A method for predicting gas components during high-temperature gasification of multi-component organic waste
Through the optimization of BP neural network model and genetic algorithm, combined with principal component analysis and box graph method to process data, the accuracy of gas production component prediction and stability of gasifier temperature measurement during high-temperature gasification of multi-component organic waste was solved. Improved prediction accuracy and productivity.
Patent Information
- Application Number
- CN202211152612.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-09-21
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2042-09-21
AI Technical Summary
In the high-temperature gasification process of multi-component organic waste, it is difficult for the prior art to accurately predict gas-producing components, and long-term stable measurement of gasifier temperature is difficult, resulting in difficulty in controlling gas-producing components and affecting production efficiency and resource utilization efficiency.
The BP neural network model is used to combine principal component analysis and genetic algorithm, and optimize data through box graph method and normalized processing to construct a BP neural network model, predict the gasifier temperature and optimize the network weight and threshold value to realize the gas production component Efficient prediction.
It improves the accuracy and production efficiency of gas-producing components prediction, reduces network computing volume and operating costs, avoids the trap of local optimality, and realizes long-term stable prediction of gasifier temperature.
Smart Images

Figure CN115762664B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the fields of energy industry and chemical engineering, and specifically is a method for predicting gasification gas production components in a multi-component organic waste treatment process. Background Art
[0002] Industrial organic hazardous wastes are solid or liquid organic items and substances that have lost their original value or have not lost their value but have been discarded or abandoned in industrial production, and they have one or more hazardous characteristics. Industrial organic wastes mainly include solid wastes and liquid wastes such as distillation residues, waste activated carbon, waste catalysts, industrial sludge, waste tires, waste printed circuit boards, and waste organic solvents. Industrial organic hazardous wastes are solid or liquid organic items and substances that have lost their original value or have not lost their value but have been discarded or abandoned in industrial production, and they have one or more hazardous characteristics. Industrial organic hazardous wastes are solid or liquid organic items and substances that have lost their original value or have not lost their value but have been discarded or abandoned in industrial production, and they have one or more hazardous characteristics.
[0003] Multi-component organic waste treatment is the use of a gasifier to dispose of multiple organic wastes in coordination with other raw materials or fuels. While meeting the normal production requirements of the enterprise and ensuring product quality and environmental safety, it achieves harmless disposal and resource utilization of organic wastes. At the same time, it reduces the consumption of fossil fuels and achieves carbon emission reduction.
[0004] Gasifiers mainly include Shell, GSP, Prenflo, Eagle and other gasifiers for dry coal feeding and Texaco, E-GAS, multi-nozzle opposed type and other gasifiers for water-coal slurry feeding, mainly producing synthetic ammonia and CO, H 2 In the gasifier, the mixed slurry composed of multi-component organic waste undergoes pyrolysis and gasification under high temperature conditions to produce CO 2 , CO, H 2 The main synthesis gas is CO and H 2 It is called effective gas. Effective gas is the target product of the gasification process and can be used as the raw material of other industrial products in subsequent production. Therefore, how to stably obtain gas with a higher proportion of effective gas is one of the research focuses of the multi-component organic waste treatment process.
[0005] The gasification process of multi-component organic waste is complicated. A variety of different gasification reactions will occur simultaneously under high temperature and high pressure. As a result, in actual production, it is difficult for slurry batching personnel responsible for slurrying with various hazardous wastes to effectively estimate the final gas-producing components of the composite slurry. Due to the lack of scientific guidance on slurry batching, it is difficult for operators to achieve stable control of the gasification gas components by changing the gasifier operating parameters.
[0006] The prediction of gas production components by experimental research using pilot plants or large-scale production plants is not only cumbersome and time-consuming, but also has high operating costs. Relatively speaking, modeling is an excellent alternative method to save money and time. By establishing a relevant neural network model, the gas components of gasification gas can be predicted in advance, thereby achieving guidance on slurry matching and optimization of gasifier operating parameters, improving plant production efficiency and reducing raw material loss.
[0007] At present, there are two main models that can be used to predict the composition of gasification synthesis gas. One is a mechanism model based on physical and chemical mechanism equations such as mass conservation, energy conservation, heat transfer, and chemical reaction kinetics. Its characteristics are that each parameter has a clear physical meaning and the parameters are easy to adjust, but it requires the modeler to have a full understanding of the entire reaction process. It is not suitable for high-temperature gasification processes of multi-component organic waste with complex reactions. When studying gasification processes with complex reaction mechanisms, some researchers choose to establish data-driven models, train the models through a large amount of relevant data, and then realize the prediction of the gasification process.
