Retired fan blade pyrolysis system parameter optimization method based on dynamic weighted feature fusion algorithm control

Abstract

The invention relates to a dynamic weighted feature fusion algorithm control-based retired fan blade pyrolysis system parameter optimization method. The method comprises the following steps of: 1, constructing a pyrolysis product distribution model based on a physical information neural network; 2, constructing a dynamic weighted feature fusion model of the decommissioned fan blade pyrolysis system based on the attention mechanism; 3, carrying out weighted summation on normalized prediction results of three sub-modules of a limit gradient lifting algorithm, a multilayer sensor and a physical information neural network, and outputting a fusion prediction result; 4, on the basis of a prediction result of the dynamic weighted feature fusion model, constructing a multi-objective optimization model and solving parameters by utilizing the prediction result of the fusion model; 5, dynamically adjusting and controlling the pyrolysis system; the physical information neural network can improve the prediction precision of the decommissioned fan blade pyrolysis product distribution by revealing the internal mechanism of the influence of the process parameters on the product distribution, the prediction precision of the pyrolysis product is effectively improved, and the process parameters of the pyrolysis system are more accurately adjusted.

Description

Technical Field

[0001] This invention relates to a parameter optimization method for a decommissioned wind turbine blade pyrolysis system based on a dynamic weighted feature fusion algorithm, belonging to the field of control algorithm services for decommissioned wind turbine blade pyrolysis systems. Background Technology

[0002] With the continuous expansion of global wind power installed capacity, the disposal of retired wind turbine blades has become an increasingly prominent issue, and how to scientifically dispose of and utilize them has become an urgent problem to be solved. Wind turbine blades are mainly made of composite materials such as glass fiber and carbon fiber. While their durability and lightweight characteristics are excellent during their service life, they pose environmental and economic challenges due to their difficulty in degradation after their life cycle ends. Traditional disposal methods for retired wind turbine blades mostly involve incineration or landfill. However, landfill requires large areas of land, and incineration may release harmful substances and increase carbon emissions. With the increasing installed capacity of wind power, traditional methods are no longer suitable for handling the massive number of retired wind turbine blades.

[0003] Currently, pyrolysis has been applied to the disposal of retired wind turbine blades. Pyrolysis decomposes blade composite materials at high temperatures, which can efficiently recover glass fibers and pyrolytic oil and gas, and retain the physical properties of glass fibers such as length and strength to the greatest extent. This is far superior to the performance loss and low added value utilization of fibers caused by violent impact during mechanical crushing and recycling. From the perspective of energy consumption and resource recovery, pyrolysis is more energy-efficient and environmentally friendly than traditional methods.

[0004] Currently, the determination of process parameters for decommissioned wind turbine blade pyrolysis systems mainly relies on experimental analysis and machine learning methods. Specifically, experimental analysis uses thermogravimetric analysis (TGA) and gas chromatography-mass spectrometry (GC-MS) to directly assess the yield and composition of pyrolysis products, thereby determining the mechanism and kinetic parameters of the pyrolysis reaction. Machine learning methods utilize large-scale datasets for model training to predict the distribution of pyrolysis products, thus optimizing the process parameters of the pyrolysis system. Compared to experimental analysis methods, machine learning-based methods can establish more accurate product prediction models under different pyrolysis conditions, thus better guiding the optimization of pyrolysis system process parameters. However, existing machine learning-based optimization methods for pyrolysis system process parameters rely solely on data fitting to establish the relationship between system input and output, with limited integration and application of the physical laws governing the pyrolysis reaction process. The pyrolysis process of decommissioned wind turbine blades is a complex reaction process involving multiple coupled parameters and dynamic nonlinearity. Considering the material differences between different batches of decommissioned blades, the real-time changes in the concentration of volatile products generated during the reaction, and the impact of dynamic variables such as local temperature field inhomogeneity within the reactor on the final product distribution, this static control mode often leads to unstable product distribution, increasing the difficulty of subsequent processing and reducing the recycling value. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a method for optimizing process parameters of a decommissioned wind turbine blade pyrolysis system based on a dynamic weighted feature fusion algorithm. A distribution prediction model for pyrolysis products of decommissioned wind turbine blades is established, and a dynamic weighted feature fusion algorithm for optimizing process parameters of a decommissioned wind turbine blade pyrolysis system is further proposed.

