A method for identifying combustion state of a biomass combined heat and power system

Abstract

The present application relates to a kind of biomass cogeneration system combustion state identification method, define the combustion state of biomass cogeneration system, obtain the data of biomass cogeneration system under different combustion states, select input variable based on knowledge and maximum mutual information coefficient method, the data obtained are reduced dimension processing based on selected input variable;With improved generative adversarial network model, data augmentation is carried out to the combustion state data of unbalanced data amount, to generate data evaluation index system to evaluate the data after data augmentation, qualified data enters the finally generated data set, otherwise delete;Repeat until generating the data of preset amount;Based on convolutional neural network, establish combustion state identification model, identify the combustion state of biomass cogeneration system.The present application is simple in design, easy to understand, strong in practicality, wide in practicality, high in identification accuracy, good in generalization performance and robust performance of model, and has strong interpretability.

Description

Technical Field

[0001] This invention relates to the field of electrical digital data processing, and in particular to a method for identifying the combustion state of a biomass cogeneration system based on knowledge and data augmentation. Background Technology

[0002] Biomass energy, as a renewable energy source, has become an important vehicle for zero-carbon and negative-carbon emission pathways due to its environmental friendliness, low cost, and carbon neutrality. It can both meet the energy needs of daily life and truly achieve zero carbon emissions.

[0003] However, biomass, as a renewable energy source, is limited by the diversity and intermittency of its resources, leading to frequent and rapid load changes or prolonged low-load operation in combined heat and power (CHP) systems. This results in unstable combustion. Unstable combustion not only reduces energy efficiency and increases pollutant emissions but can also cause uneven heat distribution and increased thermal stress on the furnace walls. Accurate identification of the combustion status allows for timely feedback of abnormal information, ensuring the system's efficient, high-quality, and stable operation. Therefore, comprehensive identification of the combustion status of biomass CHP systems has significant theoretical and practical implications.

[0004] Currently, there is no unified definition for the combustion state of biomass cogeneration systems. Combustion state identification is generally performed using flame combustion image recognition. However, acquiring flame images of large boiler equipment in biomass cogeneration systems is often very difficult, and dedicated flame imaging cameras are extremely expensive. Furthermore, biomass cogeneration systems primarily operate under normal combustion conditions, while abnormal combustion states are short-lived. Therefore, an imbalance exists between normal and abnormal combustion state data, leading to severe overfitting in the identification model and reducing the accuracy of combustion state identification. Summary of the Invention

[0005] This invention addresses the problems existing in the prior art and provides a method for identifying the combustion state of a biomass cogeneration system. Based on knowledge and data enhancement, the method can simultaneously enhance data of different combustion states, solve problems such as fuzzy definitions of combustion states and imbalanced collected combustion state data, and establish an automatic data filter to ensure the accuracy and diversity of generated data in a timely manner, thereby achieving rapid and accurate identification of the combustion state of a biomass cogeneration system.

[0006] The technical solution adopted in this invention is a method for identifying the combustion state of a biomass cogeneration system, the method comprising the following steps:

[0007] S1 defines the combustion state of the biomass cogeneration system and acquires data on the biomass cogeneration system under different combustion states.

[0008] S2 selects input variables based on knowledge and the maximum mutual information coefficient method, and performs dimensionality reduction on the data of S1 based on the selected input variables;

[0009] S3 uses an improved generative adversarial network model to augment the combustion state data with an imbalance in data volume. In fact, the biomass cogeneration system is in normal working condition for most of the time, with few abnormal states. Therefore, there is a serious imbalance between the normal and abnormal state data. Thus, it is necessary to augment the state with less data volume before identification.

[0010] S4 uses a data evaluation index system to evaluate the data augmented in S3. Qualified data is included in the final generated dataset, otherwise it is deleted. Repeat step S3 until the preset amount of data is generated.

[0011] S5 establishes a combustion state recognition model based on a deep convolutional neural network to identify the combustion state of a biomass cogeneration system.

[0012] Preferably, in S1, the combustion state is related to the pollutant emission concentration and the bed temperature of the circulating fluidized bed.

[0013] Specifically, the main pollutants in biomass cogeneration systems are nitrogen oxides (NOx), with varying upper limits under different standards. Compared to coal, biomass fuel has lower carbon content, higher volatile matter content, lower calorific value, and higher moisture content. It absorbs heat during combustion, resulting in significant heat loss in flue gas. Consequently, the bed temperature of biomass circulating fluidized beds is relatively low, typically maintained at 780-850℃. Studies have shown that when the bed temperature of a biomass circulating fluidized bed is below 750℃, combustion becomes unstable and fluctuates significantly. Based on the pollutant emission concentration C... NOx The combined analysis of the two variables, namely the bed temperature of the circulating fluidized bed, defines the combustion state of the biomass cogeneration system. Based on the classification criteria, several datasets of biomass cogeneration systems under different combustion states are obtained. That is, the biomass cogeneration system data obtained by S1 has labels.

