A training method of a net power prediction model for a power plant generator set

Abstract

The present application relates to a kind of training methods for the net power prediction model of power plant generator unit, belong to the optimization control technical field of thermal power plant, solve the problem of poor prediction accuracy of net power prediction model in prior art.The training method includes: based on preset clustering algorithm, according to the third operating parameter of each sample in steady-state operation sample total set, steady-state operation sample total set is divided into multiple steady-state operation sample subsets by clustering;Wherein, each steady-state operation sample subset corresponds to a working condition type;According to the fourth operating parameter of each sample in each steady-state operation sample subset, the BP neural network model of preset structure is trained, and the net power prediction model corresponding to each working condition type is obtained;Wherein, fourth operating parameter includes back pressure and net power.Improve the prediction accuracy of net power prediction model.

Description

Technical Field

[0001] This invention relates to the field of optimization control technology for thermal power plants, and in particular to a training method for a net power prediction model of power plant generator sets. Background Technology

[0002] In applications such as operation optimization, performance analysis, and fault diagnosis of thermal power plants, it is necessary to establish a net power prediction model based on historical data.

[0003] However, in actual operation, generator sets experience numerous transient processes such as start-up and shutdown, load changes, equipment debugging, and operating condition switching, as well as situations like sensor failures, abnormal data transmission, and abnormal equipment operation. These non-steady-state and abnormal data can severely affect the accuracy of the net power prediction model.

[0004] Meanwhile, the operating parameters used for each time step in training the net power prediction model in traditional methods are unreasonable, resulting in a large discrepancy between the predicted net power and the actual net power, making it difficult to meet the needs of practical applications.

[0005] Therefore, there is an urgent need for a technical solution to train a net power prediction model for power plant generator units. Summary of the Invention

[0006] Based on the above analysis, the present invention aims to provide a training method for a net power prediction model for power plant generator sets, in order to solve the problem of poor prediction accuracy of net power prediction models in the prior art.

[0007] This invention provides a training method for a net power prediction model of power plant generator sets, the training method comprising:

[0008] Based on a pre-defined clustering algorithm, the total set of steady-state operating samples is divided into multiple subsets based on the third operating parameter of each sample in the total set of steady-state operating samples; each subset of steady-state operating samples corresponds to a type of operating condition. The BP neural network model with a preset structure is trained based on the fourth operating parameter of each sample in each steady-state operating sample subset to obtain the net power prediction model corresponding to each operating condition type; wherein, the fourth operating parameter includes back pressure and net power.

[0009] Based on further improvements to the above training method, the third running parameter includes one or more of the following: Reheat steam pressure; Reheat steam temperature; Main steam flow rate; Main steam temperature; Main steam pressure; Ambient temperature; Generator active power; Active power on the high-voltage side of the excitation transformer; Active power on the high-voltage side; Back pressure.

[0010] Based on further improvements to the above training method, the fourth running parameter may also include one or more of the following: Reheat steam pressure; Reheat steam temperature; Main steam flow rate; Main steam temperature; Main steam pressure; Ambient temperature.

[0011] Based on further improvements to the above training method, the BP neural network model with the preset structure includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output layer connected in sequence. The first hidden layer contains 128 neurons, the second hidden layer contains 64 neurons, and the third hidden layer contains 32 neurons.

[0012] Based on further improvements to the above training method, the steps for obtaining the steady-state operating sample set include: Collect a large amount of continuous operating data of the generator set at a preset historical time period; Based on the steady-state thresholds corresponding to multiple pre-set first operating parameters in multiple value ranges, determine whether the operating data at each time point belongs to steady-state data, and obtain the judgment result of the operating data at each time point corresponding to each first operating parameter; The running data at the time point where the judgment result corresponding to all first running parameters belongs to steady-state data is taken as the first steady-state running data, and the running data at all time points in the first steady-state running data is taken as the sample of the total steady-state running sample set.

[0013] Based on a further improvement to the above training method, the step of using the running data at the time point where the judgment results corresponding to all first running parameters belong to steady-state data as the first steady-state running data includes: The theoretical main steam pressure at each time point is calculated based on the main steam temperature at each time point in the first steady-state operating data. The steam pressure difference between the theoretical steam pressure and the main steam pressure at each time point is determined. Determine whether the steam pressure difference at each time point exceeds the steam pressure threshold corresponding to the main steam temperature at each time point; if the steam pressure difference at a time point exceeds the steam pressure threshold corresponding to the main steam temperature at that time point, then delete the operating data of that time point from the first steady-state operating data.

[0014] Based on a further improvement to the above training method, the step of determining whether the running data at each time point belongs to steady-state data based on the steady-state thresholds corresponding to multiple preset first running parameters in multiple value ranges, and obtaining the judgment result of the running data at each time point for each first running parameter, includes: Obtain the running data corresponding to each first running parameter for each time point and a preset number of consecutive time points before that time point, and use it as the first data; The difference between the maximum and minimum first operating parameter values ​​in the first data is calculated and used as the second data. By using the steady-state threshold corresponding to multiple value ranges of the first operating parameter, the second data, and the operating data at the time point corresponding to the first operating parameter, it is determined whether the operating data at the time point is steady-state data when corresponding to the first operating parameter.

