A method for real-time prediction of main steam flow in urban solid waste incineration process

Abstract

A method for real-time prediction of main steam flow rate in urban solid waste incineration processes comprises the following steps: 1. Preprocessing historical data of the urban solid waste incineration process, identifying and deleting outliers using a method combining sliding window and Relevance Vector Machine (RVM), and using time-continuous data as a case library; 2. Retrieving data samples using the Just-In-Time Learning (JITL) method based on Euclidean distance and covariance as similarity indicators, obtaining n*K cases with the highest similarity to the time to be predicted; 3. Building a model on the retrieved n*K cases using a mechanism-based improved Support Vector Machine (ISVM), deriving the main steam flow rate value at the target time, and finally storing the result and the operating parameters at the target time as a case in the case library; 4. Repeating steps 2-3 until the incineration process ends, thereby achieving real-time prediction of the main steam flow rate in urban solid waste incineration processes. This invention enables accurate detection of thermal energy conversion status and improves thermal energy utilization efficiency.

Description

Technical Field

[0001] This invention relates to the field of predicting parameters in urban solid waste incineration processes. More specifically, this invention is a method for real-time prediction of the main steam flow rate in urban solid waste incineration processes. Background Technology

[0002] With the continuous advancement of urbanization in my country, the amount of solid waste generated in cities is increasing year by year, and the environmental pollution and land resource occupation caused by solid waste have attracted widespread attention. Incineration is a treatment method for reducing and harmlessly disposing of urban solid waste that is currently being vigorously promoted by the state. Utilizing the high-temperature flue gas generated during solid waste incineration to produce high-temperature, high-pressure steam in a waste heat boiler to drive the operation of subsequent equipment is a common method for utilizing the heat energy generated from solid waste incineration, which can further realize the resource utilization of urban solid waste. Accurate real-time values ​​of steam flow rate are crucial for improving heat energy utilization and monitoring the incineration process. Therefore, the research results of this invention have broad application prospects.

[0003] Currently, most methods for directly measuring the main steam flow rate in urban solid waste incineration processes use orifice plate or nozzle sensors. However, directly installing the sensors inside the steam pipes leads to problems such as throttling losses causing steam pressure drops and sensor malfunctions resulting in abnormal measurements. Furthermore, due to the unstable calorific value of solid waste in my country at present, parameters fluctuate significantly during incineration, and the sensor-measured main steam flow rate lags considerably behind the incineration process within the furnace. Therefore, sensor-based measurement methods are insufficient for accurately calculating the overall thermal system of solid waste incinerators. Thus, real-time prediction of the main steam flow rate during solid waste incineration is of great significance.

[0004] Case-based reasoning is an important problem-solving method. Its core idea is to find similar or identical cases in an existing case library by analogy with previous cases, and then adjust these older cases to solve the current problem. The case-based reasoning process includes four stages: case retrieval, case reuse, case correction, and case storage. The results of the case retrieval stage directly affect the accuracy of the solution, and most of the retrieved similar cases are not suitable for direct application to new problems and require appropriate correction. Therefore, this paper focuses on the accuracy of the case library in the case retrieval stage, the determination of the degree of similarity during the retrieval process, and the correction methods in the case correction stage to predict the main steam flow rate in real time. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a method for real-time prediction of main steam flow in urban solid waste incineration processes. This method can obtain accurate real-time values ​​of the main steam flow in solid waste incinerators, enabling accurate detection of thermal energy conversion and improving thermal energy utilization efficiency.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for real-time prediction of main steam flow rate in urban solid waste incineration processes is characterized by the following steps: 1. Preprocessing historical data of urban solid waste incineration processes, identifying and deleting outliers using a method combining sliding window and Relevance Vector Machine (RVM), and using time-continuous data as a case library; 2. Retrieving data samples using the Just-In-Time Learning (JITL) method based on Euclidean distance and covariance as similarity indicators, obtaining n*K cases with the highest similarity to the time to be predicted; 3. Building a model for the retrieved n*K cases using a mechanism-based improved Support Vector Machine (ISVM), deriving the main steam flow rate value at the target time, and finally storing the result and the operating parameters at the target time as a case in the case library; 4. Repeating steps 2 to 3 until the incineration process ends, to achieve real-time prediction of main steam flow rate in urban solid waste incineration processes. Further specific steps include the following:

[0008] I. Preprocessing of historical data on urban solid waste incineration processes: Using a sliding window-based method combined with Real-Time Visualization (RVM) to identify and remove outliers, and using time-continuous data as a case library; This process consists of four steps: Step 1: Normalizing existing historical data to form an initial case library; Step 2: Initializing outlier detection parameters; Step 3: Performing outlier detection on samples within the window; Step 4: Moving the window until outlier detection of all historical data is completed.

[0009] Step 1: Record the following characteristic variables: x1, average primary combustion chamber temperature x2, average economizer inlet flue gas temperature x3, economizer outlet flue gas temperature x4, economizer outlet water temperature x5, average evaporator inlet flue gas temperature x6, average tertiary superheater inlet flue gas temperature x7, total economizer feedwater volume x8, total desuperheating water volume for primary and secondary superheaters x9, and tertiary superheater outlet steam temperature x1. 10 Boiler drum pressure x 11 3rd stage superheater outlet steam pressure x 12 The historical data and the corresponding main steam flow rate value y1 are normalized and represented as a vector form, forming N source cases, each case denoted as CASE. i It is represented in the following form:

[0010] CASE i =(X i ;Y i ), i = 1, 2, ..., N (1)

[0011] Where N is the total number of source cases; CASE i This is the i-th source case; X iY is the normalized value of the feature variable of the i-th source case; i X is the normalized economizer feedwater flow rate corresponding to the i-th source case and its characteristic variables; i This can be further expressed as:

[0012] X i =(x 1,i ,…,x λ,i ,…,x 12,i (2)

[0013] Where, x λ,i (λ = 1, ..., 12) represents CASE i The value of the λth feature variable after normalization.

[0014] Step 2: Set a relatively large sliding window length L based on the computer's processing power; set an upper bound sigma1 for the kernel function width based on the allowable upper bound of the anomaly severity; aiming for a high anomaly detection rate on the test set with added anomalous samples, repeat the process more than 30 times using trial and error or binary search methods. Select a lower bound sigma2 for the kernel function width between 0.1 and sigma1, a base of the anomaly attribute probability between 100,000 and 200,000, an anomaly attribute probability threshold TH between 0.5 and 1, and a step size Δ for decreasing the kernel function width between 0.01 and 0.2. sigma ; sigma2, a, TH, Δ, whose values ​​change simultaneously sigma The parameter combination with the highest anomaly detection rate on the test set is selected from different combinations.

[0015] Step 3: Outlier detection for the L samples within the window mainly consists of four parts, detailed as follows:

[0016] a. To detect all samples within the detection window, set the initial abnormal attribute probability P of all samples to 1, and set the initial kernel function width sigma of the RVM model to sigma1.

[0017] b. Select L samples within the window as training samples based on the probability of abnormal attributes, and build an RVM model using the kernel function width sigma.

[0018] c. For each of the L samples within the window, use the established RVM model to make predictions, and calculate the relative deviation between the sample values ​​and the predicted values. i

[0019]

[0020] Where i = 1, ..., L represents the sample index within the window, and y i For the actual sample value, The model predicts the values ​​and updates the probabilities of anomalous attributes using the following formula:

[0021]

[0022] in It is the probability value of the new abnormal attribute of sample i.

[0023] d. Repeat parts b and c, each time using a kernel function with a width of sigma. * =sigma-Δ sigma sigma * The width of the new kernel function, up to sigma. * =sigma2, which sets the probability of abnormal attributes. CASE samples below the anomaly attribute probability threshold TH i These samples are marked as abnormal and deleted.

