CN-122018447-A - Parameter optimization method for retired fan blade pyrolysis system based on dynamic weighting feature fusion algorithm control

CN122018447ACN 122018447 ACN122018447 ACN 122018447ACN-122018447-A

Abstract

The invention relates to a parameter optimization method of a retired fan blade pyrolysis system based on dynamic weighting characteristic fusion algorithm control, which comprises the steps of firstly, constructing a pyrolysis product distribution model based on a physical information neural network, secondly, constructing a dynamic weighting characteristic fusion model of the retired fan blade pyrolysis system based on an attention mechanism, thirdly, carrying out weighted summation on normalized prediction results of three sub-modules of a limit gradient lifting algorithm, a multi-layer sensor and the physical information neural network, outputting fusion prediction results, fourthly, constructing a multi-objective optimization model and solving parameters by utilizing the prediction results of the fusion model based on the prediction results of the dynamic weighting characteristic fusion model, and fifthly, dynamically adjusting and controlling the pyrolysis system.

Inventors

LU QIANG
ZHANG QILIANG
JIANG DEKAI
ZHANG BAOXIN
XU MINGXIN
HU BIN
LI KAI

Assignees

华北电力大学

Dates

Publication Date: 20260512
Application Date: 20251203

Claims (8)

1. A retired fan blade pyrolysis system parameter optimization method based on dynamic weighting feature fusion algorithm control is characterized by comprising the following steps: S1, constructing a pyrolysis product distribution model based on a physical information neural network, and providing a physical constraint basic model of a pyrolysis process; S2, determining an objective function and constraint conditions, constructing a dynamic weighting feature fusion model of a retired fan blade pyrolysis system based on an attention mechanism, and constructing a fusion frame to realize multi-source model cooperation, wherein the multi-source model cooperation comprises three sub-modules of a limit gradient lifting algorithm, a multi-layer sensor and a physical information neural network, wherein the three sub-modules are complemented with each other through a statistical fitting algorithm, a nonlinear modeling algorithm and a physical constraint dimension forming algorithm, a sub-module outputs a normalization and splicing comprehensive data set to realize cross-module data linkage, and finally the attention mechanism dynamically distributes the weights of all the modules and performs weighted summation to output a prediction result with fitting precision and physical consistency; S3, carrying out weighted summation on the normalized prediction results of the limit gradient lifting algorithm, the multi-layer perceptron and the physical information neural network according to the weight output by the attention mechanism, outputting a fusion prediction result, and completing the input, training and output of a dynamic weighting feature fusion model; S4, based on a prediction result of the dynamic weighting characteristic fusion model, taking the yield and quality of a pyrolysis product and the energy consumption of a pyrolysis process as optimization targets, taking key process parameters of a pyrolysis system as parameter variables to be optimized, establishing a multi-target optimization model of the retired fan blade pyrolysis system, solving the model by adopting a particle swarm optimization algorithm to obtain a process parameter solution set, and constructing the multi-target optimization model and solving parameters by utilizing the fusion model prediction result; and S5, dynamically adjusting and controlling the pyrolysis system according to the optimized technological parameter solution set by monitoring key parameters of the pyrolysis process in real time and according to a prediction result of the dynamic weighting characteristic fusion model.
2. The parameter optimization method for the retired fan blade pyrolysis system based on dynamic weighted feature fusion algorithm control according to claim 1, wherein the construction of the pyrolysis product distribution model based on the physical information neural network in S1 is specifically as follows: The physical information neural network embeds a differential equation for controlling the reaction rate into a loss function, so that network output is not only dependent on data fitting, but also is constrained by pyrolysis reaction dynamics, therefore, arrhenius law and an n-level reaction rate equation representing a pyrolysis mechanism are substituted into the physical information neural network as physical constraint terms, and the obtained equation is shown as follows: wherein E a represents activation energy, R (T) represents reaction rate, T represents temperature, R represents gas constant, A represents factor before referring, beta represents reaction index for controlling nonlinear behavior of conversion rate, and alpha (T) represents conversion rate of retired fan blade material at time T; the change rate of the conversion rate of the material in unit time is represented by the quantification of the pyrolysis reaction rate; The total loss function