Implementation method and system for online correction of wind-coal-water reference function

Abstract

The invention discloses an implementation method and system for online correction of a wind-coal-water reference function, and the method comprises the steps: obtaining the historical operation data of a thermal power generating unit, dividing the historical operation data through a first steady-state detection algorithm, and obtaining steady-state subsets; performing first preprocessing on the historical operation data by using a first anomaly detection algorithm, and obtaining a first data set in combination with the steady-state subset; dividing the first data set into mutually exclusive subsets in equal proportion by using a cross validation method, and performing second preprocessing on the mutually exclusive subsets to obtain first mutually exclusive subsets; training on the first mutual exclusion subset by using a weak learning algorithm to obtain a first corrector; performing weighted summation on the output result of the first corrector by using a first ensemble learning algorithm to obtain a second corrector; and the first ensemble learning algorithm parameters are optimized on the second corrector based on an optimization strategy to obtain a wind-coal-water reference function correction result, so that good effects can be achieved under different unit and different load working conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial process control, and particularly relates to a method and system for online calibration of a wind-coal-water reference function. Background Art

[0002] Under the dual background of China's realization of the "dual carbon" goal and the construction of a new power system, coal-fired power units are undergoing a transformation stage from a supporting power source to a regulating power source. During this stage, the power grid highly integrates new energy, and coal-fired power units need to undertake new tasks of frequent peak shaving and frequency modulation, which puts higher requirements on the flexibility of coal-fired power units. Wind, coal, and water are three key variables in the analog control of coal-fired power units and the core quantities of the coordinated control system (CCS). The degree of coordination among the three directly determines the economy and flexibility of the unit operation. During the process of load change, as the load changes, the fuel master control basically slides along the load-coal quantity reference function; the air supply and water supply quantities slide along the coal quantity-air supply / water supply quantity functions. Usually, the reference functions of the three are given by the commissioning party during commissioning based on tests. The corresponding coal quantity for different loads is calculated according to the calorific value of standard coal, and then the corresponding required air volume and water volume are calculated / tested according to different coal quantities. When the functions are accurate, the unit has good wind-coal-water coordination, strong load change ability, high operation flexibility and economy. However, after the unit has been operating for a long time, the combustion characteristics of the boiler shift, and the operating coal type deviates from the designed coal type. At this time, the original wind-coal-water reference function is no longer applicable, and the economy and flexibility of the unit operation decrease.

[0003] However, the coordinated calibration of the wind-coal-water curve still relies on manual adjustment at present. A large number of experiments need to be carried out on the unit, which not only relies heavily on human experience, but only a few commissioning units have the commissioning qualifications, and the cost is quite high. During the peak load period, there is often no time for adjustment, resulting in the unit being in an uncoordinated state of wind-water-coal for a long time, which not only increases the operation intensity of the operating personnel, but also bears considerable risks and potential hazards. Moreover, with the continuous improvement of the informatization level of power plants and the continuous improvement of the big data monitoring and integration system, a large amount of production data is generated during the production of power plants. Properly using data mining and machine learning algorithms to mine and use these data is beneficial to improving the digital and intelligent levels of power plants. Therefore, it is quite feasible and has important engineering application value to conduct data mining on the historical database of thermal power units and study a method for realizing online calibration of the wind-coal-water reference function using machine learning. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a method and system for online calibration of the air-coal-water reference function to solve the problems that the coordinated calibration of the current air-coal-water curve depends on manual adjustment, with high costs, and during the peak load period, the unit is in an uncoordinated state of air-water-coal for a long time, posing risks and potential hazards.

[0006] To solve the above technical problems, the present invention provides the following technical solutions:

[0007] In the first aspect, the present invention provides a method for online calibration of the air-coal-water reference function, including:

[0008] Obtain the historical operation data of the thermal power unit, and use the first steady-state test algorithm to divide the historical operation data to obtain a steady-state subset;

[0009] Use the first anomaly detection algorithm to perform the first preprocessing on the historical operation data, and combine it with the steady-state subset to obtain a first data set;

[0010] Use the cross-validation method to divide the first data set into mutually exclusive subsets in equal proportions, and perform the second preprocessing on the mutually exclusive subsets to obtain a first mutually exclusive subset;

[0011] Use the weak learning algorithm to train on the first mutually exclusive subset to obtain a first corrector;

[0012] Use the first ensemble learning algorithm to perform weighted summation on the output results of the first corrector to obtain a second corrector;

[0013] Optimize the parameters of the first ensemble learning algorithm on the second corrector based on the optimization strategy to obtain the calibration result of the air-coal-water reference function.

