CEMS carbon monitoring method and system and storage medium

Abstract

The invention relates to the technical field of emission monitoring, and discloses a CEMS carbon monitoring method and system and a storage medium. The method comprises the steps of obtaining a carbon monitoring data set; preprocessing the carbon monitoring data set to obtain labeled and unlabeled training data; building a CEMS carbon monitoring model, wherein the CEMS carbon monitoring model comprises a sample enhancement module, a time-frequency feature extraction module, a medium-long distance time-frequency information extraction module based on self-attention convolution and a double-end structure; defining loss functions including an unsupervised loss weighting function, an unsupervised loss function and a supervised loss function, and performing network training; real-time data of the unit are input into the CEMS carbon monitoring model, and carbon emission parameters are output in real time through the model. According to the technical scheme, the dependence on manual labeling data is remarkably reduced, and the problem of high label collection cost in practical application is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of emission monitoring, and particularly to a CEMS carbon monitoring method, system and storage medium. Background Art

[0002] With the increasingly serious global climate change problem, carbon emission monitoring has become one of the important means to address climate change. CEMS (Continuous Emission Monitoring System) is an advanced technology that can real-time monitor the concentration of greenhouse gases such as carbon dioxide in industrial emissions. Its importance is reflected in multiple aspects: First, the CEMS system can accurately monitor the concentration of pollutants in industrial emissions, helping enterprises and governments to timely discover and solve emission problems, thereby reducing pollution to the atmospheric environment and improving air quality. Second, governments around the world are increasingly strict in regulating greenhouse gas emissions, and the CEMS system provides an effective monitoring tool for enterprises to help them comply with environmental protection regulations and avoid being punished for excessive emissions. In addition, by real-time monitoring emission data, enterprises can optimize production processes and equipment operations, improve energy utilization efficiency, reduce production costs, and enhance the market competitiveness of enterprises. Finally, the accurate data provided by the CEMS system provides an important basis for governments and enterprises to formulate emission reduction strategies and policies, contributing to scientific management and sustainable development.

[0003] Currently, many power plants have not installed CEMS carbon emission monitoring systems. Since the carbon emissions of thermal power units are closely related to multiple unit parameters, including fuel-related parameters, unit operation parameters, and environmental and operation parameters, etc. Therefore, these easily obtainable parameters can be used to achieve soft measurement of CEMS carbon emissions. Its essence lies in establishing a mapping relationship or model between unit operation parameters and carbon emission parameters (such as concentration). The core of this model construction lies in two points: one is the design of the model itself, and the other is the training of the model. Model design needs to consider the time characteristics of unit operation and carbon emissions, and model training needs to be based on a small amount of artificial flue gas sampling calibration data (i.e., a small amount of carbon emission labels). This work currently lacks research. Summary of the Invention

[0004] To solve the above technical problems, the present invention provides a CEMS carbon monitoring method, system and storage medium.

[0005] To solve the above technical problems, the present invention adopts the following technical solutions:

[0006] A CEMS carbon monitoring method, the carbon monitoring model used can output corresponding carbon emission parameters according to the real-time operation data of a thermal power unit. The training process of the carbon monitoring model includes:

[0007] Obtain the historical operation data of the thermal power unit as input data, manually label some of the data to generate labeled data, and use the remaining data as unlabeled data;

[0008] Use wavelet transform to extract time-frequency features from the labeled data and the unlabeled data;

[0009] Perform strong enhancement and weak enhancement on the time-frequency features of the unlabeled data respectively;

[0010] Input the time-frequency features of the labeled data and the time-frequency features of the enhanced unlabeled data into two time-frequency information extraction modules with shared parameters respectively, and extract the distance information in the time-frequency domain of the input data based on self-attention convolution to obtain the self-attention feature map corresponding to the labeled data and the self-attention feature map corresponding to the unlabeled data;

[0011] Input the self-attention feature map corresponding to the labeled data into the linear layer E1, output the corresponding carbon emission parameters, and calculate the supervised loss; input the self-attention feature map corresponding to the unlabeled data into the linear layers E1 and E2, calculate the consistency loss after the strongly enhanced unlabeled data and the weakly enhanced unlabeled data are output by the carbon monitoring model, and construct the unsupervised loss based on the calculation of the consistency loss and the gradient density; train the carbon monitoring model based on the supervised loss and the unsupervised loss.

[0012] In one embodiment, the obtaining of the historical operation data of the thermal power unit as unlabeled data and the manual labeling of some of the unlabeled data to generate labeled data specifically includes:

[0013] Obtain a dataset of the historical operation data of the unit, and the operation data includes fuel-related parameters, unit operation parameters, and environmental and operation parameters;

[0014] Select the historical operation data at multiple time points from a part of the dataset of the historical operation data, perform the labeling of carbon emission labels to obtain multiple labeled data, and form a labeled dataset; randomly select the historical operation data at multiple time points from another part of the multi-dimensional time series to obtain multiple unlabeled data, and form an unlabeled dataset.

