On-line prediction system and prediction method for flue gas oxygen content in thermal power plants
Through the PSO-Elman model and error feedback correction mechanism, the problems of low accuracy and model mismatch in flue gas oxygen content measurement are solved, and efficient online prediction of flue gas oxygen content in thermal power plants are achieved, reducing measurement costs and model updates are performed. frequency.
Patent Information
- Application Number
- CN202011179665.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-10-29
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2040-10-29
AI Technical Summary
Among the existing flue gas oxygen content measurement methods, the direct measurement method is easy to damage and has high cost. The offline prediction model of the soft measurement method is reduced when applied online, making it difficult to achieve fast and accurate online prediction.
The online prediction system using the PSO-Elman model combined with error feedback correction is established through signal acquisition, data preprocessing, expert knowledge base and intelligent modeling modules, and the prediction model of the flue gas oxygen content is established through Elman neural network and particle swarm algorithm, and the prediction model of the flue gas oxygen content is established through The weighted correction mechanism optimizes the prediction results to ensure the real-time accuracy of the model.
It improves the online prediction accuracy of flue gas oxygen content, reduces the replacement frequency of zirconia sensors, reduces the measurement cost, and maintains the stability and generalization capabilities of the model under various operating conditions.
Smart Images

Figure CN112308311B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the cross-technical field of boiler thermal technology and computer monitoring, and specifically is an online prediction system and a prediction method for flue gas oxygen content in a thermal power plant. Background Art
[0002] In recent years, with the increase in energy conservation and emission reduction efforts, the country has attached more and more importance to the relevant technical research on the economical and efficient operation of thermal power plants. As an important operating equipment in thermal power plants, the real-time evaluation of the combustion state of the boiler is an important link in realizing online monitoring and optimization of combustion, which is of great significance to the safe and economical stable operation of thermal power units. In engineering practice, the excess air coefficient is used to judge the real-time combustion state of the boiler. However, since this parameter cannot be directly measured, the excess air coefficient at the corresponding moment is usually calculated indirectly by detecting the oxygen content at the outlet of the air preheater at the rear of the boiler. Based on this, the amount of air entering the furnace is adjusted to optimize the combustion air-coal ratio of the boiler, so that the boiler is in an efficient and safe operating state. Excessive flue gas oxygen content will cause waste gas heat loss and a decrease in boiler efficiency. On the contrary, it will cause coking in the furnace, seriously threatening the safety of boiler operation. Therefore, the real-time and accurate measurement of flue gas oxygen content in thermal power plants has always been a key technology that the power industry at home and abroad attaches great importance to.
[0003] At present, the main methods for measuring flue gas oxygen content at home and abroad are divided into direct measurement method and soft measurement method: (1) The direct measurement method uses a zirconium oxide sensor to directly measure the oxygen content at the tail of the boiler. This method is simple to operate, but there is NO in the actual measurement environment. x and SO xCorrosive gases such as argon and argon are easy to damage the sensor, resulting in a decrease in the measurement accuracy of the sensor and frequent replacement, which is not conducive to the accurate measurement of flue gas oxygen content and increases the measurement cost; (2) In recent years, a large number of scholars have used soft measurement technology to predict the flue gas oxygen content that is difficult to measure directly and accurately with the help of relevant auxiliary variables that are easy to measure in the measurement environment. These include ① establishing a prediction model using an intelligent algorithm model ② establishing a prediction model using mechanism analysis ③ hybrid method modeling, that is, a method that combines mechanism analysis with an intelligent algorithm model. The document CNKI:SUN:RNWS.0.2010-03-020 "Soft measurement of flue gas oxygen content based on LS-SVM and simplex" proposes a method for predicting flue gas oxygen content in thermal power plants based on the least squares support vector machine. This method has good offline prediction accuracy, but the training and modeling process has a large amount of calculation, which is not conducive to the rapid and accurate establishment and update of the model during online prediction; Chinese patent CN2017101083999 "Prediction system and method for flue gas oxygen content at the outlet of the furnace of a circulating fluidized bed domestic waste incineration boiler" uses PSO-GA-SVM to establish an offline prediction model, which accelerates the convergence speed of network training iterations and has good offline prediction accuracy. A model update module is added, that is, when the prediction error is large during online testing, the model can be retrained, so that the model has a certain robustness. However, in actual operation, the offline prediction model is directly used for online prediction, which is easily affected by complex working conditions and disturbances, resulting in a decrease in accuracy, which will lead to frequent training and updating of the prediction model. This will lead to a decrease in prediction efficiency and is not conducive to accurate online measurement. Summary of the invention
[0004] The invention discloses an online prediction system and method for flue gas oxygen content in a thermal power plant, which solve the problems of low precision in existing soft measurement methods and model mismatch in online testing of offline prediction models.