of the pyrolysis product distribution model based on the physical information neural network consists of data loss L data and physical constraint loss L phys ; the data loss L data quantifies the difference between the predicted yield and the experimental yield based on the mean square error, and the calculation formula is as follows: Wherein, the Representing predicted values of the physical information neural network, and y data represents true values of the data; Physical constraint loss L phys is used for measuring pyrolysis reaction rate of retired fan blade predicted by neural network If the network prediction result deviates from the Arrhenius law, the loss is increased, the network can be made to approach a solution meeting the physical rule through training, and the calculation formula is as follows: the total loss function is obtained by weighting and summing the data loss and the physical constraint loss by a certain proportion lambda as follows: L total ＝λL data +(1-λ)L phys And lambda represents an adjustable weight parameter, is used for balancing the fitting precision of data and the conformity of a physical rule, and is determined through cross verification, so that lambda is increased to improve the fitting capacity of the data when experimental data are large in noise, and lambda is reduced to improve the interpretability of the model when samples are insufficient or physical constraints are required to be enhanced.
3. The method for optimizing parameters of a retired fan blade pyrolysis system based on dynamic weighted feature fusion algorithm control according to claim 1, wherein the objective function in S2 comprises training loss and regularization terms, and is expressed by the following formula: Wherein, the Is a regularization loss function, is an optimization objective function of model training, integrates prediction error and regularization constraint, n is the number of samples of the training data set, λ is a coefficient for controlling regularization strength, W j is a weight matrix of the jth sub-network layer, b j is a bias vector of the jth sub-network layer, Is a model predicted value, y i is a data true value; Constraint conditions include process parameter range constraint and system safety constraint; the technical parameter range constraint comprises pyrolysis temperature, heating rate and reaction time, and is preset according to pyrolysis equipment capacity, material thermal stability and experimental conditions; In order to construct a dynamic weighting characteristic fusion model of the retired fan blade pyrolysis system based on an attention mechanism, firstly taking the allowable range of pyrolysis process parameters as constraint conditions, including the upper limit and the lower limit of pyrolysis temperature, heating rate and reaction time, and further setting safety constraints of the safety temperature, the safety pressure, the tail gas emission concentration and the like of the reactor, so as to ensure that model training and optimization processes are only carried out in a feasible and safe process interval; On the basis, three sub-models of a limit gradient lifting algorithm, a multi-layer perceptron and a physical information neural network are adopted to train constrained sample data to obtain prediction results of product distribution respectively, then process parameters, prediction output of the three sub-models and experimental real yield are taken as comprehensive input characteristics to be introduced into a fully-connected attention network for learning, the attention network adaptively determines contribution weights among the three sub-models through analysis of the comprehensive characteristics, so that the weights can be dynamically adjusted according to different working conditions, a fusion prediction value is obtained by weighting the prediction results of the three sub-models according to the attention weights, a training target of the model is composed of a prediction error item and a regularization item, wherein the prediction error item is used for measuring difference between fusion prediction and experimental yield, the regularization item is used for limiting the scale of network parameters and preventing overfitting, in the optimization process of the target function, all training samples are required to meet the constraint of the process parameter range and the system safety constraint, meanwhile, the output of the physical information neural network is required to meet the pyrolysis reaction dynamics, a feasible solution space is limited through constraint conditions, the target function pushes the attention network to converge to optimal weight distribution, and accordingly the three sub-models are enabled to achieve the final dynamic and consistent running of the three sub-models and high-accuracy prediction and high-performance of the prediction and the final running of the model.