[0015] In one embodiment, the using of wavelet transform to extract time-frequency features from the labeled data and the unlabeled data specifically includes:

[0016] Denote the labeled data and the unlabeled data together as the input data x(t), and perform wavelet transform on the input data x(t):

[0017]

[0018] where WT x (α,τ) represents the wavelet transform result, is a wavelet basis function, α is the dilation factor for realizing the stretching of the basic wavelet, and τ is the translation factor reflecting the displacement of the wavelet function in the transformation.

[0019] In one embodiment, the time-frequency features of the unlabeled data are respectively subjected to strong enhancement and weak enhancement, specifically including:

[0020] Weak enhancement is performed by adding Gaussian noise to the time-frequency features of the unlabeled data to obtain weakly enhanced samples; strong enhancement of the time-frequency features of the unlabeled data is achieved by setting different sampling rates and translations to obtain strongly enhanced samples.

[0021] In one embodiment, the time-frequency features of the labeled data and the time-frequency features of the enhanced unlabeled data are respectively input into two time-frequency information extraction modules with shared parameters, and the distance information in the time-frequency domain of the input data is extracted based on self-attention convolution to obtain the self-attention feature map corresponding to the labeled data and the self-attention feature map corresponding to the unlabeled data, specifically including:

[0022] The input time-frequency features are respectively passed through three convolutional structures after generating feature maps using a convolutional network as the query, key, and value of the time-frequency features:

[0023] K = FW k , Q = FW q , V = FW v ;

[0024]

[0025] A = SoftMax(S);

[0026] O = AV;

[0027] where F is the feature map generated after the time-frequency features are input into the convolutional network, K, Q, and V are the key, query, and value respectively, and W k , W q and W v are the parameter matrices corresponding to the convolutional kernels of the three convolutional structures respectively; S is the similarity score matrix, is the scaling factor, A is the attention weight, and O is the self-attention feature map.

[0028] In one embodiment, the self-attention feature map corresponding to the labeled data is input into the linear layer E1, the corresponding carbon emission parameters are output, and the supervised loss is calculated, specifically including:

[0029] The supervised loss is directly calculated by combining the result output by the linear layer E1 with the gradient balance mechanism:

[0030] g = |E1(O) - p* |;

[0031]

[0032] Among them, g is the gradient norm, O is the self-attention feature map corresponding to the labeled data, and p * is the true carbon emission label of the labeled data, and B l represents the set number of labeled data, GD(·) is the gradient density, and l ∈ (g) is the length of the effective interval for normalization; g k is the gradient norm of the k-th input, and a function δ k related to g ∈ (g k , g) is defined; when , δ ∈ (g k , g) = 1, otherwise δ ∈ (g k , g) = 0;

[0033] Assign a weight to each labeled data through the gradient density: β i is the weight of the i-th labeled data, and g i is the gradient norm of the i-th labeled data;

[0034] When the carbon emission parameter output by the carbon monitoring model is the carbon emission level classification, the supervision loss L s is:

[0035]

[0036] H is the cross-entropy;

[0037] When the carbon emission parameter output by the carbon monitoring model is the carbon dioxide concentration classification, the supervision loss L s is:

[0038]

[0039] In one embodiment, the self-attention feature map corresponding to the unlabeled data is input into the linear layers E1 and E2, the consistency loss of the strongly augmented unlabeled data and the weakly augmented unlabeled data after being output by the carbon monitoring model is calculated, and the unsupervised loss is constructed based on the calculation of the consistency loss and the gradient density, specifically including:

[0040] Calculate the consistency loss of the strongly augmented unlabeled data and the weakly augmented unlabeled data after being output by the carbon monitoring model

[0041]

[0042] O s,i is the self-attention feature map corresponding to the i-th unlabeled data in a batch of training data after strong augmentation, O w,i is the self-attention feature map corresponding to the i-th unlabeled data in a batch of training data after weak augmentation, and H refers to cross-entropy;

[0043] Introduce gradient density GD to measure the density of gradient norm g in the sample distribution:

[0044] g = |E1(O s,i ) - E1(O w,i )|;

[0045]

[0046] where B u represents the number of unlabeled data in a batch of training data, l ∈ (g) is the length of the effective interval for normalization; g k is the gradient norm of the k-th input, and define a function δ k related to g ∈ (g k , g); when is true, δ ∈ (g k , g) = 1, otherwise δ ∈ (g k , g) = 0;

[0047] Assign a weight to each unlabeled data through gradient density GD(g):

[0048]

[0049] Calculate the unsupervised loss by combining the results output by E1 and the weights of each unlabeled data;

[0050] When the carbon emission parameter output by the carbon monitoring model is carbon emission level classification, the unsupervised loss L GHM is:

[0051]

[0052] Calculate the unsupervised loss L CE by combining the results output by E2 and based on cross-entropy:

[0053]

[0054] When the carbon emission parameter output by the carbon monitoring model is carbon dioxide concentration classification, the unsupervised loss L after gradient balancingGHM is:

[0055]

[0056] Combined with the result output through E2, calculate the unsupervised loss L based on cross-entropy CE :

[0057]

[0058] The final unsupervised loss L u is:

[0059] L u = L GHM + L CE .

