Urban solid waste incineration hearth temperature prediction method based on self-organizing Cauchy fuzzy neural network

Abstract

The invention relates to an urban solid waste incineration hearth temperature prediction method based on a self-organizing Cauchy fuzzy neural network, which realizes accurate prediction of hearth temperature, and comprises the following steps: firstly, collecting data; preprocessing the collected data, and performing feature selection on a data set; then, establishing a prediction model by adopting a self-organizing Cauchy fuzzy neural network; self-organizing optimization adjustment is carried out on the structure and parameters of the prediction model; and finally, testing data is used as the input of the model, and the effectiveness of the model is verified. According to the method, the problems of weak non-linear adaptive capacity, insufficient real-time prediction precision and the like of an existing incineration hearth temperature prediction method are solved, and the method has important theoretical significance and application value for achieving the win-win situation of environmental protection benefits and economic benefits.

Description

Technical Field

[0001] This invention relates to technology for predicting the furnace temperature of municipal solid waste incineration. It proposes a method for predicting the furnace temperature of municipal solid waste incineration based on a self-organizing Cauchy Fuzzy Neural Network (CFNN). This method addresses the problems of existing furnace temperature prediction methods, such as weak nonlinear adaptability and insufficient real-time prediction accuracy. The method relates to both the field of municipal solid waste treatment and the field of temperature prediction. Background Art

[0002] Municipal solid waste incineration technology, as an important solid waste treatment method, not only significantly reduces waste storage but also generates new energy. During the municipal solid waste incineration process, excessively low furnace temperatures can lead to reduced resource utilization and increased pollutant emissions. Excessively high furnace temperatures can compromise the safety of materials within the furnace and cause accidents. Therefore, to ensure that the furnace temperature during municipal solid waste incineration remains within the specified range, accurate furnace temperature prediction is necessary to achieve both environmental and economic benefits.

[0003] Currently, methods for predicting furnace temperature in municipal solid waste incineration are mainly divided into two categories: mechanism models and data-driven models. The mechanism model predicts furnace temperature through mathematical modeling based on the physical and chemical principles of the incineration process. The mechanism model has poor adaptability to the complex changes in solid waste composition and has high computational complexity, making it difficult to make real-time predictions. In recent years, with the advancement of sensing technology and computing power, data-driven models have been increasingly used in furnace temperature prediction. Data-driven models rely on a large amount of historical operating data and use statistical and machine learning algorithms to extract the patterns of furnace temperature changes. However, traditional data-driven models have limitations in processing nonlinear and high-dimensional data, and the prediction accuracy needs to be improved. Summary of the Invention

[0004] The present invention aims to solve the shortcomings of existing furnace temperature prediction methods in terms of prediction precision and accuracy. By introducing the Cauchy function into the traditional fuzzy neural network model, a self-organizing CFNN model is constructed to achieve accurate prediction of the furnace temperature of municipal solid waste incineration.

[0005] To achieve the above objectives, the present invention adopts the following technical solutions and implementation steps:

[0006] 1. Data collection: Collect and record data related to the municipal solid waste incineration process through temperature, pressure and other sensors and the ACC system;

[0007] 2. Data preprocessing: Cleaning and simplifying the collected raw data set. Data cleaning involves dealing with outliers and missing values ​​in the data; data simplification involves reducing the size of the data set based on expert knowledge.

[0008] 3. Feature selection: The Lasso algorithm is used to extract features from the 32 feature variables after data preprocessing. The method is as follows:

[0009] The Lasso algorithm is short for the least absolute shrinkage and selection operator. Its loss function consists of two terms: the sum of the squares of the Euclidean distances of all sample points to the hyperplane (i.e., the sum of the squares of the residuals), and the product of the L1-norm of the parameter vector and the penalty coefficient.

