A biomass fuel dryness degree estimation method based on machine learning

Abstract

The present application relates to a kind of biomass fuel dry-wet degree estimation method based on machine learning.The present method includes collecting boiler combustion condition signal data, sensor signal selection, generating Epoch data, using CSP algorithm to extract features to Epoch data, using LDA algorithm to feature classification to feature vector, obtain the classification prediction label of Epoch data, using classification prediction label to calculate the wetness degree score value of fuel and the like steps.The present method uses machine learning strategy, can automatically adjust model and parameter according to the change of data to adapt to different combustion conditions, to provide more accurate results, the present method can effectively process large-scale data, can also automatically make decision or prediction, cope with real-time decision demand, compared with artificial logic judgment model, the process of the present method is simpler, decision process is more efficient, classification accuracy and running speed and the like performance significantly improve, very suitable for biomass power plant use.

Description

Technical Field

[0001] This invention belongs to the technical field of biomass fuel moisture content estimation methods, and particularly relates to a biomass fuel moisture content estimation method based on machine learning. Background Technology

[0002] Biomass power plants use a wide variety of fuels, varying in shape and material. Furthermore, the weight of fuel fed into the boiler for combustion is difficult to accurately estimate due to weather and temperature variations. However, during boiler combustion, the same weight of fuel can lead to different load changes in the unit. When the fuel is relatively wet, a larger amount of fuel is needed to maintain the unit load within a certain range. Therefore, determining the current dryness or wetness of the fuel helps determine the amount of fuel to add or subtract. Appropriate fuel addition / subtraction instructions contribute to more stable and vigorous combustion, thereby generating greater power output.

[0003] The current manual logic for judging fuel dryness and wetness during production is as follows: when it is found that the flue gas volume in the furnace increases or the load is difficult to carry (such as when the unit load is low or the main steam pressure rise rate is very low after the grate vibration with the same amount of fuel and air volume), it indicates that the fuel moisture content has increased. At this time, after manually adjusting the air volume, it will be reflected that the outlet steam temperature of the second and third stage superheaters increases and the flue gas temperature of the second and third stage superheaters increases.

[0004] The above-mentioned manual methods for judging the dryness and wetness of fuel have many drawbacks: First, manual logic or formula calculations are usually fixed and cannot flexibly adapt to different combustion conditions, resulting in inaccurate judgments; second, manual logic or formula calculations are inefficient or infeasible when processing large amounts of data; and finally, manual logic or formula calculations cannot automatically make decisions or predictions, and a series of rules need to be manually written, which is time-consuming, inefficient, and has high labor costs. Summary of the Invention

[0005] To overcome the aforementioned shortcomings of manual methods for judging fuel moisture content, this invention proposes a novel machine learning-based method and model for estimating the moisture content of biomass fuel. This method and model are particularly suitable for judging the moisture content of fuel during biomass power generation. This method employs a machine learning strategy, which can automatically adjust the model and parameters according to changes in data to adapt to different combustion conditions, thereby providing more accurate results. Simultaneously, machine learning methods can effectively process large-scale data and extract useful information and patterns. Furthermore, machine learning methods can automatically make decisions or predictions without the need for manually writing a series of rules, making the decision-making process more efficient and faster, and capable of handling real-time decision-making needs.

[0006] Terminology Explanation

[0007] CSP: Common Spatial Pattern (CSP) is an algorithm for signal processing and pattern recognition. Its goal is to extract spatial features with maximum variance from multi-channel signals. The core idea of ​​CSP is to transform the original multi-channel signal into a new set of feature vectors through a linear transformation, such that these feature vectors have the maximum variance difference between different categories.

[0008] Linear Discriminant Analysis (LDA) is a classic pattern recognition and statistical analysis method, primarily used for dimensionality reduction and classification problems. The goal of LDA is to project high-dimensional data into a low-dimensional space while maximizing inter-class differences. The basic idea of ​​LDA is to use a linear transformation to map the original high-dimensional data to a new low-dimensional subspace, maximizing the inter-class distance between samples from different classes in the new space, while minimizing the variance within the same class.