[0008] When dealing with large amounts of data or complex relationships, the back propagation (BP) neural network can simplify the estimation process and accurately find the complex internal relationships between the data. However, the BP neural network has two disadvantages:
[0009] First, the reaction of multi-component organic waste gasification process is complicated, and the influencing factors of gas production components are numerous and highly correlated. If the influencing factors are directly used as input variables of the neural network, the network structure of the prediction model will be very complex and the prediction effect will be poor;
[0010] Second, the weights and thresholds of the BP network may remain at the local optimum during the training process, which will limit the prediction effect of the network.
[0011] In addition, the gasifier temperature is an important related variable of the gas production components. Since the thermocouple is easily damaged during the high-temperature gasification process, it is difficult to measure the gasifier temperature stably for a long time. It is necessary to obtain approximate gasifier temperature data as an input variable for the prediction of gas production components. Summary of the invention
[0012] The invention provides a method for predicting gas production components in a high-temperature gasification process of multi-component organic waste, which can efficiently and accurately predict the gas production components.
[0013] A method for predicting gas production components during high-temperature gasification of multi-component organic wastes, comprising:
[0014] (1) obtaining variable data and target data, and performing box plot method, normalization processing and principal component analysis on the variable data to obtain a first training sample set and a second training sample set, respectively; performing box plot method and normalization processing on the target data to obtain a gasifier temperature target data set and a gas production component target data set;
[0015] (2) constructing a BP neural network model, the BP neural network model comprising a first input layer, a first hidden layer, a second input layer, a second hidden layer, and an output layer, inputting the first training sample into the first input layer and the first hidden layer in sequence to obtain a predicted gasifier temperature, using the error between the predicted gasifier temperature and the gasifier temperature target data as a first loss function, and training based on the first training sample set through the first loss function to obtain a first threshold of the first input layer and the first hidden layer, as well as a first weight between the first input layer and the first hidden layer;
[0016] Setting weights among the second input layer, the second hidden layer and the output layer, as well as thresholds of the second input layer, the second hidden layer and the output layer, and using a genetic algorithm to update the set weights and thresholds to obtain a second threshold and a second weight; constructing an initial BP neural network model through the first threshold, the first weight, the second threshold and the second weight;
[0017] (3) The first training sample set is input into the first input layer of the initial BP neural network model and the second training sample set is input into the second input layer of the initial BP neural network model to obtain gas production component prediction data, a second loss function is constructed through the gas production component prediction data and the gas production component target data set, and the initial BP neural network model is trained through the second loss function based on the first training sample set and the second training sample set to obtain a gas production component network model.
[0018] The variable data includes a first characteristic variable and a second characteristic variable, and the target data includes a gas production component target variable and a gasifier temperature target variable;
[0019] Among them, the first characteristic variable is the cooling water inlet flow rate, the cooling water inlet temperature, the cooling water outlet flow rate, the cooling water outlet temperature, the cooling water pressure, the quench chamber temperature and the syngas temperature out of the quench chamber;
[0020] The second characteristic variable is slurry volume flow, slurry mass flow, slurry pressure, slurry concentration, slurry ash, slurry calorific value, waste liquid volume flow, waste liquid mass flow, waste liquid inlet gasifier pressure, waste liquid moisture, waste liquid calorific value, main pipe oxygen flow, center pipe oxygen flow, oxygen inlet pressure, waste liquid inlet slurry pipe flow and gasifier pressure;
[0021] The target variable of gas production component is gas production CO 2 ratio, gas production CO ratio and gas production H 2 Proportion;
[0022] The gasifier temperature target variable is the temperature at the top of the gasifier measured by a thermocouple.
[0023] The obtained variable data are subjected to box plot method and normalization processing in turn, including:
[0024] Arrange the original data from small to large, divide the arranged original data into n equal parts, and obtain the values of n-1 segmentation points, delete the data larger than the upper edge and smaller than the lower edge to obtain the optimized data, and normalize the optimized data into preprocessed variables in the interval [-1,1];
[0025] Among them, the lower edge L d for:
[0026] L d =Q 1 -1.5IQR
[0027] IQR=Q n-1 -Q 1
[0028] Upper edge L u for:
[0029] L u =Q n-1 +1.5QIR
[0030] Among them, Q 1 is the value of the first split point, Q n-1 is the n-1th segmentation point.
[0031] The preprocessing variables include performing box plot method and normalization processing on the first eigenvector in sequence to obtain the first preprocessing variable, and performing box plot method and normalization processing on the second eigenvector in sequence to obtain the second preprocessing variable, performing principal component analysis on the first preprocessing variable to obtain the first training sample set, and performing principal component analysis on the second preprocessing variable to obtain the second training sample set.
[0032] The first preprocessing variable is subjected to principal component analysis to obtain a first training sample set, including:
[0033] Calculate the covariance matrix A of the preprocessed variables:
[0034]
[0035] Where n is the number of samples of the first preprocessing variable, and x is the first preprocessing variable data set.
