Method for predicting CO emission in MSWI process

Abstract

The invention provides an MSWI process CO emission prediction method, and relates to the technical field of MSWI. Comprising the following steps: acquiring a historical data set of CO emission; constructing an initial CO emission prediction model and an initial kernel feature space detection model according to the historical data set; performing CO prediction on to-be-detected input data by using the initial CO emission prediction model to obtain initial prediction data; performing concept drift identification on the feature space of the to-be-detected input data by using an unsupervised kernel principal component analysis method and the initial kernel feature space detection model to obtain a drift sample, and merging the drift sample and the historical data set to obtain an updated data set; and updating the initial CO emission prediction model and the initial kernel feature space detection model by using the updated data set, and performing prediction by using the updated models to obtain final prediction data. According to the method, the problems that the CO prediction model in the MSWI process does not consider that the true value of the model cannot be provided when the CEMS system fails and the prediction model cannot be updated when the concept drift occurs, so that the CO prediction is not accurate are solved.

Description

Technical Field

[0001] The invention relates to the technical field of MSWI, and in particular to a method for predicting CO emissions in a MSWI process. Background Art

[0002] Municipal solid waste (MSW) refers to various types of solid waste generated in people's daily production, life and consumption. Long-term accumulation and lack of management can easily cause pollution, occupation and virus transmission in urban living spaces. The latest statistics from the World Bank show that if no action is taken, the annual global MSW production will increase by 70% to 3.4 billion tons in 2050.

[0003] Carbon monoxide (CO), as one of the toxic gases produced during the MSWI process, is difficult to be perceived by the senses. However, when it combines with hemoglobin, it will hinder the oxygen delivery of the blood. In severe cases, it may even cause myocardial infarction in the human body. Therefore, it must be strictly controlled.

[0004] Due to the unknown uncertainty fluctuations of MSW components, the high-frequency intervention and experience differences of domain experts in operation, the maintenance and aging of incineration equipment, etc., the operating condition drift of the MSWI process has complex dynamic time-varying characteristics. This kind of operating condition fluctuation that is difficult to overcome and predict in industrial processes is reflected in the distribution change of data streams over time, which is called concept drift in academia. Therefore, building a high-precision CO prediction model requires not only determining a suitable modeling algorithm, but also solving how to detect concept drift and using samples that can characterize drift to update the historical model. Although conventional pollutants including CO, NOx, SO2, etc. can be continuously detected by the CEMS system located at the tail of the chimney, the system has unavoidable inaccuracy and failure phenomena. The existing CO prediction models for the MSWI process include methods based on simplified deep features and LSTM optimization, methods based on fixed window drift detection, and heterogeneous integrated prediction model methods driven by virtual and real data. However, these methods do not consider the update of the prediction model when the CEMS system fails and cannot provide the model truth value and concept drift occurs. Therefore, when updating the CO emission concentration prediction model, it is also necessary to pay attention to how to effectively update the model in the absence of truth value. Summary of the invention

[0005] In order to overcome the shortcomings of the prior art, the purpose of the present invention is to provide a method for predicting CO emissions in a MSWI process. The present invention solves the problem that the CO prediction model for the MSWI process in the prior art does not take into account the inability to provide the model truth value when the CEMS system fails and the concept drift occurs, resulting in inaccurate CO prediction.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A method for predicting CO emissions from a MSWI process, comprising:

[0008] Obtain historical datasets of CO emissions;

[0009] Constructing an initial CO emission prediction model and an initial core feature space detection model based on the historical data set;

[0010] Using the initial CO emission prediction model to perform CO prediction on the input data to be tested to obtain initial prediction data;

[0011] Using an unsupervised kernel principal component analysis method and the initial kernel feature space detection model to identify concept drift in the feature space of the input data to be tested, if drift occurs, constructing a pseudo-labeling model and obtaining a pseudo-true value, and confirming the drift sample based on a PH test to obtain a drift sample, if concept drift does not occur, CO emissions are predicted based on the initial CO prediction model;

[0012] Merging the drift sample and the historical data set to obtain an updated data set;

[0013] Using the updated data set to update the initial CO emission prediction model and the initial kernel feature space detection model to obtain a final CO emission prediction model and a final kernel feature space detection model;

[0014] The final CO emission prediction model is used to perform CO prediction on the input data to be tested to obtain final prediction data.