[0009] Interpolation and resampling: Interpolation is the process of inferring or estimating data points at unknown locations among known data points. It estimates values ​​at intermediate locations based on existing relationships between data points. Interpolation can be linear or more complex nonlinear. Resampling is the process of redistributing existing data points at different sampling rates or time intervals. This may involve shifting data points from one time axis or spatial axis to another and redistributing the corresponding values. Resampling can also be used to adjust the sampling rate of data, such as decreasing it from a high frequency to a low frequency or increasing it from a low frequency to a high frequency.

[0010] Specifically, this invention provides a machine learning-based method for estimating the wet / dry properties of biomass fuels. This method includes:

[0011] S1. Collect boiler combustion condition signal data;

[0012] S2. Select sensor signals based on human logical judgment experience, and filter out sensor signals that are closely related to the dryness or wetness of fuel;

[0013] S3. The time-series data of the sensor signals selected in the previous step are segmented and resampled to generate Epoch data;

[0014] S4. Use the CSP algorithm to extract features from the Epoch data generated in the previous step, extracting feature vectors that are easy to distinguish from the two types of Epoch data and can be fed into a linear classifier.

[0015] S5. Use the LDA algorithm to classify the feature vectors obtained in the previous step to obtain the classification prediction labels for the Epoch data;

[0016] S6. Calculate the fuel dryness / wetness using the classification prediction labels of the Epoch data obtained in the previous step, and obtain the fuel wetness score.

[0017] Furthermore, the boiler combustion condition signal data mentioned in step S1 of the biomass fuel dryness and wetness estimation method based on machine learning of the present invention refers to the data collected in real time by various sensors during the boiler combustion process. In this method, the boiler combustion condition signal data collected is the boiler combustion condition signal data within a certain period before the test time point.

[0018] Preferably, the biomass fuel dryness and wetness estimation method based on machine learning in this invention collects boiler combustion condition signal data within 30 minutes before the test time point.

[0019] Furthermore, in step S3 of the biomass fuel dryness and wetness estimation method based on machine learning of the present invention, the time-series data of the sensor signals selected in the previous step are segmented and resampled to generate Epoch data, including:

[0020] S31. First, the time-series data of multiple sensor signals are segmented according to the grate vibration signal. The grate vibration signal is segmented into segments according to the start and end times of the grate vibration signal, and the data between two adjacent grate vibrations is retained, that is, the data between the end of the previous grate vibration and the start of the next grate vibration.

[0021] S32. Filter the data segments obtained from the segmentation. If the time length corresponding to a certain data segment exceeds the preset threshold, discard the next data segment of that data segment.

[0022] S33. Using the interpolation resampling method, all data segments obtained in the previous step are grouped into the same length to generate a two-dimensional array with the same shape, which is the Epoch data.

[0023] Furthermore, in step S4 of the biomass fuel dryness and wetness estimation method based on machine learning of the present invention, the CSP algorithm is used to extract features from the Epoch data generated in the previous step. The algorithm is as follows:

[0024] Design a spatial filter to distinguish between two sets of operating condition signals related to the combustion of dry and wet fuels in the boiler with optimal variance;

[0025] The CSP filter is labeled as:

[0026] s(t)=W T x(t),

[0027] Where W∈R m×d Let S be the spatial filter matrix, where S = (s1, s2, ..., s...). d ) T ∈Rd×times Let x ∈ R be the filtered signal matrix. m×times The original Epoch signal matrix is ​​given by t, where t is time.

[0028] At this point, the optimization objective of CSP is:

[0029] maximize tr W T ∑1W,

[0030] subject to W T (∑1+∑2)W=I,

[0031] The above formula indicates that condition W is satisfied. T Given (∑1+∑2)W=I, maximize trW T ∑1W, where, Both are sample covariance matrices of the original Epoch signal matrix. Exp(.) calculates the exponent over all elements of the input array, tr(.) sums the diagonal elements of the input square matrix array, and W... T This is the transpose of matrix W;

[0032] Finally, the generalized eigenvalue decomposition method is used to optimize CSP.