[0014] Preferably, S2 includes the following steps:

[0015] S2.1 Obtain several initial input variables using sensors;

[0016] S2.2 Preliminary screening of input variables based on pollutant formation mechanisms;

[0017] S2.3 The variables after the initial screening are further screened using the maximum mutual information coefficient method to determine the final input variables.

[0018] Preferably, in S2.1, the initial input variables include generator active power, instantaneous value of main steam flow, boiler main steam header outlet temperature, primary air main pipe flow, primary hot air outlet temperature, secondary air main pipe flow, secondary hot air outlet temperature, dual-sided flue gas temperature, dual-sided combustion chamber subboiling temperature, dual-sided combustion chamber mid-boiling temperature, dual-sided combustion chamber mid-temperature, dual-sided furnace outlet temperature, dual-sided low-temperature superheater outlet oxygen content, dual-sided cyclone outlet temperature, and dual-sided air preheater outlet flue gas oxygen content.

[0019] S2.1 involves 23 input variables. However, directly inputting these variables into the subsequent recognition model without processing may lead to an overly complex model, reducing accuracy and increasing computation time. Therefore, it is necessary to screen the input variables to achieve accurate model recognition. In S2.2, input variables are coarsely screened based on the NOx formation mechanism. Here, NOx is mainly fuel-type, with the majority being NO. Its conversion pathway mainly involves two stages: nitrogen conversion in the primary pyrolysis of fuel and the formation of nitrogen-containing pollutants in secondary reactions. In the primary pyrolysis stage, fuel nitrogen is converted into volatile nitrogen and semi-coke nitrogen. Volatile nitrogen includes gaseous nitrogen and tar nitrogen. This stage is closely related to temperature, which is a key factor directly affecting the pyrolysis products. Therefore, temperature variables at six different locations in the combustion chamber are initially selected as input variables (bottom boiling temperature in both combustion chambers, middle boiling temperature in both combustion chambers, and middle temperature in both combustion chambers). As the temperature rises, the primary pyrolysis products undergo secondary reactions between homogeneous and heterogeneous phases, generating gaseous nitrogen substances such as HCN, NH3, HNCO, NO, and N2. A small portion of these nitrogen-containing substances polymerizes to form coke nitrogen. Subsequently, both gaseous nitrogen and coke nitrogen are oxidized to NO and N2. The conversion of different types of gaseous nitrogen to NO mainly depends on the excess air coefficient, which is directly determined by the instantaneous main steam flow rate, the primary air main flow rate, the secondary air main flow rate, the oxygen content at the outlet of the dual-sided low-temperature superheaters, and the oxygen content in the flue gas at the outlet of the dual-sided air preheaters. In summary, S2.2, based on the pollutant formation mechanism, provides 13 variables as input variables through coarse screening, laying a solid foundation for further refined screening and improving identification accuracy.

[0020] Preferably, in S2.3, historical running data of the input variables after preliminary screening are collected, the data space is divided into grids, the mutual information value of each grid data is calculated and normalized to obtain the maximum mutual information coefficient (MIC), and one of the variables with a correlation higher than the preset value is discarded based on the maximum mutual information coefficient, until the pairwise correlation of all remaining variables is less than the preset value.

[0021] Considering the actual operation of industrial systems, it is assumed that when the MIC value exceeds 0.4, there is a significant correlation between the two variables, and only one variable can be selected as the input variable. Therefore, the variables after coarse screening are further refined by the MIC method, and the following variables are finally selected as the input variables of the subsequent model: instantaneous value of main steam flow, primary air main flow, secondary air main flow, subboiling temperature of the front combustion chamber, and oxygen content at the outlet of the left (or right) low-temperature superheater. Based on this screening criterion, the datasets under five different combustion states are subjected to dimensionality reduction processing.

[0022] Preferably, in S3, the improved generative adversarial network model includes:

[0023] A generator is used to receive random noise and combustion status labels, and output generated samples related to the specified labels; its main task is to generate samples that are as realistic as possible, making it difficult for the discriminator to distinguish between generated samples and real samples.