[0015] Preferably, if drift occurs, a pseudo-label model is constructed and a pseudo-true value is obtained, and drift sample confirmation is performed based on a PH test to obtain a drift sample, including:

[0016] Obtain historical pseudo-label datasets;

[0017] According to the historical pseudo-label data set, determine whether the detection value of the input data to be tested exceeds the KPCA control limit. If so, store the sample in the cache window to be labeled. When the number of samples in the window reaches the preset window capacity, build a pseudo-label model in real time and mark the drift sample with a pseudo-true value.

[0018] The Page-Hinkley detection method is used to determine whether the sample has drifted for the marked pseudo-true value. If so, a drifted sample is obtained.

[0019] Preferably, it also includes:

[0020] Updating the historical pseudo-label dataset to obtain an updated historical pseudo-label dataset;

[0021] Resetting the cache window to be marked to obtain the reset cache window to be marked;

[0022] The final CO emission prediction model and the final nuclear feature space detection model are updated in real time according to the reset to-be-annotated cache window and the updated historical pseudo-label data set.

[0023] Preferably, the expression of the initial CO emission prediction model is:

[0024]

[0025] in, is the initial CO emission prediction model, It is an LSTM model built based on historical data. It is also the prediction model used when no drift occurs. The nth input data to be tested is used as the input of the LSTM model.

[0026] Preferably, the expression of the initial kernel feature space detection model is:

[0027]

[0028] in, is the drift index of the sample in the first window, SPE k is the drift index of the samples in the second window, P t is the nonlinear principal component matrix, is a KPCA model built based on historical data. Construct the initial kernel feature space detection model.

[0029] Preferably, the expression of the final CO emission prediction model is:

[0030]

[0031] in, To update the dataset, is the updated LSTM model, It is the final CO emission prediction model.

[0032] The present invention discloses the following technical effects:

[0033] The present invention provides a method for predicting CO emissions in a MSWI process, comprising: obtaining a historical data set of CO emissions; constructing an initial CO emission prediction model and an initial kernel feature space detection model according to the historical data set; using the initial CO emission prediction model to perform CO prediction on input data to be tested to obtain initial prediction data; using an unsupervised kernel principal component analysis method and the initial kernel feature space detection model to perform concept drift recognition on the feature space of the input data to be tested, if drift occurs, constructing a pseudo-labeling model and obtaining a pseudo-true value and performing drift sample confirmation based on a PH test to obtain a drift sample, if concept drift does not occur, performing CO emission prediction based on the initial CO prediction model; merging the drift sample and the historical data set to obtain an updated data set; using the updated data set to update the initial CO emission prediction model and the initial kernel feature space detection model to obtain a final CO emission prediction model and a final kernel feature space detection model; using the final CO emission prediction model to perform CO prediction on the input data to be tested to obtain final prediction data. The present invention can not only use historical data for prediction, but also continuously learn and adapt to new data. When the CEMS system fails and cannot provide the model truth value and concept drift occurs, the prediction model can continue to be updated, thereby improving the effectiveness in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0035] Figure 1 A flow chart of a method for predicting CO emissions from a MSWI process provided by an embodiment of the present invention;

[0036] Figure 2 A modular schematic diagram of a CO emission prediction strategy for a MSWI process provided by an embodiment of the present invention;

[0037] Figure 3 A schematic diagram of the LSTM structure provided in an embodiment of the present invention;

[0038] Figure 4 The absolute value error curve of the training set provided by the embodiment of the present invention;

[0039] Figure 5 A true value difference curve diagram of a historical data set of a pseudo-label model provided in an embodiment of the present invention;

[0040] Figure 6 The T provided in the embodiment of the present invention 2Graphs;

[0041] Figure 7 The SPE curve diagram provided by the embodiment of the present invention;