[0014] As a preferred scheme of the method for online calibration of the air-coal-water reference function of the present invention, wherein: using the first steady-state test algorithm to divide the historical operation data includes:

[0015] For the data at a certain moment, obtain the test index for steady-state determination by calculating the variance estimate value with the filtered data of the previous moment and the variance estimate value with the sampled data of the previous moment;

[0016] If the test index for steady-state determination is less than or equal to the set threshold, then the data is in a steady state.

[0017] As a preferred scheme of the method for online calibration of the air-coal-water reference function of the present invention, wherein: using the first anomaly detection algorithm to perform the first preprocessing on the historical operation data includes:

[0018] Use the first anomaly detection algorithm to eliminate the abnormal working conditions in the historical operation data to obtain a first data set.

[0019] As a preferred embodiment of the method for realizing online calibration of the air-coal-water reference function according to the present invention, wherein: dividing the first data set into mutually exclusive subsets with equal proportions by using the cross-validation method includes:

[0020] Dividing the first data set into N mutually exclusive subsets with equal proportions of sample categories, where the sample categories in the N mutually exclusive subsets are the same as the category proportions in the first data set, and N is a natural number greater than or equal to 3.

[0021] As a preferred embodiment of the method for realizing online calibration of the air-coal-water reference function according to the present invention, wherein: training on the first mutually exclusive subset by using a weak learning algorithm includes:

[0022] Selecting N - 1 first mutually exclusive subsets as the training sets of N - 1 weak learning algorithms respectively to obtain N - 1 first correctors;

[0023] Multiplying the N - 1 first correctors by their corresponding weights respectively and adding the products to obtain a second corrector, where N is a natural number greater than or equal to 3.

[0024] As a preferred embodiment of the method for realizing online calibration of the air-coal-water reference function according to the present invention, wherein: optimizing the parameters of the first ensemble learning algorithm based on an optimization strategy on the second corrector includes:

[0025] The optimization operation is to select initial parameters, calculate the objective function value, and take the objective function value as the observed data;

[0026] According to the observed data and the prior probability distribution, using Bayes to update the posterior probability distribution of the objective function;

[0027] According to the posterior probability distribution, select sampling points to obtain the parameter combination that maximally improves the objective function value;

[0028] Repeat the optimization operation until the termination condition is met.

[0029] As a preferred embodiment of the method for realizing online calibration of the air-coal-water reference function according to the present invention, wherein: it further includes:

[0030] During the optimization process, record the objective function values of all sampling points, and select the parameter combination with the minimum objective function value as the optimal parameter;

[0031] Based on the optimal parameters, obtain the final calibration result of the air-coal-water reference function.

[0032] In a second aspect, the present invention provides a system for realizing online calibration of the air-coal-water reference function, including:

[0033] A partitioning module, configured to obtain historical operation data of a thermal power unit, and partition the historical operation data by using a first steady-state test algorithm to obtain a steady-state subset;

[0034] A first preprocessing module, configured to perform first preprocessing on the historical operation data by using a first anomaly detection algorithm, and combine with the steady-state subset to obtain a first data set;

[0035] A second preprocessing module, configured to partition the first data set into mutually exclusive subsets with equal proportions by using a cross-validation method, and perform second preprocessing on the mutually exclusive subsets to obtain a first mutually exclusive subset;

[0036] A training module, configured to train on the first mutually exclusive subset by using a weak learning algorithm to obtain a first corrector;

[0037] A summing module, configured to perform weighted summation on the output results of the first corrector by using a first ensemble learning algorithm to obtain a second corrector;

[0038] An optimization module, configured to optimize the parameters of the first ensemble learning algorithm based on an optimization strategy on the second corrector to obtain a correction result of a wind-coal-water benchmark function.

[0039] In a third aspect, the present invention provides a computing device, including:

[0040] A memory and a processor;

[0041] The memory is configured to store computer-executable instructions, and the processor is configured to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the implementation method for online correction of a wind-coal-water benchmark function are implemented.