[0024] A discriminator and an auxiliary classifier are set up side by side; the discriminator is used to receive generated samples from the generator output and, based on real samples, determine the probability that the input sample is a real sample; the auxiliary classifier is used to receive generated samples from the generator output and, compared with real samples and their corresponding labels, determine the probability that the input sample belongs to a category; the input samples here include real samples and generated samples.

[0025] The main task of the discriminator and the auxiliary classifier is to learn to distinguish between real samples and generated samples, and to provide corresponding category predictions, thereby guiding the generator to generate more realistic samples.

[0026] In this invention, an improved Generative Adversarial Network (GAN) model is used to augment the relatively small amount of category data, which is generally high emission state and abnormal emission state data. Data augmentation is used to obtain more rich and high-quality data, providing a good foundation for subsequent accurate identification. High-quality samples under different combustion states are generated through the game between three generators, discriminators and auxiliary classifiers.

[0027] Preferably, the generator includes a fully connected layer and four two-dimensional deconvolutional layers arranged in sequence, with a self-attention layer between the last two two-dimensional deconvolutional layers;

[0028] The discriminator includes four sequentially arranged two-dimensional convolutional layers, a self-attention layer, and a fully connected layer;

[0029] The auxiliary classifier comprises four two-dimensional convolutional layers, one self-attention layer, and three fully connected layers arranged in sequence.

[0030] In this invention, the improved generative adversarial network model consists of a generator, a discriminator, and an auxiliary classifier. The generator is composed of stacked two-dimensional deconvolutional layers, self-attention layers, and fully connected layers. The discriminator and the auxiliary classifier are both composed of stacked two-dimensional convolutional layers, self-attention layers, and fully connected layers. Random noise and labels are input to the generator, and the generator outputs generated samples and their labels. Real samples and generated samples are input to the discriminator, and the discriminator outputs the discrimination result of the input sample, i.e., the probability that the input sample is a real sample. The auxiliary classifier takes real samples, generated samples, and their labels as input, and outputs the probability of the category to which the input sample belongs.

[0031] This invention separates the discriminator's functions in discriminating and classifying samples using an independent auxiliary classifier, improving the compatibility between discrimination and classification accuracy and enhancing the ability to generate high-quality multimodal data. Self-attention modules are used in the generator, discriminator, and auxiliary classifier, enabling the model to automatically learn important features in the samples, thus solving the problem of large fluctuations in the quality of generated samples. Finally, the loss function is redesigned and Wasserstein distance is introduced to effectively solve the problems of model collapse and gradient vanishing.

[0032] Preferably, the loss function of the auxiliary classifier satisfies the following condition:

[0033]

[0034] The auxiliary classifier and the loss function for generating samples satisfy the following:

[0035]

[0036] Where x is the real sample, z is random noise, and G(z,c) g The generated sample, more precisely, represents the sample with label c generated after random noise z passes through the generator G. g The sample, P(c=c r |x) is an auxiliary classifier for x with label c. r The probability, P(c=c g |G(z,c g )) is an auxiliary classifier for G(z,c g ) with tag c g The probability; E[·] is the expected value;

[0037] The loss function of the generator is:

[0038]

[0039] The loss function of the discriminator is:

[0040] l D =E[D(x)]-E[D(G(z,c)] g))]

[0041] The loss function of the auxiliary classifier is:

[0042]

[0043] Where D(x) is the probability that x is judged as a real sample, D(G(z,c)) g )) is G(z,c g ) is the probability of being identified as a real sample, and λ1 and λ2 are the parameters of the classification loss of the real sample and the classification loss of the generated sample, respectively; the values ​​of λ1 and λ2 are generally between 0 and 1.

[0044] Preferably, the objective function of the data generation evaluation index system in S4 is,

[0045] p(x i =True)≥w1

[0046] MMD(x i ,x t )≤w2

[0047] KL(x i ,x j )≤w3

[0048] Where, x i and x j For the generated data samples, w1, w2, and w3 are the evaluation thresholds for discriminant probability, KL divergence, and maximum mean difference (MMD), respectively. t The values ​​w1 and w2 limit the accuracy of the generated data, while w3 limits the diversity of the synthesized signal. The larger w1 is and the smaller w2 is, the closer the generated signal is to the real signal; the smaller w3 is, the better the diversity of the generated data.

[0049] P(x i =True) represents the probability that the input data sample is true.

[0050] In this invention, data is considered qualified only if it meets all three criteria, and is then included in the final dataset. Otherwise, it is considered unqualified and deleted. Training ends when the number of qualified generated data reaches the initially set preset value.