[0042] Figure 8 A comparison curve diagram of pseudo-true value labeling results and actual true values ​​provided in an embodiment of the present invention;

[0043] Fig. 9 A curve chart of pseudo-marker drift detection results provided by an embodiment of the present invention;

[0044] Fig.10 This is a prediction performance change curve diagram provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0046] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0047] like Figure 1 As shown, the present invention provides a method for predicting CO emissions in a MSWI process, comprising:

[0048] Step 100: Obtain a historical data set of CO emissions;

[0049] Step 200: constructing an initial CO emission prediction model and an initial core feature space detection model according to the historical data set;

[0050] Step 300: using the initial CO emission prediction model to perform CO prediction on the input data to be tested to obtain initial prediction data;

[0051] Step 400: using an unsupervised kernel principal component analysis method and the initial kernel feature space detection model to identify concept drift in the feature space of the input data to be tested; if drift occurs, constructing a pseudo-labeling model and obtaining a pseudo-true value, and performing drift sample confirmation based on a PH test to obtain a drift sample; if concept drift does not occur, performing CO emission prediction based on the initial CO prediction model;

[0052] Step 500: merging the drift sample and the historical data set to obtain an updated data set;

[0053] Step 600: using the updated data set to update the initial CO emission prediction model and the initial kernel feature space detection model to obtain a final CO emission prediction model and a final kernel feature space detection model;

[0054] Step 700: Use the final CO emission prediction model to perform CO prediction on the input data to be tested to obtain final prediction data.

[0055] Specifically, MSWI includes storage and transportation fermentation, solid waste combustion, waste heat exchange, steam power generation, flue gas treatment and flue gas emission. In the solid waste combustion stage, due to insufficient oxygen supply (part of the MSW cannot be completely oxidized), uneven combustion (some areas of the MSW cannot be fully exposed to oxygen) or insufficient temperature (incineration temperature is insufficient, the combustion reaction rate is reduced), the MSW will be incompletely burned, resulting in the generation of CO. Therefore, changes in factors such as furnace temperature and oxygen content will change the CO generation relationship and produce concept drift.

[0056] The traditional MSWI process can detect the CO emission concentration in real time through the exhaust gas automatic monitoring system (CEMS), which is usually measured by complete extraction or diluted extraction. However, when the complete extraction is in a positive pressure environment or the extraction volume is too large, the suction port is prone to blockage. When the diluted extraction is used, the measurement response is too long and the purity of the dry compressed air is high. In addition, the CEMS requires regular maintenance by qualified technicians. Therefore, it is necessary to annotate the pseudo-true value to analyze the concept drift phenomenon in the process when the full true value cannot be obtained.

[0057] Specifically, for the prediction model, first, a CO emission prediction model and a kernel feature space detection model are constructed based on historical data, and a historical data set is obtained for real-time construction of a pseudo-labeling model; secondly, drift detection for the CO emission prediction model is performed based on real-time data, and unsupervised kernel principal component analysis (KPCA) is used to identify concept drift in the feature space. If drift occurs, a pseudo-labeling model is constructed and pseudo-true values ​​are obtained, and drift samples are confirmed based on the PH test. If no concept drift occurs, CO emission prediction is performed based on the historical prediction model; then, an updated data set for the CO emission prediction model and kernel feature space detection is obtained by combining historical samples and drift samples; finally, the historical model is updated with a new data set when the preset conditions are met.

[0058] Furthermore, if drift occurs, a pseudo-label model is constructed and a pseudo-true value is obtained, and drift sample confirmation is performed based on a PH test to obtain a drift sample, including:

[0059] Obtain historical pseudo-label datasets;

[0060] According to the historical pseudo-label data set, determine whether the detection value of the input data to be tested exceeds the KPCA control limit. If so, store the sample in the cache window to be labeled. When the number of samples in the window reaches the preset window capacity, build a pseudo-label model in real time and mark the drift sample with a pseudo-true value.

[0061] The Page-Hinkley detection method is used to determine whether the sample has drifted for the marked pseudo-true value. If so, a drifted sample is obtained.