[0033] Furthermore, in step S4 of the biomass fuel dryness and wetness estimation method based on machine learning of the present invention, the feature vectors that are easily distinguishable from the two types of Epoch data and can be fed into a linear classifier are extracted. The feature vectors fed into the linear classifier are:

[0034] f = (f1, f2, ..., f d ) T ,

[0035] in,

[0036]

[0037] Here, log(.) calculates the natural logarithm over all elements in the input array, and var[.] calculates the variance over the elements of the input one-dimensional flat array. The goal is to sum the d elements in the input array.

[0038] Furthermore, in step S5 of the biomass fuel moisture estimation method based on machine learning of the present invention, the LDA algorithm is used to classify the feature vector obtained in the previous step. The feature data fed into the LDA classifier comes from the Epoch data matrix obtained by processing the boiler combustion condition signal within a certain period before the test time point. The number of features is exactly the same as the number of Epoch data matrices. The LDA classifier estimates all the input feature data to obtain the classification prediction label of the Epoch data. The obtained classification prediction label is used to calculate the fuel moisture score.

[0039] Furthermore, in step S6 of the biomass fuel moisture estimation method based on machine learning of the present invention, the calculation of fuel moisture content using classification prediction labels of Epoch data is performed by the proportion of samples with classification prediction labels of wet category in the total test samples, and the obtained fuel moisture content score ranges from 0 to 1.

[0040] Furthermore, the biomass fuel moisture content estimation method based on machine learning in this invention can be used in conjunction with the fuel addition / reduction prediction model. The moisture content of the fuel currently fed into the furnace is calculated by the fuel moisture content estimation method. This value is then used as a coefficient and processed with the current fuel addition / reduction value to obtain a more accurate fuel addition / reduction value.

[0041] In addition, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described machine learning-based biomass fuel moisture estimation method.

[0042] In summary, this invention mainly designs and improves a method for real-time determination of the dryness and wetness of fuel in the furnace of a biomass power plant. This method has the following advantages:

[0043] (1) This method designs pre- and post-processing methods for the working condition signal data of furnace combustion. The pre-processing includes steps such as segmentation, resampling and variance calculation to convert the time series data of sensor signals into features that can be directly input into the machine learning classifier. The post-processing includes converting the model output results into the score of fuel wetness.

[0044] (2) This method uses the Common Space Pattern (CSP) algorithm to filter the signal data to obtain signal data that is easy to classify, and uses the Linear Discriminant Analysis (LDA) algorithm to classify the feature vectors.

[0045] (3) This method and model are particularly suitable for judging the dryness and wetness of fuel during the power generation process of biomass power plants. This method adopts a machine learning strategy, which can automatically adjust the model and parameters according to the changes in data to adapt to different combustion conditions, thereby providing more accurate results. At the same time, the machine learning method can effectively process large-scale data and extract useful information and patterns from it. Moreover, the machine learning method can automatically make decisions or predictions without manually writing a series of rules, making the decision-making process more efficient and faster, and can meet real-time decision-making needs. Compared with the manual logic judgment model, this method has a simpler process and significantly improved performance (classification accuracy and running speed).

[0046] (4) This method can work in conjunction with the fuel addition and reduction prediction model to output more accurate and reasonable fuel addition and reduction instructions. Attached Figure Description

[0047] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments of the present invention will be briefly introduced below. Obviously, the following drawings are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a schematic diagram illustrating the operation of the fuel dryness / wetness estimation model of the present invention.

[0049] Figure 2 This is a flowchart illustrating the overall implementation process of the method of the present invention.

[0050] Figure 3 This is a schematic diagram illustrating the generation of Epoch data in the method of the present invention.

[0051] Figure 4 This is a schematic diagram of CSP-based feature extraction in the method of the present invention.

[0052] Figure 5 This is a schematic diagram illustrating the coordinated operation of the fuel dryness / wetness judgment model and the fuel addition / reduction prediction model in the method of this invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. This invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this invention.

[0054] At the same time, it should be understood that the scope of protection of the present invention is not limited to the specific embodiments described below; it should also be understood that the terminology used in the embodiments of the present invention is for describing specific embodiments and not for limiting the scope of protection of the present invention.