[0042] In a fourth aspect, the present invention provides a computer-readable storage medium, which stores computer-executable instructions. When the computer-executable instructions are executed by a processor, the steps of the implementation method for online correction of the wind-coal-water benchmark function are implemented.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention realizes the online correction of the air-coal-water benchmark curve, can automatically correct the curve in cases such as the migration of combustion characteristics after long-term operation of the unit and the deviation of the operating coal type from the designed coal type, enables the unit to match its energy potential, maintain water-coal balance, and ensure stable combustion, improves the unit's load-changing ability, reduces the power supply coal consumption. The establishment of a high-quality dataset fully exploits the potential of the thermal power operation historical database data, can be used by similar units, and can also assist in the training of large thermal power models. In the future, it can form digital assets to increase the revenue of the power plant; the online correction of the air-coal-water benchmark curve enables the power plant to not rely on frequent experiments by the commissioning unit, saving commissioning costs and commissioning time. Introducing Bagging ensemble learning enables the model to balance robustness, generalization ability, and fitting ability, and adapt to different units and different load conditions; and through Bayesian optimization to achieve parameter self-optimization, good effects can be achieved under different units and different load conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0045] Figure 1 It is a schematic diagram of the overall process logic for the implementation method of online correction of the air-coal-water benchmark function according to an embodiment of the present invention;

[0046] Figure 2 It is the determination result of the threshold for the implementation method of online correction of the air-coal-water benchmark function according to an embodiment of the present invention;

[0047] Figure 3 It is the process of establishing a high-quality dataset for the implementation method of online correction of the air-coal-water benchmark function according to an embodiment of the present invention;

[0048] Figure 4 It is a schematic diagram of the Bagging algorithm for the implementation method of online correction of the air-coal-water benchmark function according to an embodiment of the present invention;

[0049] Figure 5 It is a schematic diagram of the load-changing situation before correction for the implementation method of online correction of the air-coal-water benchmark function according to an embodiment of the present invention;

[0050] Figure 6 It is a schematic diagram of the load-changing situation after correction for the implementation method of online correction of the air-coal-water benchmark function according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0051] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following provides a detailed description of the specific embodiments of the present invention in conjunction with the accompanying drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of the present invention.

[0052] Embodiment 1

[0053] Referring to Figure 1-4 , for an embodiment of the present invention, a method for realizing online calibration of the air-coal-water reference function is provided, including:

[0054] S100: Obtain the historical operation data of the thermal power unit, and use the first steady-state test algorithm to divide the historical operation data to obtain a steady-state subset;

[0055] S200: Use the first anomaly detection algorithm to perform the first preprocessing on the historical operation data, and combine it with the steady-state subset to obtain a first data set;

[0056] S300: Use the cross-validation method to divide the first data set into mutually exclusive subsets in equal proportions, and perform the second preprocessing on the mutually exclusive subsets to obtain a first mutually exclusive subset;

[0057] S400: Use the weak learning algorithm to train on the first mutually exclusive subset to obtain a first corrector;

[0058] S500: Use the first ensemble learning algorithm to perform weighted summation on the output results of the first corrector to obtain a second corrector;

[0059] S600: Optimize the parameters of the first ensemble learning algorithm on the second corrector based on the optimization strategy to obtain the calibration result of the air-coal-water reference function.

[0060] It should be noted that by screening the steady-state non-abnormal condition data, the quality and accuracy of the future curve can be improved. The Bagging ensemble can make the obtained strong learner take into account robustness, generalization ability, and fitting ability, and can adapt to various types of units and various load conditions, and can obtain good fitting ability. Through Bayesian optimization to achieve parameter self-optimization, not only can more optimized parameters be obtained, but also the optimal parameters can be adjusted according to different unit characteristics and different load conditions, and further liberate human resources, making it more practical.

[0061] In the embodiment of the present application, the above step S100 includes the following sub-steps A1 - A2;

[0062] In A1: For the data at a certain moment, by calculating the variance estimation value of the filtered data at the previous moment and the variance estimation value of the sampled data at the previous moment, a test index for steady-state determination is obtained.

[0063] In A2: If the test index for steady-state determination is less than or equal to the set threshold, then the data is in a steady state.

[0064] In an alternative embodiment, the first steady-state test algorithm can be a steady-state detection method based on mechanism analysis. A mechanism model is established. Based on the mechanism model, the conditions for steady state are defined. The operation data of the thermal power unit is obtained and input into the mechanism model to verify whether the output of the model conforms to the actual operation situation. If the deviation between the actual value and the theoretical value is within the preset range, it is in a steady state. However, this method is an idealized assumption premise based on understanding the mechanism characteristics of the detection object, with strong object dependence and low practicality in industrial applications.