[0036] The eigenvalue decomposition method is used to obtain the eigenvalues and corresponding eigenvectors of the covariance matrix, and the eigenvalues are arranged from large to small;
[0037] Get the cumulative contribution rate K of the first p principal components p for:
[0038]
[0039] Where M is the number of eigenvalues, j is the index of the eigenvalue; select the cumulative contribution rate K p The first a principal components that are above the threshold, that is, the first a eigenvalues, and their corresponding eigenvectors are used as row vectors to form an eigenvector matrix P; the first preprocessing variable and the corresponding eigenvector matrix are multiplied to obtain the first training sample set.
[0040] The BP neural network model includes a first input layer, a first hidden layer, a second input layer, a second hidden layer, and an output layer;
[0041] The number of nodes in the first input layer is the number of principal components a that meet the contribution rate threshold in the eigenvalues of the first preprocessing variable;
[0042] The number of nodes b in the first hidden layer is:
[0043]
[0044] Where c is the number of principal components that meet the contribution rate threshold in the eigenvalues of the second preprocessed variable, h 1 To adjust the constant, the first input layer is fully connected to the first hidden layer;
[0045] The number of nodes in the second input layer is c+1, wherein the second training sample set is input into the c neurons in the second input layer, and the c+1th neuron in the second input layer is fully connected to the first hidden layer;
[0046] The number of nodes d in the second hidden layer is:
[0047]
[0048] Among them, e is the number of nodes in the output layer, h 2 To adjust the constant, the second input layer is fully connected to the second hidden layer;
[0049] The output layer has multiple nodes, and the output layer is fully connected to the second hidden layer;
[0050] The transfer function from the first input layer to the first hidden layer and the transfer function from the second input layer to the second hidden layer are both f 1 (m) is:
[0051]
[0052] Where m is the input value of the transfer function.
[0053] The transfer function from the first hidden layer to the c+1th neuron in the second input layer, and the linear transfer function from the second hidden layer to the output layer f 2 (n) is:
[0054] f 2 (n) = n
[0055] Where n is the input value of the linear transfer function;
[0056] The output Pe of the e-th neuron in the first hidden layer is .
[0057] The output of the e-th neuron in the first hidden layer is B e for:
[0058]
[0059] Where f is the index of the neuron in the first input layer, e is the index of the neuron in the first hidden layer, and A f is the output of the fth neuron in the first input layer, w fe is the weight value between the fth neuron in the first input layer and the eth neuron in the first hidden layer, k e is the threshold of the e-th neuron in the first hidden layer;
[0060] The output C of the c+1th neuron in the second input layer c+1 for:
[0061]
[0062] Among them, k c+1 C c+1 The threshold of the neuron, w e is the eth neuron in the first hidden layer and the Cth neuron in the C layer c+1 The weights between neurons;
[0063] The output D of the qth neuron in the second hidden layer q for:
[0064]
[0065] Among them, l is the neuron index of the second input layer, q is the neuron index of the second hidden layer, and C p is the output of the pth neuron in the second output layer, w lq is the weight value between the lth neuron of the second output layer and the qth node of the second hidden layer, k q is the threshold of the qth neuron in the second hidden layer;
[0066] The output E of the rth neuron in the output layer r for:
[0067]
[0068] Among them, r is the index of the neuron node in the output layer, w qr is the weight value between the qth neuron in the second hidden layer and the rth node in the output layer, k r is the threshold of the rth neuron in the output layer.
[0069] The first loss function O 1 for:
[0070]
[0071] Where g is the index of the first training sample, G is the total number of the first training samples, is the predicted value of the gasifier temperature of the g-th training sample, t s is the target data of gasifier temperature for the g-th training sample;
[0072] The second loss function O 2 for:
[0073]
[0074] Where, r is the index of the neuron node of the output layer, R is the number of neurons in the output layer, s is the index of the second training sample, S is the number of the second training samples, is the predicted value of gas production components, Z rs It is the target data of gas production components.
[0075] Compared with the prior art, the present invention has the following beneficial effects:
[0076] (1) The present invention selects BP neural network as the implementation method of the prediction model of gas production components in the high-temperature gasification process of multi-component organic waste. Since the multi-component organic waste collaborative treatment process is complicated due to the complex source of hazardous waste, the chemical reactions in the gasifier are complex and changeable, and the neural network does not need to have too much understanding of the entire reaction process. As long as sufficient data is provided for training, accurate calculation results can be obtained.
[0077] (2) The present invention uses the principal component analysis method to process the input data. Principal component analysis is a data dimension reduction technology that replaces a large number of original variables with a small number of unrelated principal components, thereby reducing the dimension of the characteristic variables while retaining as much information as possible in the original data. After processing by principal component analysis, the number of characteristic variables input to the BP neural network is reduced, and the number of neurons in the network structure is reduced accordingly, which simplifies the network structure and reduces the amount of network calculation.
