water temperature data and inlet steam temperature data.
  • the superheater and the water spray desuperheater in the thermal power boiler are equipped with data acquisition sensors for collecting relevant parameter data; in general, the thermal power boiler includes multiple superheaters and multiple water spray desuperheaters. Thermostat, install the corresponding data acquisition sensor at the corresponding position in each superheater and water spray desuperheater to collect the corresponding data information.
  • the current superheat parameter data includes at least steam volume data, steam outlet density data, Steam inlet mass flow data, steam outlet mass flow data, pipe working fluid inlet specific enthalpy data, pipe working fluid outlet specific enthalpy data, pipe working fluid and metal pipe wall heat exchange data, pipe working fluid inlet density data, pipe work Mass inlet pressure data, working fluid outlet pressure data in the tube, flue gas inlet temperature data outside the tube, flue gas outlet temperature outside the tube, tube wall temperature data, heat transfer data from flue gas to metal, heat transfer area data, and tube wall temperature data Quality data; current water spray parameter data includes at least: inlet steam mass flow data, inlet desuperheating water mass flow data, outlet steam mass flow data, inlet desuperheating water specific enthalpy data, inlet steam specific enthalpy data, outlet steam specific enthalpy data, inlet Desuperheating water temperature data and inlet steam temperature data.
  • the collected noise is firstly reduced in the data acquisition sensor.
  • signal filtering is used to reduce noise.
  • the collected data information is converted into data information.
  • S12 Perform normalization processing on the current superheat parameter data and the current water spray parameter data, respectively, to obtain normalized current superheat parameter data and normalized current water spray parameter data;
  • the current superheat parameter data and the current spray parameter data are respectively normalized to obtain the normalized current superheat parameter data and the normalized current spray parameter data.
  • the water parameter data includes: normalizing the current overheating parameter data based on the maximum and minimum value normalization method to obtain the normalized current overheating parameter data;
  • the water spray parameter data is normalized to obtain the current water spray parameter data after the normalization.
  • the current overheating parameter data and the current spraying parameter data are separately performed by normalization.
  • the maximum and minimum normalization method is used for normalization, and the current overheating parameters are respectively adjusted according to the maximum and minimum normalization method.
  • the data and the current spray parameter data are normalized, and finally the normalized current superheat parameter data and the normalized current spray parameter data are obtained.
  • S13 Use the normalized current overheating parameter data and the normalized current water spray parameter data to construct a data feature matrix, where the data feature matrix is a matrix of M rows and N columns, and one row consists of the normalized current overheating It is composed of parameter data, and the other row is composed of the current water spray parameter data after normalization;
  • using the normalized current superheat parameter data and the normalized current spray parameter data to construct a data characteristic matrix includes: using the normalized current superheat parameter data and normalized The current water spray parameter data after the transformation constructs a data characteristic matrix according to the preset data sorting position.
  • the number of rows of the constructed data feature matrix is the sum of the number of superheaters and water spray desuperheaters, generally according to the superheater and the water spray desuperheater.
  • the order in which the water spray desuperheater is set in the thermal power boiler determines the location of the collected data, that is, which row; the collected data is arranged in a preset order to form a matrix of columns; because each superheater and water spray
  • the types of data that may be collected by the desuperheater are inconsistent.
  • a certain superheater or water spray desuperheater does not collect the corresponding type of data, its position is represented by 0; thus, a data characteristic matrix is constructed.
  • the training step of training the convergent neural network model includes: obtaining historical superheat parameter data of the superheater in the thermal power boiler and the historical water spray of the water spray desuperheater in the thermal power boiler Parameter data; construct a training data feature matrix for training according to the historical overheating parameter data and the historical water spray parameter data; set the initial loss function in the neural network model to be trained, and compare the initial loss function based on the regularization term Update; input the training data feature matrix into the neural network model to be trained after the updated loss function, and perform feature learning training until convergence, so as to obtain a neural network model for training convergence.
  • the neural network model to be trained is any one of a deep neural network model, an incomplete neural network model, or a convolutional neural network model.
  • the updating the initial loss function based on the regularization term includes:
  • the compressed neural network model Perform compression processing on the output nodes of each layer of the neural network model to be trained according to a preset compression ratio, the preset compression ratio being one-half; after the compression processing, the compressed neural network model Perform regularization processing on all layer parameter vectors in to obtain regularization terms of all layer parameter vectors; use the regularization terms of all layer parameter vectors to update the initial loss function; the initial loss function is cross-entropy loss Function or negative log likelihood loss function.
  • the training data feature matrix is input into the neural network model to be trained after the loss function is updated to perform feature learning training until convergence; the training data feature matrix and the test sample are selected according to the proportions in the training data feature matrix Data feature matrix; input the training sample data feature matrix into the neural network model to be trained after the updated loss function for feature learning and training to obtain the trained neural network model; input the test sample data feature matrix into the trained neural network Network model, input test correction prediction results; judge whether the probability of the input test correction result matching the actual correction result is greater than the preset threshold, if it is greater, the neural network model after training is determined to be converged, and the training ends; if not, use
  • the back propagation algorithm resets the coefficients of all layer parameter vectors of the trained neural network model, and uses the training sample data feature matrix for retraining.
  • the neural network model is selected as the neural network model to be trained, which can be in the deep neural network model, the incomplete neural network model or the convolutional neural network model Choose one of the arbitrarily; after determining the neural network model to be trained, the internal initial loss function needs to be updated accordingly.
  • the initial loss function generally selects either the cross-entropy loss function or the negative log-likelihood loss function; specifically;
  • the update method of using the regularization item in the neural network model to be trained is updated; after the update is completed, the training data feature matrix is input into the neural network model to be trained after the updated loss function for feature learning training until convergence, and the training is obtained Convergent neural network model.
  • the output nodes of each layer of the neural network model to be trained need to be compressed according to a preset compression ratio.
  • the preset compression ratio is two points. One, after compression, it is more conducive to the convergence of model training and reduces the phenomenon of overfitting that may occur during model training.
  • all layer parameter vectors in the neural network model after the compression processing are subjected to corresponding regularization processing, so as to obtain the regularization terms of all layer parameter vectors; use the regularization terms of all layer parameter vectors to reduce the initial loss
  • the function is updated accordingly; the specific update is to form a new loss function from the initial loss function plus the regularization terms of all layer parameter vectors.
  • the training sample data feature matrix and the test sample data feature matrix in the training data feature matrix are selected according to the ratio; in general, the ratio of training and testing data is 9:1;
  • the application also chooses a 9:1 ratio to set the training sample data feature matrix and the test sample data feature matrix; input the training sample data feature matrix into the neural network model to be trained after the updated loss function for feature learning and training to obtain the trained nerve Network model; input the test sample data feature matrix into the trained neural network model, and input the test correction prediction result; then judge whether the probability of the input test correction result matching the actual correction result is greater than the preset threshold, if it is greater, the trained
  • the neural network model converges and the training ends; if not, the backpropagation algorithm is used to reset the coefficients of all layer parameter vectors of the trained neural network model, and the training sample data feature matrix is used for retraining until convergence or Until the threshold of training times is reached.
  • the corrected prediction result can be obtained by inputting the data feature matrix into the neural network model that has been trained and converged to correct the water spray volume of the thermal power boiler spray desuperheater.
  • the thermal power boiler water spray desuperheater after obtaining the modified prediction result, performs water spray adjustment according to the modified prediction result.
  • real-time correction and prediction of the water injection volume of the water injection desuperheater in the thermal power boiler can be realized, so that the steam heat and pressure output by the superheater are within a preset range, and the operating efficiency of the thermal power boiler is guaranteed And operational safety.
  • FIG. 2 is a schematic diagram of the structural composition of a water spray adjusting device of a water spray desuperheater based on a neural network model in an embodiment of the present invention.
  • a water spray adjusting device of a water spray desuperheater based on a neural network model the device includes:
  • Data obtaining module 21 used to obtain the current superheat parameter data of the superheater in the thermal power boiler and the current water spray parameter data of the water spray desuperheater in the thermal power boiler;
  • the current superheat parameter data includes steam volume data, steam outlet density data, steam inlet mass flow data, steam outlet mass flow data, pipe working fluid inlet specific enthalpy data, pipe working fluid outlet ratio Enthalpy data, heat transfer data of the working fluid in the tube and the metal tube wall, the inlet density data of the working fluid in the tube, the inlet pressure data of the working fluid in the tube, the outlet pressure data of the working fluid in the tube, the flue gas inlet temperature data outside the tube, the flue gas outlet outside the tube Temperature, pipe wall temperature data, heat transfer data of flue gas to metal, heat exchange area data and pipe wall quality data;