[0015] Based on a further improvement to the above training method, the step of using a pre-set steady-state threshold corresponding to multiple value intervals of the first operating parameter, second data, and operating data at the time point corresponding to the first operating parameter to determine whether the operating data at the time point belongs to steady-state data when corresponding to the first operating parameter includes: Determine the value range of the running data corresponding to the first running parameter at this time point, and use it as the filter value range; Determine whether the second data is within the range of the steady-state threshold corresponding to the selected value interval; if the second data is within the range of the steady-state threshold corresponding to the selected value interval, then the running data at that time point belongs to the steady-state data under the first running parameter.

[0016] Based on the further improvement of the above training method, the step of obtaining the steady-state operating sample set also includes: Based on the preset statistical threshold and the running data of all time points in the first steady-state running data corresponding to multiple second running parameters, it is determined whether the running data of each time point in the first steady-state running data belongs to steady-state data at each second running parameter, and the judgment result of the running data of each time point in the first steady-state running data corresponding to each second running parameter is obtained. The running data at the time point where the judgment result corresponding to all second running parameters belongs to the steady-state data is taken as the second steady-state running data, and the running data at all time points in the second steady-state running data is taken as the sample of the steady-state running sample set.

[0017] Based on the further improvement of the above training method, the step of obtaining the steady-state operating sample set also includes: Input the running data corresponding to all second running parameters for all time points in the second steady-state running data into the pre-constructed random tree ensemble model to obtain the average path length of each time point in the second steady-state running data in the random tree; The average path lengths of all time points in the random tree in the second steady-state operating data are sorted in descending order to obtain the sorted second steady-state operating data. From the sorted second steady-state operating data, the operating data at the time points with the preset abnormal pollution ratio are deleted in descending order to obtain the third steady-state operating data. At the same time, the operating data at all time points in the third steady-state operating data are used as samples of the total steady-state operating sample set.

[0018] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: 1. Based on the third operating parameters of each sample in the total steady-state operating sample set, the total steady-state operating sample set is divided into multiple steady-state operating sample subsets, and each steady-state operating sample subset corresponds to a type of operating condition. The fourth operating parameters of each sample in each steady-state operating sample subset are used to train the BP neural network model with a preset structure to obtain the net power prediction model corresponding to each type of operating condition, which greatly improves the prediction accuracy of the net power prediction model for net power. 2. By using the steady-state thresholds corresponding to multiple value ranges for each first operating parameter in advance, the steady-state characteristics of the generator set under different operating conditions are fully considered, and steady-state data is extracted efficiently. At the same time, the operating data at the time point when the judgment result corresponding to all first operating parameters belongs to steady-state data is used as the extracted steady-state operating data, which improves the data quality of the total steady-state operating sample set and provides a foundation for training a net power prediction model with higher accuracy. 3. Based on the preset statistical threshold and the running data corresponding to multiple second running parameters at all time points in the first steady-state running data, the first steady-state running data is further filtered and extracted, and outliers in the first steady-state running data are removed, which further improves the data quality of the total steady-state running sample set and provides a foundation for training a net power prediction model with higher accuracy. 4. By using a pre-built random tree ensemble model to filter and extract the second steady-state operating data, data that violates the laws of physical parameters is removed, further improving the data quality of the total steady-state operating sample set and providing a foundation for training a net power prediction model with higher accuracy.

[0019] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0020] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0021] Figure 1 A flowchart illustrating a training method for a net power prediction model of a power plant generator set, provided in an embodiment of the present invention. Figure 2 A flowchart illustrating the steps for obtaining the steady-state operation sample set provided in an embodiment of the present invention; Figure 3 This is a schematic diagram showing the steady-state thresholds corresponding to different value ranges of unit load provided in an embodiment of the present invention. Detailed Implementation

[0022] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0023] A specific embodiment of the present invention discloses a training method for a net power prediction model of power plant generator sets, such as... Figure 1 As shown, the training method includes: Step S1: Based on the preset clustering algorithm, the total set of steady-state operating samples is divided into multiple steady-state operating sample subsets according to the third operating parameter of each sample in the total set of steady-state operating samples; wherein, each steady-state operating sample subset corresponds to a type of operating condition; Step S2: Train the BP neural network model with a preset structure based on the fourth operating parameters of each sample in each steady-state operating sample subset to obtain the net power prediction model corresponding to each operating condition type; wherein, the fourth operating parameters include back pressure and net power.

[0024] Specifically, traditional technologies have the following shortcomings in identifying and handling outliers in generator set steady-state operation data: 1. Steady-state identification methods are limited and have low accuracy: Traditional methods use fixed thresholds for determination, such as setting "load fluctuation < 5MW" as a steady state. However, the dynamic characteristics of generator sets vary significantly under different load conditions. A 5MW fluctuation at a 300MW load may be a normal steady state, but a 5MW fluctuation at a 100MW load is already a significant load change. Fixed thresholds cannot adapt to the multi-load conditions of generator sets.

[0025] Meanwhile, traditional methods often rely on a single parameter to determine steady state, without considering the multi-parameter coupling characteristics of the steam turbine thermodynamic system. This can easily lead to misjudging non-steady-state processes such as control valve operation and steam temperature and pressure fluctuations as steady states.