[0055] Example: A machine learning-based method for estimating the wet / dry properties of biomass fuels

[0056] like Figure 1 and Figure 2 As shown, the overall process of this method includes: collecting boiler combustion condition signal data, which is real-time data collected by various sensors during boiler combustion. In this method, the data collected and used to estimate fuel dryness / wetness is boiler combustion condition signal data from a past period (e.g., 30 minutes before the current moment). Since there are numerous sensors detecting boiler combustion (more than 100), the first step is to filter out sensor signals closely related to fuel dryness / wetness based on the experience of skilled workers. Then, the time-series data of the selected sensor signals is segmented according to the start and end points of the grate vibration signal. These segments are then interpolated and resampled to generate a two-dimensional array with a consistent shape, thus obtaining Epoch data. Finally, the CSP algorithm is used to extract features from the Epoch data, extracting easily distinguishable feature vectors that can be fed into a linear classifier from the two types of Epoch data. Finally, these feature vectors will be fed into LDA for classification, yielding classification prediction labels for multiple epochs of data. These classification prediction labels will be used to calculate the fuel wetness score, specifically the proportion of samples with the wet category in the total test samples. This score ranges from 0 to 1. The following describes each step in this method.

[0057] like Figure 3 As shown, after signal selection, the next step is to generate Epoch data from the time-series data of the multiple sensor signals. First, the time-series data of the multiple sensor signals (multi-feature) are segmented based on the grate vibration signal, retaining the data between two adjacent grate vibrations, i.e., the data between the end of one grate vibration and the start of the next. For the segmented data segments, if the time length of a certain data segment exceeds a preset threshold, such as 4 minutes, the next data segment will be discarded because the fuel in the furnace has been baked for too long, and the combustion signal after the next grate vibration will lose its reference value. After data segmentation, because the lengths of these data segments are not consistent, a resampling method is needed to group each data segment to the same length. Here, an interpolation resampling method is used to extract the first two minutes of data points from all data segments, forming a batch of Epoch data.

[0058] like Figure 4 As shown, the CSP algorithm is used to extract features from the generated Epoch data. The specific algorithm is as follows: First, a spatial filter is designed to distinguish the two sets of operating condition signals related to the combustion of dry and wet fuels in the boiler with optimal variance. Second, the CSP filter is labeled as follows:

[0059] s(t)=W T x(t),

[0060] Where W∈R m×d Let S be the spatial filter matrix, where S = (s1, s2, ..., s...). d ) T ∈R d×times Let x ∈ R be the filtered signal matrix. m×times t represents the original Epoch signal matrix, where t is time.

[0061] At this point, the optimization objective of CSP is:

[0062] maximize tr W T ∑1W,

[0063] subject to W T (∑1+∑2)W=I,

[0064] The above formula indicates that condition W is satisfied. T Given (∑1+∑2)W=I, maximize tr W T ∑1W, where, Both are sample covariance matrices of the original Epoch signal matrix. This CSP optimization problem can be solved using the generalized eigenvalue decomposition method.

[0065] Finally, the feature vector fed into the classifier is f = (f1, f2, ..., f...). d ) T ,in

[0066] In this function, log(.) calculates the natural logarithm for all elements in the input array, var[.] calculates the variance for the elements in the one-dimensional flat array, and Exp(.) calculates the exponent for all elements in the input array. The function `tr(.)` sums the d elements of the input array, and `W` sums the diagonal elements of the input square array. T Let W be the transpose of matrix W.

[0067] The classifier here will use LDA. Note that the feature data fed into the classifier comes from the Epoch data matrix obtained by processing the operating conditions time series signal 30 minutes before the model is called. The number of features is the same as the number of Epoch data matrices. The classifier will predict all the input feature data, and the resulting labels will be used to calculate the fuel wetness score.

[0068] One important function of the fuel dryness / wetness assessment model is to assist the fuel addition / subtraction prediction model in working in real time. Figure 5 An example of two models operating in tandem is provided. The predicted values ​​output by the fuel addition / reduction prediction model are fed back into the DCS system, directly affecting the instantaneous speed of the conveyor belt, which in turn determines the amount of fuel fed into the boiler per unit time, thus directly impacting the boiler's power generation. When the wetness score of the fuel currently fed into the furnace is obtained by calling the fuel dryness / wetness judgment model, this value can be used as a coefficient to calculate a more reasonable and accurate addition / reduction value, thereby making boiler combustion more stable and efficient.