[0078] (3) The initial weights and thresholds of the traditional BP neural network are randomly generated, and the subsequent network training optimization is gradually performed through gradient descent based on the initial weights and thresholds. In this case, the gradient descent algorithm for weight adjustment can easily cause the network to fall into a local optimal point. The present invention uses a genetic algorithm to optimize the initial weights and thresholds of the second input layer, the second hidden layer, and the output layer of the BP neural network. As a random global search optimization method, the genetic algorithm can optimize the initial weights and thresholds of the BP neural network, and then gradually optimize the optimized weights and thresholds through the gradient descent of the loss function, which is conducive to the network converging to a position with a smaller prediction error in subsequent training and avoiding falling into a local optimal point.
[0079] (4) The gasifier temperature in the high-temperature gasification process of multi-component organic waste disclosed in the prior art is measured by thermocouples, but thermocouples are easily damaged and it is impossible to obtain the gasifier temperature stably for a long time. The present invention obtains an approximate value of the gasifier temperature by training the first input layer and the first hidden layer of the BP neural network through a loss function based on limited gasifier temperature measurement values, and then inputs the approximate value of the gasifier temperature and the second training sample set into the second input layer for subsequent prediction of the gas production components. The gasifier temperature has a great influence on the gas production components. The Net2 part of the above-mentioned BP neural network is used to predict the gasifier temperature, so that the predicted gasifier temperature can be obtained stably and more accurately for a long time. Using it as an input variable of the Net1 part of the BP neural network improves the prediction accuracy of the gas production component prediction model. BRIEF DESCRIPTION OF THE DRAWINGS
[0080] Figure 1 A schematic diagram of the process of multi-component organic waste coordinated treatment provided for a specific method;
[0081] Figure 2 A flowchart of a method for predicting gas components during high-temperature gasification of multi-component organic waste provided as a specific method;
[0082] Figure 3 The block diagram of the BP neural network model Net1 provided for the specific method;
[0083] Figure 4 The block diagram of the BP neural network model Net2 provided for the specific method;
[0084] Figure 5 PCA-GA-BP model predicts gas production CO 2 Comparison of proportions with real data;
[0085] Figure 6 Comparison of the CO ratio of gas production predicted by the PCA-GA-BP model with the actual data;
[0086] Figure 7PCA-GA-BP model predicts gas production H 2 Comparison of proportions with real data. DETAILED DESCRIPTION
[0087] The present invention is further described and illustrated below in conjunction with specific embodiments. The technical features of each embodiment of the present invention can be combined accordingly without conflicting with each other.
[0088] Figure 1 This is a process diagram of the coordinated treatment of multi-component organic waste. The process flow of this project is as follows: the recycled hazardous waste is divided into solid waste and waste liquid according to its state. After the solid waste is ground, it is mixed with waste liquid with low calorific value, water, coal and additives to form a mixed slurry. The waste liquid with high calorific value is mixed to form a mixed waste liquid. The slurry, waste liquid and oxygen undergo multiple gasification reactions at a high temperature of 1200-1400℃. The gas products of the reaction are washed with quenching water, dust removed and cooled in the quenching chamber to obtain CO 2 ,CO,H 2 The main syngas is the inorganic matter, which will melt and solidify into glassy slag. The four-channel burner of the gasifier is directly extended into the high-temperature gasifier, and the required temperature is reduced by the cooling water jacket. In addition, the temperature of the gasifier at the top of the gasifier is measured by a thermocouple.
[0089] The present invention provides a method for predicting gas components during high-temperature gasification of multi-component organic waste. Figure 2 As shown, including:
[0090] (1) Collecting data obtained during the operation of an actual multi-component organic waste high-temperature gasification furnace to obtain variable data and target data, wherein the variable data includes a first characteristic variable and a second characteristic variable, and the target data includes a gas production component target variable and a gasification furnace temperature target variable;
[0091] The first characteristic variable is the cooling water inlet flow rate, cooling water inlet temperature, cooling water outlet flow rate, cooling water outlet temperature, cooling water pressure, quench chamber temperature and quench chamber syngas temperature, recorded as x 1 -x 7 ;
[0092] The second characteristic variable is the slurry volume flow rate, slurry mass flow rate, slurry pressure, slurry concentration, slurry ash content, slurry calorific value, waste liquid volume flow rate, waste liquid mass flow rate, waste liquid inlet gasification furnace pressure, waste liquid moisture, waste liquid calorific value, main pipe oxygen flow rate, central pipe oxygen flow rate, oxygen inlet furnace pressure, waste liquid inlet slurry pipe flow rate and gasification furnace pressure, recorded as y 1 -y 17 ;
[0093] The target variable of gas production component is gas production CO 2ratio, gas production CO ratio and gas production H 2 Ratio, recorded as z 1 -z 3 ;
[0094] Gasifier Temperature The target variable is the gasifier temperature, recorded as t.