[0005] In order to achieve the above object, the technical solution of the present invention is:
[0006] An online prediction system for flue gas oxygen content in a thermal power plant is composed of a host computer and a data communication interface, and is connected to a distributed control system of the thermal power plant. A flue gas oxygen content prediction model is constructed and updated in the host computer, and the prediction results that meet the accuracy requirements are transmitted to the distributed control system through the data communication interface. The host computer includes:
[0007] (1) Signal acquisition module: This module is used to collect boiler operating status parameters from the boiler distributed control system, and select the main steam pressure, main steam flow, total air volume, air preheater outlet flue gas temperature, total coal volume, boiler load, and blower current phase as input variables of the flue gas oxygen content prediction model by calculating the Pearson correlation coefficient.
[0008] (2) Data preprocessing module: This module uses the improved wavelet double-layer threshold denoising method to remove singular points and noise from the prediction model input data. It also establishes training set data and online test set data through stratified sampling, and then normalizes the data and maps it to the range of [-1,1].
[0009] (3) Expert knowledge base: This module uses a rolling time window method to continuously update training samples so that the training samples are always kept in the latest state. The rolling time window method refers to starting from the current time and looking back to a time dimension of length L, where L is in seconds;
[0010] (4) Intelligent modeling module: The Elman neural network is used to establish a relationship model between relevant auxiliary variables and flue gas oxygen content, and the PSO algorithm is introduced to optimize the weights and thresholds in the Elman model to obtain the optimal weights and thresholds, which are then assigned to the Elman network model to complete the construction of the PSO-Elman prediction model. The specific process is as follows:
[0011] 4.1) Establish the Elman structure and train it: The input is 7 parameters, the output is 1 parameter, the Elman network structure is established as 7-15-1, the hidden layer transfer function uses the tansig function, the output layer transfer function uses the linear function, and the training function is trainlm, that is, the formula for correcting the network weights using the Levenberg-Marquardt (LM) algorithm is:
[0012] ω(n+1)=ω(n)-[J T J+μI] -1 J T e
[0013] Where e is the error performance function, and J is the Jacobian matrix of the first-order derivative of the error performance function with respect to the network weights. By obtaining the training error, the mean absolute error is calculated. Where N is the number of training times.
[0014] 4.2) The PSO optimization algorithm is introduced to optimize all weights and thresholds in the Elman network structure. The prediction error is first defined as the fitness value in the PSO algorithm, and its calculation formula is as follows: Then, the population is randomly initialized, including setting the population size of the particle swarm to 40, the particle length to 105, the maximum number of iterations to 100, the learning factor C1=2.3, C2=1.3, the maximum speed of particle movement vmax=1, the maximum inertia weight wmax=0.9, and the minimum value wmin=0.3. Then, the fitness of the individual is calculated, and the local optimal position vector of the individual and the global optimal position vector of the population are selected; then the number of iterations, the speed vector and position vector of each individual are set; then the speed of each individual is updated: Vid =ωV id +C1r1(P id -X id )+C2r2(P gd -X id ) and displacement update: X id =X id +V id , where V id is the movement speed of the example; X id is the current position of the particle; ω is the inertia factor; C1, C2 are learning factors, generally C1 = C2 = 2; r1, r2 are random numbers on [0, 1]; Pid is the individual extreme value of the i-th example; Pgd is the global extreme value. Finally, determine whether the maximum number of iterations or the global optimal position has been reached. If it has been reached, output the optimal solution. If it has not been reached, return to continue iteration. Get the optimal weights and thresholds, and assign them to the Elman model to complete the construction of the PSO-Elman model.