[0006] The present invention adopts the following technical solution:

[0007] This invention discloses a parameter optimization method for a decommissioned wind turbine blade pyrolysis system based on a dynamically weighted feature fusion algorithm, comprising the following steps:

[0008] S1. Construct a pyrolysis product distribution model based on a physical information neural network to provide a physical constraint model for the pyrolysis process;

[0009] S2. Determine the objective function and constraints, construct a dynamic weighted feature fusion model for the pyrolysis system of decommissioned wind turbine blades based on the attention mechanism, and build a fusion framework to achieve multi-source model collaboration. Multi-source model collaboration includes three sub-modules: extreme gradient boosting algorithm, multilayer perceptron, and physical information neural network. The algorithms complement each other from the dimensions of statistical fitting, nonlinear modeling, and physical constraints. Then, the sub-module outputs are normalized and the comprehensive dataset is spliced ​​to achieve cross-module data linkage. Finally, the attention mechanism dynamically allocates the weights of each module and performs weighted summation to output a prediction result that has both fitting accuracy and physical consistency.

[0010] S3. The normalized prediction results of the three sub-modules, namely the extreme gradient boosting algorithm, the multilayer perceptron and the physical information neural network, are weighted and summed according to the weights output by the attention mechanism to output the fused prediction result, thus completing the input, training and output of the dynamic weighted feature fusion model.

[0011] S4. Based on the prediction results of the dynamic weighted feature fusion model, with the yield and quality of pyrolysis products and the energy consumption of the pyrolysis process as optimization objectives, the key process parameters of the pyrolysis system are used as the parameters to be optimized. A multi-objective optimization model for the pyrolysis system of decommissioned wind turbine blades is established, and the particle swarm optimization algorithm is used to solve the model to obtain the solution set of process parameters. The multi-objective optimization model is constructed and the parameters are solved using the prediction results of the fusion model.

[0012] S5. Based on the optimized process parameter solution set, the pyrolysis system is dynamically adjusted and controlled by real-time monitoring of key parameters of the pyrolysis process and prediction results based on the dynamic weighted feature fusion model.

[0013] The specific steps in S1 of this invention to construct the pyrolysis product distribution model based on a physical information neural network are as follows:

[0014] The physical information neural network embeds the differential equation controlling the reaction rate into the loss function, making the network output not only dependent on data fitting but also constrained by the pyrolysis reaction kinetics. Therefore, the Arrhenius law representing the pyrolysis mechanism and the nth-order reaction rate equation are substituted into the physical information neural network as physical constraint terms; the resulting equation is shown below:

[0015]

[0016] Among them, E a Let r(T) represent the activation energy, r(T) represent the reaction rate, T represent the temperature, R represent the gas constant, A represent the pre-exponential factor, β represent the reaction exponent that controls the nonlinear behavior of the conversion rate, and α(t) represent the conversion rate of the decommissioned wind turbine blade material at time t. It represents the rate of change of material conversion rate per unit time and is a quantitative expression of the pyrolysis reaction rate;

[0017] The total loss function of the pyrolysis product distribution model based on the physical information neural network is composed of the data loss L. data and physical constraint loss L phys composition;

[0018] Data loss L data The difference between the predicted yield and the experimental yield is calculated based on the mean squared error quantification formula, as shown below:

[0019]

[0020] in, t represents the predicted value of the physical information neural network. data Represents the true value of the data;

[0021] Physical constraint loss L phys Used to measure the pyrolysis reaction rate of decommissioned wind turbine blades predicted by neural networks. Does it satisfy Arrhenius's law? If the network prediction deviates from Arrhenius's law, the loss increases. Through training, the network can be made to approximate a solution that satisfies the physical law. The calculation formula is as follows:

[0022]

[0023] By weighting and summing the data loss and physical constraint loss according to a certain proportion λ, the total loss function is obtained as follows:

[0024] L total =λL data +(1-λ)L phys

[0025] Here, λ represents an adjustable weight parameter used to balance the accuracy of data fitting with the conformity to physical laws. The value of λ is determined through cross-validation. When the experimental data is noisy, λ is increased to improve the data fitting ability. When the sample is insufficient or physical constraints need to be strengthened, λ is decreased to improve the interpretability of the model.