[0010] This algorithm achieves parameter compression and feature selection through L1 regularization. The formula is as follows:

[0011]

[0012] Among them, β is the coefficient variable, and ||β||1 is the L1-norm of the coefficient variable, that is, the sum of the absolute values ​​of the coefficients. i and x i are the prediction targets and features for feature selection respectively;

[0013] λ is a regularization hyperparameter used to adjust the penalty strength of ||β||1. When λ is 0, Lasso is equivalent to least squares. When λ is too large, the model is overly simple and has poor predictive ability. Its value was determined to be 0.7 using 10-fold cross-validation.

[0014] The specific steps are:

[0015] ① Initialize weight β:

[0016] Randomly generate an initial value for each coefficient to provide a starting point for the subsequent optimization process;

[0017] ②Perform iterative optimization:

[0018] In each iteration, all β are traversed;

[0019] For the currently traversed β, it is regarded as a variable (that is, the object that needs to be optimized at present), and all other weight coefficients are temporarily fixed to the values ​​at the end of the previous iteration, that is, regarded as constants;

[0020] Under this condition where there is only one β as a variable, the optimized coordinate descent algorithm is used to calculate and update the value of the weight coefficient so that the performance of the model under the current configuration is close to the optimal;

[0021] ③ Iteration termination condition: The iteration process will continue until any of the following conditions are met:

[0022] The weight change is small: that is, the change of all weight coefficients in several consecutive iterations is very small, indicating that the model has approached or reached the optimal solution, and further iteration may not bring significant performance improvement;

[0023] The maximum number of iterations is reached: in order to avoid infinite loop, the maximum number of iterations is set to 500. When the number of iterations reaches this upper limit, the iteration is stopped regardless of the change of weight coefficients;

[0024] After the above steps, the final extracted feature variables are 7 variables including feed rate, grate speed, furnace pressure, flue gas temperature, low heat value of garbage, primary air temperature and oxygen content in flue gas;

[0025] 4. Design a CFNN model for predicting the furnace temperature:

[0026] (1) Input layer

[0027] The function is to accept external input data or features (i.e. the above 7 feature variables), and pass them to the hidden layer of the network. Each neuron of the input layer represents an input data, and this layer has 7 neurons. Let the input of each neuron be x(t),i=1,2,...,7, then the output C(t) of each neuron is represented as:

[0028] C i (t)=x i (t),i=1,2,...,7 (2)

[0029] Where C i (t) represents the output of the i-th neuron, and x i (t) represents the input of the i-th neuron.

[0030] (2) Membership degree layer

[0031] This layer introduces the Cauchy function as the membership function to better handle non-ideal data (such as multicollinearity) and further exclude the interference of outliers on the modeling strategy;

[0032] Each input neuron corresponds to multiple membership functions, and the membership degree layer has m (m=10, same below) membership functions. Then this layer has 7xm output values, which are represented as:

[0033]

[0034] Where C ij (t), l ij and θ ij (t) are the output value, center and width of the j-th membership function corresponding to the i-th input of the CFNN membership degree layer at time t, respectively.

[0035] (3) Rule layer

[0036] Each neuron in the rule layer matches a fuzzy rule by connecting the fuzzy operation and the membership layer, the input signal of each neuron in the rule layer is the output signal of the corresponding membership layer;

[0037] The output value R(t) of each neuron in the rule layer at time t is the product of all input signals C(t) of the neuron, which is expressed as:

[0038]

[0039] Where, R j (t) is the output variable of the jth neuron in the rule layer at time t.

[0040] (4) Normalization layer

[0041] The normalization layer has a total of 10 neurons, and the rule layer output G(t) is expressed as:

[0042]

[0043] Where, G j (t) is the output variable of the jth neuron in the normalization layer at time t

[0044] (5) Output layer

[0045] The function of this layer is to perform weighted summation on the output of the normalization layer to realize defuzzification. W(t) is the weight coefficient matrix of the rule neurons and the output neurons at time t, and the predicted value y(t) of the furnace temperature is expressed as:

[0046]

[0047] Where, ω j (t) represents the jth sub-vector in the matrix W(t). Similarly, g j (t) is expressed as the jth sub-vector of the G(t) matrix.