[0051] Preferably, in S5, the combustion state recognition model includes N sets of two-dimensional convolutional layers and max pooling layers, and N fully connected layers are sequentially set after the last set of max pooling layers, where N≥2.

[0052] This invention uses datasets of five different combustion states to train the above-mentioned recognition model, then inputs the test set into the recognition model, and calculates the recognition accuracy. The trained recognition model can then be used to identify different combustion states.

[0053] This invention relates to a method for identifying the combustion state of a biomass cogeneration system. The method defines the combustion state of the biomass cogeneration system, acquires data on the biomass cogeneration system under different combustion states, selects input variables based on knowledge and the maximum mutual information coefficient method, and performs dimensionality reduction processing on the acquired data based on the selected input variables. An improved generative adversarial network model is used to augment the imbalanced combustion state data. A data evaluation index system is used to evaluate the augmented data; qualified data is included in the final generated dataset, otherwise it is deleted. This process is repeated until a preset amount of data is generated. Finally, a combustion state identification model is established based on a convolutional neural network to identify the combustion state of the biomass cogeneration system.

[0054] The beneficial effects of this invention are as follows:

[0055] (1) The design is simple, easy to understand, highly practical, and widely applicable;

[0056] (2) It has high recognition accuracy and good generalization and robustness of the model;

[0057] (3) It has strong interpretability. Attached Figure Description

[0058] Figure 1 This is a flowchart of the method of the present invention;

[0059] Figure 2 This is a schematic block diagram of the improved generative adversarial network in this invention;

[0060] Figure 3 This is a schematic diagram of the structure of the recognition model in this invention. Detailed Implementation

[0061] The present invention will be further described in detail below with reference to embodiments, but the scope of protection of the present invention is not limited thereto.

[0062] This invention relates to a method for identifying the combustion state of a biomass cogeneration system. The main execution part of this invention runs on a process control computer for identifying the combustion state of a biomass cogeneration system. The network is configured with a loss function, and the Adam optimizer is used with a learning rate of 0.001. All other parameters are default values.

[0063] Parameter initialization: Set appropriate initial values ​​for the hyperparameters in the proposed method. In the import interface of the control computer model, import the variable data of the biomass cogeneration system after dimensionality reduction, as well as the corresponding combustion state of the biomass cogeneration system.

[0064] Offline training: First, the algorithm is written in code and input into the control computer to initially build the model framework. Then, the variable data of the biomass cogeneration system after dimensionality reduction and the corresponding combustion state of the biomass cogeneration system are input into the algorithm to train the algorithm to obtain the final suitable hyperparameter values. Finally, the trained algorithm model can be obtained.

[0065] Online prediction: The CPU in the computer is activated to read parameter values, and by measuring the variable data of the biomass cogeneration system online and executing the corresponding algorithm program, the corresponding combustion state can be identified in real time.

[0066] like Figure 1 As shown, this invention relates to a method for identifying the combustion state of a biomass cogeneration system, the method comprising the following steps:

[0067] S1 defines the combustion state of the biomass cogeneration system and acquires data on the biomass cogeneration system under different combustion states.

[0068] In S1, the combustion state is related to the pollutant emission concentration and the bed temperature of the circulating fluidized bed;

[0069] Based on a comprehensive analysis of two variables—pollutant emission concentration and circulating fluidized bed temperature—the combustion state of a biomass cogeneration system is defined as follows:

[0070] (1) Stopped state (C) NOx =0mg / m 3 The system is in a shut-down or shutdown state, and the NOx emission concentration is 0 or very low, indicating that the system is in a normal maintenance or non-operational phase.

[0071] (2) High emission status (C NOx >50mg / m 3 NOx emission concentrations exceeding the set limits may be due to unstable combustion processes, excessively high combustion temperatures, or other reasons. This condition may cause environmental problems and is not permitted.

[0072] (3) Normal combustion state (10mg / m³) 3 <C NOx <50mg / m 3 NOx emission concentrations meet standards, and the bed temperature of the circulating fluidized bed fluctuates within the normal range, indicating that the biomass cogeneration system is operating efficiently and in an environmentally friendly manner.

[0073] (4) Low emission status (5.5 mg / m³) 3 <C NOx <10mg / m 3 NOx emission concentrations meet standards, but bed temperature fluctuates between 770℃ and 850℃, with some instances below the standard temperature of 780℃. Combustion is relatively stable, but timely monitoring is necessary to ensure continuous system stability.

[0074] (5) Abnormal emission status (0 mg / m³) 3 <C NOx <5.5mg / m 3 Although the emission standards are met, excessively low emissions may be caused by abnormal combustion in the system. In this state, the bed temperature varies greatly (730℃-810℃), indicating that the combustion state is unstable and the system is in a stage of violent fluctuations. The entire system needs to be inspected and maintained immediately.