[0015] (5) Error correction feedback module: Use the online test set to run the PSO-Elman flue gas oxygen content prediction model, calculate the prediction error of each step, and feed back the weighted prediction error of each step to the next prediction output to complete the feedback correction of the online prediction output. The specific method is:
[0016] Using the prediction error e(k-1) at the k-1th second, that is
[0017]
[0018] Weighted correction of the k-th second prediction result: Set the weighted coefficient matrix h with a dimension of 1×800. The formula for correcting the k-th second prediction output is:
[0019]
[0020] Set the initial value of all elements in the weight coefficient matrix h to 0.5, and the value interval is [0.1], then the weight coefficient update algorithm principle is:
[0021] (Step 1) Calculate the prediction error
[0022] (Step 2) Calculate the weighted coefficient compensation increment
[0023] (Step 3) Update weighting coefficient h(1,k)=0.5+delta(k)
[0024] (Step 4) Limit the updated weighting coefficients. If h(1, k) ≥ 1, then output h(1, k) = 1. If h(1, k) ≤ 0, then output h(1, k) = 0.
[0025] (Step 5) k = k + 1, and determine whether k ≤ the maximum number of test steps. If so, return to step 1
[0026] in is the mean absolute error of the model training calculated in step 2. This allows the prediction error of the previous step to be used to correct the prediction output of the next prediction model.
[0027] (6) Model evaluation module: Obtain the prediction error of each model step and determine whether the error is greater than or equal to 5%. If so, the current prediction model is considered mismatched and returns to the intelligent modeling module to train the model again using the latest training set data from the expert knowledge base.
[0028] (7) Communication module: This module transmits the flue gas oxygen content prediction results that meet the accuracy requirements to the distributed control system of the thermal power plant.
[0029] A prediction method for an online prediction system of flue gas oxygen content in a thermal power plant is implemented by combining neural network modeling with a feedback correction link. The specific steps are as follows:
[0030] Step 1: Read historical data from the DCS (distributed control system) of the thermal power plant, select relevant auxiliary variables for modeling, and generate training and test data sets after preprocessing the data;
[0031] Step 2: Use the Elman neural network to establish a relationship model between relevant auxiliary variables and flue gas oxygen content, and perform training to calculate the mean absolute error of the prediction.
[0032] Step 3: Introduce the PSO optimization algorithm to optimize all weights and thresholds in the Elman network structure, obtain the optimal weights and thresholds, and assign them to the Elman model to complete the construction of the PSO-Elman model.
[0033] Step 4: Use the online test set to run the PSO-Elman flue gas oxygen content prediction model, calculate the prediction error of each step, and feed back the weighted prediction error of each step to the next prediction output to complete the feedback correction of the online prediction output.
[0034] Furthermore, in step 1, the Pearson correlation coefficient is used to calculate the correlation between other parameters in the operation history data set and the flue gas oxygen content. Parameters higher than 0.8 are considered to have a strong correlation with the flue gas oxygen content.
[0035] Furthermore, in step 1, main steam pressure, main steam flow, total air volume, flue gas temperature at air preheater outlet, total coal volume, boiler load, and blower current are established as relevant auxiliary variables as inputs of the prediction model, and flue gas oxygen content is used as output of the prediction model, forming a data set for training the prediction model.
[0036] Furthermore, in step 1, the wavelet threshold denoising method is improved by setting a double-layer threshold, so as to remove singular points and noise in the input data of the prediction model. Based on the traditional noise estimation and general threshold calculation, the wavelet coefficients greater than the first threshold T1 are first shrunk into T1, that is, |d j (n)|>T1, then re-estimate the noise to form T2, and use T2 to perform soft threshold processing on the global signal. The value function is:
[0037]
[0038] in are the wavelet coefficients after two threshold adjustments.
[0039] Furthermore, in step 1, in order to ensure that the data characteristics of operation at all time periods throughout the day are covered, the training set data is selected using a stratified sampling method. Since the mapping relationship between the input data during the daily peak power consumption period and the flue gas oxygen content is more complex, the daily data is divided into two layers (i.e., 7:00-23:00; 00:00-07:00 and 23:00-24:00) according to the daily peak power consumption period of 7:00 to 23:00. Since the system sampling frequency is 15S / time, the sample size of the first layer of data (7:00-23:00) is 3840, and the sampling ratio of this layer is calculated: The sample size of the second layer of data (00:00-07:00 and 23:00-24:00) is 1920, so the sampling ratio of this layer is calculated as follows: Therefore, the sampling ratios of the two layers of data are set to 2 / 3 and 1 / 3 respectively, and 900 groups of data are extracted to form the training set for training the model. And 800 consecutive groups of data are selected as the test set for online verification.