[0062] Furthermore, it also includes:

[0063] Updating the historical pseudo-label dataset to obtain an updated historical pseudo-label dataset;

[0064] Resetting the cache window to be marked to obtain the reset cache window to be marked;

[0065] The final CO emission prediction model and the final nuclear feature space detection model are updated in real time according to the reset to-be-annotated cache window and the updated historical pseudo-label data set.

[0066] Specifically, Figure 2 As shown, each step is modularly understood, including: 1) Construction of CO prediction model and drift detection model based on historical data: historical samples are used to build a prediction model and a kernel feature space detection model to obtain a historical data set to support the pseudo-labeling model.

[0067] 2) Drift detection of CO prediction model based on real-time data: KPCA (unsupervised kernel principal component analysis) is used to detect drift in the feature space of new samples. When the detection value exceeds the KPCA control limit, the sample is considered to have the possibility of drift. At this time, the sample is stored in the cache window to be labeled. When the number of samples in the window reaches the preset window capacity, a pseudo-labeling model is constructed in real time and the drifting samples are labeled with pseudo-true values. The Page-Hinkley detection method is used to determine whether the sample has drifted for the labeled pseudo-true values. If no concept drift occurs, CO concentration prediction is performed based on the historical prediction model.

[0068] 3) Update data set acquisition: The samples confirmed to have drifted in the current cache window are recorded as drift sample data sets, which are merged with the historical data sets to determine the data set used for model update based on experience.

[0069] 4) Historical model update: The acquired updated data set is used to retrain the CO prediction model and the core feature space detection model, and the historical data set of the pseudo-labeling model is updated, and the cache window to be labeled is reset.

[0070] Furthermore, the expression of the initial CO emission prediction model is:

[0071]

[0072] in, is the initial CO emission prediction model, It is an LSTM model built based on historical data. It is also the prediction model used when no drift occurs. The nth input data to be tested is used as the input of the LSTM model.

[0073] Specifically, the goal of model learning is to determine the prediction model with the minimum prediction error. The LSTM-based model construction consists of Figure 3 shown.

[0074] Figure 3 middle, represents the nth sample input, h n-1 and c n-1 Respectively represent n-1 The sample LSTM hidden layer output and state information are used as the input of the next sample, h n and c n Represent the nth sample LSTM output and state information respectively.

[0075] 1) In the initial sample calculation, the parameters are set to n = 1 and h n-1 =0, the memory cell c n Set to an empty set, forget gate f n Determine memory unit c n The n-1th sample output value is output to c n The nth sample is the ratio of the input this time, and its output The calculation of is as follows:

[0076]

[0077] Among them, U f and W f Respectively represent h n-1 and The weight of f represents bias; σ(·) represents the sigmoid activation function; ⊙ represents the Hadamard product.

[0078] Candidate value for the current input state The calculation of is as follows:

[0079]

[0080] Among them, U c and W c Respectively represent h n-1 and The weight of c represents bias; tanh(·) represents the tanh activation function.

[0081] Input Gate i n Used to control the nth sample input to be stored in c n The ratio of its output The calculation of is as follows:

[0082]

[0083] Among them, U i and W i Indicates h n-1 and The weight of b i Indicates bias.

[0084] The value obtained by the above calculation is stored in the memory cell c n As shown below:

[0085]

[0086] also, Indicates the first sample output c n The value of and

[0087] Output gate o n Used to control memory unit c n The value stored in can be output as a ratio, and the formula for obtaining it and the hidden layer output of the nth sample are as follows:

[0088]

[0089] h n =o n ⊙tanh(c n ) (8)

[0090] Among them, U o and W o Respectively represent h n-1 and The weight of b o Indicates bias.

[0091] The output value of the nth sample The calculation of is as follows:

[0092]

[0093] Among them, W out Represents the weight corresponding to the hidden layer.

[0094] Similarly, the above process is repeated for the remaining samples. Represents the output corresponding to all samples, as shown below:

[0095]

[0096] The loss function e′ of the LSTM model is calculated as follows:

[0097]

[0098] Among them, y n and represents the actual value and predicted value of the nth sample, and e′ represents the error between the two.