[0060] In one embodiment, training the carbon monitoring model based on the supervised loss and the unsupervised loss specifically includes:

[0061] Construct an unsupervised loss weighting function; use the weakly augmented unlabeled data as pseudo-labels, and assume that the confidence max(p) of the pseudo-labels follows a dynamically changing truncated Gaussian distribution; in the r-th iteration, the mean and variance of the confidence distribution of the pseudo-labels are μ r and The weighting function is defined as:

[0062]

[0063] where λ(·) represents the weight of the unsupervised loss, λ max is the maximum weight of the unlabeled data, p is the output of the carbon monitoring model, μ r and are estimated through the predictions of the model on historical batches:

[0064] μ r = mμ r-1 + (1 - m)μ c ;

[0065]

[0066] where m is the momentum parameter of exponential moving average, μ c and are the mean and variance of the current batch;

[0067] Define the pseudo-label distribution as the expectation of the carbon monitoring model's prediction for unlabeled data Normalize each pseudo-label through the following formula:

[0068]

[0069] Among them, UA(·) represents the unified alignment operation, and u(C) is the uniform distribution of the target category. is the current pseudo-label distribution estimated by EMA, and Normalize(·) ensures that the probabilities are normalized to 1; the pseudo-labels after normalization through the unified alignment operation are used to recalculate the sample weights, so as to balance the quantity and quality of the pseudo-labels.

[0070] Construct the overall loss function L to train the carbon monitoring model; the overall loss function consists of the supervised loss L s and the unsupervised loss L u as follows:

[0071] L = L s + λ(E2(O w,i ))L u ;

[0072] O w,i is the self-attention feature map corresponding to the i-th unlabeled data in a batch of training data after weak augmentation.

[0073] A CEMS carbon monitoring system includes:

[0074] Data acquisition module: Obtain the historical operation data of the thermal power unit as unlabeled data, and manually annotate part of the unlabeled data to generate labeled data.

[0075] Time-frequency feature extraction module: Use wavelet transform to extract time-frequency features from the labeled data and unlabeled data.

[0076] Data augmentation module: Perform strong augmentation and weak augmentation on the time-frequency features of the unlabeled data respectively.

[0077] Time-frequency information extraction module: Input the time-frequency features of the labeled data and the time-frequency features of the augmented unlabeled data into two parameter-sharing time-frequency information extraction modules respectively, and extract the distance information in the time-frequency domain of the input data based on self-attention convolution to obtain the self-attention feature map corresponding to the labeled data and the self-attention feature map corresponding to the unlabeled data.

[0078] Model training module: Input the self-attention feature map corresponding to the labeled data into the linear layer E1, output the corresponding carbon emission parameters, and calculate the supervised loss; input the self-attention feature map corresponding to the unlabeled data into the linear layers E1 and E2, calculate the consistency loss of the strongly augmented unlabeled data and the weakly augmented unlabeled data after being output by the carbon monitoring model, and construct the unsupervised loss based on the calculation of the consistency loss and the gradient density; Train the carbon monitoring model based on the supervised loss and the unsupervised loss.

[0079] A computer-readable storage medium stores a computer program thereon, and when the computer program is executed by a processor, the steps of the method in any one of the embodiments are implemented.

[0080] Compared with the prior art, the beneficial technical effects of the present invention are as follows:

[0081] 1. The present invention realizes the prediction of carbon emission parameters (such as concentration) through multiple unit parameters, including fuel-related parameters, unit operation parameters, and environmental and operation parameters, etc. At the same time, wavelet transform is used to convert the real-time data of the unit into a time-frequency image, and the data in multiple dimensions is analyzed from two aspects of time domain and frequency domain, improving the anti-interference ability and accuracy of the monitoring method.

[0082] 2. The present invention utilizes the self-attention-convolution mechanism to help the model capture long-distance dependencies, improve the fine-grained recognition ability, and realize the accurate monitoring and evaluation of fixed-source carbon emissions.

[0083] 3. The present invention has a small demand for labeled data. It uses labeled data for supervised learning. By performing weak augmentation and strong augmentation on unlabeled data before input respectively, the data after weak augmentation is output as pseudo-labels, and the model is trained through the consistency loss of the two augmentations. The semi-supervised method can utilize unlabeled data to improve the model effect under the condition of a small amount of labeled data. It solves the problem of high cost of collecting data labels in reality.

[0084] 4. The present invention adopts the SoftMatch framework, and trains the CEMS carbon emission prediction model based on the semi-supervised algorithm by using both unlabeled data and labeled data. It introduces a Gaussian weighting function and a unified alignment technique to balance the quantity and quality of pseudo-labels. It improves the prediction effect of the algorithm and reduces the dependence on labeled data.

[0085] 5. The present invention introduces a gradient equalization mechanism, which improves the recognition ability of the algorithm for carbon parameters under difficult-to-separate and easily confused working conditions. At the same time, in order to ensure that the overall effect does not decline, an innovation is made. A dual-head structure is adopted, and the cross-entropy loss is retained to ensure that the semi-supervised algorithm can maintain high-quality pseudo-labels during the learning process, improving the prediction accuracy of the algorithm for carbon parameters. Description of the Drawings

[0086] Figure 1 is the flowchart of the method in the embodiment of the present invention;

[0087] Figure 2 is the model architecture diagram in the embodiment of the present invention. Detailed Embodiments

[0088] The following is a detailed description of a preferred embodiment of the present invention with reference to the drawings.