[0065] In an alternative embodiment, the first steady-state test algorithm can be a steady-state detection method based on statistical theory, including the combined statistical test method. The operation data of the thermal power unit is obtained, and statistical indicators are calculated. For the data at each time point, mean, variance, and autocorrelation tests are respectively performed. If both the mean and variance of the load are within the set threshold, it is in a steady state. This method is widely used because it does not rely on an accurate model, has a simple process, and good accuracy.

[0066] In an alternative embodiment, the first steady-state test algorithm can also be a steady-state detection method based on trend extraction. The operation data of the thermal power unit is obtained, the time series trends of each variable are extracted and fitted to obtain a trend curve. Weights are assigned to different variables, and the weighted sum of the trend change rates of each variable is obtained to get a comprehensive trend index. It is judged whether the comprehensive trend index is less than the set threshold. If it is less, it is in a steady state. However, for the multi-variable data of thermal power, the determination of the index weight value relies on experience and lacks theoretical support, and the determination of parameters such as the detection threshold depends on historical steady-state data, reducing the reliability of the method.

[0067] In the embodiment of the present application, the first steady-state test algorithm is the R-value algorithm in the steady-state detection method based on statistical theory.

[0068] Specifically, for the data X at a certain moment i , by calculating and a test index R for steady-state determination is constructed, expressed as:

[0069] X f,i =λ1X i +(1 - λ1)X i-1

[0070]

[0071] Among them, λ1, λ2, and λ3 can take 0.2, 0.1, 0.1, X f,i is X i the filtered data, is the variance estimation value of the filtered data at the previous moment, is the variance estimation value of the sampled data at the previous moment;

[0072] When the test index R for steady-state determination is greater than the set threshold R α at this time, the data is non-steady state, otherwise, it is in a steady state. A steady-state label is marked for each data, 0 is steady-state data, and 1 is non-steady-state data;

[0073] At the significance level, when α = 0.5, R α = 1; when α = 0.25, R α = 1.15; when α = 0.1, R α = 1.32; when α = 0.05, R α = 1.44;

[0074] It should be noted that the operation data of the unit is huge and disorderly. Adopting all of them will not only affect the quality of online correction, but also greatly increase the operation computing power cost and calculation time. For the non-steady-state transient processes such as unit load change, peak shaving and frequency modulation, the numerical fluctuations of air, coal and water are large, there are many adjustments by the operation personnel, and the data is distorted and cannot truly reflect; under abnormal conditions such as unit MFT and RB, the data seriously deviates from the true value. By screening the steady-state non-abnormal condition data, the quality and accuracy of future curves can be improved.

[0075] In the embodiment of the present application, the above step S200 includes the following sub-steps B1 - B2;

[0076] In B1: Use the first anomaly detection algorithm to eliminate the abnormal conditions in the historical operation data to obtain the first data set;

[0077] Specifically, the first data set is a high-quality data set;

[0078] In an optional embodiment, the first anomaly detection algorithm can be the Isolation Forest algorithm. Randomly select features and data points, divide the data into two parts, recursively repeat the process for each part until all data points are isolated. For each data point, calculate the average path length in all the isolated numbers, calculate the anomaly score, set a threshold according to the distribution of the anomaly scores, and the data points with scores higher than the threshold are anomaly points. Eliminate the anomaly points to obtain a high-quality normal operation data set;

[0079] In an alternative embodiment, the first anomaly detection algorithm may be the random outlier selection algorithm. For each data point in the dataset, calculate its similarity to all other points to obtain a similarity matrix. Based on the similarity matrix, construct a similarity graph. Randomly select an initial point and propagate the anomaly score in the similarity graph through random walks. According to the distribution of the anomaly scores, set a threshold, and the data points with scores higher than the threshold are considered outliers. Remove the outliers from the steady-state dataset to obtain a high-quality normal operating dataset;

[0080] In an alternative embodiment, the first anomaly detection algorithm may also be the 3σ statistical test method. For each feature variable, calculate its mean and standard deviation. For each data point, determine whether it meets the condition of being a normal point. If it does not meet the condition, it is an outlier, and all outliers are removed from the steady-state dataset to obtain a high-quality normal operating dataset.