[0026] Furthermore, there is a lack of verification of the consistency of parameter change trends. For example, the main steam temperature and pressure should change synchronously. If only the temperature fluctuates while the pressure remains stable, it is often due to interfering factors such as desuperheating water regulation, but traditional methods cannot identify such situations as steady state.

[0027] 2. Outlier detection techniques are simplistic and have a high omission rate: Traditional methods, which employ statistical approaches such as the 3σ criterion, can only identify extreme outliers of a single parameter and cannot identify complex anomalies involving multiple parameters, such as the overall shift of multiple parameters caused by sensor drift.

[0028] Meanwhile, traditional methods are mostly offline batch processing, lacking a hierarchical detection mechanism tailored to the characteristics of power plant operation data. This results in a mixture of different types of anomalies, such as sensor failures, equipment malfunctions, and communication interruptions, making it difficult to accurately eliminate them.

[0029] 3. The processing logic for steady-state identification and anomaly detection is missing: Traditional techniques typically separate steady-state determination based on physical thresholds from anomaly detection based on statistical regularities into isolated steps. Data cleaning lacks process-specificity, often leading to "false deletion" and "false omission". The method provided in this invention has a process of two-way verification of physical and data, ensuring the rationality and purity of the data.

[0030] Specifically, such as Figure 1 As shown, in step S1, a clustering algorithm is preset and a total set of steady-state operating samples is obtained. Based on the preset clustering algorithm, the third operating parameters of each sample in the total set of steady-state operating samples are used for clustering, and all samples in the total set of steady-state operating samples are divided into multiple steady-state operating sample subsets. Each steady-state operating sample subset corresponds to a type of operating condition.

[0031] It is worth noting that when clustering each sample in the total set of steady-state operating samples, the third operating parameter of each sample in the total set of steady-state operating samples is standardized to determine the mean and standard deviation of each third operating parameter, so that each third operating parameter of each sample is normalized to a standard distribution with a mean of 0 and a standard deviation of 1, thereby eliminating the influence of dimensional differences on the clustering results.

[0032] Preferably, the third operating parameter includes one or more of the following: Reheat steam pressure; Reheat steam temperature; Main steam flow rate; Main steam temperature; Main steam pressure; Ambient temperature; Generator active power; Active power on the high-voltage side of the excitation transformer; Active power on the high-voltage side; Back pressure.

[0033] Specifically, reheat steam pressure, reheat steam temperature, main steam flow rate, main steam temperature, and main steam pressure are used to characterize the turbine's thermodynamic parameters, ambient temperature is used to characterize the external environmental parameters of the cooling system, and generator active power, excitation transformer high-voltage side active power, and high-voltage side active power are used to characterize load and power consumption parameters.

[0034] Understandably, the back pressure of an air-cooled unit is affected by multiple factors, including fan speed, exhibiting strong nonlinear and time-varying characteristics. Excessive back pressure reduces the turbine expansion ratio, decreases turbine work, and leads to a decrease in power generation. While excessively low back pressure is beneficial for turbine work, it requires increasing the cooling fan speed, increasing fan power consumption, and ultimately reducing the unit's net power generation.

[0035] Specifically, such as Figure 1 As shown, in step S1, the set of steady-state operating samples is clustered by the coupling characteristics of multiple parameters in the third operating parameters.

[0036] Preferably, the pre-set clustering algorithm is the K-Means clustering algorithm.

[0037] Specifically, the K-Means clustering algorithm is used to divide all samples in the steady-state sample set into clusters with a number of clusters from K=1 to K=10, and the within-cluster sum of squares (WCSS) is calculated when the number of clusters is from K=1 to K=10.

[0038] Specifically, after obtaining the WCSS corresponding to each cluster size, the WCSS reduction rate for different cluster size K values ​​is calculated using the following formula: ; in, Indicates the number of clusters. WCSS decline rate Indicates the number of clusters. WCSS at that time.

[0039] Specifically, the WCSS decrease rate corresponding to the number of all clusters is compared, and the number of clusters corresponding to the maximum WCSS decrease rate is selected as the number of operating condition types. For example, if the number of clusters corresponding to the maximum WCSS decrease rate is 5, then 5 is selected as the number of operating condition types, that is, the total set of steady-state operating samples is divided into 5 steady-state operating sample subsets, and each steady-state operating sample subset corresponds to one operating condition type.

[0040] It is worth noting that the cluster center corresponding to this type of operating condition can be calculated based on the third operating parameter of each sample in each steady-state operating sample subset.

[0041] Specifically, such as Figure 1 As shown, in step S1, it is also necessary to determine the total set of steady-state operating samples.

[0042] Preferably, the step of obtaining the steady-state operating sample set includes: Step S51: Collect a large number of continuous time points of operation data of the generator set within a preset historical time period; Step S52: Based on the steady-state thresholds corresponding to multiple preset first operating parameters in multiple value ranges, determine whether the operating data at each time point belongs to steady-state data, and obtain the judgment result of the operating data at each time point corresponding to each first operating parameter; Step S53: Take the running data at the time point where the judgment result corresponding to all first running parameters belongs to steady-state data as the first steady-state running data, and take the running data at all time points in the first steady-state running data as samples of the steady-state running sample set.