[0075] Based on the above classification criteria, datasets of biomass cogeneration systems under five different combustion states were obtained.

[0076] S2 selects input variables based on knowledge and the maximum mutual information coefficient method, and performs dimensionality reduction on the data of S1 based on the selected input variables;

[0077] S2.1 Obtain several initial input variables using sensors;

[0078] The main instruments and equipment of the biomass cogeneration system are equipped with sensors to collect operational data in real time. After preliminary processing, a total of 23 input variables were obtained, including generator active power, instantaneous main steam flow rate, boiler main steam header outlet temperature, primary air main flow rate, primary hot air outlet temperature, secondary air main flow rate, secondary hot air outlet temperature, flue gas temperature (left), flue gas temperature (right), combustion chamber sub-boiling temperature before, combustion chamber sub-boiling temperature after, combustion chamber mid-boiling temperature (left), combustion chamber mid-temperature (right), combustion chamber mid-temperature (left), combustion chamber mid-temperature (right), furnace outlet temperature (left), furnace outlet temperature (right), low-temperature superheater outlet oxygen content (left), low-temperature superheater outlet oxygen content (right), cyclone outlet temperature (left), cyclone outlet temperature (right), air preheater outlet flue gas oxygen content (left), air preheater outlet flue gas oxygen content (right).

[0079] Input variables need to be screened to ensure accurate model identification. To maximize the useful information extracted from the input variables of the biomass cogeneration system, a two-step process is adopted to screen the input variables.

[0080] S2.2 Preliminary screening of input variables based on pollutant formation mechanisms;

[0081] The NOx generated by biomass combined heat and power systems is mainly fuel-based, with the vast majority being NO. The NO conversion pathway mainly involves two stages: nitrogen conversion during primary pyrolysis of fuel and the formation of nitrogen-containing pollutants through secondary reactions.

[0082] In the initial pyrolysis stage, fuel nitrogen is converted into volatile nitrogen and semi-coke nitrogen, where volatile nitrogen includes gaseous nitrogen and tar nitrogen. Studies show that the entire pyrolysis stage is closely related to temperature. Temperature is a key factor directly affecting the pyrolysis products. Therefore, temperature variables at six different locations in the combustion chamber were initially selected as input variables.

[0083] Secondary reaction stage of pyrolysis products: As the temperature rises, the primary pyrolysis products undergo secondary reactions between homogeneous and heterogeneous phases, including cracking, remodeling, dehydration, condensation, polymerization, oxidation, and gasification. Tar, semi-coke, and large molecular gaseous substances further crack to generate gaseous nitrogen substances such as HCN, NH3, HNCO, NO, and N2, while a small portion of nitrogen-containing substances polymerize to form coke nitrogen. Subsequently, gaseous nitrogen and coke nitrogen are oxidized to NO and N2. For the conversion pathway of gaseous nitrogen, the oxidation law of gaseous nitrogen during combustion is usually studied using homogeneous reactions. Although the composition of gaseous nitrogen is complex, under high-temperature conditions, it needs to undergo the process of stripping C and H. Whether it is subsequently oxidized to NO or reduced to N2 depends on the competitive reaction between OH / O2 (generating NO) and NO (generating N2) and N atoms. The conversion of different types of gaseous nitrogen to NO mainly depends on the excess air coefficient. The excess air coefficient is directly determined by the main steam flow rate, primary air main flow rate, secondary air main flow rate, low-temperature superheater main flow rate (left and right), and oxygen content in the flue gas at the air preheater outlet (left and right).

[0084] In summary, 13 variables were initially selected from the variables that can be collected from the biomass cogeneration system as input variables, laying a good foundation for further fine screening and improving the identification accuracy.

[0085] S2.3 The variables after the initial screening are further screened using the maximum mutual information coefficient method to determine the final input variables;

[0086] Historical data for the 13 coarsely selected input variables were collected. Based on this data, the data space was divided into grids, and the mutual information value of each grid was calculated and normalized to obtain the maximum information coefficient (MIC). The specific calculation formula is as follows:

[0087]

[0088]

[0089] In the formula, I(X;Y) is the mutual information between variables X and Y, p(x,y) is the joint probability density between variables, MIC[X;Y] is the maximum information coefficient between variables, and B is the number of grids, which is generally taken as 0.6 times the total amount of data.