[0099] When updating LSTM parameters, there is also a hidden state c n and h n , which is defined as and The corresponding calculation formula is as follows:

[0100]

[0101] Furthermore, the weights and biases of the forget gate, input gate, and output gate are calculated. Taking the output gate as an example, its weight learning formula is as follows:

[0102]

[0103] The corresponding weight update formula is as follows:

[0104]

[0105] in, Represents the updated output gate weight.

[0106] The output of the LSTM-based CO concentration prediction model built using offline samples can be expressed as follows:

[0107]

[0108] Furthermore, the expression of the initial kernel feature space detection model is:

[0109]

[0110] Specifically, for high-dimensional process variables, the present invention uses a KPCA model to calculate a drift index control limit for determining whether a MSWI process has changed.

[0111] 1) Get training data

[0112] Get the dataset

[0113] 2) Calculate the kernel matrix

[0114] Will The vector x in Htd Mapped into high-dimensional space F (denoted as D dimension):

[0115] φ(x Htd ): R M →R D , D ≥ M (16)

[0116] Among them, φ(·) is a nonlinear mapping function. D is the dimension of the data after mapping.

[0117] Then the high-dimensional mapping space is as follows:

[0118]

[0119] A kernel function needs to be introduced to calculate the kernel matrix K. The present invention adopts the Sigmoid kernel function:

[0120]

[0121] Among them, f(·) uses the tanh function. β0 and β1 are user-defined. Generally, β1 defaults to 0.

[0122] 3) Core Matrix Centralization

[0123] Define a 1 matrix 1 N , and centralize it as:

[0124]

[0125] Then the centralized kernel matrix K c The calculation is as follows:

[0126] K c =K N×N -U N×N K N×N -K N×N U N×N +U N×N K N×N U N×N (20)

[0127] Among them, K c After calculation, the size is still N×N.

[0128] 4) Eigenvalue decomposition and eigenvector standardization

[0129] The matrix eigenvalues ​​and eigenvectors are solved by eigenvalue decomposition method, and the calculation is as follows:

[0130]

[0131] Where Q is the eigenvector matrix, Σ=diag(λ1,λ2,…,λ N ) is a diagonal matrix, and each element on the diagonal is an eigenvalue.

[0132] The eigenvalues ​​are normalized by the following formula and calculated as follows:

[0133]

[0134] Among them, Q s is the normalized matrix of Q, and λ is the eigenvector.

[0135] 5) Selection of the number of principal components

[0136] Using the cumulative feature contribution rate η and PCA contribution threshold δ PCA Perform dimensionality reduction:

[0137]

[0138] Among them, p is the number of principal components selected, M is X Htd The number of features, p is less than M.

[0139] Based on this, the drift control limit δ used to monitor whether the operating conditions have changed is determined by p PCA .

[0140] 6) Calculate nonlinear principal components

[0141] The nonlinear principal component is calculated by the number of principal components obtained, as follows:

[0142] P=K c Q s_p (twenty four)

[0143] Among them, Q s_p Q s Take a matrix of size N×p, P is the nonlinear principal component matrix, a matrix of size N×p.

[0144] 7) Calculate statistical control limits

[0145] According to Hotelling's T 2 The SPE statistical index is used as the control limit for the identification of operating condition drift. The statistical index can be expressed as:

[0146]

[0147] Among them, F α (p,Np) is an F distribution with p and (Np) degrees of freedom; c αrepresents the normal deviation from the upper (1-α) percentile; θ i The calculation of and h0 is as follows:

[0148]

[0149] Thus, a historical concept drift detection model is constructed.

[0150] Furthermore, we calculate the historical sample set The first-order difference between the output of each sample and the feature space is expressed as For example, the calculation is as follows:

[0151]

[0152] in, and Respectively represent the feature space of the samples; and They represent the true values ​​of the samples respectively.

[0153] According to the above formula, the first-order difference components of the historical sample output and the feature space are calculated as Δy Htd and ΔX Htd .