[0095] (2) The variable data are subjected to box plot method in turn and normalized to obtain preprocessed variables; including:
[0096] Arrange the original data from small to large, and divide the arranged original data into 4 equal parts. The value of the first division point is called the lower quartile, denoted by Q 1 , the second split point is called the median, denoted by Q 2 , the third split point is called the upper quartile, denoted by Q 3 ; Lower quartile Q 1 and the upper quartile Q 3 The gap is called the interquartile range (IQR), and the formula for calculating IQR is:
[0097] IQR=Q 3 -Q 1
[0098] The minimum value of normal data, that is, the lower edge is:
[0099] Lower edge = Q 1 -1.5IQR The maximum value of normal data, that is, the upper edge is:
[0100] Upper edge = Q 3 +1.5IQR
[0101] Data greater than the upper edge or less than the lower edge are considered abnormal data points and are deleted to obtain optimized data. The optimized data are normalized and converted into preprocessing variables in the interval [-1, 1]. The preprocessing variables include the first eigenvector, which is processed by the box plot method and normalization in turn to obtain the first preprocessing variable And the second eigenvector is processed by box plot method and normalization in turn to obtain the second preprocessing variable The gas production component target variables are processed by box plot method and normalization in turn to obtain the gas production component target data set Z rs , and the gasifier temperature target variable is processed by box plot method and normalization in turn to obtain the gasifier temperature target data set.
[0102] Among them, the normalization function X used new for:
[0103]
[0104] Among them, Xi is the optimized data, X min is the minimum value of the optimized data, X max is the maximum value of the optimized data.
[0105] Using the method in step (2), the target data are normalized by the box plot method to obtain the gasifier temperature target data set and the gas production component target data set.
[0106] (3) Performing principal component analysis (PCA) on the first preprocessed variables to obtain a first training sample set, including:
[0107] 3.1: First preprocessing variable The first preprocessing variable data set x is composed and the covariance matrix A of the preprocessing variables is calculated:
[0108]
[0109] Where n is the number of samples of the first preprocessing variable, and x is the first preprocessing variable data set.
[0110] 3.2: Use the eigenvalue decomposition method to obtain the eigenvalue λ of the covariance matrix 1 ,λ 2 , ..., λ 7 And the corresponding eigenvectors, and arrange the eigenvalues from large to small;
[0111] 3.3: Get the cumulative contribution rate K of the first p principal components p for:
[0112]
[0113] Where M is the number of eigenvalues and j is the index of the eigenvalue;
[0114] 3.4: Select the cumulative contribution rate K p The first a principal components that reach more than 95%, that is, the first a eigenvalues, and their corresponding eigenvectors are used as row vectors to form the eigenvector matrix P;
[0115] 3.5: Multiply the first preprocessing variable and the corresponding eigenvector matrix to obtain the first training sample set X:
[0116] X=Px.
[0117] 3.6: The feature variable y obtained after preprocessing in step (2) 1 -y 17 The data is data set y. Perform the same processing as steps 3.1-3.5 on y and select the cumulative contribution rate K p The first k principal components that reach more than 95% are used as the input second data set Y.
[0118] (4) Figure 3 As shown, a BP neural network model Net1 is constructed. The BP neural network model Net1 includes a first input layer (layer A), a first hidden layer (layer B), a second input layer (layer C), a second hidden layer (layer D), and an output layer (layer E).
[0119] The number of nodes in the first input layer is the number of principal components a that meet the contribution rate threshold in the eigenvalues of the first preprocessing variable;
[0120] The number of nodes b in the first hidden layer is:
[0121]
[0122] Where c is the number of principal components that meet the contribution rate threshold in the eigenvalues of the second preprocessed variable, h 1 To adjust the constant, the first input layer is fully connected to the first hidden layer;
[0123] The number of nodes in the second input layer is c+1, wherein the second training sample set is input into the c neurons in the second input layer, and the c+1th neuron in the second input layer is fully connected to the first hidden layer;
[0124] The number of nodes d in the second hidden layer is:
[0125]
[0126] Among them, e is the number of nodes in the output layer, h 2 To adjust the constant, the second input layer is fully connected to the second hidden layer;
[0127] The output layer has multiple nodes, and the output layer is fully connected to the second hidden layer;
[0128] The transfer function from the first input layer to the first hidden layer and the transfer function from the second input layer to the second hidden layer are both f 1 (m) is:
[0129]
[0130] Where m is the input value of the transfer function.