[0040] Furthermore, in step 2, the Elman structure is established: the input is 7 parameters, the output is 1 parameter, the Elman network structure is established as 7-15-1, the hidden layer transfer function uses the tansig function, the output layer transfer function uses the linear function, and the training function is trainlm, that is, the formula for correcting the network weights using the Levenberg-Marquardt (LM) algorithm is:
[0041] ω(n+1)=ω(n)-[J T J+μI] -1 J T e
[0042] Where e is the error performance function, and J is the Jacobian matrix of the first-order derivative of the error performance function with respect to the network weights.
[0043] Furthermore, the prediction error is defined in step 2 as the fitness value in the PSO algorithm, and its calculation formula is as follows:
[0044]
[0045] Furthermore, in step 3, the particle swarm optimizes the optimal weight and threshold process of the Elman neural network: first, the population is randomly initialized, and the fitness of the individual is calculated, and the local optimal position vector of the individual and the global optimal position vector of the population are selected; then the number of iterations, the velocity vector and position vector of each individual are set; then the velocity of each individual is updated: V id =ωV id +C1r1(P id -X id )+C2r2(P gd -X id ) and displacement update: X id =X id +V id , where V id is the movement speed of the example; X id is the current position of the particle; ω is the inertia factor; C1, C2 are learning factors, generally C1=C2=2; r1, r2 are random numbers on [0,1]; P id is the individual extreme value of the ith example; P gd is the global extreme value. Finally, determine whether the maximum number of iterations or the global optimal position has been reached. If reached, output the optimal solution. If not reached, return to continue iteration.
[0046] Furthermore, the feedback correction link in step 4 uses the prediction error e(k-1) at the k-1th second, that is,
[0047]
[0048] Weighted correction of the k-th second prediction result: Set the weighted coefficient matrix h with a dimension of 1×800. The formula for correcting the k-th second prediction output is:
[0049]
[0050] Set the initial value of all elements in the weight coefficient matrix h to 0.5, and the value interval is [0.1], then the weight coefficient update algorithm principle is:
[0051] (Step 1) Calculate the prediction error
[0052] (Step 2) Calculate the weighted coefficient compensation increment
[0053] (Step 3) Update weighting coefficient h(1,k)=0.5+delta(k)
[0054] (Step 4) Limit the updated weighting coefficients. If h(1, k) ≥ 1, then output h(1, k) = 1. If h(1, k) ≤ 0, then output h(1, k) = 0.
[0055] (Step 5) k = k + 1, and determine whether k ≤ the maximum number of test steps. If so, return to step 1
[0056] in is the mean absolute error of the model training calculated in step 2. Thus, the prediction output of the next prediction model can be predicted by using the prediction error of the previous step.
[0057] Compared with the prior art, the present invention has the following beneficial effects:
[0058] 1. By setting a double-layer threshold for denoising based on the wavelet algorithm, the singular point coefficients are first shrunk multiple times to return them to the normal range, and then the global signal is soft-thresholded, which can better solve the singular point problem. The traditional soft threshold method only shrinks the coefficients greater than the threshold once, so the effect of removing singular points is not ideal.
[0059] 2. The soft measurement method is used to predict the oxygen content of flue gas which is difficult to measure directly and accurately, avoiding the frequent replacement of zirconia sensors and saving measurement costs.
[0060] 3. The prediction model is established using the dynamic neural network Elman network (i.e., PSO-Elman model) optimized by the particle swarm algorithm, which can not only better approximate the nonlinear model of boiler combustion, but also avoid falling into the local minimum during the algorithm iteration operation, thereby accelerating the convergence speed of the iterative operation.
[0061] 4. The training set data used for model training is selected from operating conditions under various working conditions, so that the trained prediction model is applicable to a variety of working conditions and has better generalization ability.
[0062] 5. The addition of an error correction feedback link can adjust the prediction model online to avoid model mismatch caused by disturbances encountered during online testing, thereby ensuring the accuracy and stability of the prediction model. BRIEF DESCRIPTION OF THE DRAWINGS
[0063] Figure 1 It is the principle diagram of the PSO-Elman offline prediction model algorithm of the present invention;
[0064] Figure 2 It is a principle diagram of the feedback correction PSO-Elman online prediction model algorithm of the present invention;
[0065] Figure 3It is a diagram of the iterative error variation process of the PSO-Elman model of the present invention;
[0066] Figure 4 It is a comparison diagram of the feedback correction PSO-Elman model and the online test effect of the PSO-Elman model of the present invention;
[0067] Figure 5 It is a prediction error diagram of the online test of the feedback correction PSO-Elman model of the present invention;
[0068] Figure 6 It is a graph of the online test prediction relative error of the feedback correction PSO-Elman model of the present invention;
[0069] Figure 7 It is the overall system structure diagram of the present invention;
[0070] Figure 8 It is the working principle diagram of the host computer of the present invention; DETAILED DESCRIPTION
[0071] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0072] The present invention is described in detail below in conjunction with specific implementation cases. The following examples will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be pointed out that for those of ordinary skill in the art, several changes and improvements can be made without departing from the concept of the present invention. These all belong to the scope of protection of the present invention.