[0081] In the embodiment of the present application, the first anomaly detection algorithm is the LOF algorithm;

[0082] Specifically, for each data point, calculate its distance to all other points and sort them from near to far. Assume data point P, and the distance between the k-th point Q closest to P in terms of Euclidean distance and P is denoted as the K-nearest distance of P, expressed as: d k (P) = d(P, Q);

[0083] Taking point P as the center and the K-nearest distance d k (P) as the radius to draw a circle, the points within the circle are the K-nearest neighbors of P, expressed as:

[0084] N k (P) = {d(P, O ′ ) ≤ d k (P)}

[0085] Define the K-reachability distance between two data points P and O as the maximum value between the distance between P and O and the distance between P and d k (O), expressed as:

[0086] reach_dist k (O, P) = max{d k (O), d(O, P)}

[0087] The higher the value of K, the more similar the reachability distances of the points within the same neighborhood;

[0088] For each data point, find its K nearest neighbors and calculate the LOF score. The local reachability density lrd of a data point is the reciprocal of the average reachability distance based on its nearest neighbors. The larger the distance, the smaller the density. The LOF (Local Outlier Factor) of a point is the ratio of the average local reachability density of points in the point's neighborhood to the local reachability density of the data point, expressed as:

[0089]

[0090] Judge whether it is an outlier according to the LOF value, as Figure 2 shown. The threshold can be 1.3. If the LOF value is less than 1.3, it is normal; if the LOF value is greater than 1.3, it is abnormal;

[0091] Label each data as a normal label. 0 is normal data and 1 is abnormal data;

[0092] As Figure 3 shown, filter out those with both the steady-state label and the normal label being 0 as the high-quality data set;

[0093] The data in the high-quality data set takes into account 8 characteristics of the unit, including load, feed water flow rate, total air volume, total coal quantity, main steam temperature, main steam pressure, Nox concentration, and Sox concentration.

[0094] It should be noted that the LOF algorithm can effectively identify and remove abnormal data points by calculating the local outlier factor of data points, thereby screening out a high-quality data set. The high-quality data set is helpful for subsequent benchmark function correction and model training, reduces the impact of abnormal data on the model, and enhances the robustness and reliability of the model.

[0095] In the embodiment of the present application, the above step S300 includes the following sub-step C1;

[0096] In C1: Divide the first data set into N mutually exclusive subsets with equal proportions of sample categories. The sample categories in the N mutually exclusive subsets are the same as the category proportions in the first data set, where N is a natural number greater than or equal to 3;

[0097] In the embodiment of the present application, when N takes 4, divide the high-quality data set into 4 mutually exclusive subsets D N 1 - 4 with equal proportions of sample categories according to the Stratified KFold cross-validation method for the second preprocessing;

[0098] The second preprocessing includes normalization and missing value filling;

[0099] Specifically, normalization changes a column of features to a certain fixed interval range. The interval range can be a decimal between [0, 1] or (-1, 1), and maps the data to the range of 0 to 1 for processing, expressed as:

[0100]

[0101] Among them, X i is the i-th data in a feature, X min is the minimum value in this feature, X max is the maximum value in this feature, x i is the data after normalizing X i ;

[0102] Missing values refer to the situation where the values of some variables in the dataset are missing. This will cause the system to lose a large amount of useful information, making the uncertainty shown in the system more significant and making it more difficult to grasp the deterministic components contained in the system; data containing null values will make the data mining process fall into chaos, resulting in unreliable outputs.

[0103] The nearest neighbor imputation method is adopted, which is expressed as:

[0104]

[0105] Among them, X i is the i-th data in a feature and is a missing value, X i-1 is the (i - 1)-th value, X i+1 is the (i + 1)-th value, x i is the value after imputing X i ;

[0106] It should be noted that through hierarchical partitioning, it is ensured that the sample distribution of each subset is consistent with the original dataset, avoiding model bias caused by uneven data partitioning. This partitioning method can better simulate the real scenario and improve the generalization ability of the model; the feature values are scaled to a unified interval to avoid some features having a dominant effect on model training due to differences in dimension or numerical range. After normalization, the model converges faster and the training efficiency is higher.

[0107] In the embodiments of the present application, the above steps S400 and S500 include the following sub-steps D1 - D2;

[0108] In D1: Select N - 1 first mutually exclusive subsets as the training sets of N - 1 weak learning algorithms respectively to obtain N - 1 first correctors;

[0109] In D2: Multiply the N - 1 first correctors by the corresponding weights respectively and add the products to obtain a second corrector, where N is a natural number greater than or equal to 3.