[0090] Numerous studies have shown a relationship between the MIC (Minimum Indicator) values ​​of variables and their correlation. Specifically, a MIC between 0 and 0.2 indicates a very weak or no correlation between variables; a MIC between 0.2 and 0.4 indicates a weak correlation; a MIC between 0.4 and 0.6 indicates a moderate correlation; a MIC between 0.6 and 0.8 indicates a significant correlation; and a MIC between 0.8 and 1.0 indicates an extremely significant and strong correlation.

[0091] Considering the actual operation of industrial systems, a MIC value exceeding 0.4 indicates a significant correlation between the two variables. The variables after initial screening were further refined using the MIC method, ultimately selecting the following variables as input variables for the subsequent model: main steam flow rate, primary air main flow rate, secondary air main flow rate, combustion chamber subboiling temperature before temperature, and low-temperature superheater main flow rate (left). Based on the above input variable selection criteria, dimensionality reduction was performed on the datasets under five different combustion conditions.

[0092] S3 uses an improved generative adversarial network model to augment combustion state data with imbalanced data volume.

[0093] After dimensionality reduction of the datasets for five different combustion states, a detailed comparison was made of the proportion of data for each combustion state to the total data volume. It was found that the data volume for the high-emission and abnormal-emission states was very limited and could not be fully utilized for training the subsequent identification model. An improved Generative Adversarial Network (GAN) model was used to augment the data for these two categories, aiming to obtain more abundant and high-quality data to provide a solid foundation for accurate identification.

[0094] After selecting the input features for different combustion state datasets, the proportion of data for each combustion state to the total data is statistically analyzed. For combustion state datasets with a data proportion of less than 10%, data augmentation is performed based on an improved generative adversarial network model.

[0095] like Figure 2 As shown, the improved generative adversarial network model consists of a generator, a discriminator, and an auxiliary classifier, specifically including:

[0096] A generator is used to receive random noise and combustion status labels, and output generated samples related to the specified labels;

[0097] A discriminator and an auxiliary classifier are set up in parallel; the discriminator is used to receive generated samples from the generator output and, based on real samples, determine the probability that the input sample is a real sample; the auxiliary classifier is used to receive generated samples from the generator output and, together with real samples and their corresponding labels, determine the probability that the input sample belongs to a category.

[0098] The generator consists of stacked 2D deconvolutional layers, self-attention layers, and fully connected layers. The discriminator and auxiliary classifier both consist of stacked 2D convolutional layers, self-attention layers, and fully connected layers. More precisely, the generator includes one fully connected layer and four 2D deconvolutional layers arranged sequentially, with a self-attention layer between the last two deconvolutional layers. The discriminator includes four 2D convolutional layers, one self-attention layer, and one fully connected layer arranged sequentially. The auxiliary classifier includes four 2D convolutional layers, one self-attention layer, and three fully connected layers arranged sequentially. The generator's input is random noise and labels indicating the combustion state, and its output is the generated sample and its label. The discriminator's input is real samples and generated samples, and its output is the probability that the data is a real sample. The auxiliary classifier's input is real samples and generated samples, along with their labels, and its output is the probability of the sample's class.

[0099] For the improved generative adversarial network, the parameters are first configured, including the learner's learning rate, batch size, number of samples, number of classes, number of iterations, etc. Then, multidimensional random noise and sample labels are input into the generator. The real samples that need data augmentation and their labels are input into the discriminator and the auxiliary classifier for training, and finally, labeled samples are generated.

[0100] The advantages of the improved generative adversarial network model in this invention are:

[0101] (1) The discriminator plays the role of classification and discrimination at the same time, which can easily lead to problems such as inaccurate feedback, learning bias, and unstable training. In this embodiment, an independent auxiliary classifier is designed to separate the discriminator's functions in discriminating samples and classifying samples, improve the compatibility between discrimination and classification accuracy, and enhance the ability to generate high-quality multi-mode data.

[0102] (2) Traditional GAN ​​models usually use convolutional layers to extract features of samples. However, the local receptive field of convolutional layers is limited and cannot obtain global information. If fully connected layers are used to obtain global information, there are too many parameters to calculate. Therefore, the network of this invention uses self-attention modules in the generator, discriminator and auxiliary classifier, so that the model can automatically learn important targets in the samples and solve the problem of large fluctuations in the quality of generated samples.

[0103] (3) Most generative adversarial networks use JS divergence as the loss function. Due to its discrete nature, it is prone to instability and gradient vanishing during training. Therefore, by redesigning the loss function and introducing Wasserstein distance, the problems of model collapse and gradient vanishing can be effectively solved.

[0104] For the improved generative adversarial network, the loss function of the auxiliary classifier and the real sample satisfy...