[0154] Furthermore, 1) obtain test data

[0155] Get the dataset

[0156] 2) Calculate the kernel matrix

[0157] Will Vector in Map it into the high-dimensional space F, and introduce the kernel function to calculate the kernel matrix K t (Size is N k ×N). Using Sigmoid kernel function:

[0158]

[0159] Among them, f(·) uses the tanh function.

[0160] 3) Core Matrix Centralization

[0161] Define a size of N k ×N 1 matrix Centralize it as:

[0162]

[0163] Then the centralized kernel matrix K t_c The calculation is as follows:

[0164]

[0165] Among them, K t_c The calculated size is N k ×N.

[0166] 4) Calculate nonlinear principal components

[0167] The nonlinear principal component is calculated by the number of principal components obtained, as follows:

[0168] P t =K t_c Q s_p (34)

[0169] Among them, Q s_p is a matrix of size N×p. t is the size N k ×p matrix.

[0170] 5) Calculate statistical index values

[0171] The drift value of the new sample is calculated based on the following formula:

[0172]

[0173] SPE k =∑(K t_c Q s ) 2 -∑P t (36)

[0174] Based on the samples in the window (drift indicator of samples in the first window) and SPE k Whether (drift index of samples in the second window) satisfies the following conditions is determined to determine whether drift occurs.

[0175]

[0176] Furthermore, after KPCA screening and the cache window to be labeled is filled, the sample set in the window is recorded as Where w is the preset cache window sample capacity, and the specific labeling strategy is:

[0177] Construct a new set of first-order difference components as:

[0178]

[0179] in, and are the true value and feature space respectively; and for The first-order difference calculated together with the last sample in the current training set.

[0180] from , calculate its The feature space The first-order difference of is:

[0181]

[0182] Based on the nearest neighbor idea, we can use the Euclidean distance from ΔX Htd ′Select The ε feature space difference components with the smallest distance are recorded together with their corresponding output space difference components as follows:

[0183]

[0184] Using Ω nearest Establish a new LSTM prediction model for The output spatial difference To take a measurement:

[0185]

[0186] Then The pseudo-truth value of Marked as:

[0187]

[0188] Repeat the above process until all samples in the window are labeled with pseudo-true values.

[0189] Furthermore, a reasonable analysis of the difference between the sample pseudo-true value and the measured value is the key to confirming the final concept drift of the sample. In the Page-Hinkley detection method, given a series of observations [l1,l2,…l m ], the likelihood ratio statistic of the alternative hypothesis (there is a drift point γ in the observed value, that is, 1<γ<m) to the null hypothesis (there is no drift in the observed value, that is, γ>m) is calculated as:

[0190]

[0191] in, f D (·) represents the distribution density function of the standard normal distribution N(0,1); δ means that the drift sample obeys the normal distribution with mathematical expectation δ.

[0192] The above formula can be expressed logarithmically as:

[0193]

[0194] Based on this, the log-likelihood ratio statistic of the alternative hypothesis (with drift) to the null hypothesis (without drift) is:

[0195]

[0196] By setting the threshold and Z m By making a comparison, you can determine whether there is concept drift within the current series of observations.

[0197] When all samples in the cache window to be labeled have completed the pseudo-true value labeling, the Page-Hinkley detection method is used to detect concept drift in the output space of these samples. Taking the observation value Obs(t) at time t as an example, the detection process is as follows:

[0198] First, calculate the cumulative variable about Obs(t)

[0199]

[0200] in, Represents the mean of all historical observations before time t-1; cumulative variable Represents the difference between the current observation value Obs(t) and the mean of historical observation values.

[0201] Then, by calculating the change index PH t Determine whether the current observation value Obs(t) is abnormal:

[0202]

[0203] Among them, φ t Indicates the minimum cumulative variable value recorded from all current moments; PH t Indicates the accumulated variable at the current time t The difference with the minimum cumulative variable value. When the condition PH is met t When >λ, the observed value Obs(t) is considered abnormal, where λ is the empirical threshold.