[0024] Step 4: Move the window to the next position until all samples are detected. Increasing the length of the window movement can improve the efficiency of outlier detection.

[0025] 2. The Just-In-Time (JITL) learning method, based on Euclidean distance and covariance as similarity metrics, is used to retrieve data samples and obtain the n*K cases with the highest similarity to the time to be predicted. The case retrieval process consists of two parts, detailed below:

[0026] a. Calculate the Euclidean distance between the normalized feature variables of the data to be predicted and all samples in the case library, and select the first 2*K samples with the smallest distance. These samples are cases where the incinerator operating conditions at a specific time are similar to the operating conditions at the time to be predicted.

[0027] b. Obtain the mean vector of the primary combustion chamber temperature TEMP from the first n cases arranged in chronological order at the time t0 to be predicted. t0 , can be represented as

[0028] TEMP t0 =(T t0-Δt ,…,T t0-j*Δt ,…,T t0-n*Δt ) T ,j=1,2,...n (5)

[0029] Where T t0-j*Δt Let t be the average combustion chamber temperature of the j-th case before time t0, and Δt be the data recording time interval; then, obtain the vector group of average combustion chamber temperatures of the first n cases in the first n cases of the 2*K samples in part a, ordered by time.

[0030] TEMP t1 ,...,TEMP t2*K (6)

[0031] Find TEMP t0 The covariance of each vector in the vector group is used, and the n*K cases corresponding to the top K vectors with the largest covariance are used as the case retrieval results. These samples are cases where the change process of the incinerator's operating conditions at a specific time is similar to the change process of the operating conditions at the time to be predicted.

[0032] Third, use the improved ISVM based on mechanism analysis to build models for the retrieved n*K cases, and obtain the main steam flow rate value at the target time. Finally, store the results and the operating parameters at the target time as a case in the case library. This step includes three stages, detailed as follows:

[0033] a. Improve the SVM based on mechanism analysis; since the water level in the steam drum and the steam temperature and pressure at the superheater outlet remain basically constant during the actual operation of the incinerator, the main process affecting the steam flow is the heating process of the water in the steam drum, that is, the amount of water consumed to reach the boiling temperature per unit time is equal to the amount of cold water added to the steam drum; based on the incinerator's instruction manual, the volume parameters of the waste heat boiler are obtained, the heat transfer area A of the water in the boiler is calculated, and then the heat transfer rate Φ in the boiler is calculated.

[0034]

[0035] Where T w The average temperature of the primary combustion chamber, T f Let h be the outlet water temperature of the economizer, and h be the convective heat transfer coefficient of the water; simultaneously calculate the change in internal energy ΔE.

[0036] ΔE=mC(T b -T f (8)

[0037] Where ΔE represents the change in internal energy, and in the process of heating water, it represents the heat absorbed, T b Let T be the boiling point of water in the waste heat boiler, m be the mass of water added to the steam drum per unit time, and C be the specific heat capacity of water. Finally, calculate the temperature when the average temperature of the primary combustion chamber stabilizes at T. w The time Δt required for the cold water added to the steam drum to boil per unit time is considered to be 1.2-1.5 times the time required for the water temperature to rise at a decelerating rate during the actual process.

[0038]

[0039] The meaning of Δt is: the operating conditions of the internal incinerator before a certain time t0 have a relatively "important" impact on the steam flow rate at time t0, while the operating conditions of the incinerator before t0-Δt are not very important. Therefore, in the existing SVM model, the global penalty coefficient c is changed to

[0040]

[0041] Where t0 represents the time of the K samples that are to be predicted or whose Euclidean distance is close to that of the predicted working condition, t i Let t be the time of the i-th sample, Δt0 be the time required to heat the cold water to boiling at the time to be predicted, and c be the original penalty coefficient.