[0043] Specifically, such as Figure 2 As shown, high-quality steady-state data can be extracted from the massive historical operating data of power plant generator sets, which can be used to provide a reliable data foundation for the performance analysis of generator sets and the optimization of power generation control models.

[0044] Specifically, such as Figure 2 As shown, in step S51, when implementing the training method for a net power prediction model of a power plant generator set provided in this embodiment of the invention, a large number of continuous time points of operation data of the generator set within a preset historical time period before the current time point can be selected.

[0045] Specifically, the preset historical time period can be selected as 2 years.

[0046] It is understandable that within a preset historical time period, the data recorded at each point in time includes a large amount of operational data corresponding to various operational parameters, and the time difference between adjacent points in time is fixed.

[0047] Specifically, such as Figure 2 As shown, in step S51, a large number of continuous time points of the generator set are collected within a preset historical time period, that is, the time difference between two adjacent time points is the same.

[0048] Specifically, such as Figure 2 As shown, in step S52, multiple first operating parameters are preset, and a steady-state threshold corresponding to each first operating parameter in different value ranges is set. For example, unit load is used as a first operating parameter, and steady-state thresholds corresponding to unit load in different value ranges are set, such as... Figure 3 As shown, the load range of the generating unit is divided into 0-300MW, 300-450MW, 450-550MW, 550-600MW and 600-800MW. At the same time, steady-state thresholds for the load of the generating unit are set in different ranges. For example, the steady-state threshold is 15MW for 0-300MW, 12MW for 300-450MW, 8MW for 450-550MW, 5MW for 550-600MW and 3MW for 600-800MW.

[0049] Specifically, such as Figure 2 As shown, in step S52, preferably, the plurality of first operating parameters include: Unit load or active power; Main steam temperature; Main steam pressure.

[0050] Specifically, either the unit load or the active power can be selected, while the main steam temperature and main steam pressure are mandatory options as multiple primary operating parameters.

[0051] It is understandable that the main steam temperature and main steam pressure are used as initial parameters to characterize the state of the working medium at the turbine inlet. As key parameters of the generator set, they are selected as the first operating parameters required for judging steady-state data.

[0052] It is worth noting that by combining the three parameters of unit load or active power, main steam temperature and main steam pressure, the multi-parameter coupling characteristics of the turbine thermodynamic system are fully considered. Compared with the single-parameter determination method, the steady-state identification accuracy is effectively improved.

[0053] Specifically, such as Figure 2As shown, in step S52, the steady-state thresholds corresponding to multiple first operating parameters in multiple value ranges are determined, and the operating data at each time point is judged to determine whether the operating data at each time point belongs to steady-state data.

[0054] It is worth noting that in step S52, the steady-state threshold corresponding to multiple value ranges of a first operating parameter is used to judge the operating data at each time point, thereby determining the judgment result corresponding to the first operating parameter for each time point. When there are multiple first operating parameters, the operating data at each time point corresponds to multiple judgment results among the multiple first operating parameters.

[0055] For example, multiple first operating parameters are unit load, main steam temperature, and main steam pressure. If the operating data at a certain time point is determined to be steady-state data based on the unit load, it does not prevent the operating data at that time point from being determined to be non-steady-state data based on the main steam temperature. In other words, the determination of whether the operating data at the same time point is steady-state data based on different first operating parameters does not interfere with each other and is independent of each other.

[0056] Specifically, such as Figure 2 As shown, by executing step S52, the judgment result corresponding to each first operating parameter can be obtained for the operating data at each time point. For example, when the multiple first operating parameters are unit load, main steam temperature, and main steam pressure, the judgment results for the operating data at each time point in terms of unit load, main steam temperature, and main steam pressure can be obtained respectively.

[0057] Preferably, the step of determining whether the operating data at each time point belongs to steady-state data based on a plurality of pre-set first operating parameters corresponding to steady-state thresholds in multiple value ranges, and obtaining the determination result of the operating data at each time point for each first operating parameter, includes: Obtain the running data corresponding to each first running parameter for each time point and a preset number of consecutive time points before that time point, and use it as the first data; The difference between the maximum and minimum first operating parameter values ​​in the first data is calculated and used as the second data. By using the steady-state threshold corresponding to multiple value ranges of the first operating parameter, the second data, and the operating data at the time point corresponding to the first operating parameter, it is determined whether the operating data at the time point is steady-state data when corresponding to the first operating parameter.

[0058] Specifically, the preset quantity is generally set to 20-30. When determining the operating data at the current time point in relation to any first operating parameter, the operating data corresponding to any first operating parameter at the current time point and a preset number of consecutive time points before the current time point are selected. For example, if the preset quantity is 25 and the first operating parameter is the unit load, then the operating data corresponding to the unit load at the current time point and 25 time points before the current time point are selected as the first data.

[0059] Specifically, the maximum and minimum values, i.e. the maximum unit load value and the minimum unit load value, are statistically determined from the first data, and the difference between the maximum unit load value and the minimum unit load value is used as the second data.