[0204] When the cache window to be labeled is filled for the dth time and all samples are labeled, the sample set in the window is Calculate the average measurement error Aveero of samples in the current window d as follows:

[0205]

[0206] in, and They represent the true value and false value of the mth sample in the window respectively.

[0207] On this basis, the observation value Obs(t) is selected as the cumulative average measurement error of the samples in the window when the window is filled for the dth time, that is:

[0208]

[0209] At this time, the cumulative variable Represents the difference between the current cumulative average measurement error and the mean of the historical cumulative average measurement error; φ t Indicates the minimum value of the current record value.

[0210] In addition, when the cache window is filled for the first time, that is, d = 1, At this time, there is a lack of drift judgment basis in the sample output space, so φ t It is expressed as:

[0211]

[0212] Among them, φ0 is the benchmark cumulative average measurement error, which will be obtained based on the average measurement error of the verification sample. At the same time, set λ = 0, that is, when It means that when the cumulative average measurement error in this window is significantly higher than that of historical samples, it is considered that the samples in the window can represent concept drift and are used to construct a new training set.

[0213] Furthermore, when the samples in the cache window are confirmed to have drifted, the present invention constructs a new training set based on the historical samples and the samples in the current window to update the measurement model. As an example, construct a new training set as follows:

[0214]

[0215] in, and They are the feature space and true value set of the updated data set; and They are the feature space and pseudo-true value set of samples in the current window respectively.

[0216] Furthermore, the expression of the final CO emission prediction model is:

[0217]

[0218] in, To update the dataset, is the updated LSTM model, It is the final CO emission prediction model.

[0219] Specifically, the prediction model and feature space drift detection model are constructed using the above-mentioned updated data set.

[0220] based on The new prediction model output of CO concentration can be expressed as follows:

[0221]

[0222] The KPCA detection model (final kernel feature space detection model) can be expressed as follows:

[0223]

[0224] Use the new model to replace the old model and update the pseudo-labeled model historical dataset at the same time, as follows

[0225]

[0226] After the above updates, the proposed prediction model can better adapt to the dynamic changes of the MSWI process. More specifically, the CO prediction model adopts the LSTM model with 128 hidden layer units, 210 training times, 0.022 learning rate, 0.2 learning drop factor and 0.4 dropout rate.

[0227] from Figure 4 It can be seen that the LSTM model has a certain effect on the training samples in the process data set, and the error is very small. Therefore, LSTM can be used to effectively build a prediction model.

[0228] The principal component contribution rate threshold of KPCA is set to 0.91, and 5 principal components are selected as the principal space for running status detection. 2 and SPE control limits were 9.4707 and 0.4246, respectively.

[0229] Δy in the pseudo-labeled model historical dataset Htd The curve is as Figure 5 As shown,

[0230] Depend on Figure 5 It can be seen that most of the true value differences are small, which shows that this method has a certain degree of reliability.

[0231] T during KPCA test 2 And SPE results such as Figure 6 and Figure 7 shown.

[0232] The window capacity set in the present invention is 3, the parameters of the LSTM model constructed for drift sample pseudo-labeling are the same as those of the LSTM model, the number of historical samples selected by Euclidean distance is 15, and the comparison between the pseudo-true value labeling results and the actual true value based on TD learning is as follows: Figure 8 shown.

[0233] from Figure 8 It can be seen that the change trend of the pseudo-true value is similar to the sample true value. Therefore, when the sample true value is difficult to obtain completely, the pseudo-true value can be used to approximate the drift of the sample output space. After the pseudo-true value is marked for the sample with drift in the kernel feature space, the Page-Hinkley method is used to detect the pseudo-label drift. Fig. 9 shown.

[0234] Fig. 9 This is the change of the cumulative average measurement error of samples in the window after each time the cache window to be labeled is filled and the samples in it are labeled with pseudo-true values. Among them, the cache window to be labeled is filled 10 times, and the cumulative average measurement error of samples in the window increases significantly when the drift occurs, and tends to be stable as the model is continuously updated, indicating that the proposed algorithm can effectively detect the concept changes in the sample output space.