[0042] b. Set ISVM parameters; aiming to minimize the mean square error of the test set, repeat the process more than 30 times using trial and error or bisection method, selecting the original penalty coefficient c between 1 and 100, the RBF kernel function parameter g between 0.001 and 0.1, and the loss function parameter p between 0.001 and 0.1; among the different combinations of c, g, and p with simultaneously varying values, select the parameter combination that minimizes the mean square error of the test set.

[0043] c. Use the cases retrieved in step two to build an ISVM model, input the normalized feature variables of the time to be predicted into the model, calculate the main steam flow rate of the time to be predicted, and store the feature variables and prediction results of this time in vector form in the case library.

[0044] Fourth, repeat steps two to three until the incineration process ends, so as to achieve real-time prediction of the main steam flow rate in the urban solid waste incineration process.

[0045] Compared with existing technologies, this invention has the following advantages: 1. This invention uses historical data of the operating conditions during solid waste incineration and utilizes the normal distribution characteristics of the results of the correlation vector machine to determine whether the data is abnormal, avoiding the arbitrariness of manually setting abnormal thresholds. At the same time, the use of a sliding window accelerates the detection efficiency, making this method time-saving, accurate, and easy to operate; 2. The numerical prediction model based on case reasoning has strong interpretability. In addition, the real-time updating of the case library and the case retrieval based on instant learning can cope with the variability of incinerator operating conditions, giving the model a certain degree of adaptability; 3. An improved support vector machine is used in the case correction stage, which reduces the overfitting degree of the model, thereby improving the accuracy of main steam flow prediction. Attached Figure Description

[0046] Figure 1 This is a schematic diagram illustrating the principle of the method for predicting the main steam flow rate in the urban solid waste incineration process according to the present invention. Detailed Implementation

[0047] The sample data consists of 10,000 data points generated during the incineration process at a solid waste incineration plant. These were randomly divided into 9,000 source cases and 1,000 test cases. The following section combines... Figure 1 The specific embodiments of the present invention will be further described below.

[0048] A method for real-time prediction of main steam flow in urban solid waste incineration processes, characterized by the following steps:

[0049] I. Preprocessing of historical data on urban solid waste incineration processes: Using a sliding window-based method combined with Real-Time Visualization (RVM) to identify and remove outliers, and using time-continuous data as a case library; This process consists of four steps: Step 1: Normalizing existing historical data to form an initial case library; Step 2: Initializing outlier detection parameters; Step 3: Performing outlier detection on samples within the window; Step 4: Moving the window until outlier detection of all historical data is completed.

[0050] Step 1: Record the following characteristic variables: x1, average primary combustion chamber temperature x2, average economizer inlet flue gas temperature x3, economizer outlet flue gas temperature x4, economizer outlet water temperature x5, average evaporator inlet flue gas temperature x6, average tertiary superheater inlet flue gas temperature x7, total economizer feedwater volume x8, total desuperheating water volume for primary and secondary superheaters x9, and tertiary superheater outlet steam temperature x1. 10 Boiler drum pressure x 11 3rd stage superheater outlet steam pressure x 12 The historical data and the corresponding main steam flow rate value y1 are normalized and represented as a vector form, forming 9000 source cases, each case denoted as CASE. i It is represented in the following form:

[0051] CASE i =(X i ;Y i ), i = 1, 2, ..., 9000 (1)

[0052] Of these, 9000 represents the total number of source cases; CASE i This is the i-th source case; X i Y is the normalized value of the feature variable of the i-th source case; i X is the normalized economizer feedwater flow rate corresponding to the i-th source case and its characteristic variables; i This can be further expressed as:

[0053] X i =(x 1,i ,…,x λ,i ,…,x 12,i (2)

[0054] Where, x λ,i (λ = 1, ..., 12) represents CASE i The value of the λth feature variable after normalization.