[0048] 5. Dynamic adaptive algorithm:

[0049] In the modeling process, the dynamic self-organizing algorithm is used to adjust the model parameters to continuously improve the model accuracy;

[0050] (1) Self-organizing mechanism

[0051] The activity of a neuron is directly proportional to its importance, and the more active the neuron is, the higher its importance is;

[0052] ① Growth: The growth index I g (t) is defined as:

[0053]

[0054] In the formula, NA j (t) is the neuronal activity of the jth regular neuron, is the normalized neuronal activity of the jth regular neuron, r(t) is the number of existing regular neurons at time t;

[0055] When I g (t) is greater than β1 (a predefined growth threshold value of 0.8) and the root mean square error of the current moment is greater than the root mean square error of the previous moment, the g-th rule neuron is split;

[0056] Initialize the parameters of the newly generated m+1th regular neuron:

[0057]

[0058] θ m+1 (t) = θ g (t) (10)

[0059]

[0060] Among them, it can be seen from formula 3 that l g (t) and θ g (t) are the center and width of the g-th regular node, l m+1 (t) and θ m+1 (t) are the center and width of the newly generated rule node, g m+1 (t) and w m+1 (t) are the output value and weight of the newly generated rule node, y d (t) is the predicted value of the dth rule neuron.

[0061] ② Pruning: Pruning Index I d is defined as:

[0062]

[0063] Among them, ξ s (t) and ξ j (t) is the importance of the neurons of the sth rule and the jth rule at time t. When the growth index is less than β2 (a predefined pruning threshold of 0.2) and the root mean square error of the current moment is less than the root mean square error of the previous moment, the redundant neuron of the dth rule is pruned;

[0064] In addition, in order to ensure the convergence of CFNN, the parameters of the qth regular node with the largest singular value are adjusted as follows:

[0065] l' q (t) = lq (t) (13)

[0066] θ′ q (t) = θ q (t) (14)

[0067]

[0068] Among them, l' q (t) and θ q '(t) are the center widths after the dth rule node is deleted, w q (t) and w' q (t) are the connection weights of the qth rule node before and after the dth rule node is deleted, G s (t) and G q (t) are the output values ​​of the sth and qth rule nodes before the dth rule node is deleted;

[0069] After executing the pruning operator, the parameters of the d-th rule node are set to zero, while the center and width of the q-th rule node remain unchanged;

[0070] (2) Adaptive learning algorithm:

[0071] The Levenberg-Marquardt (LM) algorithm combines the Gauss-Newton algorithm and the gradient descent algorithm. The second-order LM algorithm can use the second-order derivative information of the objective function to update the parameters of the fuzzy neural network and realize the adaptive learning of the algorithm.

[0072] For the pth sample, the expected output y d p (t) and actual output y p The error between (t) p (t) is expressed as:

[0073] e p (t) = y d p (t)-y p (t),p=1,2,...,P (16)

[0074] According to the derivation of LM algorithm, the update rule of the algorithm is:

[0075] Θ(t+1)=Θ(t)+(Ψ(t)+η(t)×I) -1 ×Ω(t) (17)

[0076] η(t)=β3||e(t)||+(1-β3)||Ω(t)|| (18)

[0077]

[0078] Θ(t)=[w(t),l(t),θ(t)] (21)

[0079] Wherein, Θ(t+1) and Θ(t) are the to-be-optimized parameter vectors at t+1 time and t time respectively, I is a unit matrix, η(t) is an adaptive learning rate, β3 is a user-defined constant and is 0.5, e(t)=[e1(t),e2(t),...,e p (t)] T Error vector at t time;

[0080] Ψ(t) is quasi-Hessian matrix at t time, Ω(t) is gradient vector at t time, and ψ p (t) and ω p (t) are accumulations of sub-matrix ψ p (t) and sub-vector ω p (t) of the Jacobian matrix, and the definitions of sub-matrix ψ p (t) and sub-vector ω

[0081] j p (t) are as follows:

[0082]