  • the current water spray parameter data includes: inlet steam mass flow data, inlet desuperheating water mass flow data, Outlet steam mass flow data, inlet desuperheating water specific enthalpy data, inlet steam specific enthalpy data, outlet steam specific enthalpy data, inlet desuperheating water temperature data and inlet steam temperature data.
  • the superheater and the water spray desuperheater in the thermal power boiler are equipped with data acquisition sensors for collecting relevant parameter data; in general, the thermal power boiler includes multiple superheaters and multiple water spray desuperheaters. Thermostat, install the corresponding data acquisition sensor at the corresponding position in each superheater and water spray desuperheater to collect the corresponding data information.
  • the current superheat parameter data includes at least steam volume data, steam outlet density data, Steam inlet mass flow data, steam outlet mass flow data, pipe working fluid inlet specific enthalpy data, pipe working fluid outlet specific enthalpy data, pipe working fluid and metal pipe wall heat exchange data, pipe working fluid inlet density data, pipe work Mass inlet pressure data, working fluid outlet pressure data in the tube, flue gas inlet temperature data outside the tube, flue gas outlet temperature outside the tube, tube wall temperature data, heat transfer data from flue gas to metal, heat transfer area data, and tube wall temperature data Quality data; current water spray parameter data includes at least: inlet steam mass flow data, inlet desuperheating water mass flow data, outlet steam mass flow data, inlet desuperheating water specific enthalpy data, inlet steam specific enthalpy data, outlet steam specific enthalpy data, inlet Desuperheating water temperature data and inlet steam temperature data.
  • the collected noise is firstly reduced in the data acquisition sensor.
  • signal filtering is used to reduce noise.
  • the collected data information is converted into data information.
  • Normalization module 22 used to perform normalization processing on the current superheat parameter data and the current water spray parameter data, respectively, to obtain normalized current superheat parameter data and normalized current water spray parameter data data;
  • the current superheat parameter data and the current spray parameter data are respectively normalized to obtain the normalized current superheat parameter data and the normalized current spray parameter data.
  • the water parameter data includes: normalizing the current overheating parameter data based on the maximum and minimum value normalization method to obtain the normalized current overheating parameter data;
  • the water spray parameter data is normalized to obtain the current water spray parameter data after the normalization.
  • the current overheating parameter data and the current spraying parameter data are separately performed by normalization.
  • the maximum and minimum normalization method is used for normalization, and the current overheating parameters are respectively adjusted according to the maximum and minimum normalization method.
  • the data and the current spray parameter data are normalized, and finally the normalized current superheat parameter data and the normalized current spray parameter data are obtained.
  • Matrix construction module 23 used to construct a data characteristic matrix using normalized current superheat parameter data and normalized current water spray parameter data, wherein the data characteristic matrix is a matrix of M rows and N columns, and one row is normalized It is composed of the current superheat parameter data after normalization, and the other row is composed of the current water spray parameter data after normalization;
  • using the normalized current superheat parameter data and the normalized current spray parameter data to construct a data characteristic matrix includes: using the normalized current superheat parameter data and normalized The current water spray parameter data after the transformation constructs a data characteristic matrix according to the preset data sorting position.
  • the number of rows of the constructed data feature matrix is the sum of the number of superheaters and water spray desuperheaters, generally according to the superheater and the water spray desuperheater.
  • the order in which the water spray desuperheater is set in the thermal power boiler determines the location of the collected data, that is, which row; the collected data is arranged in a preset order to form a matrix of columns; because each superheater and water spray
  • the type of data that may be collected by the desuperheater is inconsistent.
  • a certain superheater or water spray desuperheater does not collect the corresponding type of data, its position is represented by 0; thus, a data characteristic matrix is constructed.
  • Correction prediction module 24 used to input the data feature matrix into the neural network model for training convergence to correct and predict the water injection volume of the water spray desuperheater of the thermal power boiler, and obtain a correction prediction result;
  • the training step of training the convergent neural network model includes: obtaining historical superheat parameter data of the superheater in the thermal power boiler and the historical water spray of the water spray desuperheater in the thermal power boiler Parameter data; construct a training data feature matrix for training according to the historical overheating parameter data and the historical water spray parameter data; set the initial loss function in the neural network model to be trained, and compare the initial loss function based on the regularization term Update; input the training data feature matrix into the neural network model to be trained after the updated loss function, and perform feature learning training until convergence, so as to obtain a neural network model for training convergence.
  • the neural network model to be trained is any one of a deep neural network model, an incomplete neural network model, or a convolutional neural network model.
  • the updating the initial loss function based on the regularization term includes:
  • the compressed neural network model Perform compression processing on the output nodes of each layer of the neural network model to be trained according to a preset compression ratio, the preset compression ratio being one-half; after the compression processing, the compressed neural network model Perform regularization processing on all layer parameter vectors in to obtain regularization terms of all layer parameter vectors; use the regularization terms of all layer parameter vectors to update the initial loss function; the initial loss function is cross-entropy loss Function or negative log likelihood loss function.
  • the training data feature matrix is input into the neural network model to be trained after the loss function is updated to perform feature learning training until convergence; the training data feature matrix and the test sample are selected according to the proportions in the training data feature matrix Data feature matrix; input the training sample data feature matrix into the neural network model to be trained after the updated loss function for feature learning and training to obtain the trained neural network model; input the test sample data feature matrix into the trained neural network Network model, input test correction prediction results; judge whether the probability of the input test correction result matching the actual correction result is greater than the preset threshold, if it is greater, the neural network model after training is determined to be converged, and the training ends; if not, use
  • the back propagation algorithm resets the coefficients of all layer parameter vectors of the trained neural network model, and uses the training sample data feature matrix for retraining.
  • the neural network model is selected as the neural network model to be trained, which can be in the deep neural network model, the incomplete neural network model or the convolutional neural network model Choose one of the arbitrarily; after determining the neural network model to be trained, the internal initial loss function needs to be updated accordingly.
  • the initial loss function generally selects either the cross-entropy loss function or the negative log-likelihood loss function; specifically;
  • the update method of using the regularization item in the neural network model to be trained is updated; after the update is completed, the training data feature matrix is input into the neural network model to be trained after the updated loss function for feature learning training until convergence, and the training is obtained Convergent neural network model.
  • the output nodes of each layer of the neural network model to be trained need to be compressed according to a preset compression ratio.
  • the preset compression ratio is two points. One, after compression, it is more conducive to the convergence of model training and reduces the phenomenon of overfitting that may occur during model training.
  • all layer parameter vectors in the neural network model after the compression processing are subjected to corresponding regularization processing, so as to obtain the regularization terms of all layer parameter vectors; use the regularization terms of all layer parameter vectors to reduce the initial loss
  • the function is updated accordingly; the specific update is to form a new loss function from the initial loss function plus the regularization terms of all layer parameter vectors.
  • the training sample data feature matrix and the test sample data feature matrix in the training data feature matrix are selected according to the ratio; in general, the ratio of training and testing data is 9:1;
  • the application also chooses a 9:1 ratio to set the training sample data feature matrix and the test sample data feature matrix; input the training sample data feature matrix into the neural network model to be trained after the updated loss function for feature learning and training to obtain the trained nerve Network model; input the test sample data feature matrix into the trained neural network model, and input the test correction prediction result; then judge whether the probability of the input test correction result matching the actual correction result is greater than the preset threshold, if it is greater, the trained
  • the neural network model converges and the training ends; if not, the backpropagation algorithm is used to reset the coefficients of all layer parameter vectors of the trained neural network model, and the training sample data feature matrix is used for retraining until convergence or Until the threshold of training times is reached.
  • the corrected prediction result can be obtained by inputting the data feature matrix into the neural network model that has been trained and converged to correct the water spray volume of the thermal power boiler spray desuperheater.
  • the adjustment module 25 is used for the thermal power boiler water spray desuperheater to adjust the water spray according to the corrected prediction result.
  • the thermal power boiler water spray desuperheater after obtaining the modified prediction result, performs water spray adjustment according to the modified prediction result.
  • real-time correction and prediction of the water injection volume of the water injection desuperheater in the thermal power boiler can be realized, so that the steam heat and pressure output by the superheater are within a preset range, and the operating efficiency of the thermal power boiler is guaranteed And operational safety.
  • the program can be stored in a computer-readable storage medium, and the storage medium can include: Read only memory (ROM, Read Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk, etc.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Mechanical Engineering (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Feedback Control In General (AREA)