[0055] Step 2: Set the sliding window length to 400, the upper bound of the kernel function width to 2.5, the lower bound of the kernel function width to 1.5, the kernel function width reduction step size to 0.1, the abnormal attribute probability base to 104875, and the abnormal attribute probability threshold to 0.9.

[0056] Step 3: Outlier detection for the 400 samples within the window is mainly divided into four parts, detailed as follows:

[0057] a. To detect all samples within the window, the initial abnormal attribute probability P of all samples within the window is set to 1, and the initial kernel function width sigma of the RVM model is set to 2.5.

[0058] b. Select 400 samples within the window as training samples based on the probability of abnormal attributes, and build an RVM model using the kernel function width sigma.

[0059] c. Perform predictions on each of the 400 samples within the window using the established RVM model, and calculate the relative deviation between the sample values ​​and the predicted values. i

[0060]

[0061] Where i = 1, ..., 400 represents the sample index within the window, and y i For the actual sample value, The model predicts the values ​​and updates the probabilities of anomalous attributes using the following formula:

[0062]

[0063] in It is the probability value of the new abnormal attribute of sample i.

[0064] d. Repeat parts b and c, each time using a kernel function with a width of sigma. * =sigma-Δ sigma sigma * The width of the new kernel function, up to sigma. * =1.5, which sets the probability of abnormal attributes. Cases with anomaly attribute probability below 0.9 i These samples are marked as abnormal and deleted.

[0065] Step 4: Move the window to the next position until all samples are detected. Increasing the length of the window movement can improve the efficiency of outlier detection.

[0066] 2. Using the Just-In-Time (JITL) learning method based on Euclidean distance and covariance as similarity metrics, data samples were retrieved to obtain the 5*8 cases with the highest similarity to the time to be predicted. The case retrieval process consists of two parts, detailed below:

[0067] a. Calculate the Euclidean distance between the normalized feature variables of the data to be predicted and all samples in the case library, and select the first 2*8 samples with the smallest distance. These samples are cases where the incinerator operating conditions at a specific time are similar to the operating conditions at the time to be predicted.

[0068] b. Obtain the mean vector of primary combustion chamber temperature TEMP from the first 5 cases arranged in chronological order at the time t0 to be predicted. t0 , can be represented as

[0069] TEMP t0 =(T t0-Δt ,…,T t0-j*Δt ,…,T t0-n*Δt ) T ,j=1,2,...5 (5)

[0070] Where T t0-j*Δt Let t be the mean combustion chamber temperature of the j-th case before time t0, and Δt be the data recording time interval; then, obtain the vector group of the mean combustion chamber temperatures of the first 5 cases in the 2*8 samples of part a, ordered by time.

[0071] TEMP t1 ,...,TEMP t2*8 (6)

[0072] Find TEMP t0 The covariance of each vector in the vector group is used, and the 5*8 cases corresponding to the top 8 vectors with the largest covariance are used as the case retrieval results. These samples are cases where the change process of the incinerator's operating conditions at a specific time is similar to the change process of the operating conditions at the time to be predicted.

[0073] Third, models are built using ISVM based on mechanism analysis to identify 5*8 cases, and the main steam flow rate at the target time is obtained. Finally, the results and the operating parameters at the target time are stored as a case in the case library. This step includes three stages, detailed as follows:

[0074] a. Improve the SVM based on mechanism analysis; since the water level in the steam drum and the steam temperature and pressure at the superheater outlet remain basically constant during the actual operation of the incinerator, the main process affecting the steam flow is the heating process of the water in the steam drum, that is, the amount of water consumed to reach the boiling temperature per unit time is equal to the amount of cold water added to the steam drum; based on the incinerator's instruction manual, the volume parameters of the waste heat boiler are obtained, and the heat transfer area of ​​the water in the boiler is calculated to be 40m².2 Then calculate the heat transfer rate Φ inside the boiler.