[0235] According to the above test results, after the model is updated with a new training set consisting of concept drift samples and historical samples, the measured performance changes are as follows: Fig.10 shown.

[0236] The above results show that after the prediction model adopts the proposed drift detection algorithm, its prediction error is significantly reduced compared with the original model. The proposed algorithm can significantly improve the measurement performance of the model for concept drift samples when the true values ​​of most drift samples are not labeled, and can effectively improve the prediction accuracy of the MSWI process CO concentration prediction model in a drift environment.

[0237] More specifically, the present invention adopts the strategy of combining KPCA and Page-Hinkley detection method to fully reflect the concept drift behavior of new samples in feature space and sample output space; 2) the semi-supervised mechanism based on TD learning is used for pseudo-true value labeling of drift samples in feature space, which provides a new method for semi-supervised concept drift detection for industrial regression problems; 3) the feasibility of the proposed method in practical application is verified by using a real MSWI process data set, and it is shown that it has better performance than existing methods. Further research work includes: improving the pseudo-true value labeling algorithm to improve the accuracy of labeling; considering the mechanism of screening historical modeling samples and mixed pseudo-labeled modeling samples to improve the accuracy of the prediction model.

[0238] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0239] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for predicting CO emissions from a MSWI process, characterized in that: include: Obtain historical datasets of CO emissions; Constructing an initial CO emission prediction model and an initial core feature space detection model based on the historical data set; Using the initial CO emission prediction model to perform CO prediction on the input data to be tested to obtain initial prediction data; Using an unsupervised kernel principal component analysis method and the initial kernel feature space detection model to identify concept drift in the feature space of the input data to be tested, if drift occurs, constructing a pseudo-labeling model and obtaining a pseudo-true value, and confirming the drift sample based on a PH test to obtain a drift sample, if concept drift does not occur, CO emissions are predicted based on the initial CO prediction model; Merging the drift sample and the historical data set to obtain an updated data set; Using the updated data set to update the initial CO emission prediction model and the initial kernel feature space detection model to obtain a final CO emission prediction model and a final kernel feature space detection model; The final CO emission prediction model is used to perform CO prediction on the input data to be tested to obtain final prediction data.

2. The method for predicting CO emissions from a MSWI process according to claim 1, characterized in that: If drift occurs, a pseudo-label model is constructed and a pseudo-true value is obtained, and drift sample confirmation is performed based on PH test to obtain a drift sample, including: Obtain historical pseudo-label datasets; According to the historical pseudo-label data set, determine whether the detection value of the input data to be tested exceeds the KPCA control limit. If so, store the sample in the cache window to be labeled. When the number of samples in the window reaches the preset window capacity, build a pseudo-label model in real time and mark the drift sample with a pseudo-true value. The Page-Hinkley detection method is used to determine whether the sample has drifted for the marked pseudo-true value. If so, a drifted sample is obtained.

3. The method for predicting CO emissions from a MSWI process according to claim 2, characterized in that: Also includes: Updating the historical pseudo-label dataset to obtain an updated historical pseudo-label dataset; Resetting the cache window to be marked to obtain the reset cache window to be marked; The final CO emission prediction model and the final nuclear feature space detection model are updated in real time according to the reset to-be-annotated cache window and the updated historical pseudo-label data set.

4. The method for predicting CO emissions from a MSWI process according to claim 2, characterized in that: The expression of the initial CO emission prediction model is: in, is the initial CO emission prediction model, It is an LSTM model built based on historical data. It is also the prediction model used when no drift occurs. The nth input data to be tested is used as the input of the LSTM model.

5. The method for predicting CO emissions from a MSWI process according to claim 4, characterized in that: The expression of the initial kernel feature space detection model is: in, is the drift index of the sample in the first window, SPE k is the drift index of the samples in the second window, P t is the nonlinear principal component matrix, is a KPCA model built based on historical data. Historical data for constructing the initial kernel feature space detection model.

6. The method for predicting CO emissions from a MSWI process according to claim 2, characterized in that: The expression of the final CO emission prediction model is: in, To update the dataset, is the updated LSTM model, It is the final CO emission prediction model.

Images