[0089] As Figure 1 shown, the present invention provides a CEMS carbon monitoring method. The carbon monitoring model adopted can output corresponding carbon emission parameters according to the real-time operation data of a thermal power unit. The training process of the carbon monitoring model includes:

[0090] Step S1: Obtain the historical operation data of the thermal power unit as input data, manually label part of the data to generate labeled data, and use the remaining data as unlabeled data;

[0091] Step S2: Use wavelet transform to extract time-frequency features from the labeled data and the unlabeled data;

[0092] Step S3: Perform strong enhancement and weak enhancement on the time-frequency features of the unlabeled data respectively;

[0093] Step S4: Input the time-frequency features of the labeled data and the time-frequency features of the enhanced unlabeled data into two time-frequency information extraction modules with shared parameters respectively, extract the distance information in the time-frequency domain of the input data based on self-attention convolution, and obtain the self-attention feature map corresponding to the labeled data and the self-attention feature map corresponding to the unlabeled data;

[0094] Step S5: Input the self-attention feature map corresponding to the labeled data into the linear layer E1, output the corresponding carbon emission parameters, and calculate the supervised loss; input the self-attention feature map corresponding to the unlabeled data into the linear layers E1 and E2, calculate the consistency loss after the strongly enhanced unlabeled data and the weakly enhanced unlabeled data are output by the carbon monitoring model, and construct the unsupervised loss based on the calculation of the consistency loss and the gradient density; train the carbon monitoring model based on the supervised loss and the unsupervised loss.

[0095] By combining wavelet transform time-frequency analysis and semi-supervised learning methods, the present invention effectively improves the accuracy and robustness of carbon emission monitoring of thermal power units. Specifically, wavelet transform is used to convert multi-dimensional operation data into time-frequency images, extract features from both the time domain and the frequency domain, enhancing the anti-interference ability; pseudo-labels are generated by weakly and strongly enhancing the data, and a consistency loss function is designed to improve the model performance under the condition of only a small amount of manually labeled data; the self-attention-convolution module is innovatively introduced to capture long-distance dependence relationships, combined with the bilinear layer structure and the gradient equalization mechanism, which not only improves the fine-grained recognition ability for complex working conditions but also ensures the quality of pseudo-labels. The overall solution significantly reduces the dependence on manually labeled data and solves the problem of high label acquisition cost in practical applications.

[0096] In one embodiment, obtaining the historical operation data of the thermal power unit as unlabeled data and manually labeling part of the unlabeled data to generate labeled data in step S1 specifically includes:

[0097] Obtain a data set of the historical operation data of the unit, where the operation data includes fuel-related parameters, unit operation parameters, and environmental and operation parameters;

[0098] Select the historical operation data at multiple time points from a part of the data set of the historical operation data, perform the annotation of carbon emission labels, obtain multiple labeled data, and form a labeled data set; randomly select a sequence with a fixed length in another part of the multi-dimensional time series as unlabeled data to obtain an unlabeled data set.

[0099] Specifically, prepare the data set required for training and testing the carbon monitoring model: Obtain the historical operation data of the unit, including fuel-related parameters, unit operation parameters, and environmental and operation parameters. In a preferred embodiment, the fuel-related parameters include fuel consumption, fuel type (such as coal, natural gas, etc.), fuel quality (such as carbon content per unit calorific value), and carbon oxidation rate, and these factors directly affect the amount of carbon dioxide generated by fuel combustion. Unit operation parameters such as unit load, boiler efficiency, boiler air intake, and mill separator speed will affect the combustion efficiency and carbon emission concentration. In addition, environmental temperature and operation optimization measures (such as adjusting boiler air intake, optimizing coal blending plan, etc.) will also have a certain impact on the carbon emission concentration. The multi-dimensional time series composed of these historical operation data is denoted as X.

[0100] Select multiple time points from X for artificial flue gas sampling and annotation, assign carbon emission labels to it, obtain labeled data, and form a labeled data set; randomly select another sequence with a fixed length as unlabeled data to obtain an unlabeled data set; denote a multi-dimensional time series with a fixed length as x(t), and use y to represent the carbon emission label; the carbon emission label can be continuously valued, such as CO2 concentration, etc.; it can also be discretely valued, such as emission level, reliability level, etc.

[0101] In one embodiment, the time-frequency feature extraction of the labeled data and the unlabeled data by using wavelet transform in step S2 specifically includes:

[0102] Denote the labeled data and the unlabeled data as the input data x(t) uniformly, and perform wavelet transform on the input data x(t):

[0103]

[0104] Among them, WT x (α,τ) represents the wavelet transform result, is the wavelet basis function, α is the dilation factor for realizing the stretching of the basic wavelet, and τ is the translation factor reflecting the displacement of the wavelet function in the transform.