[0083] Using the update rule of the gradient descent learning algorithm, the row vectors of the Jacobian matrix j p (t) are as follows:

[0084]

[0085] The present application has the following obvious advantages and beneficial effects, compared with the prior art, the beneficial effects of the present application mainly reflect in the following aspects:

[0086] The present application combines expert knowledge and Lasso algorithm to select key feature variables, reduces the redundancy of input data, improves the prediction efficiency and accuracy of the model, introduces Cauchy function as the membership function of the traditional fuzzy neural network model, improves the processing ability of the model to non-ideal data and outliers, enhances the robustness and prediction accuracy of the model, adopts LM algorithm and dynamic pruning strategy to realize adaptive optimization of the model, and improves the convergence speed and prediction performance of the model;

[0087] The experimental results show that the present application has high precision and accurate real-time performance in predicting the furnace temperature, and provides strong support for the optimization control of the urban solid waste incineration process. BRIEF DESCRIPTION OF DRAWINGS

[0088] Figure 1 is a flow chart of the present invention;

[0089] Figure 2 Flowchart for feature extraction;

[0090] Figure 3 This is the CFNN neural network structure diagram;

[0091] Figure 4 This is the CFNN training result diagram;

[0092] Figure 5 is the CFNN training error graph;

[0093] Figure 6 This is the CFNN test result diagram;

[0094] Figure 7 This is the CFNN test error graph. DETAILED DESCRIPTION

[0095] The present invention uses training data to establish a self-organizing Cauchy fuzzy neural network model for predicting furnace temperature; and uses a test data set to verify the accuracy of the real-time furnace temperature output by the self-organizing Cauchy fuzzy neural network model.

[0096] As an example, the effectiveness of the proposed method was verified using real-world data from a municipal solid waste incineration plant in Beijing. After removing significantly abnormal data, a total of 1,600 sets of 31-dimensional experimental data were obtained. Based on this data sample, a Lasso algorithm was used for feature selection, selecting seven variables with a high correlation with furnace temperature as input variables for the prediction model. See Table 1 for details.

[0097] Table 1

[0098]

[0099] Of the 1600 data sets after dimensionality reduction, 1200 sets of data were used to build the prediction model, and the remaining 400 sets of data were used to test the model performance;

[0100] (1) Based on 1200 sets of training data, Figure 4 、 Figure 5 They are respectively the training result graph and the training error graph of the prediction model, X-axis: number of training samples, unit is / sample, Y-axis: furnace temperature, unit is ℃;

[0101] (2) Based on 400 sets of test data, the furnace temperature is predicted by self-organizing Cauchy fuzzy neural network. The test results are as follows Figure 6 As shown, the X axis is the number of test samples, the unit is / sample, the Y axis is the furnace temperature, the unit is ℃; the test error is as follows Figure 7As shown, X axis: training sample number, unit: individual / sample, Y axis: furnace temperature, unit: ℃;

[0102] (3) The test accuracy is quantitatively evaluated by using the root mean square error RMSE and the mean absolute percentage error MAPE, and the calculation result is RMSE = 4.0623 and MAPE = 0.2796%.

[0103] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. It should be noted that the above description is only a specific embodiment of the present application and does not limit the present application. Any modification and optimization made within the spirit and principles of the present application shall be covered by the scope of the claims of the present application.

Claims

1. A method for predicting the furnace temperature of municipal solid waste incineration based on a self-organizing Cauchy fuzzy neural network, characterized in that: The following steps are involved: S1: Collect and record data related to the municipal solid waste incineration process in temperature, pressure sensors and ACC system; S2: Clean and simplify the collected original data set; S3: Use the Lasso algorithm to extract features from the data set after data preprocessing; S4: Design of a Cauchy fuzzy neural network model for predicting furnace temperature; S5: During the modeling process, dynamic self-organizing algorithms are used to adjust model parameters and continuously improve model accuracy.