[0026] The objective function in S2 of this invention includes training loss and regularization term, as expressed by the following formula:

[0027]

[0028] in, is the regularization loss function, which is the optimization objective function for model training, integrating prediction error and regularization constraints. n is the number of samples in the training dataset, λ is a coefficient used to control the strength of regularization, and W... j b is the weight matrix of the j-th sub-network layer. j It is the bias vector of the j-th sub-network layer. It is the model prediction value, y i It is the actual value of the data;

[0029] Constraints include process parameter range constraints and system safety constraints;

[0030] The constraints on the range of process parameters include pyrolysis temperature, heating rate, and reaction time, which are preset based on the capacity of the pyrolysis equipment, the thermal stability of the material, and experimental conditions.

[0031] System safety constraints are designed to ensure the safe and stable operation of the reaction system, including the maximum allowable temperature of the reactor, the maximum operating pressure of the system, the upper limit of the tail gas emission concentration, and the safe value of the heating power. To construct a dynamic weighted feature fusion model of the pyrolysis system of decommissioned wind turbine blades based on the attention mechanism, the allowable range of pyrolysis process parameters is first used as a constraint condition, including the upper and lower limits of pyrolysis temperature, heating rate, and reaction time. Furthermore, safety constraints such as safe reactor temperature, safe pressure, and tail gas emission concentration are set to ensure that the model training and optimization process is carried out only within a feasible and safe process range.

[0032] Based on this, three sub-models—the extreme gradient boosting algorithm, the multilayer perceptron, and the physical information neural network—are used to train the constrained sample data, respectively obtaining the predicted results of the product distribution. Subsequently, the process parameters, the predicted outputs of the three sub-models, and the actual experimental yield are used as comprehensive input features and introduced into a fully connected attention network for learning. The attention network analyzes the comprehensive features and adaptively determines the contribution weights among the three sub-models, enabling the weights to be dynamically adjusted according to different operating conditions. The fused predicted value is obtained by weighting the prediction results of the three sub-models according to the attention weights. The training objective of the model consists of a prediction error term and a regularization term, where the prediction error term is used to measure the difference between the fused prediction and the experimental yield, and the regularization term is used to limit the network parameter size and prevent overfitting. During the optimization process of the objective function, all training samples must meet the constraints of the process parameter range and the system safety constraints, and the output of the physical information neural network must also conform to the pyrolysis reaction kinetics. By limiting the feasible solution space through constraints, the objective function drives the attention network to converge to the optimal weight distribution, thereby realizing the dynamic fusion of the advantages of the three sub-models and finally obtaining a pyrolysis product prediction model that combines high accuracy, physical consistency, and operational safety.

[0033] The process by which the three sub-modules of the S3 of this invention—the extreme gradient boosting algorithm, the multilayer perceptron, and the physical information neural network—obtain the training results is as follows:

[0034] The extreme gradient boosting algorithm is based on gradient boosting decision trees, and its prediction formula is expressed as follows:

[0035]

[0036] In the formula, α is the total number of trees, and g k (x i ) represents the prediction for the k-th tree. This is the predicted value from the Limit Gradient Boosting algorithm; the objective function L(θ) of the Limit Gradient Boosting algorithm includes training loss and regularization terms, and can be expressed as follows:

[0037]

[0038]

[0039] In the formula, This represents the loss function, used to measure the true value y of the i-th sample. i and model predictions The difference between them, Ω(g) k ) represents the regularization term penalizing model complexity, γ represents the tree complexity coefficient, T represents the number of leaf nodes in the k-th decision tree (more leaf nodes result in higher output complexity), ε represents the regularization coefficient, θ represents the number of leaves, and w jThis represents the weight of each leaf node;

[0040] A multilayer perceptron is a feedforward neural network that can model nonlinear relationships. The output of its l-th layer is expressed by the following formula:

[0041] h (l) =σ(A (l) h (l-1) +b (l) )

[0042] In the formula, h (l) Let A represent the output vector of the l-th layer, σ represent the non-linear activation function, and A represent the output vector of the l-th layer. (l) Let h represent the weight matrix. (l-1) This represents the output vector of layer l-1, which is the input vector of layer l. (l) The deviation vector is represented by the following formula:

[0043]

[0044] In the formula, This represents the model's predicted output, where L represents the total number of layers in the network, and A represents the predicted output. (L) h represents the weight matrix of the output layer. (L-1) This represents the output vector of layer L-1, and is the input vector of layer L. (L) Represents the bias vector of the output layer.