[0110] In an alternative embodiment, the weak learning algorithm can be a decision tree, which belongs to a tree-based model and divides data into different categories or predicts target values through a series of rules; the weak learning algorithm can be a random forest, which is an ensemble learning method based on decision trees. By constructing multiple decision trees and voting or averaging, it can improve the prediction performance, handle high-dimensional data, and has strong robustness to noise and outliers.

[0111] In an alternative embodiment, the weak learning algorithm can also be the K-nearest neighbor algorithm, which predicts the category or value of the target point by finding the K nearest points to the target point in the training set and is used in scenarios where the sample distribution in the operating data of thermal power units is relatively uniform and the local similarity is obvious.

[0112] In the embodiment of the present application, the weak learning algorithm includes three weak learning algorithms: BP neural network, SVM, and OLD.

[0113] Specifically, the first corrector is a weak corrector, and the second corrector is a strong corrector.

[0114] As Figure 4 shown, the strong corrector formed by Bagging integration is a weighted sum of the outputs of the weak corrector D N 1-3, which is expressed as:

[0115] C N 4 = aC N 1 + bC N 2 + cC N 3

[0116] Among them, a, b, and c are the corresponding weight parameters, C N 1 is the weak corrector obtained by training the BP neural network, C N 2 is the weak corrector obtained by training SVM, C N 3 is the weak corrector obtained by training OLD.

[0117] It should be noted that the three weak learning algorithms of BP neural network, SVM, and OLD are preferably selected. The BP neural network has strong ability to process a large amount of data, adapts to the massive characteristics of the data of coal-fired units, and makes up for the insufficient processing performance of the SVM algorithm for a large amount of data; the BP neural network has strong learning and reasoning abilities, and its prediction accuracy is the best.

[0118] The SVM algorithm can effectively process high-dimensional data sets, adapt to the multi-dimensional characteristics of the data of coal-fired units; has strong ability to process non-linear features, adapts to the strong non-linear characteristics of the data of coal-fired units; and compared with the above BP neural network, SVM is not easily trapped in local minima and can make up for its disadvantage of poor generalization ability.

[0119] The principle of the OLD method is the simplest and most rigorous, with high computational efficiency and strong stability. It can make up for the problem of large computational resource occupation of the above two algorithms and ensure the convergence of the results.

[0120] Different weak learners have different characteristics respectively. Some have good robustness, some have strong generalization ability, and some have strong fitting ability. Through Bagging integration, the obtained strong learner can take into account robustness, generalization ability and fitting ability, and can adapt to various types of units and various load conditions, and can obtain good fitting ability.

[0121] In the embodiment of the present application, the above step S600 includes the following sub-steps E1-E6;

[0122] In E1: The optimization operation is to select initial parameters, calculate the objective function value, and use the objective function value as the observed data;

[0123] In E2: According to the observed data and the prior probability distribution, use Bayes to update the posterior probability distribution of the objective function;

[0124] In E3: According to the posterior probability distribution, select sampling points to make the parameter combination with the largest increase in the objective function value;

[0125] In E4: Repeat the optimization operation until the termination condition is met.

[0126] In E5: During the optimization process, record the objective function values of all sampling points, and select the parameter combination with the smallest objective function value as the optimal parameter;

[0127] In E6: Obtain the final correction result of the air-coal-water reference function based on the optimal parameters.

[0128] Specifically, the objective function f is evaluated based on Bayes' formula to find the next set of parameters that can maximize the global improvement, so as to quickly approach the optimal solution and find the best hyperparameter combination. Bayes' formula is expressed as:

[0129]

[0130] Among them, f is the objective function, D is the observed set, p(f|D) is the posterior probability distribution of f, p(f) is the prior probability distribution of f, that is, the assumption of the state of the unknown objective function, and p(D|f) is the marginal likelihood distribution of f;

[0131] f is used as the objective function and is expressed as:

[0132] minf=∑(y i -y prediction )

[0133] Among them, y i is CN the data in 4, y prediction is the output obtained by the strong corrector;

[0134] Taking p(f) as the prior distribution and using the Gaussian distribution as the initialization probability, it is expressed as:

[0135]

[0136] where σ is the data standard deviation and μ is the data mean;

[0137] It should be noted that the adjustment and optimization of the weights often need to be continuously adjusted. By realizing parameter self-optimization through Bayesian optimization, not only can more optimized parameters be obtained, but the optimal parameters can be adjusted according to the characteristics of different units and different load conditions. Moreover, it further liberates human resources and makes it more practical.