[0075]

[0076] Where T w The average temperature of the primary combustion chamber, T f Let be the outlet water temperature of the economizer, and h be the convective heat transfer coefficient of water, which is 700 here; simultaneously calculate the internal energy change ΔE.

[0077] ΔE=mC(T b -T f (8)

[0078] Where ΔE represents the change in internal energy, and in the process of heating water, it represents the heat absorbed, T b Let T be the boiling point of water in the waste heat boiler, m be the mass of water added to the steam drum per unit time, and C be the specific heat capacity of water. Finally, calculate the temperature when the average temperature of the primary combustion chamber stabilizes at T. w The time Δt required for the cold water added to the steam drum to boil per unit time is considered to be 1.2-1.5 times the time required for the water temperature to rise at a decelerating rate during the actual process.

[0079]

[0080] The meaning of Δt is: the operating conditions of the internal incinerator before a certain time t0 have a relatively "important" impact on the steam flow rate at time t0, while the operating conditions of the incinerator before t0-Δt are not very important. Therefore, in the existing SVM model, the global penalty coefficient c is changed to

[0081]

[0082] Where t0 represents the time of the K samples that are to be predicted or whose Euclidean distance is close to that of the predicted working condition, t i Let t be the time of the i-th sample, Δt0 be the time required to heat the cold water to boiling at the time to be predicted, and c be the original penalty coefficient.

[0083] b. Set ISVM parameters; set RBF kernel function parameter to 0.01, loss function to 0.001, and original penalty coefficient to 100.

[0084] c. Use the cases retrieved in step two to build an ISVM model, input the normalized feature variables of the time to be predicted into the model, calculate the main steam flow rate of the time to be predicted, and store the feature variables and prediction results of this time in vector form in the case library.

[0085] Fourth, repeat steps two to three until the incineration process ends, so as to achieve real-time prediction of the main steam flow rate in the urban solid waste incineration process.

[0086] Table 1

[0087]

[0088] Traditionally, main steam flow data is obtained from sensors, which is prone to equipment malfunctions causing abnormal measurement data. Furthermore, this data is significantly delayed compared to the incinerator's operating conditions, thus affecting subsequent steam utilization. This invention first utilizes the sparsity of the Restricted Virtual Machine (RVM) model to identify and remove abnormal data from the original dataset. Secondly, by combining real-time learning, ISVM, and case-based reasoning, a real-time prediction method for main steam flow in solid waste incineration (referred to as RVM-ISVMCBR) is derived, exhibiting good adaptability, low overfitting, and high accuracy. To further verify the effectiveness of this method in real-time prediction of main steam flow in solid waste incineration, comparative experiments were conducted using other prediction methods and field-collected data. The prediction results are shown in Table 1. Experimental data shows that, in 10-fold cross-validation on a given set of 10,000 data points, the prediction algorithm proposed in this invention achieves a mean absolute error (MAE) of 0.1465, a mean square error (MSE) of 0.0327, and a mean absolute percentage error (MAPE) of 0.0020, all of which outperform the experimental results of other prediction methods. This demonstrates that the RVM-ISVMCBR algorithm can effectively reduce model overfitting and achieves high prediction accuracy.

Claims

1. A method for real-time prediction of main steam flow rate in urban solid waste incineration process, characterized in that... Includes the following steps:

1. Preprocess historical data of urban solid waste incineration processes, using a sliding window-based method combined with Relevance Vector Machine (RVM) to identify and remove outliers, and using time-continuous data as a case library; 2. Use the Just-In-Time (JITL) method, based on Euclidean distance and covariance as similarity indicators, to retrieve data samples and obtain n*K cases with the highest similarity to the time to be predicted; 3. Use an improved support vector machine (ISVM) based on mechanism analysis to build a model for the retrieved n*K cases and derive the main steam flow rate value at the target time. Finally, store the result and the operating parameters at the target time as a case in the case library; 4. Repeat steps 2-3 until the incineration process ends to achieve real-time prediction of the main steam flow rate of urban solid waste incineration processes; specifically including the following steps: I. Preprocessing of historical data on urban solid waste incineration processes, using a sliding window-based method combined with Regression-Version Method (RVM) to identify and remove outliers, and using time-continuous data as a case library; this process consists of four steps: Step 1: Normalizing existing historical data to form an initial case library; Step 2: Initializing outlier detection parameters; Step 3: Performing outlier detection on samples within the window; Step 4: Move the window until outlier detection of all historical data is complete; Step 1: Record the following characteristic variables: x1, average primary combustion chamber temperature x2, average economizer inlet flue gas temperature x3, economizer outlet flue gas temperature x4, economizer outlet water temperature x5, average evaporator inlet flue gas temperature x6, average tertiary superheater inlet flue gas temperature x7, total economizer feedwater volume x8, total desuperheating water volume for primary and secondary superheaters x9, and tertiary superheater outlet steam temperature x1. 10 Boiler drum pressure x 11 3rd stage superheater outlet steam pressure x 12 The historical data and the corresponding main steam flow rate value y1 are normalized and represented as a vector form, forming N source cases, each case denoted as CASE. i It is represented in the following form: CASE i =(X i ;Y i ),i=1,2,L,N (1) Where N is the total number of source cases; CASE i This is the i-th source case; X i Y is the normalized value of the feature variable of the i-th source case; i X is the normalized economizer feedwater flow rate corresponding to the i-th source case and its characteristic variables; i Expressed as: X i =(x 1,i ,L,x λ,i ,L,x 12,i ) (2) Where, x λ,i , indicating CASE i The normalized value of the λth feature variable in the equation; where λ = 1, ..., 12; Step 2: Set a relatively large sliding window length L based on the computer's processing power; set an upper bound sigma1 for the kernel function width based on the allowable upper bound of the anomaly severity; aiming for a high anomaly detection rate on the test set with added anomalous samples, repeat the process more than 30 times using trial and error or binary search methods. Select a lower bound sigma2 for the kernel function width between 0.1 and sigma1, a base of the anomaly attribute probability between 100,000 and 200,000, an anomaly attribute probability threshold TH between 0.5 and 1, and a step size Δ for decreasing the kernel function width between 0.01 and 0.

2. s igma; sigma2, a, TH, Δ, whose values ​​change simultaneously. s Among different combinations of igma, select the parameter combination with the highest anomaly detection rate on the test set; Step 3: Outlier detection for the L samples within the window is divided into four parts, detailed as follows: a. To detect all samples within the detection window, set the initial abnormal attribute probability P of all samples to 1, and set the initial kernel function width sigma of the RVM model to sigma1; b. Select L samples within the window as training samples based on the probability of abnormal attributes, and use the kernel function width sigma to build an RVM model; c. For each of the L samples within the window, use the established RVM model to make predictions, and calculate the relative deviation between the sample values ​​and the predicted values. i Where i = 1, ..., L represents the sample index within the window, and y i For the actual sample value, The model predicts the values ​​and updates the probabilities of anomalous attributes using the following formula: in It is the probability value of the new abnormal attribute of sample i; d. Repeat parts b and c, each time using a kernel function with a width of sigma. * =sigma-Δ sigma sigma * The width of the new kernel function, up to sigma. * =sigma2, which sets the probability of abnormal attributes. CASE samples below the anomaly attribute probability threshold TH i These samples are marked as abnormal and deleted. Step 4: Move the window to the next position until all samples are detected. Increasing the length of the window movement can improve the efficiency of outlier detection.