[0045] The multilayer perceptron is trained by minimizing the mean squared error (MSE). Specifically, the weights and biases of the network are iteratively optimized. First, the input data is passed through the input layer, undergoing linear transformations and non-linear activations in the hidden layers. Finally, the prediction result is obtained through linear calculation in the output layer. The MSE between the predicted and true values ​​is then calculated as the loss. Next, the chain rule is used to derive the loss from the output layer backwards, calculating the gradient of the loss with respect to the weights and biases of each layer, thus determining the direction and magnitude of parameter adjustments. Then, the step size is controlled according to the learning rate, and the parameters of all layers are updated along the gradient descent direction to reduce the loss. This process is repeated until the MSE converges to a stable minimum or the preset number of training iterations is reached, completing the model training. The loss function is expressed as follows:

[0046]

[0047] In the formula, Represents the loss function. y represents the model's predicted output. i This represents the actual value.

[0048] The specific construction of the dynamic weighted feature fusion model for the pyrolysis system of decommissioned wind turbine blades based on the attention mechanism in this invention is as follows:

[0049] The min-max normalization method is used to normalize the prediction outputs of the three sub-modules: the extreme gradient boosting algorithm, the multilayer perceptron, and the physical information neural network, mapping the output values ​​to the [0,1] interval. The min-max normalization formula used in the normalization process is as follows:

[0050]

[0051] In the formula, y norm This represents the normalization result, where y represents the original data value that needs to be standardized, specifically the predicted output value of the submodule. min y represents the minimum value in the original dataset. max The maximum value in the original dataset is represented by the sum of the predictions from the normallyd sub-models trained separately, the original input data, and the actual experimental yield, which are then concatenated to construct a comprehensive training dataset. The comprehensive dataset is represented as follows:

[0052] Data fusion =[X,y XGB|norm ,y MLP|norm ,y PINN|norm ,y data ]

[0053] In the formula, X represents the original input parameter, y XGB|norm y represents the predicted value after normalization by the extreme gradient boosting algorithm. MLP|norm y represents the normalized prediction value of the multilayer perceptron. PINN|norm y represents the normalized predicted value of physical information from the neural network. data Indicates the original output value;

[0054] The comprehensive dataset is used to train a fully connected attention network. The attention network outputs three weights ω1, ω2, and ω3, where ω1, ω2, and ω3 represent the contributions of the extreme gradient boosting algorithm, the multilayer perceptron, and the physical information neural network, respectively. The attention network evaluates the correlation between the sub-models based on input features and performance. Its original output a j Normalization is performed using the softmax function, a j and ω j The following formula can be used to calculate:

[0055]

[0056] In the formula, X represents the input parameter, W j and b j This represents the trainable weight matrix and bias vector;

[0057] Based on this, the Adam optimizer is used to update the parameter W. j and b jThe learning rate was adjusted through cross-validation. Cross-validation divides the dataset into multiple subsets, and different subsets are used as validation sets in turn to evaluate the model's performance. When adjusting the learning rate, different learning rate values ​​were tested in multiple rounds of cross-validation experiments to observe the model's convergence speed and generalization effect on the validation set. Finally, the optimal learning rate was selected that allows the Adam optimizer to update W. j and b j At this time, the optimal learning rate is found to ensure balanced convergence of the model, preventing parameter oscillations and non-convergence due to an excessively large learning rate, and avoiding slow training due to an excessively small learning rate.

[0058] Fusion prediction y fusion Represented as:

[0059] y fusion =ω1y XGB|norm +ω2y MLP | norm +ω3y PINN | norm .

[0060] The process parameter optimization model for the decommissioned wind turbine blade pyrolysis system constructed in S4 of this invention specifically includes the following training steps:

[0061] S41, Optimization objective and constraints:

[0062] The yield prediction result Y(x) is obtained from the dynamic weighted feature fusion model. Other objectives can be obtained through physical relationships. Let the optimization variables be expressed as:

[0063] x = [T, r(T), t]

[0064] In the formula, T represents temperature, r(T) represents heating rate, and t represents reaction time;

[0065] The other objectives include constraints on product quality, energy consumption, and process safety.