[0131] The transfer function from the first hidden layer to the c+1th neuron in the second input layer, and the linear transfer function from the second hidden layer to the output layer f 2 (n) is:
[0132] f 2 (n) = n
[0133] Where n is the input value of the linear transfer function;
[0134] The output Pe of the e-th neuron in the first hidden layer is .
[0135] , the output of the e-th neuron in the first hidden layer is B e for:
[0136]
[0137] Where f is the index of the neuron in the first input layer, e is the index of the neuron in the first hidden layer, and A f is the output of the fth neuron in the first input layer, w fe is the weight value between the fth neuron in the first input layer and the eth neuron in the first hidden layer, k e is the threshold of the e-th neuron in the first hidden layer;
[0138] The output C of the c+1th neuron in the second input layer c+1 for:
[0139]
[0140] Among them, k c+1 C c+1 The threshold of the neuron, w e is the eth neuron in the first hidden layer and the Cth neuron in the C layer c+1 The weights between neurons;
[0141] The output D of the qth neuron in the second hidden layer q for:
[0142]
[0143] Among them, l is the neuron index of the second input layer, q is the neuron index of the second hidden layer, and C p is the output of the pth neuron in the second output layer, w lq is the weight value between the lth neuron of the second output layer and the qth node of the second hidden layer, k q is the threshold of the qth neuron in the second hidden layer;
[0144] The output E of the rth neuron in the output layer r for:
[0145]
[0146] Among them, r is the index of the neuron node in the output layer, w qr is the weight value between the qth neuron in the second hidden layer and the rth node in the output layer, k r is the threshold of the rth neuron in the output layer.
[0147] (5) First, the BP neural network Net2 is trained by the first loss function to obtain the first threshold of the first input layer and the first hidden layer, as well as the first weight between the first input layer and the first hidden layer, such as Figure 4 As shown, BP neural network Net2 is a neural network that combines the Cth layer in layer A, layer B and layer C. c+1 A neuron.
[0148] The first loss function O 1 for:
[0149]
[0150] Where g is the index of the first training sample, G is the total number of the first training samples, is the predicted value of the gasifier temperature of the g-th training sample, t g is the target data of gasifier temperature for the g-th training sample;
[0151] (6) setting weights among the second input layer, the second hidden layer, and the output layer, as well as thresholds of the second input layer, the second hidden layer, and the output layer, and using a genetic algorithm to update the set weights and thresholds to obtain a second threshold and a second weight, specifically including:
[0152] 6.1: Set the population size M to 50, randomly generate the initial population, and the chromosome of each individual in the population corresponds to a set of binary codes of the weights between the C layer and the D layer, the weights between the D layer and the E layer, the threshold of the D layer, and the threshold of the E layer in the BP neural network Net1 established in step 4;
[0153] 6.2: The fitness function F of the genetic algorithm is defined as follows:
[0154]
[0155] Among them, O is the prediction error value of the second loss function in the Net1 model with the weight and threshold set;
[0156] 6.3, based on fitness, individuals are randomly selected to form a new population, and the probability P of parent individual t being selected as offspring is calculated using the following relationship: t :
[0157]
[0158] Among them, U is the population size, u is the number of individuals in the population, and F t is the fitness of individual t;
[0159] 6.4: Generate a new population through crossover and mutation, where the crossover method is single-point crossover, which means randomly pairing the offspring selected in step 6.3, randomly setting a crossover point in the individual pairs participating in the crossover, and then exchanging the chromosomes of the two individuals on one side of the point. The crossover probability of each pair of individuals is set to 0.7; the mutation operation refers to reversing the value of the random position in the chromosome of the mutated individual, that is, 1 becomes 0, 0 becomes 1, and the mutation probability is set to 0.01;
[0160] Step 6.5, repeating steps 6.3 and 6.4 until the maximum genetic generation is reached, ending the genetic process, obtaining the optimized second input layer, second hidden layer and output layer, obtaining the second threshold and the second weight, wherein the maximum genetic generation is set to 50; constructing the initial BP neural network model through the first threshold, the first weight, the second threshold and the second weight;
[0161] (7) Inputting the first training sample set X into the first input layer of the initial BP neural network model and inputting the second training sample set Y into the second input layer of the initial BP neural network model to obtain gas production component prediction data, constructing a second loss function through the gas production component prediction data and the gas production component target data set, and training the initial BP neural network model through the second loss function based on the first training sample set and the second training sample set to obtain a gas production component network model.
[0162] The second loss function O 2 for:
[0163]
[0164] Where, r is the index of the neuron node of the output layer, R is the number of neurons in the output layer, s is the index of the second training sample, S is the number of the second training samples, is the predicted value of gas production components, Z rs It is the target data set of gas production components.