Abstract

本发明公开了一种基于神经网络模型的喷水减温器喷水调整方法及装置，其中，所述方法包括：获得火电锅炉中过热器的当前过热参数数据以及获得火电锅炉中喷水减温器的当前喷水参数数据；对当前过热参数数据和当前喷水参数数据分别进行归一化处理；利用归一化后的当前过热参数数据和归一化后的当前喷水参数数据构建数据特征矩阵；将数据特征矩阵输入训练收敛的神经网络模型中对火电锅炉喷水减温器的喷水量进行修正预测，获得修正预测结果；火电锅炉喷水减温器根据修正预测结果进行喷水调整。在本发明实施例中，可以实现对火电锅炉内的喷水减温器的喷水量的实时修正预测，使得过热器输出的蒸汽热量和压力在预设的范围内，保障火电锅炉的运行效率。

Description

一种基于神经网络模型的喷水减温器喷水调整方法及装置 技术领域
本发明涉及火电发电的喷水减温器喷水调整技术领域，尤其涉及一种基于神经网络模型的喷水减温器喷水调整方法及装置。
背景技术
锅炉喷水减温系统在大型火电机组中不可缺少，其作用是将流经锅炉的超温大流量蒸汽进行调节、降温或降压；是锅炉和汽轮机安全运行重要保障，使进入汽轮机的蒸汽温度在机组正常运行要求范围内；在机组运行工况突变时，尤其是当汽轮机机组快速甩负荷(FCB)后，锅炉所产生蒸汽的温度控制与汽轮机的协调运行有着重要的联系；现有的锅炉喷水减温系统中对于喷水减温器喷水调整的方式还是停留在出现问题之后，由人工控制的方式进行调整，比较难做到实时的调整，可能带来交大的安全隐患。
发明内容
本发明的目的在于克服现有技术的不足，本发明提供了一种基于神经网络模型的喷水减温器喷水调整方法及装置，可以实现对火电锅炉内的喷水减温器的喷水量的实时修正预测，使得过热器输出的蒸汽热量和压力在预设的范围内，保障火电锅炉的运行效率和运行安全。
为了解决上述技术问题，本发明实施例还提供了一种基于神经网络模型的炉喷水减温器喷水调整方法，所述方法包括：
获得火电锅炉中过热器的当前过热参数数据以及获得火电锅炉中喷水减温器的当前喷水参数数据；
对所述当前过热参数数据和所述当前喷水参数数据分别进行归一化处理，获得归一化后的当前过热参数数据和归一化后的当前喷水参数数据；
利用归一化后的当前过热参数数据和归一化后的当前喷水参数数据构建数据特征矩阵，其中所述数据特征矩阵为M行N列矩阵，一行由归一化后的当前过热参数数据组成，另一行由归一化后的当前喷水参数数据组成；
将所述数据特征矩阵输入训练收敛的神经网络模型中对所述火电锅炉喷水减温器的喷水量进行修正预测，获得修正预测结果；
所述火电锅炉喷水减温器根据所述修正预测结果进行喷水调整。
可选的，所述当前过热参数数据包括蒸汽的容积数据、蒸汽出口密度数据、蒸汽进口质量流量数据、蒸汽出口质量流量数据、管内工质进口比焓数据、管内工质出口比焓数据、管内工质与金属管壁的换热量数据、管内工质进口密度数据、管内工质进口压力数据、管内工质出口压力数据、管外烟气进口温度数据、管外烟气出口温度、管壁温度数据、烟气对金属的换热量数据、换热面积数据和管壁的质量数据；
所述当前喷水参数数据包括：入口蒸汽质量流量数据、入口减温水质量流量数据、出口蒸汽质量流量数据、入口减温水比焓数据、入口蒸汽比焓数据、出口蒸汽比焓数据、入口减温水温度数据和入口蒸汽温度数据。
可选的，所述对所述当前过热参数数据和所述当前喷水参数数据分别进行归一化处理，获得归一化后的当前过热参数数据和归一化后的当前喷水参数数据，包括：
基于最大最小值归一化法对所述当前过热参数数据进行归一化处理，获得归一化后的当前过热参数数据；
基于最大最小值归一化法对所述当前喷水参数数据进行归一化处理，获得一化后的当前喷水参数数据。
可选的，所述训练收敛的神经网络模型的训练步骤，包括：
获取所述火电锅炉中过热器的历史过热参数数据和所述火电锅炉中喷水减温器的历史喷水参数数据；
根据所述历史过热参数数据和所述历史喷水参数数据构建用于训练的训练数据特征矩阵；
设置待训练神经网络模型中的初始损失函数，基于正则化项对所述初始损失函数进行更新；
将所述训练数据特征矩阵输入更新损失函数后的待训练神经网络模型中进行特征学习训练直至收敛，获得训练收敛的神经网络模型。
可选的，所述待训练神经网络模型为深度神经网络模型、残缺神经网 络模型或卷积神经网络模型中的任意一种。
可选的，所述基于正则化项对所述初始损失函数进行更新，包括：
对所述待训练神经网络模型中的每一层网络的输出节点按照预设压缩比例进行压缩处理，所述预设压缩比例为二分之一；
在压缩处理之后，压缩处理后的神经网络模型中的所有层参数向量进行正则化处理，获得所有层参数向量的正则化项；
利用所述所有层参数向量的正则化项对所述初始损失函数进行更新；
所述初始损失函数为交叉熵损失函数或负对数似然损失函数。
可选的，所述将所述训练数据特征矩阵输入更新损失函数后的待训练神经网络模型中进行特征学习训练直至收敛；
在所述训练数据特征矩阵中按照比例选取训练样本数据特征矩阵和测试样本数据特征矩阵；
将所述训练样本数据特征矩阵输入更新损失函数后的待训练神经网络模型中进行特征学习训练，获得训练后的神经网络模型；
将所述测试样本数据特征矩阵输入训练后的神经网络模型，输入测试修正预测结果；
判断所述输入测试修正结果与实际修正结果相匹配的概率是否大于预设阈值，若大于，认定训练后的神经网络模型收敛，训练结束；
若否，则采用反向传播算法对所述训练后的神经网络模型所有层参数向量的系数进行重置，并采用所述训练样本数据特征矩阵进行重新训练。
可选的，所述利用归一化后的当前过热参数数据和归一化后的当前喷水参数数据构建数据特征矩阵，包括：
利用归一化后的当前过热参数数据和归一化后的当前喷水参数数据按照预设数据排序位置构建数据特征矩阵。
另外，本发明实施例还提供了一种基于神经网络模型的喷水减温器喷水调整装置，所述装置包括：
数据获得模块：用于获得火电锅炉中过热器的当前过热参数数据以及获得火电锅炉中喷水减温器的当前喷水参数数据；
归一化模块：用于对所述当前过热参数数据和所述当前喷水参数数据 分别进行归一化处理，获得归一化后的当前过热参数数据和归一化后的当前喷水参数数据；
矩阵构建模块：用于利用归一化后的当前过热参数数据和归一化后的当前喷水参数数据构建数据特征矩阵，其中所述数据特征矩阵为M行N列矩阵，一行由归一化后的当前过热参数数据组成，另一行由归一化后的当前喷水参数数据组成；
修正预测模块：用于将所述数据特征矩阵输入训练收敛的神经网络模型中对所述火电锅炉喷水减温器的喷水量进行修正预测，获得修正预测结果；