[0105] In one embodiment, the time-frequency features of the unlabeled data are strongly enhanced and weakly enhanced in step S3, which specifically includes:

[0106] Weak enhancement is performed by adding Gaussian noise to the unlabeled data to obtain weakly enhanced samples; strong enhancement of the unlabeled data is achieved by setting different sampling rates and translations to obtain strongly enhanced samples.

[0107] In the present invention, for unlabeled data, it is introduced into the training process through pseudo-labels, and the unlabeled data is strongly enhanced and weakly enhanced respectively, so as to simulate various situations that may be encountered in the actual environment.

[0108] In one embodiment, the time-frequency features of the labeled data and the time-frequency features of the enhanced unlabeled data in step S4 are respectively input into two time-frequency information extraction modules with shared parameters, and the distance information in the time-frequency domain of the input data is extracted based on self-attention convolution to obtain the self-attention feature map corresponding to the labeled data and the self-attention feature map corresponding to the unlabeled data, which specifically includes:

[0109] After the input time-frequency features are used to generate feature maps through a convolutional network, they respectively pass through three convolutional structures as the query, key, and value of the time-frequency features:

[0110] K = FW k , Q = FW q , V = FW v ;

[0111]

[0112] A = SoftMax(S);

[0113] O = AV;

[0114] where F is the feature map generated after the time-frequency features are input into the convolutional network, K, Q, and V are the obtained key, query, and value respectively, and W k , W q and W v are the parameter matrices corresponding to the convolutional kernels of the three convolutional structures respectively; S is the similarity score matrix, is the scaling factor, A is the attention weight, and O is the self-attention feature map.

[0115] The present invention constructs a medium and long-distance time-frequency information extraction module based on self-attention convolution, which can improve the model's ability to extract medium and long-distance information in the time-frequency domain, improve the model's fine-grainedness, facilitate the modeling of multi-level dependency relationships, improve the recognition ability for difficult-to-separate samples, and transform the feature extraction model using the self-attention mechanism.

[0116] In one embodiment, the self-attention feature map corresponding to the labeled data in step S5 is input into the linear layer E1, the corresponding carbon emission parameters are output, and the supervised loss is calculated, specifically including:

[0117] The supervised loss is directly calculated by combining the result output by the linear layer E1 with the gradient balance mechanism:

[0118] g = |E1(O) - p * |;

[0119]

[0120] where g is the gradient norm, O is the self-attention feature map corresponding to the labeled data, and p * is the true carbon emission label of the labeled data, B l represents the set number of labeled data, GD(·) is the gradient density, and l ∈ (g) is the length of the effective interval for normalization; when δ ∈ (g k , g) = 1, otherwise δ ∈ (g k , g) = 0.

[0121] Through the gradient density, a weight is assigned to each labeled data: β i is the weight of the i-th labeled data, and g i is the gradient norm of the i-th labeled data.

[0122] When the carbon emission parameter output by the carbon monitoring model is the carbon emission level classification, the supervised loss L s is:

[0123]

[0124] where H is the cross-entropy.

[0125] When the carbon emission parameter output by the carbon monitoring model is the carbon dioxide concentration classification, the supervised loss L s is:

[0126]

[0127] In one embodiment, the self-attention feature map corresponding to the unlabeled data in step S5 is input into the linear layers E1 and E2, the consistency loss of the strongly augmented unlabeled data and the weakly augmented unlabeled data after being output by the carbon monitoring model is calculated, and the unsupervised loss is constructed based on the calculation of the consistency loss and the gradient density, specifically including:

[0128] Calculate the consistency loss after the carbon monitoring model outputs the strongly augmented unlabeled data and the weakly augmented unlabeled data

[0129]

[0130] O s,i is the self-attention feature map corresponding to the i-th unlabeled data in a batch of training data after strong augmentation, O w,i is the self-attention feature map corresponding to the i-th unlabeled data in a batch of training data after weak augmentation, and H refers to cross-entropy.

[0131] Introduce gradient density GD to measure the density of gradient norm g in the sample distribution:

[0132] g = |E1(O s,i ) - E1(O w,i )|;

[0133]

[0134] where B u represents the number of unlabeled data in a batch of training data, l ∈ (g) is the length of the effective interval for normalization; when δ ∈ (g k , g) = 1, otherwise δ ∈ (g k , g) = 0.

[0135] Assign a weight to each unlabeled data through gradient density GD(g):

[0136]

[0137] Calculate the unsupervised loss by combining the results output by E1 and the weights of each unlabeled data.

[0138] When the carbon emission parameter output by the carbon monitoring model is carbon emission level classification, the unsupervised loss L GHM based on gradient balance is:

[0139]

[0140] Calculate the unsupervised loss L CE by combining the results output by E2 and based on cross-entropy:

[0141]

[0142] When the carbon emission parameter output by the carbon monitoring model is carbon dioxide concentration classification, the unsupervised loss L based on gradient balanceGHM is:

[0143]

[0144] Combined with the result output through E2, calculate the unsupervised loss L based on cross-entropy CE :

[0145]

[0146] The final unsupervised loss L u is:

[0147] L u = L GHM + L CE .