2. According to the method for predicting the furnace temperature of a municipal solid waste incineration system based on a self-organizing Cauchy fuzzy neural network in claim 1, the essence of the Lasso algorithm is the sum of the Euclidean distances of all sample points to the hyperplane, plus the L1-norm with a penalty coefficient λ; The steps are: (1) Coordinate descent method Initialize weight β: Randomly generate an initial value for each coefficient to provide a starting point for the subsequent optimization process; ②Perform iterative optimization: In each iteration, all β are traversed; For the currently traversed β, it is regarded as a variable, that is, the object that needs to be optimized, and all other weight coefficients are temporarily fixed to the values ​​at the end of the previous iteration, that is, they are regarded as constants; Under this condition where there is only one β as a variable, the optimized coordinate descent algorithm is used to calculate and update the value of the weight coefficient so that the performance of the model under the current configuration is close to the optimal; ③ Iteration termination condition: The iteration process will continue until any of the following conditions are met: Small weight changes: that is, the changes in all weight coefficients in several consecutive iterations are very small, indicating that the model has approached or reached the optimal solution; Reach the maximum number of iterations: Set the maximum number of iterations to 500; when the number of iterations reaches this upper limit, the iteration stops regardless of the change of the weight coefficient; Based on expert knowledge, the characteristic variables are simplified to 31 groups, and the Lasso algorithm is used to select seven variables from the 31 groups of characteristic variables, including feed rate, grate speed, furnace pressure, flue gas temperature, lower calorific value of garbage, primary air temperature and flue gas oxygen content.

3. The method for predicting the temperature of a municipal solid waste incineration furnace based on a self-organizing Cauchy fuzzy neural network according to claim 1 is characterized in that: The Cauchy function is used as the membership function, and each input neuron corresponds to multiple membership functions.

4. The method for predicting the temperature of a municipal solid waste incineration furnace based on a self-organizing Cauchy fuzzy neural network according to claim 1 is characterized in that: In step S5: During the modeling process, dynamic self-organizing algorithms are used to adjust model parameters and continuously improve model accuracy; (1) Self-organizing mechanism ① Growth; ② Pruning; (2) Adaptive learning algorithm: The Levenberg-Marquardt (LM) algorithm combines the Gauss-Newton algorithm and the gradient descent algorithm. The second-order LM algorithm can use the second-order derivative information of the objective function to update the parameters of the fuzzy neural network and realize the adaptive learning of the algorithm.

5. The method for predicting the temperature of a municipal solid waste incineration furnace based on a self-organizing Cauchy fuzzy neural network according to claim 1, characterized in that: The final extracted characteristic variables are 7 variables: feed rate, grate speed, furnace pressure, flue gas temperature, garbage lower calorific value, primary air temperature and flue gas oxygen content; Design of a CFNN model for predicting furnace temperature: (1) Input layer The function is to accept external input data or features, that is, the 7 feature variables mentioned above, and pass them to the hidden layer of the network; each neuron in the input layer represents an input data. There are 7 neurons in this layer. Let the input of each neuron be, then the output C(t) of each neuron is expressed as: C i (t)=x i (t),i=1,2,...,7 (1) Among them, C i (t) represents the output of the i-th neuron, x i (t) represents the input of the i-th neuron; (2) Membership layer This layer introduces the Cauchy function as the membership function to better handle non-ideal data (such as multicollinearity) and further eliminate the interference of outliers on the modeling strategy; Each input neuron corresponds to multiple membership functions. The membership layer has a total of m membership functions, m = 10, and this layer has a total of 7 × m output values, which can be expressed as: Among them, C ij (t), l ij and θ ij (t) are the output value, center, and width of the j-th membership function corresponding to the i-th input of the CFNN membership layer at time t; (3) Rule layer Each neuron in the rule layer matches the fuzzy rule by connecting the fuzzy operation and the membership layer. The input signal of each neuron in the rule layer is the output signal of the corresponding membership layer. The output value R(t) of each neuron in the rule layer at time t is the product of all the input signals C(t) of the neuron, expressed as: Among them, R j (t) is the output variable of the jth neuron in the regular layer at time t; (4) Normalization layer The normalization layer has 10 neurons, and the regularization layer output G(t) is expressed as: Among them, G j (t) is the output variable of the jth neuron in the normalization layer at time t (5) Output layer W(t) is the weight coefficient matrix of the rule neuron and the output neuron at time t. The predicted value y(t) of the furnace temperature is expressed as: Among them, ω j (t) represents the jth subvector in the matrix W(t), g j (t) is represented as the j-th subvector of the G(t) matrix.