调整模块：用于所述火电锅炉喷水减温器根据所述修正预测结果进行喷水调整。
在本发明实施例中，可以实现对火电锅炉内的喷水减温器的喷水量的实时修正预测，使得过热器输出的蒸汽热量和压力在预设的范围内，保障火电锅炉的运行效率和运行安全。
附图说明
为了更清楚地说明本发明实施例或现有技术中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍，显而易见的，下面描述中的附图仅仅是本发明的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其它的附图。
图1是本发明实施例中的基于神经网络模型的喷水减温器喷水调整方法的流程示意图；
图2是本发明实施例中的基于神经网络模型的喷水减温器喷水调整装置的结构组成示意图。
具体实施方式
下面将结合本发明实施例中的附图，对本发明实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例仅仅是本发明一部分实施例， 而不是全部的实施例。基于本发明中的实施例，本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其它实施例，都属于本发明保护的范围。
实施例
请参阅图1，图1是本发明实施例中的基于神经网络模型的喷水减温器喷水调整方法的流程示意图。
如图1所示，一种基于神经网络模型的炉喷水减温器喷水调整方法，所述方法包括：
S11：获得火电锅炉中过热器的当前过热参数数据以及获得火电锅炉中喷水减温器的当前喷水参数数据；
在本发明具体实施过程中，所述当前过热参数数据包括蒸汽的容积数据、蒸汽出口密度数据、蒸汽进口质量流量数据、蒸汽出口质量流量数据、管内工质进口比焓数据、管内工质出口比焓数据、管内工质与金属管壁的换热量数据、管内工质进口密度数据、管内工质进口压力数据、管内工质出口压力数据、管外烟气进口温度数据、管外烟气出口温度、管壁温度数据、烟气对金属的换热量数据、换热面积数据和管壁的质量数据；所述当前喷水参数数据包括：入口蒸汽质量流量数据、入口减温水质量流量数据、出口蒸汽质量流量数据、入口减温水比焓数据、入口蒸汽比焓数据、出口蒸汽比焓数据、入口减温水温度数据和入口蒸汽温度数据。
具体的，在火电锅炉中的过热器上和喷水减温器上均安装有用于采集相关参数数据的数据采集传感器；一般情况下，火电锅炉中包括有多个过热器和多个喷水减温器，在每个过热器和喷水减温器中的对应位置上安装对应的数据采集传感器，采集对应的数据信息，其中，当前过热参数数据至少包括蒸汽的容积数据、蒸汽出口密度数据、蒸汽进口质量流量数据、蒸汽出口质量流量数据、管内工质进口比焓数据、管内工质出口比焓数据、管内工质与金属管壁的换热量数据、管内工质进口密度数据、管内工质进口压力数据、管内工质出口压力数据、管外烟气进口温度数据、管外烟气出口温度、管壁温度数据、烟气对金属的换热量数据、换热面积数据和管壁的质量数据；当前喷水参数数据至少包括：入口蒸汽质量流量数据、入 口减温水质量流量数据、出口蒸汽质量流量数据、入口减温水比焓数据、入口蒸汽比焓数据、出口蒸汽比焓数据、入口减温水温度数据和入口蒸汽温度数据。
并且，对采集的数据在数据采集传感器内部首先对采集到的进行相应的降噪，一般利用信号滤波的方式进行降噪，在降噪之后，把采集的到数据信息转换为数据信息。
S12：对所述当前过热参数数据和所述当前喷水参数数据分别进行归一化处理，获得归一化后的当前过热参数数据和归一化后的当前喷水参数数据；
在本发明具体实施过程中，所述对所述当前过热参数数据和所述当前喷水参数数据分别进行归一化处理，获得归一化后的当前过热参数数据和归一化后的当前喷水参数数据，包括：基于最大最小值归一化法对所述当前过热参数数据进行归一化处理，获得归一化后的当前过热参数数据；基于最大最小值归一化法对所述当前喷水参数数据进行归一化处理，获得一化后的当前喷水参数数据。
具体的，利用归一化对当前过热参数数据和当前喷水参数数据分别进行，一般情况下采用最大最小值归一化法进行归一化，按照最大最小值归一化法分别对当前过热参数数据和当前喷水参数数据进行归一化处理，最终得到归一化后的当前过热参数数据和归一化后的当前喷水参数数据。
S13：利用归一化后的当前过热参数数据和归一化后的当前喷水参数数据构建数据特征矩阵，其中所述数据特征矩阵为M行N列矩阵，一行由归一化后的当前过热参数数据组成，另一行由归一化后的当前喷水参数数据组成；
在本发明具体实施过程中，所述利用归一化后的当前过热参数数据和归一化后的当前喷水参数数据构建数据特征矩阵，包括：利用归一化后的当前过热参数数据和归一化后的当前喷水参数数据按照预设数据排序位置构建数据特征矩阵。
具体的，一般情况下，火电锅炉中拥有几个过热器和几个喷水减温器，则构建的数据特征矩阵的行数是过热器和喷水减温器数量总和，一般按照 过热器和喷水减温器在火电锅炉中设置的顺序来确定其所采集的数据所在的位置，即第几行；采集的数据按照预设的顺序排列形成矩阵的列；因为每个过热器和喷水减温器可能采集的数据种类不一致，所有在某一个过热器或者喷水减温器没有采集对应种类数据时，在其位置上用0表示；由此构建成数据特征矩阵。
S14：将所述数据特征矩阵输入训练收敛的神经网络模型中对所述火电锅炉喷水减温器的喷水量进行修正预测，获得修正预测结果；
在本发明具体实施过程中，所述训练收敛的神经网络模型的训练步骤，包括：获取所述火电锅炉中过热器的历史过热参数数据和所述火电锅炉中喷水减温器的历史喷水参数数据；根据所述历史过热参数数据和所述历史喷水参数数据构建用于训练的训练数据特征矩阵；设置待训练神经网络模型中的初始损失函数，基于正则化项对所述初始损失函数进行更新；将所述训练数据特征矩阵输入更新损失函数后的待训练神经网络模型中进行特征学习训练直至收敛，获得训练收敛的神经网络模型。
进一步的，所述待训练神经网络模型为深度神经网络模型、残缺神经网络模型或卷积神经网络模型中的任意一种。
进一步的，所述基于正则化项对所述初始损失函数进行更新，包括：
对所述待训练神经网络模型中的每一层网络的输出节点按照预设压缩比例进行压缩处理，所述预设压缩比例为二分之一；在压缩处理之后，压缩处理后的神经网络模型中的所有层参数向量进行正则化处理，获得所有层参数向量的正则化项；利用所述所有层参数向量的正则化项对所述初始损失函数进行更新；所述初始损失函数为交叉熵损失函数或负对数似然损失函数。