[0073] like Figure 1 , Figure 2As shown, the present invention provides an online prediction method for flue gas oxygen content in a thermal power plant based on a PSO-Elman model. The main idea of the method is to use the PSO-Elman model to approximate the complex boiler combustion nonlinear model and error feedback correction to realize the online correction prediction model to achieve the online accurate prediction function, which mainly includes two parts: offline prediction model establishment and online model correction algorithm. The establishment of the offline prediction model includes the following three steps: first, read the historical data from the DCS (distributed control system) of the thermal power plant, select the relevant auxiliary variables and target prediction variables for modeling, divide them into training sets and online test sets, and normalize the data; then use the Elman neural network to establish a prediction model between the relevant auxiliary variables and the flue gas oxygen content, and set the fitness of the particle swarm optimization algorithm; finally, introduce the PSO optimization algorithm to optimize all the weights and thresholds in the Elman network structure, obtain the optimal weights and thresholds, and assign them to the Elman model to complete the construction of the PSO-Elman model. The online model correction algorithm uses the online test set to run the PSO-Elman flue gas oxygen content prediction model, calculates the prediction error of each step, and feeds back the weighted prediction error of each step to the next prediction output to complete the feedback correction of the online prediction output.
[0074] The details are as follows:
[0075] (1) First, the combustion data of SG-2136.5 / 17.55-M 660MW subcritical boiler was read from the DCS system of a thermal power plant in Jiaxing, Zhejiang. The correlation coefficient between each variable and the oxygen content at the outlet of the air preheater (i.e., the oxygen content of the flue gas) was calculated using the Pearson correlation coefficient method. When the correlation coefficient was higher than 8.0, it was considered to have a high correlation with the oxygen content of the flue gas. The calculation formula is as follows:
[0076]
[0077] (2) The improved wavelet double-layer threshold denoising method is used to remove noise and singular points from the data set consisting of main steam pressure, main steam flow, total air volume, air preheater outlet flue gas temperature, total coal volume, boiler load, blower current, and flue gas oxygen content. Based on traditional noise estimation and general threshold calculation, the wavelet coefficients greater than the first threshold T1 are first shrunk into T1, that is, |d j (n)|>T1, then re-estimate the noise to form T2, and use T2 to perform soft threshold processing on the previous step signal. The value function is:
[0078]
[0079] in are the wavelet coefficients after two threshold adjustments.
[0080] (3) The training set data is selected using the stratified sampling method. Since the mapping relationship between the input data during the daily peak electricity consumption period and the flue gas oxygen content is more complicated, the daily data is divided into two layers (i.e., 7:00-23:00; 00:00-07:00 and 23:00-24:00) according to the daily peak electricity consumption period of 7:00 to 23:00. Since the system sampling frequency is 15S / time, the sample size of the first layer of data (7:00-23:00) is 3840, and the sampling ratio of this layer is calculated as follows: The sample size of the second layer of data (00:00-07:00 and 23:00-24:00) is 1920, so the sampling ratio of this layer is calculated as follows: Therefore, the sampling ratios of the two layers of data are set to 2 / 3 and 1 / 3 respectively, and 900 groups of data are extracted to form the training set for training the model. And 800 consecutive groups of data are selected as the test set for online verification.
[0081] (4) Establishing the Elman structure: The input is 7 parameters, the output is 1 parameter, the Elman network structure is established as 7-15-1, the hidden layer transfer function uses the tansig function, the output layer transfer function uses the linear function, and the training function is trainlm, that is, the formula for correcting the network weights using the Levenberg-Marquardt (LM) algorithm is:
[0082] ω(n+1)=ω(n)-[J T J+μI] -1 J T e
[0083] Where e is the error performance function, and J is the Jacobian matrix of the first-order derivative of the error performance function with respect to the network weights. By obtaining the training error, the mean absolute error is calculated. Where N is the number of training times.