6. The method for predicting the temperature of a municipal solid waste incineration furnace based on a self-organizing Cauchy fuzzy neural network according to claim 1, characterized in that: The dynamic adaptive algorithm is as follows: (1) Self-organizing mechanism The activity of a neuron is proportional to its importance. The more active the neuron, the higher its importance. ① Growth: Growth Index I g (t) is defined as: In the formula, NA j (t) is the neuronal activity of the jth regular neuron, is the normalized neuronal activity of the jth regular neuron, r(t) is the number of existing regular neurons at time t; When I g (t) When the root mean square error of the current moment is greater than the root mean square error of the previous moment (t) is greater than β1 (a predefined growth threshold value of 0.8), the g-th rule neuron is split; Initialize the parameters of the newly generated m+1th regular neuron: i m+1 (t)=θ g (t) (9) Among them, it can be seen from formula 3 that l g (t) and θ g (t) are the center and width of the g-th regular node, l m+1 (t) and θ m+1 (t) are the center and width of the newly generated rule node, g m+1 (t) and w m+1 (t) are the output value and weight of the newly generated rule node, y d (t) is the predicted value of the dth rule neuron; ② Pruning: Pruning Index I d is defined as: Among them, ξ s (t) and ξ j (t) is the importance of the neurons of the sth rule and the jth rule at time t; when the growth exponent is less than β2 and the root mean square error of the current moment is less than the root mean square error of the previous moment, the redundant neuron of the dth rule is pruned; β2 is a predefined pruning threshold with a value of 0.2; In addition, in order to ensure the convergence of CFNN, the parameters of the qth regular node with the largest singular value are adjusted as follows: l' q (t)=l q (t) (12) i q '(t)=θ q (t) (13) Among them, l' q (t) and θ q '(t) are the center widths after the dth rule node is deleted, w q (t) and w' q (t) are the connection weights of the qth rule node before and after the dth rule node is deleted, G s (t) and G q (t) are the output values ​​of the sth and qth rule nodes before the dth rule node is deleted; After executing the pruning operator, the parameters of the d-th rule node are set to zero, while the center and width of the q-th rule node remain unchanged; (2) Adaptive learning algorithm: For the pth sample, the expected output y d p (t) and actual output y p The error between (t) p (t) is expressed as: e p (t)=y d p (t)-y p (t),p=1,2,...,P (15) The update rule of the Levenberg-Marquardt or LM algorithm is: Θ(t+1)=Θ(t)+(Ψ(t)+η(t)×I) -1 ×Ω(t) (16) η(t)=β3||e(t)||+(1-β3)||Ω(t)|| (17) Θ(t)=[w(t),l(t),θ(t)] (20) Where Θ(t+1) and Θ(t) are the parameter vectors to be optimized at time t+1 and time t, respectively, I is the identity matrix, η(t) is the adaptive learning rate, β3 is a user-defined constant and is set to 0.5, e(t) = [e1(t), e2(t), ..., e p (t)] T is the error vector at time t; Ψ(t) is the quasi-Hessian matrix at time t, Ω(t) is the gradient vector at time t, and Ψ(t) and Ω(t) are the submatrices ψ of P samples in the current sliding window respectively. p (t) and subvector ω p (t) accumulation; j p (t) is the pth row vector of the Jacobian matrix, then the submatrix ψ p (t) and subvector ω p (t) is defined as: Using the update rule of the gradient descent learning algorithm, the Jacobian matrix j is derived p The row vector of (t) is:

Images