4. The method for optimizing parameters of a retired fan blade pyrolysis system based on dynamic weighting feature fusion algorithm control according to claim 1, wherein the process of obtaining training results by three sub-modules of a limit gradient lifting algorithm, a multi-layer sensor and a physical information neural network in the step S3 is as follows: the limit gradient lifting algorithm is based on a gradient enhancement decision tree, and the prediction formula is expressed as follows: Where a is the total number of trees, g k (x i ) represents the prediction of the kth tree, The objective function L (θ) of the limiting gradient lifting algorithm includes training loss and regularization terms, which can be expressed as follows: In the formula, Representing a loss function for measuring the true value y i and model predictive value of the ith sample The difference between the two is that omega (g k ) represents the complexity of a regularization term penalty model, gamma represents a tree complexity coefficient, T represents the number of leaf nodes of a kth decision tree, the more the leaf nodes are, the higher the input complexity is, delta represents the regularization coefficient, θ represents the number of leaves, and w j represents the weight of each leaf node; The multi-layer perceptron is a feedforward neural network capable of modeling nonlinear relationships, and the output of the first layer is expressed by the following formula: h (l) ＝ρ(A (l) h (l-1) +b (l) ) Where h (l) denotes the output vector of the first layer, σ denotes the nonlinear activation function, a (l) denotes the weight matrix, h (l-1) denotes the output vector of the l-1 layer, which is the input of the first layer, b (l) denotes the bias vector, and the final output is expressed as follows: In the formula, Representing the predicted output of the model, L representing the total layer number of the network, A (L) representing the weight matrix of the output layer, h (L-1) representing the output vector of the L-1 layer, which is the input of the L layer, and b (L) representing the deviation vector of the output layer; The method comprises the steps of training a multi-layer perceptron by minimizing mean square error, namely, iteratively optimizing the weight and the deviation of a network, firstly enabling input data to pass through linear transformation and nonlinear activation of a hidden layer by layer from an input layer, finally obtaining a prediction result through linear calculation of an output layer, calculating the mean square error between the prediction value and a true value as loss, then reversely deducing from the output layer by utilizing a chain rule, calculating the gradient of the loss to the weight and the deviation of each layer, determining the direction and the amplitude of parameters to be adjusted, controlling the step length according to the learning rate, updating the parameters of all layers along the gradient descending direction, reducing the loss, and repeating the process until the mean square error is converged to a stable minimum value or reaches the preset training times to complete model training, wherein the loss function is expressed by the following formula: In the formula, The loss function is represented by a function of the loss, Representing the predicted output of the model, y i represents the actual value.
5. The parameter optimization method for the retired fan blade pyrolysis system based on dynamic weighting feature fusion algorithm control according to claim 1, wherein the construction of the retired fan blade pyrolysis system dynamic weighting feature fusion model based on an attention mechanism is specifically as follows: Carrying out normalization processing on the prediction output of the limit gradient lifting algorithm, the multilayer perceptron and the physical information neural network by adopting a minimum-maximum normalization method, mapping an output value into a [0,1] interval, wherein a minimum-maximum normalization formula adopted in the normalization processing process is as follows: Wherein y norm represents a normalization result, y represents an original data value to be normalized, wherein y min represents a minimum value in an original data set, y max represents a maximum value in the original data set, and the normalized prediction result of the sub-model is respectively and independently trained and spliced with the original input data and the experimental real yield to construct a comprehensive training data set, wherein the comprehensive data set is represented as follows: Data fusion ＝[X,y XGB|norm ,y MLP|norm ,y PINN|norm ,y data ] Wherein, X represents an original input parameter, y XGB|norm represents a predicted value normalized by an extreme gradient lifting algorithm, y MLP|norm represents a predicted value normalized by a multi-layer perceptron, y PINN|norm represents a predicted value normalized by a physical information neural network, and y data represents an original output value; The comprehensive data set trains a fully-connected attention network, the attention network outputs three weights omega 1 、ω 2 , omega 3 ,ω 1 、ω 2 and omega 3 to respectively represent contributions of three sub-modules of a limit gradient lifting algorithm, a multi-layer perceptron and a physical information neural network, the attention network evaluates correlation of sub-models according to input characteristics and performance, the original output a j of the attention network is normalized through a softmax function, and a j