[0138] The above is a schematic solution of a method for online correction of the wind-coal-water benchmark function in this embodiment. It should be noted that the technical solution of the system for online correction of the wind-coal-water benchmark function belongs to the same concept as the technical solution of the above method for online correction of the wind-coal-water benchmark function. For the details not described in detail in the technical solution of the system for online correction of the wind-coal-water benchmark function in this embodiment, reference can be made to the description of the technical solution of the above method for online correction of the wind-coal-water benchmark function.

[0139] The system for online correction of the wind-coal-water benchmark function in this embodiment includes:

[0140] A partitioning module, configured to obtain the historical operation data of the thermal power unit, and partition the historical operation data by using the first steady-state test algorithm to obtain a steady-state subset;

[0141] A first preprocessing module, configured to perform first preprocessing on the historical operation data by using the first anomaly detection algorithm and combine it with the steady-state subset to obtain a first data set;

[0142] A second preprocessing module, configured to partition the first data set into mutually exclusive subsets with equal proportions by using the cross-validation method, and perform second preprocessing on the mutually exclusive subsets to obtain a first mutually exclusive subset;

[0143] A training module, configured to train on the first mutually exclusive subset by using a weak learning algorithm to obtain a first corrector;

[0144] A summation module, configured to perform weighted summation on the output results of the first corrector by using the first ensemble learning algorithm to obtain a second corrector;

[0145] An optimization module, configured to optimize the parameters of the first ensemble learning algorithm on the second corrector based on an optimization strategy to obtain the correction result of the wind-coal-water benchmark function.

[0146] This embodiment also provides a computing device, which is applicable to the implementation of online calibration of the air-coal-water reference function, and includes:

[0147] A memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the method for online calibration of the air-coal-water reference function as proposed in the above embodiment.

[0148] This embodiment also provides a storage medium, on which a computer program is stored. When the program is executed by a processor, it implements the method for online calibration of the air-coal-water reference function as proposed in the above embodiment.

[0149] The storage medium proposed in this embodiment and the method for online calibration of the air-coal-water reference function proposed in the above embodiment belong to the same inventive concept. For the technical details not described in detail in this embodiment, reference can be made to the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0150] From the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software and necessary general-purpose hardware, and of course, it can also be implemented by hardware. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a floppy disk, read-only memory (ROM), random access memory (RAM), flash memory (FLASH), hard disk, or optical disc of a computer, etc., including several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of the various embodiments of the present invention.

[0151] Embodiment 2

[0152] Referring to Table 1-2, Figure 5-6 , this embodiment is different from the first embodiment and provides a verification test for the method of online calibration of the air-coal-water reference function to verify and illustrate the technical effects adopted in this method.

[0153] An adjustment was made in a 1000MW supercritical coal-fired unit in December 2023, as shown in Table 1-2;

[0154] Table 1 Reference function of original coal feeding amount (t / h) to feed water flow and total air volume (t / h)

[0155] coal 0 160 195 230 288 325 360 395 410 420 450 water 700 850 1380 1700 2160 2365 2600 2780 2860 2900 2900 wind 1000 1600 1800 2000 2300 2500 2680 2860 2940 3000 3000

[0156] Table 2 Benchmark function of coal feeding rate (t / h) to feed water flow rate and total air volume (t / h) after online calibration

[0157] coal 0 150 169 202 252 286 320 355 373 408 430 water 700 1027 1190 1473 1902 2194 2486 2786 2941 3241 3241 wind 1000 1768 1895 2110 2425 2635 2840 3040 3140 3400 3400

[0158] As Figure 5 , Figure 6 shown, the online calibration method of the air-coal-water benchmark function based on ensemble learning in this application has achieved remarkable results in optimizing the operating parameters of the unit.

[0159] The calibrated benchmark function can not only effectively reduce the deviation of main steam temperature and main steam pressure, improve the economy and flexibility of the unit, but also enhance the operation stability, reduce equipment wear, improve the environmental protection performance, and promote the transformation of thermal power units to an intelligent operation management mode.