[0060] Specifically, in combination Figure 3 As shown, based on the second data and the operating data corresponding to the unit load at the current time point, the steady-state threshold corresponding to the unit load in multiple value ranges is used to determine whether the operating data at the current time point belongs to steady-state data when the unit load is in place.

[0061] Preferably, the step of determining whether the operating data at a given time point is steady-state data when corresponding to the first operating parameter, using a pre-set steady-state threshold corresponding to multiple value ranges of the first operating parameter, second data, and operating data at the given time point corresponding to the first operating parameter, includes: Determine the value range of the running data corresponding to the first running parameter at this time point, and use it as the filter value range; Determine whether the second data is within the range of the steady-state threshold corresponding to the selected value interval; if the second data is within the range of the steady-state threshold corresponding to the selected value interval, then the running data at that time point belongs to the steady-state data under the first running parameter.

[0062] Specifically, for example, if the operating data of the unit load at the current time point is 400MW, then the value range of the operating data corresponding to the unit load at this time point is 300-450MW. At this time, the steady-state threshold corresponding to the unit load in the value range of 300-450MW is 12MW.

[0063] It is worth noting that the second data point, namely the difference between the maximum and minimum unit load values, is determined by whether it falls within the steady-state threshold range corresponding to the selected value interval. If the difference between the maximum and minimum unit load values ​​is within the steady-state threshold range corresponding to the selected value interval, then the operating data at the current time point is considered steady-state data under unit load. If the difference between the maximum and minimum unit load values ​​exceeds the steady-state threshold range corresponding to the selected value interval, then the operating data at the current time point is not considered steady-state data under unit load.

[0064] Specifically, such as Figure 2 As shown, in step S52, the judgment result corresponding to each first running parameter is obtained from the running data at each time point.

[0065] Specifically, such as Figure 2 As shown, in step S53, the judgment result corresponding to each first operating parameter is statistically analyzed for the running data at each time point. If the judgment result corresponding to all first operating parameters belongs to steady-state data, then the running data at that time point is used as the filtered steady-state data, that is, as the running data at one time point in the first steady-state running data.

[0066] It is worth noting that by adopting a dynamic threshold division method, the dynamic characteristics of the generator set under different operating states of the first operating parameter are automatically adapted to the steady-state judgment standard, avoiding the problem that the fixed threshold method in the traditional method is too lenient at low load and too strict at high load.

[0067] Preferably, the step of using the operating data at the time point where the judgment results corresponding to all first operating parameters belong to steady-state data as the first steady-state operating data includes: The theoretical main steam pressure at each time point is calculated based on the main steam temperature at each time point in the first steady-state operating data. The steam pressure difference between the theoretical steam pressure and the main steam pressure at each time point is determined. Determine whether the steam pressure difference at each time point exceeds the steam pressure threshold corresponding to the main steam temperature at each time point; if the steam pressure difference at a time point exceeds the steam pressure threshold corresponding to the main steam temperature at that time point, then delete the operating data of that time point from the first steady-state operating data.

[0068] Specifically, the consistency of changes in main steam temperature and main steam pressure is checked, and the theoretical main steam pressure at each time point is calculated using the following formula: ; The steam pressure difference between the theoretical steam pressure and the main steam pressure at each time point is then calculated using the following formula: .

[0069] It is worth noting that the steam pressure threshold is set differently for different main steam temperatures. For example, when the main steam temperature is greater than 540℃, the steam pressure threshold is set to 0.8; when the main steam temperature is less than or equal to 540℃, the steam pressure threshold is set to 1.3.

[0070] Specifically, after obtaining the steam pressure difference between the theoretical steam pressure and the main steam pressure at each time point, the steam pressure threshold corresponding to the main steam temperature at each time point is determined. The steam pressure difference at each time point is compared with the steam pressure threshold. If the steam pressure difference at a time point exceeds the steam pressure threshold corresponding to the main steam temperature at that time point, the operating data at that time point is deleted from the first steady-state operating data.

[0071] Preferably, such as Figure 2 As shown, the steps for obtaining the steady-state operating sample set further include: Step S54: Based on the preset statistical threshold and the running data of all time points in the first steady-state running data at multiple second running parameters, determine whether the running data of each time point in the first steady-state running data belongs to steady-state data at each second running parameter, and obtain the judgment result of the running data of each time point in the first steady-state running data at each second running parameter; Step S55: Take the running data at the time point where the judgment result corresponding to all the second running parameters belongs to the steady-state data as the second steady-state running data, and take the running data at all time points in the second steady-state running data as the sample of the steady-state running sample set.

[0072] Specifically, in order to further optimize the first steady-state operating data obtained in step S53 and remove operating data that deviates too much from the average level, in step S54, a statistical threshold and multiple second operating parameters are preset.

[0073] Specifically, the statistical threshold is set to 3.

[0074] Preferably, the plurality of second operating parameters includes a plurality of first operating parameters, and further includes one or more of the following: Automatic power generation control commands for the generating unit; Main steam flow rate; Reheat steam temperature; Reheat steam pressure; Coal feed rate of the coal feeder; Furnace negative pressure; Inlet NOx concentration; Vibration amplitude of the fan bearing.