[0066] The product quality Q(x) is related to the yield and temperature conditions, and can be estimated using an empirical function, as shown in the following formula:

[0067] Q(x) = θ1Y(x) - θ2(T - T0) 2

[0068] In the formula, θ1 and θ2 are fitting coefficients, and T0 is the optimal temperature range;

[0069] Energy consumption E(x) is expressed as the product of heat input power and time, and its formula is as follows:

[0070]

[0071] Where c is the system correlation coefficient;

[0072] The objective function can then be expressed as:

[0073]

[0074] The constraints (including process parameter range constraints and system safety constraints) are as follows:

[0075]

[0076] In the formula, g j (x) represents a safety constraint;

[0077] S42, Particle Swarm Optimization Solution:

[0078] Each particle represents a set of process parameters x = [T, r(T), t]. Each particle updates its position based on its own historical best solution and the global best solution, gradually approaching the optimal process conditions. The update formula is as follows:

[0079]

[0080] Where ω represents the inertia weight, which controls the balance of particle exploration or development; c1 and c2 represent learning factors, which control the influence of individuals and groups; and r1 and r2 represent random numbers, which increase randomness. Finally, the particle swarm optimization outputs a Pareto optimal solution set P, which reflects different trade-offs between high yield, good quality, and low energy consumption, and can then be selected from the solution set according to actual needs.

[0081] The present invention relates to a pyrolysis system for decommissioned wind turbine blades controlled by a dynamic weighted feature fusion algorithm. The system includes a pyrolysis reactor, a carbon storage tank and a condenser tower group connected to the outlet of the pyrolysis reactor, a gas collecting bottle connected to the gas outlet at the top of the condenser tower group, a dry packed tower connected to the side of the condenser tower group, an adsorption tower connected to the outlet of the dry packed tower, and an exhaust fan connected to the outlet of the adsorption tower. The bottom condensate collection sections of the condenser tower group and the dry packed tower are both connected to an oil collecting tank.

[0082] The condenser tower group of the present invention includes a pre-condenser tower, a middle condenser tower, and a post-condenser tower; the feed inlet of the pre-condenser tower is connected to and communicates with the discharge outlet of the pyrolysis reactor, the discharge outlet of the pre-condenser tower is connected to and communicates with the feed inlet of the middle condenser tower, and the discharge outlet of the middle condenser tower is connected to and communicates with the feed inlet of the post-condenser tower, forming a staged condensation flow path. The gas outlets of the pre-condenser tower, the middle condenser tower, and the post-condenser tower are connected to and communicates with the gas inlet of the gas collecting bottle through a connecting pipeline. The inlet of the dry packed tower is connected to and communicates with the side of the post-condenser tower. The bottom condensate collection section of the pre-condenser tower, the middle condenser tower, the post-condenser tower, and the dry packed tower are all connected to an oil collecting tank.

[0083] The positive effects of this invention are as follows: The physical information neural network of this invention can improve the prediction accuracy of the pyrolysis product distribution of decommissioned wind turbine blades by revealing the intrinsic mechanism of the influence of process parameters on product distribution; the dynamic weighted feature fusion model based on the attention mechanism can make full use of the advantages of the three sub-modules of the extreme gradient boosting algorithm, multilayer perceptron, and physical information neural network, dynamically and adaptively adapting the input data features to provide accurate and physically consistent prediction of the pyrolysis product distribution of decommissioned wind turbine blades and process parameter optimization schemes; the multi-objective optimization takes yield, quality, and energy consumption as the core objectives, quantifies the priority of objectives by using empirical functions and physical relationships, limits the feasible solution space by combining constraints, and solves the optimal solution set by the particle swarm optimization algorithm to meet the personalized needs of high yield, high quality, and low energy consumption in different scenarios; effectively improving the prediction accuracy of pyrolysis products and more accurately adjusting the process parameters of the pyrolysis system. Attached Figure Description

[0084] Figure 1 This is a process flow diagram of the rapid pyrolysis of decommissioned wind turbine blades according to the present invention;

[0085] Figure 2 This is a diagram of the dynamic weighted feature fusion framework based on the attention mechanism of this invention;

[0086] Figure 3 This is a scatter plot of tar yield prediction from an embodiment of the present invention;

[0087] Figure 4 This is a scatter plot of carbon yield prediction from an embodiment of the present invention;

[0088] Figure 5 This is a scatter plot of the syngas yield prediction in an embodiment of the present invention;

[0089] Figure 6 This is a bar chart of the Digital Micrograph (DM) results of this invention. Detailed Implementation

[0090] To make the technical means, objectives and effects of this invention easy to understand, the invention will be specifically described below in conjunction with embodiments and accompanying drawings.