进一步的，所述将所述训练数据特征矩阵输入更新损失函数后的待训练神经网络模型中进行特征学习训练直至收敛；在所述训练数据特征矩阵中按照比例选取训练样本数据特征矩阵和测试样本数据特征矩阵；将所述训练样本数据特征矩阵输入更新损失函数后的待训练神经网络模型中进行特征学习训练，获得训练后的神经网络模型；将所述测试样本数据特征矩阵输入训练后的神经网络模型，输入测试修正预测结果；判断所述输入测 试修正结果与实际修正结果相匹配的概率是否大于预设阈值，若大于，认定训练后的神经网络模型收敛，训练结束；若否，则采用反向传播算法对所述训练后的神经网络模型所有层参数向量的系数进行重置，并采用所述训练样本数据特征矩阵进行重新训练。
具体的，首先提取将火电锅炉中过热器的历史过热参数数据和电锅炉中喷水减温器的历史喷水参数数据；并进去归一化，在归一化之后，构建用于训练的训练数据特征矩阵，具体的构建方式，与上述构建数据特征矩阵方式一致；然后，选取神经网络模型作为待训练神经网络模型，具体可以在深度神经网络模型、残缺神经网络模型或卷积神经网络模型中的任意选取一种；在确定待训练神经网络模型之后，需要对其内部的初始损失函数进行相应的更新，初始损失函数一般选取交叉熵损失函数或负对数似然损失函数任意一种；具体的更新方式，利用待训练神经网络模型中的正则化项来进行更新；在完成更新之后，将训练数据特征矩阵输入更新损失函数后的待训练神经网络模型中进行特征学习训练直至收敛，获得训练收敛的神经网络模型。
在对待训练神经网络模型中的初始损失函数进行更新之前，需要对待训练神经网络模型中的每一层网络的输出节点按照预设压缩比例进行压缩处理，在本发明中预设压缩比例为二分之一，压缩之后，更有利于模型训练的收敛，减少模型训练中可能发生过度拟合的现象。
在完成压缩之后，在压缩处理后的神经网络模型中的所有层参数向量都进行相应的正则化处理，从而获得所有层参数向量的正则化项；利用所有层参数向量的正则化项对初始损失函数进行相应的更新；具体的更新是由初始损失函数加上所有层参数向量的正则化项形成新的损失函数。
在完成待训练神经网络模型的损失函数更新之后，在训练数据特征矩阵中按照比例选取训练样本数据特征矩阵和测试样本数据特征矩阵；一般情况下训练和测试数据的比例为9:1；在本申请中也选择9:1的比例设置训练样本数据特征矩阵和测试样本数据特征矩阵；将训练样本数据特征矩阵输入更新损失函数后的待训练神经网络模型中进行特征学习训练，获得训练后的神经网络模型；将测试样本数据特征矩阵输入训练后的神经网络模 型，输入测试修正预测结果；然后判断输入测试修正结果与实际修正结果相匹配的概率是否大于预设阈值，若大于，认定训练后的神经网络模型收敛，训练结束；若否，则采用反向传播算法对训练后的神经网络模型所有层参数向量的系数进行重置，并采用所述训练样本数据特征矩阵进行重新训练，直至收敛或者达到训练次数的阈值为止。
通过将数据特征矩阵输入训练收敛的神经网络模型中对所述火电锅炉喷水减温器的喷水量进行修正预测，即可获得修正预测结果。
S15：所述火电锅炉喷水减温器根据所述修正预测结果进行喷水调整。
在本发明具体实施过程中，在获得修正预测结果之后，火电锅炉喷水减温器根据该修正预测结果进行喷水调整。
在本发明实施例中，可以实现对火电锅炉内的喷水减温器的喷水量的实时修正预测，使得过热器输出的蒸汽热量和压力在预设的范围内，保障火电锅炉的运行效率和运行安全。
实施例
请参阅图2，图2是本发明实施例中的基于神经网络模型的喷水减温器喷水调整装置的结构组成示意图。
如图2所示，一种基于神经网络模型的喷水减温器喷水调整装置，所述装置包括：
数据获得模块21：用于获得火电锅炉中过热器的当前过热参数数据以及获得火电锅炉中喷水减温器的当前喷水参数数据；
在本发明具体实施过程中，所述当前过热参数数据包括蒸汽的容积数据、蒸汽出口密度数据、蒸汽进口质量流量数据、蒸汽出口质量流量数据、管内工质进口比焓数据、管内工质出口比焓数据、管内工质与金属管壁的换热量数据、管内工质进口密度数据、管内工质进口压力数据、管内工质出口压力数据、管外烟气进口温度数据、管外烟气出口温度、管壁温度数据、烟气对金属的换热量数据、换热面积数据和管壁的质量数据；所述当前喷水参数数据包括：入口蒸汽质量流量数据、入口减温水质量流量数据、出口蒸汽质量流量数据、入口减温水比焓数据、入口蒸汽比焓数据、出口蒸汽比焓数据、入口减温水温度数据和入口蒸汽温度数据。
具体的，在火电锅炉中的过热器上和喷水减温器上均安装有用于采集相关参数数据的数据采集传感器；一般情况下，火电锅炉中包括有多个过热器和多个喷水减温器，在每个过热器和喷水减温器中的对应位置上安装对应的数据采集传感器，采集对应的数据信息，其中，当前过热参数数据至少包括蒸汽的容积数据、蒸汽出口密度数据、蒸汽进口质量流量数据、蒸汽出口质量流量数据、管内工质进口比焓数据、管内工质出口比焓数据、管内工质与金属管壁的换热量数据、管内工质进口密度数据、管内工质进口压力数据、管内工质出口压力数据、管外烟气进口温度数据、管外烟气出口温度、管壁温度数据、烟气对金属的换热量数据、换热面积数据和管壁的质量数据；当前喷水参数数据至少包括：入口蒸汽质量流量数据、入口减温水质量流量数据、出口蒸汽质量流量数据、入口减温水比焓数据、入口蒸汽比焓数据、出口蒸汽比焓数据、入口减温水温度数据和入口蒸汽温度数据。
并且，对采集的数据在数据采集传感器内部首先对采集到的进行相应的降噪，一般利用信号滤波的方式进行降噪，在降噪之后，把采集的到数据信息转换为数据信息。
归一化模块22：用于对所述当前过热参数数据和所述当前喷水参数数据分别进行归一化处理，获得归一化后的当前过热参数数据和归一化后的当前喷水参数数据；
在本发明具体实施过程中，所述对所述当前过热参数数据和所述当前喷水参数数据分别进行归一化处理，获得归一化后的当前过热参数数据和归一化后的当前喷水参数数据，包括：基于最大最小值归一化法对所述当前过热参数数据进行归一化处理，获得归一化后的当前过热参数数据；基于最大最小值归一化法对所述当前喷水参数数据进行归一化处理，获得一化后的当前喷水参数数据。
具体的，利用归一化对当前过热参数数据和当前喷水参数数据分别进行，一般情况下采用最大最小值归一化法进行归一化，按照最大最小值归一化法分别对当前过热参数数据和当前喷水参数数据进行归一化处理，最终得到归一化后的当前过热参数数据和归一化后的当前喷水参数数据。