and omega j can be calculated according to the following formulas: a j ＝W j X+b j Wherein X represents an input parameter, and W j and b j represent a trainable weight matrix and a bias vector; On the basis, updating parameters W j and b j by using an Adam optimizer, adjusting the learning rate through cross-validation, dividing a data set into a plurality of subsets through cross-validation, and evaluating the performance of the model by taking different subsets as validation sets in turn; when the learning rate is adjusted, different learning rate values are tested in a plurality of experiments of cross verification, the convergence speed and the generalization effect of the model on a verification set are observed, and the optimal learning rate of the model equilibrium convergence is ensured by finally selecting the optimization device of Adam to ensure that the parameter oscillation is not converged due to overlarge learning rate and the training is not too slow due to overlarge learning rate when the W j and the b j are updated; Fusion prediction y fusion is expressed as: y fusion ＝ω 1 y XGB|norm +ω 2 y MLP|norm +ω 3 y PINN|norm 。
6. The parameter optimization method for the retired fan blade pyrolysis system based on dynamic weighted feature fusion algorithm control according to claim 1, wherein the step of constructing a process parameter optimization model for the retired fan blade pyrolysis system in S4 specifically comprises the following training steps: s41, optimizing targets and constraint conditions: Obtaining a yield prediction result Y (x) by a dynamic weighted feature fusion model, wherein other targets can be obtained through physical relations, and the optimization variables are expressed as follows: x=[T,r(T),t] wherein T represents temperature, r (T) represents heating rate, and T represents reaction time; The other objectives include product quality, energy consumption, and process safety constraints; The product quality Q (x) is related to yield and temperature conditions and is estimated using empirical functions, and the formula is as follows: Q(x)=θ 1 Y(x)-θ 2 (T-T 0 ) 2 Wherein θ 1 、θ 2 is a fitting coefficient, and T 0 is an optimal temperature interval; the energy consumption E (x) is expressed by the product of the heat input power and time, and the formula is as follows: Wherein c is a system correlation coefficient; The optimization objective function may be expressed as: the constraint condition is that Wherein g j (x) is a safety constraint; s42, optimizing and solving a particle swarm: each particle represents a group of process parameters x= [ T, r (T), T ], each particle gradually approaches the optimal process condition according to the self history optimal solution and the global optimal solution updating position, and the updating formula is that Wherein ω represents inertial weight, c 1 ,c 2 represents learning factor, controlling individual and group influence, r 1 ,r 2 represents random number, increasing randomness, and finally, the particle swarm optimally outputs a group of pareto optimal solution set P, reflecting different tradeoffs among high yield, good quality and low energy consumption, and further can be selected from the solution set according to actual needs.
7. The retired fan blade pyrolysis system based on dynamic weighting characteristic fusion algorithm control is characterized by comprising a pyrolysis reaction furnace (2), a carbon storage (3) and a condensing tower group which are connected with a discharge hole of the pyrolysis reaction furnace (2), a gas collecting bottle (7) connected with a gas outlet at the top of the condensing tower group, a dry type packing tower (8) connected with the side part of the condensing tower group, an adsorption tower (9) connected with an outlet of the dry type packing tower (8) and an exhaust fan (6) connected with an outlet of the adsorption tower (9); And the bottom condensate collecting parts of the condensing tower group and the dry type packing tower (8) are connected with the oil collecting tank (5).
8. The retired fan blade pyrolysis system based on dynamic weighting characteristic fusion algorithm control according to claim 7, wherein the condensing tower group comprises a front condensing tower (41), a middle condensing tower (42) and a rear condensing tower (43), wherein a feed inlet of the front condensing tower (41) is connected and communicated with a feed inlet of a pyrolysis reaction furnace (2), a feed outlet of the front condensing tower (41) is connected and communicated with a feed inlet of the middle condensing tower (42), a feed outlet of the middle condensing tower (42) is connected and communicated with a feed inlet of the rear condensing tower (43) to form a fractional condensing flow path, a gas outlet of the rear condensing tower (43) is connected and communicated with a gas inlet of a gas collecting cylinder (7), and a dry filler tower (8) inlet is connected and communicated with the side part of the rear condensing tower (43); the bottom condensate collecting parts of the front condensing tower (41), the middle condensing tower (42), the rear condensing tower (43) and the dry type packing tower (8) are all connected with the oil collecting tank (5).