[0165] Figure 5 For CO 2 The comparison chart of the predicted value and the actual value shows that CO 2 The predicted results of the proportions are very close to the real data, and the prediction effect is good;
[0166] Figure 6 This is a comparison chart of the predicted value and the true value of the CO ratio. Similarly, it can be seen that the predicted result of the CO ratio is very close to the real data, and the model prediction accuracy is high;
[0167] Figure 7 H 2 Comparison chart of the predicted value and the true value and H 2 The relative percentage error diagram of the proportional prediction value shows that H2 The predicted effect of ratio is higher than that of CO 2 The prediction effect of CO is better, and the predicted value curve is highly consistent with the true value curve;
[0168] At the same time, for comparison, we established BP model, PCA-BP model and GA-BP model to analyze the CO 2 , CO and H 2 The proportion of is predicted, and the average relative percentage error of the prediction results is calculated. The prediction errors of different models are shown in Table 1.
[0169] Table 1 Prediction results of gas production components of different models
[0170]
[0171]
[0172] It can be seen that the prediction model of the gas production components in the high-temperature gasification process of multi-component organic waste can achieve a high accuracy. The PCA-GA-BP model is very accurate for CO 2 , CO and H 2 The prediction errors of the ratios are 2.26%, 1.53% and 0.75% respectively, which are within the acceptable range in industry. The comparison of the prediction results of each model also shows that principal component analysis and genetic algorithm have a positive effect on the prediction of BP neural network. The prediction effect of PCA-GA-BP model is the best among the four models, and the prediction errors of the three gas ratios are the lowest among all models.
[0173] In summary, the method for predicting gas production components in the high-temperature gasification process of multi-component organic waste of the present invention can accurately and effectively simulate and estimate the synthesis gas and gas components generated in the high-temperature gasification process of multi-component organic waste, which is helpful for the smooth operation of the coordinated treatment process of multi-component organic waste.
Claims
1. A method for predicting gas components during high-temperature gasification of multi-component organic waste, characterized in that: include: (1) obtaining variable data and target data, and performing box plot method, normalization processing and principal component analysis on the variable data to obtain a first training sample set and a second training sample set, respectively; performing box plot method and normalization processing on the target data to obtain a gasifier temperature target data set and a gas production component target data set; (2) constructing a BP neural network model, the BP neural network model comprising a first input layer, a first hidden layer, a second input layer, a second hidden layer, and an output layer, inputting the first training sample into the first input layer and the first hidden layer in sequence to obtain a predicted gasifier temperature, using the error between the predicted gasifier temperature and the gasifier temperature target data as a first loss function, and training based on the first training sample set through the first loss function to obtain a first threshold of the first input layer and the first hidden layer, as well as a first weight between the first input layer and the first hidden layer; Setting weights among the second input layer, the second hidden layer and the output layer, as well as thresholds of the second input layer, the second hidden layer and the output layer, and using a genetic algorithm to update the set weights and thresholds to obtain a second threshold and a second weight; constructing an initial BP neural network model through the first threshold, the first weight, the second threshold and the second weight; (3) inputting the first training sample set into the first input layer of the initial BP neural network model and the second training sample set into the second input layer of the initial BP neural network model to obtain gas production component prediction data, constructing a second loss function through the gas production component prediction data and the gas production component target data set, and training the initial BP neural network model through the second loss function based on the first training sample set and the second training sample set to obtain a gas production component network model; The variable data includes a first characteristic variable and a second characteristic variable, and the target data includes a gas production component target variable and a gasifier temperature target variable; Among them, the first characteristic variable is the cooling water inlet flow rate, the cooling water inlet temperature, the cooling water outlet flow rate, the cooling water outlet temperature, the cooling water pressure, the quench chamber temperature and the syngas temperature out of the quench chamber; The second characteristic variable is slurry volume flow, slurry mass flow, slurry pressure, slurry concentration, slurry ash, slurry calorific value, waste liquid volume flow, waste liquid mass flow, waste liquid inlet gasifier pressure, waste liquid moisture, waste liquid calorific value, main pipe oxygen flow, center pipe oxygen flow, oxygen inlet pressure, waste liquid inlet slurry pipe flow and gasifier pressure; The target variables of gas production components are the gas production CO2 ratio, gas production CO ratio and gas production H2 ratio; The gasifier temperature target variable is the temperature measured by the thermocouple at the top of the gasifier; The obtained variable data are subjected to box plot method and normalization processing in turn, including: Arrange the original data from small to large, divide the arranged original data into n equal parts, and obtain the values of n-1 segmentation points, delete the data larger than the upper edge and smaller than the lower edge