矩阵构建模块23：用于利用归一化后的当前过热参数数据和归一化后的当前喷水参数数据构建数据特征矩阵，其中所述数据特征矩阵为M行N列矩阵，一行由归一化后的当前过热参数数据组成，另一行由归一化后的当前喷水参数数据组成；
在本发明具体实施过程中，所述利用归一化后的当前过热参数数据和归一化后的当前喷水参数数据构建数据特征矩阵，包括：利用归一化后的当前过热参数数据和归一化后的当前喷水参数数据按照预设数据排序位置构建数据特征矩阵。
具体的，一般情况下，火电锅炉中拥有几个过热器和几个喷水减温器，则构建的数据特征矩阵的行数是过热器和喷水减温器数量总和，一般按照过热器和喷水减温器在火电锅炉中设置的顺序来确定其所采集的数据所在的位置，即第几行；采集的数据按照预设的顺序排列形成矩阵的列；因为每个过热器和喷水减温器可能采集的数据种类不一致，所有在某一个过热器或者喷水减温器没有采集对应种类数据时，在其位置上用0表示；由此构建成数据特征矩阵。
修正预测模块24：用于将所述数据特征矩阵输入训练收敛的神经网络模型中对所述火电锅炉喷水减温器的喷水量进行修正预测，获得修正预测结果；
在本发明具体实施过程中，所述训练收敛的神经网络模型的训练步骤，包括：获取所述火电锅炉中过热器的历史过热参数数据和所述火电锅炉中喷水减温器的历史喷水参数数据；根据所述历史过热参数数据和所述历史喷水参数数据构建用于训练的训练数据特征矩阵；设置待训练神经网络模型中的初始损失函数，基于正则化项对所述初始损失函数进行更新；将所述训练数据特征矩阵输入更新损失函数后的待训练神经网络模型中进行特征学习训练直至收敛，获得训练收敛的神经网络模型。
进一步的，所述待训练神经网络模型为深度神经网络模型、残缺神经网络模型或卷积神经网络模型中的任意一种。
进一步的，所述基于正则化项对所述初始损失函数进行更新，包括：
对所述待训练神经网络模型中的每一层网络的输出节点按照预设压缩 比例进行压缩处理，所述预设压缩比例为二分之一；在压缩处理之后，压缩处理后的神经网络模型中的所有层参数向量进行正则化处理，获得所有层参数向量的正则化项；利用所述所有层参数向量的正则化项对所述初始损失函数进行更新；所述初始损失函数为交叉熵损失函数或负对数似然损失函数。
进一步的，所述将所述训练数据特征矩阵输入更新损失函数后的待训练神经网络模型中进行特征学习训练直至收敛；在所述训练数据特征矩阵中按照比例选取训练样本数据特征矩阵和测试样本数据特征矩阵；将所述训练样本数据特征矩阵输入更新损失函数后的待训练神经网络模型中进行特征学习训练，获得训练后的神经网络模型；将所述测试样本数据特征矩阵输入训练后的神经网络模型，输入测试修正预测结果；判断所述输入测试修正结果与实际修正结果相匹配的概率是否大于预设阈值，若大于，认定训练后的神经网络模型收敛，训练结束；若否，则采用反向传播算法对所述训练后的神经网络模型所有层参数向量的系数进行重置，并采用所述训练样本数据特征矩阵进行重新训练。
具体的，首先提取将火电锅炉中过热器的历史过热参数数据和电锅炉中喷水减温器的历史喷水参数数据；并进去归一化，在归一化之后，构建用于训练的训练数据特征矩阵，具体的构建方式，与上述构建数据特征矩阵方式一致；然后，选取神经网络模型作为待训练神经网络模型，具体可以在深度神经网络模型、残缺神经网络模型或卷积神经网络模型中的任意选取一种；在确定待训练神经网络模型之后，需要对其内部的初始损失函数进行相应的更新，初始损失函数一般选取交叉熵损失函数或负对数似然损失函数任意一种；具体的更新方式，利用待训练神经网络模型中的正则化项来进行更新；在完成更新之后，将训练数据特征矩阵输入更新损失函数后的待训练神经网络模型中进行特征学习训练直至收敛，获得训练收敛的神经网络模型。
在对待训练神经网络模型中的初始损失函数进行更新之前，需要对待训练神经网络模型中的每一层网络的输出节点按照预设压缩比例进行压缩处理，在本发明中预设压缩比例为二分之一，压缩之后，更有利于模型训 练的收敛，减少模型训练中可能发生过度拟合的现象。
在完成压缩之后，在压缩处理后的神经网络模型中的所有层参数向量都进行相应的正则化处理，从而获得所有层参数向量的正则化项；利用所有层参数向量的正则化项对初始损失函数进行相应的更新；具体的更新是由初始损失函数加上所有层参数向量的正则化项形成新的损失函数。
在完成待训练神经网络模型的损失函数更新之后，在训练数据特征矩阵中按照比例选取训练样本数据特征矩阵和测试样本数据特征矩阵；一般情况下训练和测试数据的比例为9:1；在本申请中也选择9:1的比例设置训练样本数据特征矩阵和测试样本数据特征矩阵；将训练样本数据特征矩阵输入更新损失函数后的待训练神经网络模型中进行特征学习训练，获得训练后的神经网络模型；将测试样本数据特征矩阵输入训练后的神经网络模型，输入测试修正预测结果；然后判断输入测试修正结果与实际修正结果相匹配的概率是否大于预设阈值，若大于，认定训练后的神经网络模型收敛，训练结束；若否，则采用反向传播算法对训练后的神经网络模型所有层参数向量的系数进行重置，并采用所述训练样本数据特征矩阵进行重新训练，直至收敛或者达到训练次数的阈值为止。
通过将数据特征矩阵输入训练收敛的神经网络模型中对所述火电锅炉喷水减温器的喷水量进行修正预测，即可获得修正预测结果。
调整模块25：用于所述火电锅炉喷水减温器根据所述修正预测结果进行喷水调整。
在本发明具体实施过程中，在获得修正预测结果之后，火电锅炉喷水减温器根据该修正预测结果进行喷水调整。
在本发明实施例中，可以实现对火电锅炉内的喷水减温器的喷水量的实时修正预测，使得过热器输出的蒸汽热量和压力在预设的范围内，保障火电锅炉的运行效率和运行安全。
本领域普通技术人员可以理解上述实施例的各种方法中的全部或部分步骤是可以通过程序来指令相关的硬件来完成，该程序可以存储于一计算机可读存储介质中，存储介质可以包括：只读存储器(ROM，Read Only Memory)、随机存取存储器(RAM，Random Access Memory)、磁盘或光盘 等。
另外，以上对本发明实施例所提供的一种基于神经网络模型的喷水减温器喷水调整方法及装置进行了详细介绍，本文中应采用了具体个例对本发明的原理及实施方式进行了阐述，以上实施例的说明只是用于帮助理解本发明的方法及其核心思想；同时，对于本领域的一般技术人员，依据本发明的思想，在具体实施方式及应用范围上均会有改变之处，综上所述，本说明书内容不应理解为对本发明的限制。