[0091] As attached Figure 1As shown, the decommissioned wind turbine blade pyrolysis system based on dynamic weighted feature fusion algorithm control of the present invention includes a pyrolysis reactor 2, a carbon storage tank 3 and a condenser tower group connected to the outlet of the pyrolysis reactor 2, a gas collecting bottle 7 connected to the gas outlet at the top of the condenser tower group, a dry packed tower 8 connected to the side of the condenser tower group, an adsorption tower 9 connected to the outlet of the dry packed tower 8, and an exhaust fan 6 connected to the outlet of the adsorption tower 9; the bottom condensate collection parts of the condenser tower group and the dry packed tower 8 are both connected to an oil collecting tank 5. This embodiment of the condenser tower group includes a pre-condenser tower 41, a middle condenser tower 42, and a post-condenser tower 43. The inlet of the pre-condenser tower 41 is connected to and communicates with the outlet of the pyrolysis reactor 2. The outlet of the pre-condenser tower 41 is connected to and communicates with the inlet of the middle condenser tower 42. The outlet of the middle condenser tower 42 is connected to and communicates with the inlet of the post-condenser tower 43, forming a staged condensation flow path. The gas outlet of the post-condenser tower 43 is connected to and communicates with the inlet of the gas collecting bottle 7. The inlet of the dry packed tower 8 is connected to and communicates with the side of the post-condenser tower 43. The bottom condensate collection parts of the pre-condenser tower 41, the middle condenser tower 42, the post-condenser tower 43, and the dry packed tower 8 are all connected to the oil collecting tank 5.

[0092] When the system of this invention is running, the parameters of the pyrolysis reactor 2 are first set, and the pre-treated decommissioned fan blade material is sent into the reactor. High-purity nitrogen is introduced to create an oxygen-free environment. The material undergoes a pyrolysis reaction and decomposes into solid products, gaseous products, and liquid products. The solid products are directly collected in the carbon storage tank 3 through a branch pipe at the outlet of the pyrolysis reactor 2. The mixed gaseous and liquid products are then sent to three sets of staged condensers 4 connected in series for precise separation. The pre-condenser 41 is set at a higher temperature and mainly condenses and separates heavy tar. The intermediate condenser 41... The condenser 42 lowers the temperature to separate light tar, and the post-condenser 43 further cools down to capture pyrolysis water and trace amounts of residual light tar. The different liquid phase products collected by the three condenser towers are individually controlled by valves on their respective pipelines and finally collected in the oil collection tank 5. The non-condensable gas that has not condensed after three-stage condensation enters the gas collection bottle 7 for storage through the top collection pipeline of each condenser tower. The exhaust gas discharged from the post-condenser 43 needs to first pass through the dry packed tower 8 for preliminary adsorption and impurity removal, and then enter the adsorption tower 9 for deep purification. Finally, the qualified exhaust gas is discharged under the negative pressure drive of the exhaust fan 6.

[0093] This invention relates to a parameter optimization method for a decommissioned wind turbine blade pyrolysis system based on a dynamically weighted feature fusion algorithm, comprising the following steps:

[0094] S1. Construct a pyrolysis product distribution model based on a physical information neural network; achieve interpretable modeling of the pyrolysis process of decommissioned wind turbine blades considering multi-parameter coupling; specifically:

[0095] Physical Information Neural Networks (PINNs) embed the differential equations controlling the reaction rate into the loss function, ensuring that the network output depends not only on data fitting but also on the constraints of pyrolysis reaction kinetics. Therefore, by substituting the Arrhenius law representing the pyrolysis mechanism and the nth-order reaction rate equation as physical constraints into the PINN, we ensure that the network's predictions physically satisfy the laws of pyrolysis reaction kinetics. The resulting equations are shown below:

[0096]

[0097] Among them, E a Let r(T) represent the activation energy, r(T) represent the reaction rate, T represent the temperature, R represent the gas constant, A represent the pre-exponential factor, β represent the reaction exponent that controls the nonlinear behavior of the conversion rate, and α(t) represent the conversion rate of the decommissioned wind turbine blade material at time t. It represents the rate of change of material conversion rate per unit time and is a quantitative representation of the pyrolysis reaction rate.