[0105]

[0106] The auxiliary classifier and the loss function for generating samples satisfy the following:

[0107]

[0108] Where x is the real sample, z is random noise, and G(z,c) g ) is the generated sample, P(c=c r |x) Determine if the auxiliary classifier has label c for x. r The probability, P(c=c g |G(z,c g Determine the auxiliary classifier for G(z,c) g ) with tag c g The probability of;

[0109] The loss function of the generator is:

[0110]

[0111] The loss function of the discriminator is:

[0112] l D =E[D(x)]-E[D(G(z,c)] g ))]

[0113] The loss function of the auxiliary classifier is:

[0114]

[0115] Where D(x) is the probability that x is judged as a real sample, D(G(z,c)) g )) is G(z,c g ) is the probability of being identified as a real sample, and λ1 and λ2 are the parameters of the classification loss of the real sample and the classification loss of the generated sample, respectively; in practical applications, they can be adjusted from 0 to 1.

[0116] The design of the loss function can make the recognition more efficient and allow the loss to converge to stability in a very short time.

[0117] S4 uses a data evaluation index system to evaluate the data augmented in S3. Qualified data is included in the final generated dataset, otherwise it is deleted. Repeat step S3 until the preset amount of data is generated.

[0118] In the practical training phase of improved generative adversarial networks, training termination typically depends on the network's stable convergence. However, convergence of network training does not necessarily imply good quality of synthesized data. Furthermore, data generation requires not only high accuracy but also rich diversity. Therefore, it is necessary to establish an effective and timely evaluation criterion for generation, primarily including discriminant probability, KL divergence, and maximum mean difference (MMD).

[0119] After the dimensionality reduction of high-emission and abnormal-emission state data, data is generated through an improved generative adversarial network (GAN), and then automatically input into a data evaluation index system for accuracy and diversity assessment. Only data that simultaneously meets all three criteria is considered qualified and included in the final dataset; otherwise, it is considered unqualified and deleted. Training ends when the number of qualified generated data reaches a pre-set preset value. The objective function of the evaluation index system is expressed as:

[0120] p(x i =True)≥w1

[0121] MMD(x i ,x t )≤w2

[0122] KL(x i ,x j )≤w3

[0123] Where x i and x j This represents the generated sample, where w1, w2, and w3 represent the thresholds for the three evaluation metrics, respectively. t Represents the actual signal. P(x) i =True) represents the probability that an input sample belongs to the "True" class. Theoretically, w1 and w2 limit the accuracy of the generated data, while w3 limits the diversity of the synthesized signal. A larger w1 and a smaller w2 indicate that the generated signal is closer to the real signal, while a smaller w3 indicates better diversity in the generated data. The setting of these three thresholds not only controls the quality level of the generated data but also determines the time consumption of training the entire improved generative adversarial network; therefore, trade-offs and sacrifices are necessary.

[0124] After filtering the generated data, datasets for five different combustion states are constructed.

[0125] like Figure 3As shown, S5 establishes a combustion state recognition model based on a deep convolutional neural network (DCNN) to identify the combustion state of a biomass cogeneration system.

[0126] In S5, the combustion state recognition model includes N sets of two-dimensional convolutional layers and max pooling layers. After the last set of max pooling layers, N fully connected layers are sequentially set, where N≥2.

[0127] In the implementation process, N is set to 3. A deep convolutional neural network is used to construct the recognition module. First, three sets of two-dimensional convolutional layers and max pooling layers are stacked to extract features. Then, continuous fully connected layers are used as auxiliary classifiers to gradually reduce dimensionality and ensure stability. The network uses cross-entropy as the loss function, and the Adam optimizer has a learning rate of 0.001. Other parameters are set to default values. The five datasets of different combustion states are used to train the recognition model. Then, the test set is input into the recognition model, and the recognition accuracy is calculated. The optimizer adjusts the network parameters based on the loss function.

[0128] To implement the above embodiments, a computer-readable storage medium is proposed, on which a program for identifying the combustion state of a biomass cogeneration system is stored. When this program is executed by a processor, it implements the method for identifying the combustion state of a biomass cogeneration system. A computer device is also proposed, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the method for identifying the combustion state of a biomass cogeneration system as described above. Through the method, medium, device, and application for identifying the combustion state of a biomass cogeneration system, the problems of low identification accuracy, ambiguous combustion state definition, and high economic cost in existing biomass cogeneration system combustion state identification methods are overcome. These methods are easy to operate and have interpretability.