Description

Parameter optimization method for retired fan blade pyrolysis system based on dynamic weighting feature fusion algorithm control Technical Field The invention relates to a parameter optimization method for a retired fan blade pyrolysis system based on dynamic weighting feature fusion algorithm control, and belongs to the field of retired fan blade pyrolysis systems served by control algorithms. Background With the continuous expansion of the global wind power installation scale, the problem of processing the retired fan blades is increasingly remarkable, and how to scientifically treat and utilize the retired fan blades becomes a problem to be solved urgently. The fan blade is mainly made of composite materials such as glass fiber and carbon fiber, and the durability and the light weight property of the fan blade are excellent in service, but after the life cycle is finished, the fan blade is difficult to degrade and forms challenges to environment and economy, and the traditional disposal means are mostly incinerated or buried aiming at the retired fan blade. However, because the landfill occupies a large land, harmful substances can be released after incineration, carbon emission is increased, and along with the increase of the installed capacity of wind power, the traditional method is not suitable for processing the retired fan blades with huge quantity. At present, the pyrolysis method is applied to the treatment of retired fan blades, and can efficiently recycle glass fibers and pyrolysis oil gas by decomposing the blade composite material at high temperature, so that the physical properties such as the length, the strength and the like of the glass fibers are reserved to the greatest extent, the thermal decomposition method is far higher than the performance loss and low added value utilization of the fibers caused by severe impact in mechanical crushing recovery, and is more energy-saving and environment-friendly than the traditional method from the angles of energy consumption and resource recovery. At present, the determination of the technological parameters of the retired fan blade pyrolysis system is mainly carried out by an experimental analysis method and a machine learning method, specifically, the thermal gravimetric analysis and gas chromatography-mass spectrometry combined technology is adopted in the experimental analysis method to directly evaluate the yield and the composition of pyrolysis products, so that the mechanism and the kinetic parameters of pyrolysis reaction are determined, and the model training is carried out by using a large-scale data set based on the machine learning method to predict the distribution situation of the pyrolysis products, so that the technological parameter optimization of the pyrolysis system is realized. Compared with an experimental analysis method, the machine learning-based method can build a more accurate product prediction model under different pyrolysis conditions, so that the optimization of the process parameters of the pyrolysis system can be better guided. However, the existing pyrolysis system process parameter optimization method based on machine learning only relies on data fitting to establish the relationship between system input and output, and has limited application in integrating the physical law of the pyrolysis reaction process. The pyrolysis process of the retired fan blade is a complex reaction process with multi-parameter coupling and dynamic nonlinearity, and the influence of dynamic variables such as the concentration change of volatile products generated in real time in the reaction process and the local temperature field imbalance in the reactor on the distribution of final products is considered in consideration of the material difference of retired blades of different batches. The static control mode often leads to unstable product distribution, increases the subsequent treatment difficulty, and reduces the recovery value. Disclosure of Invention The technical problem to be solved by the invention is to provide a method for optimizing the technological parameters of the retired fan blade pyrolysis system based on the control of a dynamic weighting characteristic fusion algorithm, which establishes a retired fan blade pyrolysis product distribution prediction model and further provides a dynamic weighting characteristic fusion algorithm for optimizing the technological parameters of the retired wind power blade pyrolysis system. The invention adopts the following technical scheme: The invention discloses a parameter optimization method for a retired fan blade pyrolysis system based on dynamic weighting characteristic fusion algorithm control, which comprises the following steps: S1, constructing a pyrolysis product distribution model based on a physical information neural network, and providing a physical constraint basic model of a pyrolysis process; S2, determining an objective function and constraint conditions, construc