to obtain the optimized data, and normalize the optimized data into preprocessed variables in the interval [-1,1]; Among them, the lower edge L d for: <h2 style=";text-align:left;direction:ltr">L<h2 style=";text-align:left;direction:ltr"> d <h2 style=";text-align:left;direction:ltr"> =Q1-1.5IQR IQR=Q n-1 -Q1 Upper edge L u for: <h2 style=";text-align:left;direction:ltr">L<h2 style=";text-align:left;direction:ltr"> u <h2 style=";text-align:left;direction:ltr"> =Q<h2 style=";text-align:left;direction:ltr"> n-1 <h2 style=";text-align:left;direction:ltr"> +1.5IQR Among them, Q1 is the value of the first split point, Q n-1 is the n-1th segmentation point; The preprocessing variables include sequentially performing a box plot method and a normalization process on the first eigenvector to obtain a first preprocessing variable, sequentially performing a box plot method and a normalization process on the second eigenvector to obtain a second preprocessing variable, performing a principal component analysis on the first preprocessing variable to obtain a first training sample set, and performing a principal component analysis on the second preprocessing variable to obtain a second training sample set; The first preprocessing variable is subjected to principal component analysis to obtain a first training sample set, including: Calculate the covariance matrix A of the preprocessed variables: Where n is the number of samples of the first preprocessing variable, and x is the first preprocessing variable data set; The eigenvalue decomposition method is used to obtain the eigenvalues and corresponding eigenvectors of the covariance matrix, and the eigenvalues are arranged from large to small; Get the cumulative contribution rate K of the first p principal components p for: Where M is the number of eigenvalues, j is the index of the eigenvalue; select the cumulative contribution rate K p The first a principal components that reach above the threshold, i.e., the first a eigenvalues, and their corresponding eigenvectors are used as row vectors to form an eigenvector matrix P; the first preprocessing variable and the corresponding eigenvector matrix are multiplied to obtain a first training sample set; The first loss function O1 is: Where g is the index of the first training sample, G is the total number of the first training samples, is the predicted value of the gasifier temperature of the g-th training sample, t g is the target data of gasifier temperature for the g-th training sample; The second loss function O2 is: Among them, r is the index of the neuron node of the output layer, R is the number of neurons in the output layer, s is the index of the second training sample, S is the number of the second training samples, is the predicted value of gas production components, Z rs It is the target data of gas production components.
2. The method for predicting gas components during high-temperature gasification of multi-component organic waste according to claim 1, characterized in that: The BP neural network model includes a first input layer, a first hidden layer, a second input layer, a second hidden layer, and an output layer; The number of nodes in the first input layer is the number of principal components a that meet the contribution rate threshold in the eigenvalues of the first preprocessed variable; The number of nodes b in the first hidden layer is: Where c is the number of principal components that meet the contribution rate threshold in the eigenvalue of the second preprocessed variable, h1 is the adjustment constant, and the first input layer is fully connected to the first hidden layer; The number of nodes in the second input layer is c+1, wherein the second training sample set is input into the c neurons in the second input layer, and the c+1th neuron in the second input layer is fully connected to the first hidden layer; The number of nodes d in the second hidden layer is: Among them, e is the number of nodes in the output layer, h2 is the adjustment constant, and the second input layer is fully connected to the second hidden layer; The output layer has multiple nodes, and the output layer is fully connected to the second hidden layer; The transfer function from the first input layer to the first hidden layer and the transfer function from the second input layer to the second hidden layer are both f1(m), and f1(m) is: Where m is the input value of the transfer function; The transfer function from the first hidden layer to the c+1th neuron in the second input layer, and the linear transfer function f2(n) from the second hidden layer to the output layer are: f2(n)=n Where n is the input value of the linear transfer function.
3. The method for predicting gas components during high-temperature gasification of multi-component organic waste according to claim 2, characterized in that: The output of the e-th neuron in the first hidden layer is B e for: Where f is the index of the neuron in the first input layer, e is the index of the neuron in the first hidden layer, and A f is the output of the fth neuron in the first input layer, w fe is the weight value between the fth neuron in the first input layer and the eth neuron in the first hidden layer, k e is the threshold of the e-th neuron in the first hidden layer; The output C of the c+1th neuron in the second input layer c+1 for: Among them, k c+1 C c+1 The threshold of the neuron, w e is the eth neuron in the first hidden layer and the Cth neuron in the C layer c+1 The weights between neurons; The output D of the qth neuron in the second hidden layer q for: Among them, l is the neuron index of the second input layer, q is the neuron index of the second hidden layer, and C p is the output of the pth neuron in the second output layer, w lq is the weight value between the lth neuron of the second output layer and the qth node of the second hidden layer, k q is the threshold of the qth neuron in the second hidden layer; The output E of the rth neuron in the output layer r for: Among them, r is the index of the neuron node in the output layer, w qr is the weight value between the qth neuron in the second hidden layer and the rth node in the output layer, k r is the threshold of the rth neuron in the output layer.