[0148] The unsupervised loss in the present invention mainly consists of two parts. Since carbon monitoring is vulnerable to environmental interference in the open environment and forms some difficult-to-separate samples, there is a problem of difficulty imbalance in the overall carbon monitoring dataset. Therefore, the present invention introduces a gradient balance mechanism to solve such problems.

[0149] In one embodiment, training the carbon monitoring model based on the supervised loss and the unsupervised loss in step S5 specifically includes:

[0150] Construct an unsupervised loss weighting function; use the weakly augmented unlabeled data as pseudo-labels, and assume that the confidence max(p) of the pseudo-labels follows a dynamically changing truncated Gaussian distribution; in the r-th iteration, the mean and variance of the confidence distribution of the pseudo-labels are μ r and The weighting function is defined as:

[0151]

[0152] where λ(·) represents the weight of the unsupervised loss, λ max is the maximum weight of the unlabeled data, p is the output of the carbon monitoring model, μ r and are estimated through the predictions of the model on historical batches:

[0153] μ r = mμ r-1 + (1 - m)μ c ;

[0154]

[0155] where m is the momentum parameter of the exponential moving average, μ c and are the mean and variance of the current batch.

[0156] Define the pseudo-label distribution as the expectation of the carbon monitoring model's prediction for unlabeled data. Normalize each pseudo-label using the following formula:

[0157]

[0158] where UA(·) represents the unified alignment operation, u(C) is the uniform distribution of the target class, is the current pseudo-label distribution estimated by EMA, and Normalize(·) ensures that the probabilities are normalized to 1; the pseudo-labels normalized through the unified alignment operation are used to recalculate the sample weights to balance the quantity and quality of the pseudo-labels;

[0159] Construct the overall loss function L to train the carbon monitoring model; the overall loss function consists of the supervised loss L s and the unsupervised loss L u as follows:

[0160] L = L s + λ(E2(O w,i ))L u ;

[0161] O w,i is the self-attention feature map corresponding to the i-th unlabeled data in a batch of training data after weak augmentation.

[0162] After the carbon monitoring model is trained, write the carbon monitoring model into the relevant device, input the real-time operation data of the thermal power unit, and the corresponding carbon emission parameters can be output in real time through E1.

[0163] It should be understood that although the steps in the flowchart of the accompanying drawings of the specification are shown in sequence according to the arrows, these steps do not necessarily have to be executed in the order indicated by the arrows. Unless there is a clear indication in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowchart of the accompanying drawings of the specification may include multiple steps or multiple stages, and these steps or stages do not necessarily have to be executed at the same time, but can be executed at different times, and the execution order of these steps or stages does not necessarily have to be sequential, but can be executed alternately or in turn with at least a part of the steps or stages in other steps or other steps.

[0164] Based on the description of the above method embodiments, the present disclosure also provides a CEMS carbon monitoring system. The system may include a system (including a distributed system), software (application), module, component, server, client, etc. that uses the method described in the embodiments of this specification and combines the necessary implementation hardware. Based on the same inventive concept, the systems in one or more embodiments provided by the embodiments of the present disclosure are as described in the following embodiments. Since the implementation solutions for the system to solve problems are similar to those of the method, the implementation of the specific system in the embodiments of this specification can refer to the implementation of the foregoing method. As used hereinafter, the term "module" or "modular unit" may be a combination of software and / or hardware that can achieve a predetermined function. Although the systems described in the following embodiments are preferably implemented in software, implementation in hardware, or a combination of software and hardware is also possible and contemplated.

[0165] A CEMS carbon monitoring system, comprising:

[0166] A data acquisition module: acquiring historical operation data of a thermal power unit as unlabeled data, and performing manual annotation on part of the unlabeled data to generate labeled data;

[0167] A time-frequency feature extraction module: performing time-frequency feature extraction on the labeled data and the unlabeled data by using wavelet transform;

[0168] A data enhancement module: performing strong enhancement and weak enhancement on the time-frequency features of the unlabeled data respectively;

[0169] A time-frequency information extraction module: inputting the time-frequency features of the labeled data and the time-frequency features of the enhanced unlabeled data into two time-frequency information extraction modules with shared parameters respectively, and extracting the distance information in the time-frequency domain of the input data based on self-attention convolution to obtain a self-attention feature map corresponding to the labeled data and a self-attention feature map corresponding to the unlabeled data;

[0170] A model training module: inputting the self-attention feature map corresponding to the labeled data into a linear layer E1, outputting the corresponding carbon emission parameters, and calculating a supervised loss; inputting the self-attention feature map corresponding to the unlabeled data into linear layers E1 and E2, calculating the consistency loss after the strongly enhanced unlabeled data and the weakly enhanced unlabeled data are output by the carbon monitoring model, and constructing an unsupervised loss based on the calculation of the consistency loss and the gradient density; training the carbon monitoring model based on the supervised loss and the unsupervised loss.

[0171] In an exemplary embodiment, a computer-readable storage medium including instructions, such as a memory including the instructions, is also provided. The above instructions can be executed by a processor to complete the above method. The storage medium may be a computer-readable storage medium. For example, the computer-readable storage medium may be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, and an optical data storage device, etc.