2. The Just-In-Time (JITL) learning method, based on Euclidean distance and covariance as similarity metrics, is used to retrieve data samples and obtain the n*K cases with the highest similarity to the time to be predicted. The case retrieval process consists of two parts, detailed below: a. Calculate the Euclidean distance between the normalized feature variables of the data to be predicted and all samples in the case library, and select the first 2*K samples with the smallest distance. These samples are cases where the incinerator operating conditions at a specific time are similar to the operating conditions at the time to be predicted. b. Obtain the mean vector of the primary combustion chamber temperature TEMP from the first n cases arranged in chronological order at the time t0 to be predicted. t0 , represented as TEMP t0 =(T t0-Δt ,L,T t0-j*Δt ,L,T t0-n*Δt ) T ,j=1,2,...n (5) Where T t0-j*Δt Let t be the average combustion chamber temperature of the j-th case before time t0, and Δt be the data recording time interval; then, obtain the vector group of average combustion chamber temperatures of the first n cases in the first n cases of the 2*K samples in part a, ordered by time. TEMP t1 ,...,TEMP t2*K (6) Determine TEMP t0 The covariance of each vector in the vector group is used, and the n*K cases corresponding to the top K vectors with the largest covariance are used as the case retrieval results. These samples are cases where the change process of the incinerator's operating conditions at a specific time is similar to the change process of the operating conditions at the time to be predicted. Third, use the improved ISVM based on mechanism analysis to build models for the retrieved n*K cases, and obtain the main steam flow rate value at the target time. Finally, store the results and the operating parameters at the target time as a case in the case library. This step includes three stages, detailed as follows: a. Improve the SVM based on mechanism analysis; since the water level in the steam drum and the steam temperature and pressure at the superheater outlet remain basically constant during the actual operation of the incinerator, the main process affecting the steam flow is the heating process of the water in the steam drum, that is, the amount of water consumed to reach the boiling temperature per unit time is equal to the amount of cold water added to the steam drum; based on the incinerator's instruction manual, the volume parameters of the waste heat boiler are obtained, the heat transfer area A of the water in the boiler is calculated, and then the heat transfer rate Φ in the boiler is calculated. Where T w The average temperature of the primary combustion chamber, T f Let h be the outlet water temperature of the economizer, and h be the convective heat transfer coefficient of the water; simultaneously calculate the change in internal energy ΔE. ΔE=mC(T b -T f ) (8) Where ΔE represents the change in internal energy, and in the process of heating water, it represents the heat absorbed, T b Let T be the boiling point of water in the waste heat boiler, m be the mass of water added to the steam drum per unit time, and C be the specific heat capacity of water. Finally, calculate the temperature when the average temperature of the primary combustion chamber stabilizes at T. w The time Δt required for the cold water added to the steam drum to boil per unit time is considered to be 1.2-1.5 times the uniform heating time, since the actual water temperature rise is decelerated. The meaning of Δt is: the operating conditions of the internal incinerator before a certain time t0 have a relatively "important" impact on the steam flow rate at time t0, while the operating conditions of the incinerator before t0-Δt are not very important. Therefore, in the existing SVM model, the global penalty coefficient c is changed to Where t0 represents the time of the K samples that are to be predicted or whose Euclidean distance is close to that of the predicted working condition, t i Let t be the time of the i-th sample, Δt0 be the time required to heat the cold water to boiling at the time to be predicted, and c be the original penalty coefficient. b. Set ISVM parameters; aiming to minimize the mean square error of the test set, repeat the process more than 30 times using trial and error or bisection method, selecting the original penalty coefficient c between 1 and 100, the RBF kernel function parameter g between 0.001 and 0.1, and the loss function parameter p between 0.001 and 0.1; among the different combinations of c, g, and p with simultaneously varying values, select the parameter combination that minimizes the mean square error of the test set; c. Use the cases retrieved in step two to build an ISVM model, input the normalized feature variables of the time to be predicted into the model, calculate the main steam flow rate of the time to be predicted, and store the feature variables and prediction results of this time in vector form in the case library. Fourth, repeat steps two to three until the incineration process ends, so as to achieve real-time prediction of the main steam flow rate in the urban solid waste incineration process.

Images