[0075] Specifically, the multiple second operating parameters include unit load or active power, main steam temperature and main steam pressure, as well as one or more of the following: unit automatic power generation control command, main steam flow rate, reheat steam temperature, reheat steam pressure, coal feeder coal rate, furnace negative pressure, inlet NOx concentration and fan bearing vibration amplitude.

[0076] Specifically, the operating data at all time points in the first steady-state operating data corresponding to multiple second operating parameters are determined, and the operating data at each time point in the first steady-state operating data is determined in combination with the pre-set statistical thresholds to determine whether the operating data at each time point in the first steady-state operating data belongs to steady-state data at each second operating parameter.

[0077] It is worth noting that whether the operating data at each time point in the first steady-state operating data belongs to steady-state data may be the same or different under different second operating parameters. The judgment process is independent and does not interfere with each other. For example, if the operating data at a certain time point in the first steady-state operating data does not belong to steady-state data under the main steam flow rate, it does not affect the judgment result of the operating data at that time point in the first steady-state operating data under other second operating parameters.

[0078] Specifically, such as Figure 2 As shown, step S54 obtains the judgment result corresponding to each second operating parameter for the operating data at each time point in the first steady-state operating data. If the second operating parameters include unit load, main steam temperature, main steam pressure, main steam flow rate and furnace negative pressure, then the operating data at each time point in the first steady-state operating data corresponds to a judgment result for unit load, main steam temperature, main steam pressure, main steam flow rate and furnace negative pressure, respectively.

[0079] Preferably, the step of determining whether the operating data at each time point in the first steady-state operating data belongs to steady-state data at each second operating parameter, based on a preset statistical threshold and the operating data at all time points in the first steady-state operating data corresponding to multiple second operating parameters, includes: Obtain the operating data for each second operating parameter at all time points in the first steady-state operating data, and calculate the average and standard deviation for each second operating parameter. Based on the mean and standard deviation of each second operating parameter, and combined with the operating data of each time point for each second operating parameter, the score value of each time point for each second operating parameter is calculated. Determine whether the score value of each time point for each second operating parameter is within the preset statistical threshold; if the score value of each time point for each second operating parameter is within the preset statistical threshold, then the operating data at that time point belongs to steady-state data under that second operating parameter.

[0080] Specifically, the operating data for each time point in the first steady-state operating data corresponding to different second operating parameters are determined, and the average value and standard deviation for each second operating parameter are calculated, which will not be elaborated here.

[0081] Specifically, the score for each second running parameter at each time point is calculated using the following formula: ; in, Indicates the first The time point at the The score value of the second running parameter, Indicates the first The time point at the The running data of the second running parameter, Indicates the first The average value of the second operating parameter Indicates the first The standard deviation of the second operating parameter.

[0082] It is worth noting that after calculating the score value of each time point for each second operating parameter, it is determined whether the score value of each time point for each second operating parameter is within the preset statistical threshold. The statistical threshold is set to 3. If the score value of each time point for each second operating parameter is within the preset statistical threshold, then the operating data at that time point belongs to steady-state data for that second operating parameter.

[0083] For example, when the second operating parameter is the main steam flow rate, the average value and standard deviation of the operating data corresponding to the main steam flow rate at all time points in the first steady-state data are calculated. The score value of each time point at the main steam flow rate is calculated. If the score value is within the preset statistical threshold, the operating data at that time point belongs to the steady-state data at the main steam flow rate.

[0084] Specifically, such as Figure 2 As shown, in step S55, the running data at the time point when the judgment results corresponding to all second running parameters belong to steady-state data is taken as the second steady-state running data.

[0085] It is worth noting that if all the second operating parameters are unit load, main steam temperature, main steam pressure, main steam flow rate, and furnace negative pressure, then the operating data at a certain time point is considered as the second steady-state operating data only when the judgment results corresponding to the unit load, main steam temperature, main steam pressure, main steam flow rate, and furnace negative pressure are all steady-state data.

[0086] Specifically, after obtaining the second steady-state operating data, the original samples in the steady-state operating sample set are deleted, and the operating data at all time points in the second steady-state operating data are used as samples in the steady-state operating sample set.

[0087] Preferably, such as Figure 2 As shown, the steps for obtaining the steady-state operating sample set further include: Step S56: Input the running data corresponding to all time points in the second steady-state running data and all second running parameters into the pre-constructed random tree ensemble model to obtain the average path length of each time point in the second steady-state running data in the random tree; Step S57: Sort the average path lengths of all time points in the random tree in the second steady-state running data in descending order to obtain the sorted second steady-state running data; Step S58: Delete the operating data at the time points with the preset abnormal pollution ratio from the sorted second steady-state operating data in descending order to obtain the third steady-state operating data. At the same time, use the operating data at all time points in the third steady-state operating data as samples of the steady-state operating sample set.

[0088] Specifically, such as Figure 2 As shown, the second steady-state operating data is obtained in step S55. In order to further optimize the operating data at each time point in the second steady-state operating data, in step S56, the operating data corresponding to all time points in the second steady-state operating data is input into the pre-constructed random tree ensemble model, and the average path length of each time point in the second steady-state operating data in the random tree is calculated using the random tree ensemble model.

[0089] It is understandable that the random tree ensemble model is built based on the Isolation Forest algorithm. The average path length in the random tree at each time point in the second steady-state running data is calculated through each data point in the random tree ensemble model, which will not be elaborated here.