[0172] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and without departing from the spirit or basic characteristics of the present invention, the present invention can be implemented in other specific forms. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention, and any reference signs in the claims should not be regarded as limiting the claims involved.

[0173] In addition, it should be understood that although this specification is described according to embodiments, not every embodiment only contains an independent technical solution. This narrative way of the specification is only for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A CEMS carbon monitoring method, characterized in that: The carbon monitoring model used can output corresponding carbon emission parameters according to the real-time operation data of the thermal power unit. The training process of the carbon monitoring model includes: The historical operation data of thermal power units are obtained as input data, some data are manually annotated to generate labeled data, and the rest of the data is used as unlabeled data; Wavelet transform is used to extract time-frequency features from labeled and unlabeled data; The time-frequency features of unlabeled data are strongly enhanced and weakly enhanced respectively; The time-frequency features of the labeled data and the enhanced time-frequency features of the unlabeled data are respectively input into two parameter-sharing time-frequency information extraction modules, and the distance information in the time-frequency domain of the input data is extracted based on the self-attention convolution to obtain the self-attention feature map corresponding to the labeled data and the self-attention feature map corresponding to the unlabeled data; The self-attention feature map corresponding to the labeled data is input into the linear layer E1, the corresponding carbon emission parameters are output, and the supervised loss is calculated; the self-attention feature map corresponding to the unlabeled data is input into the linear layers E1 and E2, the consistency loss of the strongly enhanced unlabeled data and the weakly enhanced unlabeled data after being output by the carbon monitoring model is calculated, and the unsupervised loss is constructed based on the consistency loss calculation and the gradient density; the carbon monitoring model is trained based on the supervised loss and the unsupervised loss.

2. The CEMS carbon monitoring method according to claim 1, characterized in that: The obtaining of historical operation data of the thermal power unit as unlabeled data and manually labeling part of the unlabeled data to generate labeled data specifically includes: Acquire a data set of historical operating data of the unit, the operating data including fuel-related parameters, unit operating parameters, and environmental and operating parameters; Historical operation data at multiple time points are selected from one part of the historical operation data set, and carbon emission labels are annotated to obtain multiple labeled data to form a labeled data set; historical operation data at multiple time points in another part of the multidimensional time series are randomly selected to obtain multiple unlabeled data to form an unlabeled data set.

3. The CEMS carbon monitoring method according to claim 1, characterized in that: The use of wavelet transform to extract time-frequency features of labeled data and unlabeled data specifically includes: The labeled data and the unlabeled data are recorded as input data x(t), and the input data x(t) is subjected to wavelet transform: Among them, WT x (α,τ) represents the wavelet transform result, is the wavelet basis function, α is the expansion factor for realizing the basic wavelet stretching, and τ is the translation factor reflecting the displacement of the wavelet function in the transformation.

4. The CEMS carbon monitoring method according to claim 1, characterized in that: The time-frequency features of the unlabeled data are strongly enhanced and weakly enhanced respectively, specifically including: The time-frequency features of unlabeled data are weakly enhanced by adding Gaussian noise to obtain weakly enhanced samples; the time-frequency features of unlabeled data are strongly enhanced by setting different sampling rates and translation to obtain strongly enhanced samples.

5. The CEMS carbon monitoring method according to claim 1, characterized in that: The time-frequency features of the labeled data and the enhanced time-frequency features of the unlabeled data are respectively input into two parameter-shared time-frequency information extraction modules, and the distance information in the time-frequency domain of the input data is extracted based on the self-attention convolution to obtain the self-attention feature map corresponding to the labeled data and the self-attention feature map corresponding to the unlabeled data, which specifically includes: The input time-frequency features are used to generate feature maps using a convolutional network and then pass through three convolution structures as the query, key, and value of the time-frequency features: K=FW k ,Q=FW q ,V=FW v ; A = SoftMax(S); O = AV; Among them, F is the feature map generated after the time-frequency feature is input into the convolutional network, K, Q, and V are the key, query, and value respectively, and W k , W q and W v are the corresponding parameter matrices of the convolution kernels of the three convolution structures; S is the similarity score matrix, is the scaling factor, A is the attention weight, and O is the self-attention feature map.

6. The CEMS carbon monitoring method according to claim 1, characterized in that: The self-attention feature map corresponding to the labeled data is input into the linear layer E1, the corresponding carbon emission parameters are output, and the supervised loss is calculated, specifically including: The supervised loss is calculated directly using the output of the linear layer E1 combined with the gradient balancing mechanism: g=|E1(O)-p * |; Among them, g is the gradient norm, O is the self-attention feature map corresponding to the labeled data, and p * is the real carbon emission label of the labeled data, B l represents the number of labeled data, GD(·) is the gradient density, l ∈ (g) is the length of the valid interval, used for normalization; g k is the gradient norm of the kth input, and defines a k ,g-related functionδ ∈ (g k ,g); when When ∈ (g k ,g)=1, otherwise δ ∈ (g k ,g)=0; Through the gradient density, a weight is assigned to each labeled data: β i is the weight of the i-th labeled data, g i is the gradient norm of the i-th labeled data; When the carbon emission parameters output by the carbon monitoring model are carbon emission grade classifications, the supervision loss L s for: H is the cross entropy; When the carbon emission parameter output by the carbon monitoring model is the carbon dioxide concentration classification, the supervision loss L s for:

7. The CEMS carbon monitoring method according to claim 1, characterized in that: The self-attention feature map corresponding to the unlabeled data is input into the linear layers E1 and E2, and the consistency loss of the strongly enhanced unlabeled data and the weakly enhanced unlabeled data after being output by the carbon monitoring model is calculated. Based on the consistency loss calculation and the gradient density, the unsupervised loss is constructed, which specifically includes: Calculate the consistency loss of strongly enhanced unlabeled data and weakly enhanced unlabeled data after being output by the carbon monitoring model O s,i is the self-attention feature map corresponding to the i-th unlabeled data in a batch of training data after strong enhancement, O w,i is the self-attention feature map corresponding to the i-th unlabeled data in a batch of training data after weak enhancement, and H refers to the cross entropy; The gradient density GD is introduced to measure the density of the gradient norm g in the sample distribution: g=|E1(O s,i )-E1(O w,i )|; Among them, B u represents the number of unlabeled data in a batch of training data, l ∈ (g) is the length of the valid interval, used for normalization; g k is the gradient norm of the kth input, and defines a k ,g-related functionδ ∈ (g k ,g); when When ∈ (g k ,g)=1, otherwise δ ∈ (g k ,g)=0; Through the gradient density GD(g), a weight is assigned to each unlabeled data: Calculate the unsupervised loss by combining the result output by E1 and the weight of each unlabeled data; When the carbon emission parameters output by the carbon monitoring model are carbon emission grade classification, the unsupervised loss L after gradient balancing is GHM for: Combined with the results output by E2, the unsupervised loss L is calculated based on cross entropy CE : When the carbon emission parameter output by the carbon monitoring model is the carbon dioxide concentration classification, the unsupervised loss L after gradient balancing is GHM for: Combined with the results output by E2, the unsupervised loss L is calculated based on cross entropy CE : The final unsupervised loss L u for: L u =L GHM +L CE .

8. The CEMS carbon monitoring method according to claim 1, characterized in that: The training of the carbon monitoring model based on the supervised loss and the unsupervised loss specifically includes: Construct an unsupervised loss weighting function; take the weakly enhanced unlabeled data as pseudo-labels, and assume that the confidence of the pseudo-labels, max(p), follows a dynamically changing truncated Gaussian distribution; in the rth iteration, the mean and variance of the confidence distribution of the pseudo-labels are μ r and The weighting function is defined as: where λ(·) represents the weight of the unsupervised loss, λ max is the maximum weight of unlabeled data, p is the output of the carbon monitoring model, μ r and Estimated by the model's predictions on historical batches: m r =mm r-1 +(1-m)m c ; Where m is the momentum parameter of the exponential moving average, μ c and are the mean and variance of the current batch; Define pseudo-label distribution as the expectation of carbon monitoring model prediction for unlabeled data Each pseudo-label is normalized by the following formula: Where UA(·) represents the uniform alignment operation, u(C) is the uniform distribution of target categories, is the current pseudo-label distribution estimated by EMA, Normalize(·) ensures that the probability is normalized to 1; the pseudo-labels normalized by the uniform alignment operation are used to recalculate the sample weights to balance the quantity and quality of the pseudo-labels; Construct an overall loss function L to train the carbon monitoring model; the overall loss function is composed of the supervision loss L s and the unsupervised loss L u composition: O w,i It is the self-attention feature map corresponding to the i-th unlabeled data in a batch of training data after weak enhancement.

9. A CEMS carbon monitoring system, characterized in that: include: Data acquisition module: obtain historical operation data of thermal power units as unlabeled data, and manually annotate some unlabeled data to generate labeled data; Time-frequency feature extraction module: uses wavelet transform to extract time-frequency features of labeled data and unlabeled data; Data enhancement module: strong enhancement and weak enhancement are performed on the time-frequency features of unlabeled data respectively; Time-frequency information extraction module: The time-frequency features of the labeled data and the enhanced time-frequency features of the unlabeled data are respectively input into two parameter-sharing time-frequency information extraction modules, and the distance information in the time-frequency domain of the input data is extracted based on the self-attention convolution to obtain the self-attention feature map corresponding to the labeled data and the self-attention feature map corresponding to the unlabeled data; Model training module: input the self-attention feature map corresponding to the labeled data into the linear layer E1, output the corresponding carbon emission parameters, and calculate the supervised loss; input the self-attention feature map corresponding to the unlabeled data into the linear layers E1 and E2, calculate the consistency loss of the strongly enhanced unlabeled data and the weakly enhanced unlabeled data after being output by the carbon monitoring model, and construct the unsupervised loss based on the consistency loss calculation and gradient density; train the carbon monitoring model based on the supervised loss and unsupervised loss.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images