Claims

1. A method for realizing online correction of wind, coal and water reference functions, characterized in that: include: Acquire historical operation data of the thermal power unit, and divide the historical operation data by using a first steady-state test algorithm to obtain steady-state subsets; Performing a first preprocessing on the historical operation data using a first anomaly detection algorithm, and combining the steady-state subset to obtain a first data set; Dividing the first data set into mutually exclusive subsets of equal proportions by using a cross-validation method, performing a second preprocessing on the mutually exclusive subsets to obtain first mutually exclusive subsets; Use a weak learning algorithm to train on the first mutually exclusive subset to obtain a first corrector; Using a first ensemble learning algorithm to perform weighted summation on output results of the first corrector to obtain a second corrector; The parameters of the first integrated learning algorithm are optimized on the second corrector based on the optimization strategy to obtain the correction result of the wind, coal and water benchmark functions.

2. The method for realizing online correction of wind, coal and water reference functions according to claim 1, characterized in that: Using the first steady-state inspection algorithm to divide the historical operation data includes: For the data at a certain moment, the test index for steady-state determination is obtained by calculating the variance estimate of the filtered data at the previous moment and the variance estimate of the sampled data at the previous moment; If the inspection index for steady-state judgment is less than or equal to the set threshold, the data is steady-state.

3. The method for realizing online correction of wind, coal and water reference functions according to claim 1 or 2, characterized in that: Performing a first preprocessing on the historical operation data using a first anomaly detection algorithm includes: The first anomaly detection algorithm is used to eliminate abnormal operating conditions in the historical operating data to obtain a first data set.

4. The method for realizing online correction of wind, coal and water reference functions according to claim 3, characterized in that: Using a cross-validation method to divide the first data set into mutually exclusive subsets of equal proportions includes: The first data set is divided into N mutually exclusive subsets with equal proportions of sample categories, and the sample categories in the N mutually exclusive subsets have the same proportion as the categories in the first data set, where N is a natural number greater than or equal to 3.

5. The method for realizing online correction of wind, coal and water reference functions according to claim 4, characterized in that: Training on the first mutually exclusive subset using a weak learning algorithm includes: Select N-1 first mutually exclusive subsets as training sets for N-1 weak learning algorithms, and obtain N-1 first correctors; Multiply N-1 first correctors by corresponding weights respectively, and add the products to obtain a second corrector, where N is a natural number greater than or equal to 3.

6. The method for realizing online correction of wind, coal and water reference functions according to claim 5, characterized in that: Optimizing the first ensemble learning algorithm parameters based on the optimization strategy on the second corrector includes: The optimization operation is to select the initial parameters, calculate the objective function value, and use the objective function value as the observed data; Based on the observed data and the prior probability distribution, the posterior probability distribution of the objective function is updated using Bayesian. According to the posterior probability distribution, select sampling points to maximize the parameter combination that improves the objective function value; The optimization operation is repeated until the termination condition is met.

7. The method for realizing online correction of wind, coal and water reference functions according to claim 6, characterized in that: Also includes: During the optimization process, the objective function values ​​of all sampling points are recorded, and the parameter combination with the smallest objective function value is selected as the optimal parameter; The final wind, coal and water benchmark function correction results are obtained based on the optimal parameters.

8. A system using the method for implementing online correction of wind, coal and water reference functions as described in any one of claims 1 to 7, characterized in that: include: A partitioning module, used for acquiring historical operation data of the thermal power unit, and partitioning the historical operation data using a first steady-state test algorithm to obtain a steady-state subset; A first preprocessing module, configured to perform a first preprocessing on the historical operation data using a first anomaly detection algorithm, and combine the steady-state subset to obtain a first data set; A second preprocessing module, used for dividing the first data set into mutually exclusive subsets of equal proportions by using a cross-validation method, and performing a second preprocessing on the mutually exclusive subsets to obtain first mutually exclusive subsets; A training module, used for training on the first mutually exclusive subset using a weak learning algorithm to obtain a first corrector; A summing module, used for performing weighted summation on the output results of the first corrector by using a first ensemble learning algorithm to obtain a second corrector; An optimization module is used to optimize the parameters of the first integrated learning algorithm based on the optimization strategy on the second corrector to obtain the correction result of the wind, coal and water benchmark function.

9. An electronic device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the method for online correction of the wind, coal and water reference functions as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the method for online correction of the wind, coal and water reference functions as described in any one of claims 1 to 7.

Images