Claims (9)

  1. 一种基于神经网络模型的喷水减温器喷水调整方法，其特征在于，所述方法包括：
    获得火电锅炉中过热器的当前过热参数数据以及获得火电锅炉中喷水减温器的当前喷水参数数据；
    对所述当前过热参数数据和所述当前喷水参数数据分别进行归一化处理，获得归一化后的当前过热参数数据和归一化后的当前喷水参数数据；
    利用归一化后的当前过热参数数据和归一化后的当前喷水参数数据构建数据特征矩阵，其中所述数据特征矩阵为M行N列矩阵，一行由归一化后的当前过热参数数据组成，另一行由归一化后的当前喷水参数数据组成；
    将所述数据特征矩阵输入训练收敛的神经网络模型中对所述火电锅炉喷水减温器的喷水量进行修正预测，获得修正预测结果；
    所述火电锅炉喷水减温器根据所述修正预测结果进行喷水调整。
  2. 根据权利要求1所述的喷水减温器喷水调整方法，其特征在于，所述当前过热参数数据包括蒸汽的容积数据、蒸汽出口密度数据、蒸汽进口质量流量数据、蒸汽出口质量流量数据、管内工质进口比焓数据、管内工质出口比焓数据、管内工质与金属管壁的换热量数据、管内工质进口密度数据、管内工质进口压力数据、管内工质出口压力数据、管外烟气进口温度数据、管外烟气出口温度、管壁温度数据、烟气对金属的换热量数据、换热面积数据和管壁的质量数据；
    所述当前喷水参数数据包括：入口蒸汽质量流量数据、入口减温水质量流量数据、出口蒸汽质量流量数据、入口减温水比焓数据、入口蒸汽比焓数据、出口蒸汽比焓数据、入口减温水温度数据和入口蒸汽温度数据。
  3. 根据权利要求1所述的喷水减温器喷水调整方法，其特征在于，所述对所述当前过热参数数据和所述当前喷水参数数据分别进行归一化处理， 获得归一化后的当前过热参数数据和归一化后的当前喷水参数数据，包括：
    基于最大最小值归一化法对所述当前过热参数数据进行归一化处理，获得归一化后的当前过热参数数据；
    基于最大最小值归一化法对所述当前喷水参数数据进行归一化处理，获得一化后的当前喷水参数数据。
  4. 根据权利要求1所述的喷水减温器喷水调整方法，其特征在于，所述训练收敛的神经网络模型的训练步骤，包括：
    获取所述火电锅炉中过热器的历史过热参数数据和所述火电锅炉中喷水减温器的历史喷水参数数据；
    根据所述历史过热参数数据和所述历史喷水参数数据构建用于训练的训练数据特征矩阵；
    设置待训练神经网络模型中的初始损失函数，基于正则化项对所述初始损失函数进行更新；
    将所述训练数据特征矩阵输入更新损失函数后的待训练神经网络模型中进行特征学习训练直至收敛，获得训练收敛的神经网络模型。
  5. 根据权利要求4所述的喷水减温器喷水调整方法，其特征在于，所述待训练神经网络模型为深度神经网络模型、残缺神经网络模型或卷积神经网络模型中的任意一种。
  6. 根据权利要求4所述的喷水减温器喷水调整方法，其特征在于，所述基于正则化项对所述初始损失函数进行更新，包括：
    对所述待训练神经网络模型中的每一层网络的输出节点按照预设压缩比例进行压缩处理，所述预设压缩比例为二分之一；
    在压缩处理之后，压缩处理后的神经网络模型中的所有层参数向量进行正则化处理，获得所有层参数向量的正则化项；
    利用所述所有层参数向量的正则化项对所述初始损失函数进行更新；
    所述初始损失函数为交叉熵损失函数或负对数似然损失函数。
  7. 根据权利要求4所述的喷水减温器喷水调整方法，其特征在于，所述将所述训练数据特征矩阵输入更新损失函数后的待训练神经网络模型中进行特征学习训练直至收敛；
    在所述训练数据特征矩阵中按照比例选取训练样本数据特征矩阵和测试样本数据特征矩阵；
    将所述训练样本数据特征矩阵输入更新损失函数后的待训练神经网络模型中进行特征学习训练，获得训练后的神经网络模型；
    将所述测试样本数据特征矩阵输入训练后的神经网络模型，输入测试修正预测结果；
    判断所述输入测试修正结果与实际修正结果相匹配的概率是否大于预设阈值，若大于，认定训练后的神经网络模型收敛，训练结束；
    若否，则采用反向传播算法对所述训练后的神经网络模型所有层参数向量的系数进行重置，并采用所述训练样本数据特征矩阵进行重新训练。
  8. 根据权利要求1所述的喷水减温器喷水调整方法，其特征在于，所述利用归一化后的当前过热参数数据和归一化后的当前喷水参数数据构建数据特征矩阵，包括：
    利用归一化后的当前过热参数数据和归一化后的当前喷水参数数据按照预设数据排序位置构建数据特征矩阵。
  9. 一种基于神经网络模型的喷水减温器喷水调整装置，其特征在于，所述装置包括：
    数据获得模块：用于获得火电锅炉中过热器的当前过热参数数据以及获得火电锅炉中喷水减温器的当前喷水参数数据；
    归一化模块：用于对所述当前过热参数数据和所述当前喷水参数数据分别进行归一化处理，获得归一化后的当前过热参数数据和归一化后的当 前喷水参数数据；
    矩阵构建模块：用于利用归一化后的当前过热参数数据和归一化后的当前喷水参数数据构建数据特征矩阵，其中所述数据特征矩阵为M行N列矩阵，一行由归一化后的当前过热参数数据组成，另一行由归一化后的当前喷水参数数据组成；
    修正预测模块：用于将所述数据特征矩阵输入训练收敛的神经网络模型中对所述火电锅炉喷水减温器的喷水量进行修正预测，获得修正预测结果；
    调整模块：用于所述火电锅炉喷水减温器根据所述修正预测结果进行喷水调整。
PCT/CN2020/114593 2020-04-17 2020-09-10 一种基于神经网络模型的喷水减温器喷水调整方法及装置 WO2021208343A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010307415.9 2020-04-17
CN202010307415.9A CN111582472A (zh) 2020-04-17 2020-04-17 一种基于神经网络模型的喷水减温器喷水调整方法及装置

Publications (1)