[0090] Specifically, such as Figure 2 As shown, in step S57, the average path lengths of all time points in the random tree in the second steady-state operating data obtained in step S56 are sorted in descending order to obtain the second steady-state operating data sorted from high to low.

[0091] Specifically, such as Figure 2 As shown, in step S58, the preset abnormal pollution ratio is set to 0.05. For the sorted second steady-state operating data, the total number of time points is determined, and the number of time points to be deleted is calculated based on the preset abnormal pollution ratio. Deletion begins from the operating data of the time point corresponding to the maximum average path length, until the number of time points to be deleted is reached. The operating data of the remaining time points is then used as the third steady-state operating data.

[0092] Specifically, the Isolation Forest algorithm constructs a high-dimensional collaborative feature model based on generator set operating parameters, breaking through the limitations of traditional single-parameter analysis. It can keenly capture hidden abnormal patterns in seemingly normal single parameters but with logical relationships between parameters. Compared with the single Z-score statistical method, it has a significantly higher ability to detect complex logical anomalies and effectively avoids model distortion caused by the mixing of "pseudo-normal data".

[0093] It is worth noting that, such as Figure 2 As shown, through the continuous optimization of the running data at the time points in steps S51-S58, the third steady-state running data has high quality and can be used as a sample in the total set of steady-state running samples.

[0094] Specifically, after obtaining the third steady-state operating data, the original samples in the steady-state operating sample set are deleted, and the operating data at all time points in the third steady-state operating data are used as samples in the steady-state operating sample set.

[0095] Specifically, such as Figure 1 As shown, the total set of steady-state operating samples is determined in step S1.

[0096] Specifically, such as Figure 1 As shown, in step S2, the net power prediction model corresponding to each operating condition type is trained using each sample in each steady-state operating sample subset after clustering.

[0097] Specifically, the fourth operating parameter includes back pressure and net power. It can be understood that net power = generator active power - excitation transformer high-voltage side active power - high-voltage side active power.

[0098] Preferably, the fourth operating parameter further includes one or more of the following: Reheat steam pressure; Reheat steam temperature; Main steam flow rate; Main steam temperature; Main steam pressure; Ambient temperature.

[0099] Specifically, such as Figure 1 As shown, in step S2, the back pressure of each sample in each steady-state operating sample subset and other used fourth operating parameters are input into the BP neural network model with a preset structure to obtain the predicted net power of each sample. The net power and the preset loss function are used to calculate the net power loss value until the training number meets the condition, and the training of the net power prediction model corresponding to the operating condition type is completed, thus obtaining the net power prediction model corresponding to the operating condition type.

[0100] Preferably, the BP neural network model with the preset structure includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output layer connected in sequence; The first hidden layer contains 128 neurons, the second hidden layer contains 64 neurons, and the third hidden layer contains 32 neurons.

[0101] Specifically, the number of input layers is determined based on the number of selected fourth operating parameters, and the output layer outputs net power.

[0102] Specifically, when using the net power prediction model corresponding to each operating condition type, you only need to input the fourth operating parameter corresponding to the sample of that operating condition type into the net power prediction model for that operating condition type to obtain the predicted net power of the sample.

[0103] Specifically, by inputting all the fourth operating parameters of the sample to be predicted, excluding net power, into the net power prediction model, the net power corresponding to the sample to be predicted can be obtained.

[0104] Specifically, such as Figure 1 As shown, in step S2, the BP neural network model with a preset structure is trained independently by multiple steady-state operating sample subsets, so that the net power prediction model corresponding to various operating conditions can be obtained.

[0105] Compared with existing technologies, the training method for a net power prediction model of power plant generator sets provided in this embodiment of the invention clusters the total steady-state operating sample set into multiple steady-state operating sample subsets based on the third operating parameters of each sample in the total steady-state operating sample set. Each steady-state operating sample subset corresponds to a certain operating condition type. A BP neural network model with a preset structure is trained using the fourth operating parameters of each sample in each steady-state operating sample subset to obtain a net power prediction model corresponding to each operating condition type, greatly improving the prediction accuracy of the net power prediction model. Simultaneously, by using pre-set steady-state thresholds corresponding to multiple value ranges for each first operating parameter, the steady-state characteristics of the generator set under different operating states are fully considered, and steady-state data is extracted efficiently. Furthermore, the method ensures that the judgment results corresponding to all first operating parameters belong to... Using the operational data at each time point of the steady-state data set as the extracted steady-state operational data improves the data quality of the overall steady-state operational sample set, laying a foundation for training a net power prediction model with higher accuracy. Furthermore, based on preset statistical thresholds and operational data at all time points corresponding to multiple second operational parameters in the first steady-state operational data set, outliers are removed, further improving the data quality of the overall steady-state operational sample set and laying a foundation for training a net power prediction model with higher accuracy. Finally, a pre-constructed random tree ensemble model is used to filter and extract the second steady-state operational data, removing data that violates the laws of physical parameters, further improving the data quality of the overall steady-state operational sample set and laying a foundation for training a net power prediction model with higher accuracy.