[0129] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0130] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0131] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0132] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0133] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0134] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for identifying the combustion state of a biomass cogeneration system, characterized in that: The method includes the following steps: S1 defines the combustion state of the biomass cogeneration system and obtains data on the biomass cogeneration system under different combustion states; S2 selects input variables based on knowledge and the maximum mutual information coefficient method, and performs dimensionality reduction on the data of S1 based on the selected input variables; S3 uses an improved generative adversarial network (GAN) model to augment imbalanced combustion state data; the improved GAN model includes: A generator is used to receive random noise and combustion status labels, and output generated samples related to specified labels; the generator includes a fully connected layer and four two-dimensional deconvolutional layers arranged in sequence, with a self-attention layer between the last two two-dimensional deconvolutional layers. A discriminator and an auxiliary classifier are arranged in parallel. The discriminator receives generated samples from the generator and, based on real samples, determines the probability that an input sample is a real sample. The discriminator includes four sequentially arranged two-dimensional convolutional layers, one self-attention layer, and one fully connected layer. The auxiliary classifier receives generated samples from the generator and, compared with real samples and their corresponding labels, determines the probability that an input sample belongs to a category. The auxiliary classifier includes four sequentially arranged two-dimensional convolutional layers, one self-attention layer, and three fully connected layers. S4 evaluates the data augmented in S3 using a generated data evaluation index system. Qualified data is included in the final generated dataset, otherwise it is deleted. Repeat step S3 until the preset amount of data is generated. S5 establishes a combustion state recognition model based on a deep convolutional neural network to identify the combustion state of a biomass cogeneration system.

2. The method for identifying the combustion state of a biomass cogeneration system according to claim 1, characterized in that: In S1, the combustion state is related to the pollutant emission concentration and the bed temperature of the circulating fluidized bed.

3. The method for identifying the combustion state of a biomass cogeneration system according to claim 1, characterized in that: S2 includes the following steps: S2.1 Obtain several initial input variables using sensors; S2.2 Preliminary screening of input variables based on pollutant formation mechanisms; S2.3 The variables after the initial screening are further screened using the maximum mutual information coefficient method to determine the final input variables.

4. The method for identifying the combustion state of a biomass cogeneration system according to claim 3, characterized in that: In S2.1, the initial input variables include generator active power, instantaneous value of main steam flow, boiler main steam header outlet temperature, primary air main flow, primary hot air outlet temperature, secondary air main flow, secondary hot air outlet temperature, dual-sided flue gas temperature, dual-sided combustion chamber subboiling temperature, dual-sided combustion chamber mid-boiling temperature, dual-sided combustion chamber mid-temperature, dual-sided furnace outlet temperature, dual-sided low-temperature superheater outlet oxygen content, dual-sided cyclone outlet temperature, and dual-sided air preheater outlet flue gas oxygen content.

5. The method for identifying the combustion state of a biomass cogeneration system according to claim 3, characterized in that: In S2.3, historical running data of the input variables after preliminary screening are collected, the data space is divided into grids, the mutual information value of each grid data is calculated and normalized to obtain the maximum mutual information coefficient, and one of the variables with a correlation higher than the preset value is discarded based on the maximum mutual information coefficient, until the pairwise correlation of all remaining variables is less than the preset value.

6. The method for identifying the combustion state of a biomass cogeneration system according to claim 1, characterized in that: The auxiliary classifier and the loss function of the real samples satisfy the following: ， The auxiliary classifier and the loss function for generating samples satisfy the following: ， Where x is the real sample, It is random noise. To generate samples, For the auxiliary classifier to have labels for x The probability, For auxiliary classifiers With tags The probability of; The loss function of the generator is: ， The loss function of the discriminator is: ， The loss function of the auxiliary classifier is: ， in, Let x be the probability that x is classified as a real sample. for The probability of being identified as a real sample, where λ1 and λ2 are the parameters for the classification loss of the real sample and the classification loss of the generated sample, respectively.

7. The method for identifying the combustion state of a biomass cogeneration system according to claim 1, characterized in that: The objective function of the S4 generated data evaluation index system is: ， ， ， Where, x i and x j For the generated data samples, w1, w2, and w3 are the evaluation thresholds for discriminant probability, maximum mean difference (MMD), and KL divergence, respectively, and x t For real signals; P(x i =True) represents the probability that the input data sample is true.

8. The method for identifying the combustion state of a biomass cogeneration system according to claim 1, characterized in that: In S5, the combustion state recognition model includes N sets of two-dimensional convolutional layers and max pooling layers. After the last set of max pooling layers, N fully connected layers are sequentially set, where N≥2.

Images