[0084] (5) The PSO optimization algorithm is introduced to optimize all weights and thresholds in the Elman network structure. The prediction error is first defined as the fitness value in the PSO algorithm. The calculation formula is as follows: Then, the population is randomly initialized, including setting the population size of the particle swarm to 40, the particle length to 105, the maximum number of iterations to 100, the learning factor C1=2.3, C2=1.3, the maximum speed of particle movement vmax=1, the maximum inertia weight wmax=0.9, and the minimum value wmin=0.3. Then, the fitness of the individual is calculated, and the local optimal position vector of the individual and the global optimal position vector of the population are selected; then the number of iterations, the speed vector and position vector of each individual are set; then the speed of each individual is updated: V id =ωV id +C1r1(P id -Xid )+C2r2(P gd -X id ) and displacement update: X id =X id +V id , where Vid is the speed of the example; Xid is the current position of the particle; ω is the inertia factor; C1, C2 are learning factors, generally C1 = C2 = 2; r1, r2 are random numbers on [0,1]; Pid is the individual extreme value of the i-th example; Pgd is the global extreme value. Finally, determine whether the maximum number of iterations or the global optimal position has been reached. If it has been reached, output the optimal solution. If it has not been reached, return to continue iteration. Get the optimal weights and thresholds, and assign them to the Elman model to complete the construction of the PSO-Elman model.
[0085] (6) Set the test duration to 175 minutes (sampling frequency 15 seconds / time), gradually read data from the online test set, and conduct online testing of the PSO-Elman prediction model. Use the prediction error e(k-1) at the k-1th second, that is,
[0086]
[0087] Weighted correction of the k-th second prediction result: Set the weighted coefficient matrix h with a dimension of 1×800. The formula for correcting the k-th second prediction output is:
[0088]
[0089] In this way, the prediction error of the previous step is used to correct the prediction output of the next prediction model. Set the initial value of all elements in the weight coefficient matrix h to 0.5, and the value interval is [0.1]. The pseudo code of the weight coefficient update algorithm is:
[0090]
[0091] in is the mean absolute error of the model training calculated in step 2. This allows the prediction error of the previous step to be used to correct the prediction output of the next prediction model.
[0092] Figure 3 is the iterative error change rate during the evolution of the PSO algorithm. The iterative error of the PSO-Elman model is at the order of -2, and the prediction accuracy of the model is very high. After the 20th iteration, the iterative error change rate dropped below 0.04 and quickly stabilized at around 0.035.
[0093] Figure 4The feedback correction PSO-Elman model and the PSO-Elman model online test effect are compared. It can be seen that as the real flue gas oxygen content changes, the feedback correction PSO-Elman model can track the real flue gas oxygen content curve more closely than the PSO-Elman model, and the fluctuation is smaller.
[0094] Figure 5 , Figure 6 They are the error and relative error of the online test of the feedback-corrected PSO-Elman model, respectively. It can be seen that the errors are all less than ±0.05, and the relative errors are all within ±0.15.
[0095] Figure 7 , Figure 8 The present invention provides a thermal power plant flue gas oxygen content prediction system, including a thermal power plant boiler, a distributed control system for the boiler operation control, a data communication interface, a database and a host computer. The database reads data from the distributed control system through the data communication interface and is used for training, learning and testing of the host computer. The host computer exchanges data with the distributed control system through the data communication interface. The host computer includes offline learning and training, online prediction verification, error feedback correction part, and model update judgment part. The host computer specifically includes:
[0096] (1) Signal acquisition module: This module is used to collect boiler operating status parameters from the boiler distributed control system, and select the main steam pressure, main steam flow, total air volume, air preheater outlet flue gas temperature, total coal volume, boiler load, and blower current phase as input variables of the flue gas oxygen content prediction model by calculating the Pearson correlation coefficient.
[0097] (2) Data preprocessing module: This module uses the improved wavelet double-layer threshold denoising method to remove singular points and noise from the prediction model input data. It also establishes training set data and online test set data through stratified sampling, and then normalizes the data and maps it to the range of [-1,1].