[0098] The total loss function of the pyrolysis product distribution model based on physical information neural network consists of data loss and physical constraint loss.

[0099] Data loss L data The difference between the predicted yield and the experimental yield is calculated based on the mean squared error quantification formula, as shown below:

[0100]

[0101] in, y represents the predicted value of the physical information neural network. data Represents the true value of the data;

[0102] Physical constraint loss L phys Used to measure the pyrolysis reaction rate of decommissioned wind turbine blades predicted by neural networks. Does the Arrhenius law apply? If the network prediction deviates from the Arrhenius law, the loss increases. Through training, the network can be made to approximate a solution that satisfies the physical laws. Therefore, to ensure that the conversion rate prediction satisfies the dynamic equation, the calculation formula is as follows:

[0103]

[0104] By weighting and summing the data loss and physical constraint loss according to a certain ratio, the total loss function is obtained as follows:

[0105] L total =λL data +(1-λ)L phys

[0106] Here, λ represents an adjustable weight parameter used to balance the accuracy of data fitting with the conformity to physical laws. The value of λ is determined through cross-validation. When the experimental data is noisy, λ is increased to improve the data fitting ability. When the sample is insufficient or physical constraints need to be strengthened, λ is decreased to improve the interpretability of the model.

[0107] S2. Determine the objective function and constraints, and construct a dynamic weighted feature fusion model for the pyrolysis system of decommissioned wind turbine blades based on the attention mechanism; construct a fusion framework to achieve multi-source model collaboration; multi-source model collaboration includes three sub-modules: extreme gradient boosting algorithm, multilayer perceptron, and physical information neural network, forming algorithmic complementarity from the dimensions of statistical fitting, nonlinear modeling, and physical constraints, and then normalizing the output of the sub-modules, splicing the comprehensive dataset to achieve cross-module data linkage, and finally dynamically allocating the weights of each module by the attention mechanism and weighted summing to output a prediction result that has both fitting accuracy and physical consistency;

[0108] The objective function includes training loss and regularization term, as shown in the following formula:

[0109]

[0110] in, is the regularization loss function, which is the optimization objective function for model training, integrating prediction error and regularization constraints. n is the number of samples in the training dataset, λ is a coefficient used to control the strength of regularization, and W... j b is the weight matrix of the j-th sub-network layer. j It is the bias vector of the j-th sub-network layer. It is the model prediction value, y i It is the actual value of the data;

[0111] The constraints include process parameter range constraints and system safety constraints. Process parameter range constraints include pyrolysis temperature, heating rate, and reaction time, which are pre-set based on the pyrolysis equipment capacity, material thermal stability, and experimental conditions. System safety constraints are designed to ensure the safe and stable operation of the reaction system, including the reactor's maximum allowable temperature, the system's maximum operating pressure, the upper limit of tail gas emission concentration, and the safe value of heating power. To construct a dynamic weighted feature fusion model for the pyrolysis system of decommissioned wind turbine blades based on an attention mechanism, the allowable range of pyrolysis process parameters is first used as constraints, including upper and lower limits for pyrolysis temperature, heating rate, and reaction time. Further safety constraints such as reactor safety temperature, safety pressure, and tail gas emission concentration are set to ensure that model training and optimization are conducted only within a feasible and safe process range. Based on this, three sub-models—the extreme gradient boosting algorithm, the multilayer perceptron, and the physical information neural network—are used to train the constrained sample data, obtaining prediction results for product distribution. Subsequently, the process parameters, the predicted outputs of the three sub-models, and the actual experimental yield are used as comprehensive input features and introduced into a fully connected attention network for learning. The attention network adaptively determines the contribution weights among the three sub-models by analyzing comprehensive features, allowing the weights to be dynamically adjusted according to different operating conditions. The fused prediction value is obtained by weighting the prediction results of the three sub-models according to the attention weights. The training objective of the model consists of a prediction error term and a regularization term. The prediction error term measures the difference between the fused prediction and the experimental yield, while the regularization term limits the network parameter size and prevents overfitting. During the optimization of the objective function, all training samples must satisfy the aforementioned process constraints and safety constraints, and the output of the physical information neural network must also conform to the pyrolysis reaction kinetics. By limiting the feasible solution space through constraints, the objective function drives the attention network to converge to the optimal weight distribution, thereby achieving the dynamic fusion of the advantages of the three sub-models and ultimately obtaining a pyrolysis product prediction model that combines high accuracy, physical consistency, and operational safety.