Publication Number Publication Date
WO2021208343A1 true WO2021208343A1 (zh) 2021-10-21

Family

ID=72120575

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/114593 WO2021208343A1 (zh) 2020-04-17 2020-09-10 一种基于神经网络模型的喷水减温器喷水调整方法及装置

Country Status (2)

Country Link
CN (1) CN111582472A (zh)
WO (1) WO2021208343A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114004159A (zh) * 2021-11-02 2022-02-01 西安热工研究院有限公司 一种余热锅炉过热器变工况效能计算方法
CN115607719A (zh) * 2022-12-16 2023-01-17 威海中远海运重工科技有限公司 一种压载水紫外线反应器的智能控制方法
CN117180952A (zh) * 2023-11-07 2023-12-08 湖南正明环保股份有限公司 多向气流料层循环半干法烟气脱硫系统及其方法

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111582472A (zh) * 2020-04-17 2020-08-25 广西电网有限责任公司电力科学研究院 一种基于神经网络模型的喷水减温器喷水调整方法及装置
CN112685964B (zh) * 2021-03-19 2021-06-08 浙江浙能技术研究院有限公司 一种基于卷积神经网络的锅炉受热面管壁泄漏预测方法
CN113339781B (zh) * 2021-05-31 2022-11-18 烟台龙源电力技术股份有限公司 屏式过热器外壁温度计算方法及装置、温度监测系统

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107831652A (zh) * 2017-10-13 2018-03-23 国网河北能源技术服务有限公司 一种基于冷端系统储能的机组负荷智能优化控制方法
CN108805259A (zh) * 2018-05-23 2018-11-13 北京达佳互联信息技术有限公司 神经网络模型训练方法、装置、存储介质及终端设备
US20190236446A1 (en) * 2018-01-29 2019-08-01 Lawrence Livermore National Security, Llc Optimization control technology for building energy conservation
CN110925741A (zh) * 2019-12-13 2020-03-27 吉林大学 一种基于减温水量的主蒸汽各过热器进出口温度获取方法
CN111582472A (zh) * 2020-04-17 2020-08-25 广西电网有限责任公司电力科学研究院 一种基于神经网络模型的喷水减温器喷水调整方法及装置

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106016229B (zh) * 2016-05-23 2018-08-07 神华集团有限责任公司 超临界循环流化床锅炉机组的主汽温控制方法和装置
CN207112778U (zh) * 2017-08-25 2018-03-16 河北大唐国际张家口热电有限责任公司 一种主蒸汽喷水减温控制系统
CN108954294B (zh) * 2018-05-22 2019-09-24 哈尔滨工业大学 火电机组过热器/再热器汽温的前馈控制方法
CN109063359A (zh) * 2018-08-16 2018-12-21 燕山大学 一种循环流化床锅炉燃烧过程的动态建模方法
CN110531797A (zh) * 2019-05-31 2019-12-03 华电国际电力股份有限公司技术服务分公司 基于神经网络的超超临界机组高温过热器壁温预测方法
CN110515304A (zh) * 2019-09-25 2019-11-29 南京信息工程大学 基于ARX-Laguerre函数模型的过热汽温PID预测控制方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107831652A (zh) * 2017-10-13 2018-03-23 国网河北能源技术服务有限公司 一种基于冷端系统储能的机组负荷智能优化控制方法
US20190236446A1 (en) * 2018-01-29 2019-08-01 Lawrence Livermore National Security, Llc Optimization control technology for building energy conservation
CN108805259A (zh) * 2018-05-23 2018-11-13 北京达佳互联信息技术有限公司 神经网络模型训练方法、装置、存储介质及终端设备
CN110925741A (zh) * 2019-12-13 2020-03-27 吉林大学 一种基于减温水量的主蒸汽各过热器进出口温度获取方法
CN111582472A (zh) * 2020-04-17 2020-08-25 广西电网有限责任公司电力科学研究院 一种基于神经网络模型的喷水减温器喷水调整方法及装置

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
ZHANG PEIYAO: "Neural Network Inverse Modelling and Steam Temperature Optimal Control for the Superheated Water-spray Desuperheating System", CHINESE MASTER'S THESES FULL-TEXT DATABASE, TIANJIN POLYTECHNIC UNIVERSITY, CN, 15 January 2020 (2020-01-15), CN , XP055858676, ISSN: 1674-0246 *

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114004159A (zh) * 2021-11-02 2022-02-01 西安热工研究院有限公司 一种余热锅炉过热器变工况效能计算方法
CN114004159B (zh) * 2021-11-02 2024-03-01 西安热工研究院有限公司 一种余热锅炉过热器变工况效能计算方法
CN115607719A (zh) * 2022-12-16 2023-01-17 威海中远海运重工科技有限公司 一种压载水紫外线反应器的智能控制方法
CN117180952A (zh) * 2023-11-07 2023-12-08 湖南正明环保股份有限公司 多向气流料层循环半干法烟气脱硫系统及其方法
CN117180952B (zh) * 2023-11-07 2024-02-02 湖南正明环保股份有限公司 多向气流料层循环半干法烟气脱硫系统及其方法

Also Published As

Publication number Publication date
CN111582472A (zh) 2020-08-25

Similar Documents

Publication Publication Date Title
WO2021208343A1 (zh) 一种基于神经网络模型的喷水减温器喷水调整方法及装置
US20200175435A1 (en) System and method for optimizing combustion of boiler
JP2907672B2 (ja) プロセスの適応制御方法およびプロセスの制御システム
CN104482525B (zh) 超超临界机组再热汽温的控制方法和系统
CN110285403A (zh) 基于被控参数预估的主蒸汽温度控制方法
CN105955026B (zh) 模糊pid控制方法和装置及系统
CN109670629A (zh) 基于长短期记忆神经网络的燃煤锅炉热效率预测方法
CN107831652A (zh) 一种基于冷端系统储能的机组负荷智能优化控制方法
CN111102559B (zh) 一种基于双神经网络逆模型的屛式过热器汽温控制方法
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
CN115510904B (zh) 基于时序预测的锅炉受热面积灰监测方法
CN115700330A (zh) 时空融合深度神经网络的锅炉再热器温度偏差预测方法
CN111027258A (zh) 一种用于超临界机组的发电负荷与供热量的智能预测方法
KR102291800B1 (ko) 보일러 연소 모델 도출 장치 및 방법
Wang et al. Deep-learning modeling and control optimization framework for intelligent thermal power plants: A practice on superheated steam temperature
US11287126B2 (en) System and method for controlling operation of boiler
CN112396162A (zh) 一种燃煤机组屛式过热器壁温预测神经网络模型
CN115326237A (zh) 一种考虑锅炉过热器氧化皮影响的壁温测量方法与系统
JPH11242503A (ja) プラント運転制御支援システム
CN117093836A (zh) 一种基于vmd-ewoa-lstm的末级再热器汽温预警方法及系统
Chaibakhsh Modelling and long-term simulation of a heat recovery steam generator
CN112348696B (zh) 基于bp神经网络的供热机组调峰上限评估方法及系统
CN114004159A (zh) 一种余热锅炉过热器变工况效能计算方法
Tang et al. Computer Prediction Model of Heat Consumption in Thermal System of Coal-Fired Power Station Based on Big Data Analysis and Information Sorting
CN115296335B (zh) 一次调频过程中锅炉热力参数动态特性的求解方法及装置

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: 20931241

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: 20931241

Country of ref document: EP

Kind code of ref document: A1