[0098] (3) Expert knowledge base: This module uses a rolling time window method to continuously update training samples so that the training samples are always kept in the latest state. The rolling time window method refers to starting from the current time and looking back to a time dimension of length L, where L is in seconds;
[0099] (4) Intelligent modeling module: The Elman neural network is used to establish a relationship model between relevant auxiliary variables and flue gas oxygen content, and the PSO algorithm is introduced to optimize the weights and thresholds in the Elman model to obtain the optimal weights and thresholds, which are then assigned to the Elman network model to complete the construction of the PSO-Elman prediction model. The specific process is as follows:
[0100] 4.1) Establish the Elman structure and train it: The input is 7 parameters, the output is 1 parameter, the Elman network structure is established as 7-15-1, the hidden layer transfer function uses the tansig function, the output layer transfer function uses the linear function, and the training function is trainlm, that is, the formula for correcting the network weights using the Levenberg-Marquardt (LM) algorithm is:
[0101] ω(n+1)=ω(n)-[J T J+μI] -1 J T e
[0102] Where e is the error performance function, and J is the Jacobian matrix of the first-order derivative of the error performance function with respect to the network weights. By obtaining the training error, the mean absolute error is calculated. Where N is the number of training times.
[0103] 4.2) The PSO optimization algorithm is introduced to optimize all weights and thresholds in the Elman network structure. The prediction error is first defined as the fitness value in the PSO algorithm, and its calculation formula is as follows: Then, the population is randomly initialized, including setting the population size of the particle swarm to 40, the particle length to 105, the maximum number of iterations to 100, the learning factor C1=2.3, C2=1.3, the maximum speed of particle movement vmax=1, the maximum inertia weight wmax=0.9, and the minimum value wmin=0.3. Then, the fitness of the individual is calculated, and the local optimal position vector of the individual and the global optimal position vector of the population are selected; then the number of iterations, the speed vector and position vector of each individual are set; then the speed of each individual is updated: V id =ωV id +C1r1(P id -X id )+C2r2(P gd -X id ) and displacement update: X id =X id +V id , where Vid is the speed of the example; Xid is the current position of the particle; ω is the inertia factor; C1, C2 are learning factors, generally C1 = C2 = 2; r1, r2 are random numbers on [0,1]; Pid is the individual extreme value of the i-th example; Pgd is the global extreme value. Finally, determine whether the maximum number of iterations or the global optimal position has been reached. If it has been reached, output the optimal solution. If it has not been reached, return to continue iteration. Get the optimal weights and thresholds, and assign them to the Elman model to complete the construction of the PSO-Elman model.
[0104] (5) Error correction feedback module: Use the online test set to run the PSO-Elman flue gas oxygen content prediction model, calculate the prediction error of each step, and feed back the weighted prediction error of each step to the next prediction output to complete the feedback correction of the online prediction output. The specific method is:
[0105] Using the prediction error e(k-1) at the k-1th second, that is
[0106]
[0107] Weighted correction of the k-th second prediction result: Set the weighted coefficient matrix h with a dimension of 1×800. The formula for correcting the k-th second prediction output is:
[0108]
[0109] Set the initial value of all elements in the weight coefficient matrix h to 0.5, and the value interval is [0.1]. Then the pseudo code of the weight coefficient update algorithm is:
[0110]
[0111] in is the mean absolute error of the model training calculated in step 2. This allows the prediction error of the previous step to be used to correct the prediction output of the next prediction model.
[0112] (6) Model evaluation module: Obtain the prediction error of each model step and determine whether the error is greater than or equal to 5%. If so, the current prediction model is considered mismatched and returns to the intelligent modeling module to train the model again using the latest training set data from the expert knowledge base.
[0113] (7) Communication module: This module transmits the flue gas oxygen content prediction results that meet the accuracy requirements to the boiler distributed control system.
[0114] The above specific examples are used to illustrate the present invention, which is only used to help understand the present invention and is not intended to limit the present invention. Any partial modification or replacement within the technical scope disclosed by the present invention by anyone familiar with the technology should be included in the scope of the present invention.