[0106] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0107] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A training method for a net power prediction model of power plant generator sets, characterized in that, The training method includes: Based on a pre-defined clustering algorithm, the total set of steady-state operating samples is divided into multiple subsets of steady-state operating samples by clustering according to the third operating parameter of each sample in the total set of steady-state operating samples; each subset of steady-state operating samples corresponds to a type of operating condition. The BP neural network model with a preset structure is trained based on the fourth operating parameter of each sample in each steady-state operating sample subset to obtain the net power prediction model corresponding to each operating condition type; wherein, the fourth operating parameter includes back pressure and net power.

2. The training method according to claim 1, characterized in that, The third operating parameter includes one or more of the following: Reheat steam pressure; Reheat steam temperature; Main steam flow rate; Main steam temperature; Main steam pressure; Ambient temperature; Generator active power; Active power on the high-voltage side of the excitation transformer; Active power on the high-voltage side; Back pressure.

3. The training method according to claim 1, characterized in that, The fourth operating parameter also includes one or more of the following: Reheat steam pressure; Reheat steam temperature; Main steam flow rate; Main steam temperature; Main steam pressure; Ambient temperature.

4. The training method according to claim 1, characterized in that, The pre-defined BP neural network model includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output layer connected in sequence. The first hidden layer contains 128 neurons, the second hidden layer contains 64 neurons, and the third hidden layer contains 32 neurons.

5. The training method according to any one of claims 1-4, characterized in that, The steps for obtaining the total set of steady-state operating samples include: Collect a large amount of continuous operating data of the generator set at a preset historical time period; Based on the steady-state thresholds corresponding to multiple pre-set first operating parameters in multiple value ranges, determine whether the operating data at each time point belongs to steady-state data, and obtain the judgment result of the operating data at each time point corresponding to each first operating parameter; The running data at the time point where the judgment result corresponding to all first running parameters belongs to steady-state data is taken as the first steady-state running data, and the running data at all time points in the first steady-state running data is taken as the sample of the total steady-state running sample set.

6. The training method according to claim 5, characterized in that, The step of using the running data at the time point where the judgment results corresponding to all first running parameters belong to steady-state data as the first steady-state running data includes: The theoretical main steam pressure at each time point is calculated based on the main steam temperature at each time point in the first steady-state operating data. The steam pressure difference between the theoretical steam pressure and the main steam pressure at each time point is determined. Determine whether the steam pressure difference at each time point exceeds the steam pressure threshold corresponding to the main steam temperature at each time point; if the steam pressure difference at a time point exceeds the steam pressure threshold corresponding to the main steam temperature at that time point, then delete the operating data of that time point from the first steady-state operating data.

7. The training method according to claim 5, characterized in that, The process of determining whether the running data at each time point belongs to steady-state data based on the steady-state thresholds corresponding to multiple preset first running parameters in multiple value ranges, and obtaining the determination result of the running data at each time point for each first running parameter, includes: Obtain the running data corresponding to each first running parameter for each time point and a preset number of consecutive time points before that time point, and use it as the first data; The difference between the maximum and minimum first operating parameter values ​​in the first data is calculated and used as the second data. By using the pre-set steady-state threshold corresponding to multiple value ranges of the first operating parameter, the second data, and the operating data at the time point corresponding to the first operating parameter, it is determined whether the operating data at the time point belongs to steady-state data when corresponding to the first operating parameter.

8. The training method according to claim 7, characterized in that, The step of determining whether the operating data at a given time point is steady-state data when corresponding to the first operating parameter, using a pre-set steady-state threshold corresponding to multiple value ranges of the first operating parameter, second data, and the operating data at that time point corresponding to the first operating parameter, includes: Determine the value range of the running data corresponding to the first running parameter at this time point, and use it as the filter value range; Determine whether the second data is within the range of the steady-state threshold corresponding to the selected value interval; if the second data is within the range of the steady-state threshold corresponding to the selected value interval, then the running data at that time point belongs to the steady-state data under the first running parameter.

9. The training method according to claim 5, characterized in that, The steps for obtaining the steady-state operating sample set also include: Based on the preset statistical threshold and the running data of all time points in the first steady-state running data corresponding to multiple second running parameters, it is determined whether the running data of each time point in the first steady-state running data belongs to steady-state data at each second running parameter, and the judgment result of the running data of each time point in the first steady-state running data corresponding to each second running parameter is obtained. The running data at the time point where the judgment result corresponding to all second running parameters belongs to the steady-state data is taken as the second steady-state running data, and the running data at all time points in the second steady-state running data is taken as the sample of the steady-state running sample set.

10. The training method according to claim 9, characterized in that, The steps for obtaining the steady-state operating sample set also include: Input the running data corresponding to all second running parameters for all time points in the second steady-state running data into the pre-constructed random tree ensemble model to obtain the average path length of each time point in the second steady-state running data in the random tree; The average path lengths of all time points in the random tree in the second steady-state operating data are sorted in descending order to obtain the sorted second steady-state operating data. From the sorted second steady-state operating data, the operating data at the time points with the preset abnormal pollution ratio are deleted in descending order to obtain the third steady-state operating data. At the same time, the operating data at all time points in the third steady-state operating data are used as samples of the total steady-state operating sample set.

Images