[0069] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, substitutions, etc., made within the spirit and principle of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for estimating dryness of biomass fuel based on machine learning, characterized by, The method comprises: S1. Collecting boiler combustion condition signal data; S2. Selecting sensor signals according to artificial logic judgment experience, and screening out sensor signals closely related to fuel dryness and wetness; S3. Splitting and resampling the time series data of the sensor signals screened in the previous step to generate Epoch data, comprising: S31. First, split the time series data of multiple sensor signals according to the grate vibration signal, and split into segments according to the starting and ending time points of the grate vibration signal, and retain the data between adjacent two times of grate vibration, i.e. the data between the end of the last time of grate vibration and the start of the next time of grate vibration; S32. If the time length of a certain data segment exceeds the pre-set threshold, discard the next data segment of the data segment; S33. Use the interpolation resampling method to divide all the data segments obtained in the previous step into the same length to generate a two-dimensional array with consistent shape, i.e. Epoch data; S4. Use the CSP algorithm to extract features from the Epoch data generated in the previous step, and the algorithm is as follows: Design a spatial filter to distinguish the optimal variance of two groups of working condition signals related to dry and wet fuel combustion in the boiler; The CSP filter is marked as: ， wherein, is a spatial filter matrix, is a filtered signal matrix, is a raw Epoch signal matrix, t is time; At this time, the optimization target of CSP is: ， ， The above equation represents the maximization of under the condition where , are the sample covariance matrices of the original Epoch signal matrix, is the computation of the exponential of all elements in the input array, is the summation of the diagonal elements of the input array of square matrices, is the transpose of the matrix W; Finally, use the generalized eigenvalue decomposition method to complete the optimization of CSP; Then extract the feature vectors from the two types of Epoch data which are easy to distinguish and can be fed into a linear classifier, and the feature vectors are: ， Wherein, ， wherein, is the natural logarithm of all elements in the input array, is the variance of the elements in the input one-dimensional flat array, is the cumulative sum of d elements in the input array; S5. Use the LDA algorithm to classify the feature vectors obtained in the previous step to obtain the classification prediction label of the Epoch data; S6. Calculate the fuel dryness and wetness degree using the classification prediction label of the Epoch data obtained in the previous step to obtain the fuel wetness degree score value.

2. The machine learning based biomass fuel dryness level estimation method according to claim 1, wherein, The boiler combustion condition signal data in step S1 refers to the data collected by various sensors in real time during the combustion process of the boiler. In this method, the boiler combustion condition signal data in the previous period of the test time point is collected.

3. The machine learning based biomass fuel dryness level estimation method according to claim 2, wherein, The boiler combustion condition signal data in the previous period of the test time point is the boiler combustion condition signal data in the previous 30 minutes of the test time point.

4. The machine learning based biomass fuel dryness level estimation method according to claim 1, wherein, In step S5, the feature data fed into the LDA classifier comes from the Epoch data matrix obtained by processing the boiler combustion condition signal data in the previous period of the test time point. The number of features is completely consistent with the number of Epoch data matrices. The LDA classifier estimates all incoming feature data to obtain the classification prediction label of the Epoch data, and the obtained classification prediction label is used to calculate the fuel wetness degree score value.

5. The machine learning based biomass fuel dryness estimation method according to claim 4, wherein, In step S6, the fuel dryness and wetness degree is calculated using the classification prediction label of the Epoch data obtained in the previous step. The calculation method is the proportion of the number of samples with wet category in the total test samples, and the obtained fuel wetness degree score value ranges from 0 to 1.

6. The machine learning based biomass fuel dryness estimation method according to claim 1, wherein, The fuel dry-wet degree estimation method can be cooperated with a fuel addition / subtraction prediction model, and a wet degree score value of the fuel fed into the current furnace is calculated by the fuel dry-wet degree estimation method, and the value is taken as a coefficient and the current fuel addition / subtraction value is operated, so that a more accurate fuel addition / subtraction value can be obtained. 7.A computer readable storage medium, having stored thereon a computer program, which when executed by a processor implements the steps of the biomass fuel dry-wet degree estimation method based on machine learning according to any one of claims 1-6.

Images