Claims
1. The prediction method of the online prediction system for flue gas oxygen content in thermal power plants is characterized by: The system includes a host computer and a data communication interface, and the data communication interface is connected to a boiler distributed control system; a flue gas oxygen content prediction model is constructed and updated in the host computer, and the prediction results that meet the accuracy requirements are transmitted to the distributed control system through the data communication interface; the host computer includes: a signal acquisition module: the module is used to collect boiler operating status parameters from the boiler distributed control system, and select the main steam pressure, main steam flow, total air volume, air preheater outlet flue gas temperature, total coal volume, boiler load, and blower current phase as input variables of the flue gas oxygen content prediction model by calculating the Pearson correlation coefficient; a data preprocessing module: the module uses an improved wavelet double-layer threshold denoising method to remove singular points and noise from the prediction model input data; and establishes training set data and online test set data through stratified sampling, and then normalizes the data and maps it to the range of [-1,1]; an expert knowledge base: the module uses a rolling time window method to continuously update training samples, so that the training samples are always kept in the latest state, and the rolling time The time window method refers to the time dimension of L length that is traced back from the current time, where L is in seconds; Intelligent modeling module: This module uses the Elman neural network to establish a relationship model between relevant auxiliary variables and flue gas oxygen content, and then introduces the PSO algorithm to optimize the weights and thresholds in the Elman model to obtain the optimal weights and thresholds, and assign them to the Elman network model to complete the construction of the PSO-Elman prediction model; Error correction feedback module: This module uses the online test set to run the PSO-Elman flue gas oxygen content prediction model, calculates the prediction error of each step, and feeds back the weighted prediction error of each step to the next prediction output to complete the feedback correction of the online prediction output; Model evaluation module: This module obtains the prediction error of each model step and determines whether the error is greater than or equal to 5%. If so, it is considered that the current prediction model is mismatched and returns to the intelligent modeling module to train the model again using the latest training set data from the expert knowledge base; Communication module: This module transmits the prediction results of flue gas oxygen content that meet the accuracy requirements to the boiler distributed control system; In the intelligent modeling module, the online prediction of flue gas oxygen content is realized by establishing a PSO-Elman prediction model and introducing a feedback correction structure, including the following steps: Step 1: Read historical data from the distributed control system of the thermal power plant, select relevant auxiliary variables and target prediction variables for modeling, divide them into training sets and online test sets, and normalize the data; Step 2: Use the Elman neural network to establish a prediction model between relevant auxiliary variables and flue gas oxygen content, and perform training to calculate the prediction mean absolute error, specifically including: establishing the Elman structure and training: the input is 7 parameters, the output is 1 parameter, the Elman network structure is established as 7-15-1, the hidden layer transfer function uses the tansig function, the output layer transfer function uses the linear function, and the training function is trainlm, that is, the formula for correcting the network weights using the Levenberg-Marquardt (LM) algorithm is: ω(n+1)=ω(n)-[J T J+μI] -1 J T its Where e is the error performance function, J is the Jacobian matrix of the first-order derivative of the error performance function with respect to the network weights, and the mean absolute error is calculated by obtaining the training error Where N is the number of training times; Step 3: Introduce the PSO optimization algorithm to optimize all weights and thresholds in the Elman network structure, obtain the optimal weights and thresholds, and assign them to the Elman model to complete the construction of the PSO-Elman model, including: First, define the prediction error as the fitness value in the PSO algorithm, and its calculation formula is as follows: Then the population is randomly initialized, including setting the population size of the particle swarm to 40, the particle length to 105, the maximum number of iterations to 100, the learning factors C1=2.3, C2=1.3, the maximum speed of particle movement vmax=1, the maximum inertia weight wmax=0.9, and the minimum inertia weight wmin=0.3; Then calculate the fitness of the individual, select the local optimal position vector of the individual and the global optimal position vector of the population; Then set the number of iterations, each individual velocity vector and position vector; Then update the speed of each individual: V id =ωV id +C1r1(P id -X id )+C2r2(P gd -X id ) and displacement update: X id =X id +V id , where Vid is the speed of the example; Xid is the current position of the particle; ω is the inertia factor; C1, C2 are learning factors, generally C1 = C2 = 2; r1, r2 are random numbers on [0,1]; Pid is the individual extreme value of the i-th example; Pgd is the global extreme value, Finally, determine whether the maximum number of iterations or the global optimal position has been reached. If reached, output the optimal solution. If not, return to continue iteration. Get the optimal weight and threshold, and assign them to the Elman model to complete the construction of the PSO-Elman model. Step 4: Use the online test set to run the PSO-Elman flue gas oxygen content prediction model, calculate the prediction error of each step, and feed back the weighted prediction error of each step to the next prediction output to complete the feedback correction of the online prediction output, including: Set the test duration, gradually read data from the online test set, perform online testing of the PSO-Elman prediction model, and use the prediction error e(k-1) at the k-1th second, that is, Weighted correction of the k-th second prediction result: Set the weighted coefficient matrix h with a dimension of 1×800. The formula for correcting the k-th second prediction output is:
