Key parameter soft measurement method and system applied to thermal power generating unit

Abstract

The embodiment of the invention provides a key parameter soft measurement method and system applied to a thermal power generating unit. The method comprises the steps that unit operation state parameters are collected; based on a collinearity index, determining a plurality of representative features with spatial correlation between the unit operation state parameters and unit target parameters; based on different time sequences, respectively calculating time cross correlation between the plurality of representative features and the unit target parameters, and obtaining the plurality of representative features with space-time correlation; and based on a spatial attention mechanism convolutional neural network, repeatedly training the plurality of representative features with spatial-temporal correlation until convergence. According to the embodiment of the invention, time-space two-dimensional precision optimization is realized through spatial feature screening and time sequence alignment, a complete chain is further formed through attention modeling and real-time control, and collaborative improvement of thermal power generating unit parameter sensing, modeling and optimization is realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of operation control of thermal power generating units, and particularly relates to a key parameter soft measurement method and system applied to thermal power generating units. BACKGROUND

[0002] Thermal power generation is the core component of power supply in China, accounting for about 60% of the total power generation in China. Accurate measurement of key parameters of thermal power generating units is the technical basis for ensuring safe operation of the unit, improving power generation efficiency and optimizing operation and maintenance economy, but some key parameters cannot be directly monitored online relying on sensing technology alone. Soft-sensing technique realizes indirect measurement by establishing a mathematical model between measurable variables and parameters to be measured, and becomes an effective means to solve the above problems.

[0003] Traditional soft measurement methods mainly rely on thermodynamic mechanism models or small sample statistical regression, and have technical bottlenecks such as over-simplified modeling assumptions and poor working condition adaptability, which are difficult to meet the parameter prediction accuracy requirements in complex operation scenarios. SUMMARY

[0004] The application provides a key parameter soft measurement method and system applied to thermal power generating units to solve the defects of the prior art.

[0005] According to a first aspect of an embodiment of the application, a key parameter soft measurement method applied to a thermal power generating unit is provided, comprising:

[0006] Collecting unit operation state parameters;

[0007] Based on a collinearity index, a plurality of representative features having spatial correlation between the unit operation state parameters and the unit target parameters are determined;

[0008] Based on different time series, the time cross-correlation of the plurality of representative features and the unit target parameters is calculated respectively, and the plurality of representative features having spatial and temporal correlation are obtained;

[0009] Based on a spatial attention mechanism convolutional neural network, the plurality of representative features having spatial and temporal correlation are repeatedly trained until convergence.

[0010] In some embodiments, the determination of the plurality of representative features having spatial correlation between the unit operation state parameters and the unit target parameters comprises:

[0011] From the unit operation state parameters, features associated with the unit target parameters beyond a preset threshold are selected;

[0012] Based on a variance threshold, low-variance features in the plurality of representative features are removed;

[0013] Based on the collinearity index, the plurality of representative features are screened from the features exceeding the preset threshold.

[0014] In some embodiments, the time cross-correlation between the plurality of representative features and the unit target parameters is calculated based on different time series, respectively, including:

[0015] The plurality of representative features and the unit target parameters of the same length time series are selected to ensure time series alignment and data integrity;

[0016] The time series of the plurality of representative features are subjected to bidirectional sliding operation, and a plurality of groups of sliding sub-sequences are generated;

[0017] The cross-correlation coefficients of the plurality of sliding sub-sequences and the target parameter sequence are analyzed;

[0018] According to the time node difference corresponding to the maximum value of the cross-correlation coefficient, the time cross-correlation between the plurality of representative features and the unit target parameters is determined.

[0019] In some embodiments, the plurality of sliding sub-sequences includes a first sliding sub-sequence and a second sliding sub-sequence, and the bidirectional sliding operation on the time series of the plurality of representative features and the generation of a plurality of groups of sliding sub-sequences include:

[0020] The time series of the plurality of representative features are slid upward by 50 time nodes, and 50 first sliding sub-sequences are generated;

[0021] The time series of the plurality of representative features are slid downward by 50 time nodes, and 50 second sliding sub-sequences are generated;

[0022] The sliding step is set to 1 time node to generate 100 sliding sub-sequences.

[0023] In some embodiments, the spatial attention mechanism convolutional neural network repeatedly trains the plurality of representative features having spatio-temporal correlation until convergence, including:

[0024] Spatial attention mechanisms are inserted after the 7th and 13th residual blocks of the 50-layer residual network, and the output results of the residual blocks are obtained;

[0025] The spatial weight tensor of the plurality of representative features is generated by performing pooling operation and convolution dimension reduction on the output results of the residual blocks;

[0026] Based on the output results of the residual blocks and the spatial weight tensor, a 5-layer fully connected network is used to realize joint regression of the combustion chamber outlet temperature and the turbine efficiency.

[0027] In some embodiments, the pooling operation and the convolution dimension reduction on the output of the residual block generate the spatial weight tensor of the plurality of representative features, including:

[0028] Based on the spatial attention mechanism convolutional neural network, the output of the residual block is subjected to channel-level global pooling and maximum pooling, and two feature tensors with the same size as the output of the residual block are obtained.

[0029] The two feature tensors are channel spliced, and the spatial weight tensor of the plurality of representative features is obtained through an activation function.

[0030] In some embodiments, the channel splicing of the two feature tensors includes:

[0031] The number of channels is reduced to 1 using a convolution layer with a convolution kernel number of 1.

[0032] In some embodiments, the spatial attention mechanism convolutional neural network repeatedly trains the plurality of representative features with spatio-temporal correlation until convergence, further including:

[0033] Based on the mean square error loss function, the loss between the result of the joint regression and the true value is calculated.

[0034] Based on the loss, the plurality of representative features with spatio-temporal correlation are trained again until convergence.

[0035] According to a second aspect of the embodiments of the present application, a key parameter soft measurement system applied to a thermal power unit is provided, including:

[0036] A parameter acquisition module is configured to acquire unit operating state parameters.

[0037] A spatial feature determination module is configured to determine a plurality of representative features with spatial correlation between the unit operating state parameters and unit target parameters based on a collinearity index.

[0038] A time feature determination module is configured to calculate time cross-correlations between the plurality of representative features and the unit target parameters based on different time series, and obtain the plurality of representative features with spatio-temporal correlation.

[0039] A neural network training module is configured to repeatedly train the plurality of representative features with spatio-temporal correlation based on a spatial attention mechanism convolutional neural network until convergence.

[0040] According to a third aspect of the embodiment of the present application, a thermal power generating unit is provided, comprising an integrated vibration spectrum sensor, a flue gas composition analyzer and a thermal imaging camera, and is configured to execute the key parameter soft measurement method for the thermal power generating unit according to the above.

[0041] The key parameter soft measurement method and system for the thermal power generating unit according to the embodiment of the present application have at least the following beneficial effects:

[0042] According to the embodiment of the present application, the spatial dimension feature is screened and optimized through variance selection and collinearity group representative feature extraction, the input data dimension is reduced and the noise interference is reduced, the model training speed is significantly improved compared with the traditional method, the time offset of the feature-target parameter is determined based on the cross-correlation analysis, and the prediction accuracy in the variable load scene is improved, the 50-layer residual network and the spatial attention mechanism are fused, the combustion chamber outlet temperature prediction is significantly reduced compared with the full connection network, and the cross-condition generalization error is significantly reduced. Compared with the traditional mechanism model, the embodiment of the present application fully utilizes the massive parameters of the system, performs real-time control optimization on the unit operation, and enhances the multi-condition adaptability of the unit. According to the embodiment of the present application, the spatial feature screening and time sequence alignment realize the accuracy optimization in time and space dimensions, and further through attention modeling and real-time control, a complete chain is formed to realize the collaborative improvement of parameter perception, modeling and optimization of the thermal power generating unit. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 The flowchart of the key parameter soft measurement method for the thermal power generating unit according to the embodiment of the present application is shown in FIG. 1.

[0044] Figure 2 The structural diagram of the key parameter soft measurement system for the thermal power generating unit according to the embodiment of the present application is shown in FIG. 2. DETAILED DESCRIPTION

[0045] In order for those skilled in the art to better understand the technical solutions of the present application, the present application will be further described in detail below in combination with the drawings and specific embodiments.

[0046] The embodiments of the present application will be further described in detail below in combination with the drawings and embodiments. The detailed description and drawings of the following embodiments are used to exemplarily illustrate the principles of the present application, but cannot be used to limit the scope of the present application, i.e., the present application is not limited to the described embodiments.

[0047] With the development of information systems in thermal power plants, massive sensor data (including temperature, pressure, flow and other multi-dimensional time series parameters) provide a basic condition for data-driven modeling. The multi-source data fusion modeling method based on deep neural network extracts the deep correlation between features and measured parameters from high-dimensional nonlinear data, breaking through the accuracy limit of traditional methods.

[0048] Referring to the drawings Figure 1 As shown in the accompanying drawings, the embodiments of the present application disclose specific implementation steps of a key parameter soft measurement method applied to a thermal power generating unit. The method is configured in a key parameter soft measurement system applied to a thermal power generating unit, so as to ensure that the technical solution of the present application can be implemented by the skilled in the art. The method specifically comprises the following steps 110-140.

[0049] Step 110: Collecting unit operation state parameters.

[0050] Part of the key parameters of the thermal power generating unit (such as the outlet temperature of the combustion chamber of the gas turbine and the wear state of the heating surface) cannot be directly measured by sensors, and needs to be indirectly obtained based on the unit operation data to construct a soft measurement model. The key parameters described in the present application cover the whole life cycle indicators of the unit, including basic output parameters (such as rated power, minimum technical output), energy efficiency parameters (such as power supply coal consumption, boiler efficiency, steam turbine heat rate) and regulation and control parameters (such as start-up and shutdown time, regulation rate, pollution emission indicators). The accurate measurement and optimization of the above-mentioned parameters directly affect the peak regulation capacity of the power system, fuel economy and environmental protection compliance, and are the core basis for the operation optimization of the thermal power generating unit.

[0051] In addition, the unit operation state parameters of the present application include a large amount of sensor data stored in the thermal power plant information system. There is a quantitative correlation between these large amounts of sensor data and the key parameters of the unit. By constructing a data-driven regression model, the soft measurement of key parameters such as combustion efficiency and power supply coal consumption can be realized.

[0052] Step 120: Based on the collinearity index, determining a plurality of representative features having spatial correlation between the unit operation state parameters and the unit target parameters.

[0053] In some embodiments, the determination of the plurality of representative features having spatial correlation between the unit operation state parameters and the unit target parameters comprises: screening features with correlation to the unit target parameters exceeding a preset threshold from the unit operation state parameters; removing low-variance features in the plurality of representative features based on a variance threshold; and screening the plurality of representative features from the features exceeding the preset threshold based on the collinearity index.

[0054] For example: based on the analysis of whether the correlation to the unit target parameters exceeds a preset threshold, redundant features with no significant correlation to the unit target parameters are removed; based on the variance threshold, low-variance features in the plurality of representative features are removed; and then for the collinearity features, a single representative feature is screened from the features exceeding the preset threshold, so as to reduce the risk of model overfitting. Thus, the screening of the representative features of the key parameters of the unit in the spatial dimension is realized.

[0055] When constructing a regression model using massive sensor data, features with high contribution to key parameters need to be screened to avoid model accuracy decline caused by missing high-contribution features or misselecting low-contribution features. In step 120, high-contribution features of a target parameter of a thermal power generating unit are screened in a spatial dimension. Based on massive historical data in a thermal power plant information system (which is quantitatively correlated with key parameters such as the temperature at the outlet of a combustion chamber and turbine efficiency), a feature preliminary screening (for example, redundant features irrelevant to the target parameter of the unit are removed by a variance selection method), a collinearity optimization (for example, a variance inflation factor between features can be calculated, and then a single representative feature is selected from a collinearity group), and a correlation enhancement (for example, features significantly correlated with the target parameter are extracted based on a correlation coefficient as core inputs of a regression model) are used to optimize the modeling process, and then based on the process, a three-level screening strategy of "de-redundancy, collinearity reduction, and strong correlation" is used to solve the problems of insufficient modeling accuracy of a traditional small sample and overfitting of high-dimensional data.

[0056] Step 130: Based on different time series, the time cross-correlation of the plurality of representative features and the target parameter of the unit is calculated respectively, and the plurality of representative features with spatiotemporal correlation are obtained.

[0057] In some embodiments, based on different time series, the time cross-correlation of the plurality of representative features and the target parameter of the unit is calculated respectively, including: selecting time series of the same length of the plurality of representative features and the target parameter of the unit to ensure time series alignment and data integrity; performing bidirectional sliding operation on the time series of the plurality of representative features and generating a plurality of groups of sliding sub-sequences; analyzing the cross-correlation coefficients of the plurality of sliding sub-sequences and the target parameter sequence; and determining the time cross-correlation of the plurality of representative features and the target parameter of the unit according to the time node difference corresponding to the maximum value of the cross-correlation coefficients.

[0058] For example, the plurality of sliding sub-sequences includes a first sliding sub-sequence and a second sliding sub-sequence, and the bidirectional sliding operation on the time series of the plurality of representative features and the generation of a plurality of groups of sliding sub-sequences include: selecting time series data of 1000 time nodes of the target parameter of the unit, and simultaneously obtaining time series data of the same length of each representative feature. The time series of the plurality of representative features are slid upward by 50 time nodes to generate 50 first sliding sub-sequences; the time series of the plurality of representative features are slid downward by 50 time nodes to generate 50 second sliding sub-sequences; and the sliding step is set to 1 time node to generate 100 sliding sub-sequences.

[0059] The components in the operation of the thermal power generating unit have time sequence dependency, and the target parameter is affected by the characteristic value at the historical time rather than the parameter at the current time. In view of the time lag characteristic, the step 130 in the embodiment of the application calculates the correlation coefficient of each feature at different time offset sequences and the target parameter through cross correlation analysis, and determines the optimal time corresponding relationship between the feature and the target parameter. By performing feature-target parameter correlation analysis in the time dimension, the representative features that are significantly correlated with the target parameter can be screened; and the time lag relationship between the feature and the target parameter can be quantified through cross correlation analysis, and the time offset corresponding to the maximum correlation coefficient is extracted.

[0060] Step 140: repeatedly training the plurality of representative features with spatio-temporal correlation based on a spatial attention mechanism convolutional neural network until convergence.

[0061] In some embodiments, the repeatedly training the plurality of representative features with spatio-temporal correlation based on the spatial attention mechanism convolutional neural network until convergence comprises: inserting a spatial attention mechanism after the 7th and 13th residual blocks of a residual network 50 layer (ResNet50), and obtaining the output results of the residual blocks, so as to realize the backbone network building; generating a spatial weight tensor of the plurality of representative features by performing a pooling operation and a convolution dimension reduction on the output results of the residual blocks; based on the output results of the residual blocks and the spatial weight tensor, realizing joint regression of the combustion chamber outlet temperature and the turbine efficiency by a 5-layer fully connected network, for example, using linear activation at the end to output the parameter prediction values of the combustion chamber outlet temperature, the turbine efficiency and the like.

[0062] In an exemplary embodiment, the generating the spatial weight tensor of the plurality of representative features by performing a pooling operation and a convolution dimension reduction on the output results of the residual blocks comprises: based on the spatial attention mechanism convolutional neural network, performing a channel-level global pooling and a maximum pooling on the output results of the residual blocks, and obtaining two feature tensors with the same size as the output results of the residual blocks; channel splicing the two feature tensors, and obtaining the spatial weight tensor of the plurality of representative features through an activation function.

[0063] In an exemplary embodiment, the channel splicing the two feature tensors comprises: using a convolution layer with a convolution kernel number of 1 to reduce the channel number to 1.

[0064] For example, after splicing, 1*1 convolution dimension reduction is performed to a single channel, and then a spatial weight matrix is generated through an activation function, and multiplied with the representative features element by element.

[0065] In an example embodiment, the repeatedly training the plurality of representative features with the spatial correlation until convergence based on the spatial attention mechanism convolutional neural network further comprises: calculating a loss between a result of the joint regression and a true value based on a mean square deviation loss function; and training the plurality of representative features with the spatial correlation again until convergence based on the loss.

[0066] The step 140 of the example embodiment captures complex nonlinear relationships between features through a ResNet50 deep residual structure, focuses on high contribution areas by combining a spatial attention mechanism, improves feature utilization compared to a traditional fully connected network, and thus realizes feature-dependent modeling; enhances the adaptability of the model to variable operating condition data through sliding window time alignment and spatial attention weighting, and thus realizes cross-condition generalization; realizes lightweight packaging of the model based on a deep learning framework, supports online soft measurement of the combustion chamber temperature and turbine efficiency driven by real-time sensor data, and finally realizes industrial deployment. Based on this, the example embodiment realizes the construction of a thermal power unit parameter regression model based on a spatial attention mechanism.

[0067] The example embodiment realizes the screening and optimization of spatial dimension features through variance selection and collinearity group representative feature extraction, reduces the dimension of input data and reduces noise interference, so that the model training speed is significantly improved compared to traditional methods; determines the time offset of the feature-target parameter based on cross-correlation analysis, which improves the prediction accuracy under variable load conditions; based on the fusion of a 50-layer residual network and a spatial attention mechanism, the combustion chamber outlet temperature prediction is significantly reduced compared to a fully connected network, and the cross-condition generalization error is significantly reduced. Compared with the traditional mechanism model, the example embodiment makes full use of the massive parameters of the system to optimize the real-time control of the unit operation and enhances the multi-condition adaptability of the unit. The example embodiment realizes the accuracy optimization of the space-time two-dimensional by spatial feature screening and time series alignment, and further forms a complete chain through attention modeling and real-time control to realize the collaborative improvement of thermal power unit parameter perception, modeling and optimization.

[0068] Referring to the accompanying Figure 2 The example embodiment also provides a key parameter soft measurement system applied to a thermal power unit, which comprises a parameter acquisition module 210, a spatial feature determination module 220, a time feature determination module 230, and a neural network training module 240.

[0069] The parameter acquisition module 210 is configured to acquire unit operating state parameters.

[0070] The spatial feature determination module 220 is configured to determine a plurality of representative features with spatial correlation between the unit operating state parameters and the unit target parameters based on a collinearity index.

[0071] The time feature determination module 230 is configured to calculate time cross correlations between the plurality of representative features and the unit target parameters based on different time sequences, and obtain the plurality of representative features with spatiotemporal correlations.

[0072] The neural network training module 240 is configured to repeatedly train the plurality of representative features with spatiotemporal correlations based on a spatial attention mechanism convolutional neural network until convergence.

[0073] The embodiments of the present application realize precision optimization in the spatiotemporal double dimensions through spatial feature screening and time sequence alignment, further form a complete chain through attention modeling and real-time control, and realize collaborative improvement of parameter perception, modeling and optimization of a thermal power generating unit.

[0074] The embodiments of the present application also provide a thermal power generating unit including an integrated vibration frequency spectrum sensor, a flue gas composition analyzer and a thermal imaging camera, and a key parameter soft measurement system applied to the thermal power generating unit based on the above, configured to perform a key parameter soft measurement method applied to the thermal power generating unit based on the above.

[0075] It can be understood that the above embodiments are only exemplary embodiments adopted for illustrating the principles of the present application, and the present application is not limited thereto. Various modifications and improvements can be made by those skilled in the art without departing from the spirit and essence of the present application, and these modifications and improvements are also considered as the protection scope of the present application.

Claims

1. A soft measurement method for key parameters of a thermal power unit, characterized in that: include: Collect unit operating status parameters; Determining, based on a collinearity index, a plurality of representative features having spatial correlation between the unit operating state parameters and the unit target parameters; Based on different time series, respectively calculating the time cross-correlations between the plurality of representative features and the target parameters of the unit, and obtaining the plurality of representative features with spatiotemporal correlations; Based on the spatial attention mechanism convolutional neural network, the multiple representative features with spatiotemporal correlation are repeatedly trained until convergence.

2. The method according to claim 1, characterized in that The multiple representative features having spatial correlation between the unit operating state parameters and the unit target parameters are determined to include: Screening features from the unit operating status parameters, the features having a correlation with the unit target parameters exceeding a preset threshold; Eliminating low-variance features from the plurality of representative features based on a variance threshold; Based on the collinearity indicator, the plurality of representative features are screened from the features exceeding a preset threshold.

3. The method according to claim 1, characterized in that Calculating the time cross-correlations between the plurality of representative features and the target parameters of the unit based on different time series includes: Selecting the plurality of representative features and the unit target parameters of equal time series to ensure time series alignment and data integrity; Performing a bidirectional sliding operation on the time series of the multiple representative features and generating multiple groups of sliding subsequences; Analyzing cross-correlation coefficients between the plurality of sliding subsequences and a target parameter sequence; The time cross-correlations between the plurality of representative features and the target parameters of the unit are determined according to the time node differences corresponding to the maximum values ​​of the cross-correlation coefficients.

4. The method according to claim 3, characterized in that The multiple sliding subsequences include a first sliding subsequence and a second sliding subsequence, and performing a bidirectional sliding operation on the multiple time series representing features to generate multiple groups of sliding subsequences includes: Slide the time series of the multiple representative features upward by 50 time nodes, and generate 50 first sliding subsequences; Slide the time series of the multiple representative features downward by 50 time nodes, and generate 50 second sliding subsequences; The sliding step is set to 1 time node to generate 100 sliding subsequences.

5. The method according to claim 1, wherein The convolutional neural network based on the spatial attention mechanism repeatedly trains the multiple representative features with spatiotemporal correlation until convergence, comprising: Insert the spatial attention mechanism after the 7th and 13th residual blocks of the residual network layer 50, and obtain the output results of the residual blocks; Generating spatial weight tensors of the plurality of representative features by performing pooling operations and convolution dimensionality reduction on the output results of the residual block; Based on the output result of the residual block and the spatial weight tensor, a five-layer fully connected network is used to implement the joint regression of the combustion chamber outlet temperature and the turbine efficiency.

6. The method according to claim 5, characterized in that The performing of pooling operation and convolution dimensionality reduction on the output result of the residual block to generate the spatial weight tensors of the plurality of representative features includes: Based on a spatial attention mechanism convolutional neural network, channel-level global pooling and maximum pooling are performed on the output result of the residual block, and two feature tensors of the same size as the output result of the residual block are obtained; The two feature tensors are channel-joined, and the spatial weight tensor of the multiple representative features is obtained through an activation function.

7. The method according to claim 6, characterized in that Channel concatenation of the two feature tensors includes: Use a convolution layer with a convolution kernel of 1 to reduce the number of channels to 1.

8. The method according to claim 5, characterized in that The convolutional neural network based on the spatial attention mechanism repeatedly trains the plurality of representative features with spatiotemporal correlation until convergence further comprising: Calculate the loss between the result of the joint regression and the true value based on the mean square error loss function; The plurality of representative features with spatiotemporal correlation are trained again based on the loss until convergence.

9. A soft measurement system for key parameters of thermal power units, characterized in that: include: Parameter acquisition module, used to collect unit operating status parameters; a spatial feature determination module, configured to determine, based on a collinearity index, a plurality of representative features having spatial correlation between the unit operating state parameters and the unit target parameters; a time feature determination module, configured to calculate, based on different time series, the time cross-correlations between the plurality of representative features and the target parameters of the unit, and obtain the plurality of representative features having spatiotemporal correlations; The neural network training module is used to repeatedly train the multiple representative features with spatiotemporal correlation based on a spatial attention mechanism convolutional neural network until convergence.

10. A thermal power unit, characterized in that: The invention comprises an integrated vibration spectrum sensor, a flue gas composition analyzer and a thermal imaging camera, and is based on a soft measurement system for key parameters applied to a thermal power unit as described in claim 9, and is configured to execute a soft measurement method for key parameters applied to a thermal power unit as described in any one of claims 1 to 8.

Images