[0112] The construction of the dynamic weighted feature fusion model for the pyrolysis system of decommissioned wind turbine blades based on the attention mechanism is as follows:

[0113] The min-max normalization method is used to normalize the prediction outputs of the three sub-modules: the extreme gradient boosting algorithm, the multilayer perceptron, and the physical information neural network, mapping the output values ​​to the [0,1] interval. The min-max normalization formula used in the normalization process is as follows:

[0114]

[0115] In the formula, y norm This represents the normalization result, where y represents the original data value that needs to be standardized, specifically the predicted output value of the submodule. miny represents the minimum value in the original dataset. max The maximum value in the original dataset is represented by the sum of the predictions from the normallyd sub-models trained separately, the original input data, and the actual experimental yield, which are then concatenated to construct a comprehensive training dataset. The comprehensive dataset is represented as follows:

[0116] Data fusion =[X,y XGB|norm ,y MLP|norm ,y PINN|norm ,y data ]

[0117] In the formula, X represents the original input parameter, y XGB|norm y represents the predicted value after normalization by the extreme gradient boosting algorithm. MLP|norm y represents the normalized prediction value of the multilayer perceptron. PINN|norm y represents the normalized predicted value of physical information from the neural network. data The original output value is represented by the comprehensive dataset used to train a fully connected attention network. The attention network outputs three weights ω1, ω2, and ω3, where ω1, ω2, and ω3 represent the contributions of the extreme gradient boosting algorithm, the multilayer perceptron, and the physical information neural network, respectively. The attention network evaluates the correlation between the sub-models based on input features and performance. Its original output a j Normalization is performed using the softmax function, a j and ω j The following formula can be used to calculate:

[0118]

[0119] In the formula, X represents the input parameter, W j and b j Indicates trainable parameters;

[0120] The attention mechanism network is trained by minimizing the mean squared error loss function to quantify the difference between fused predictions and experimental results, and to enhance generalization ability and prevent overfitting. The L2 regularization term... The calculation formula is

[0121]

[0122] Where λ is a coefficient used to control the strength of regularization, W j b is the weight matrix of the j-th sub-network layer. j It is the bias vector of the j-th sub-network layer. It is the model prediction value, y i It is the actual value of the data;

[0123] Based on this, the Adam optimizer is used to update the parameter W.j and b j The learning rate is adjusted through cross-validation. Cross-validation divides the dataset into multiple subsets, and the model performance is evaluated by using different subsets as validation sets in turn. When adjusting the learning rate, y fusion =ω1y XGB|norm +ω2y MLP|norm +ω3y PINN|norm ;

[0124] S3. The normalized prediction results of the three sub-modules, namely the extreme gradient boosting algorithm, the multilayer perceptron and the physical information neural network, are weighted and summed according to the weights output by the attention mechanism to output the fused prediction result, thus completing the input, training and output of the dynamic weighted feature fusion model.

[0125] The training of the three sub-modules—the extreme gradient boosting algorithm, the multilayer perceptron, and the physical information neural network—is shown below:

[0126] The Extreme Gradient Boosting (XGB) algorithm is based on gradient boosting decision trees, and its prediction formula is expressed as follows:

[0127]

[0128] In the formula, α is the total number of trees, and g k (x i ) represents the prediction for the k-th tree. This is the predicted value from the Limit Gradient Boosting algorithm; the objective function L(θ) of the Limit Gradient Boosting algorithm includes training loss and regularization terms, and can be expressed as follows:

[0129]

[0130] In the formula, This represents the loss function, used to measure the true value y of the i-th sample. i and model predictions The difference between them, Ω(g) k ) represents the regularization term penalizing model complexity, γ represents the tree complexity coefficient, T represents the number of leaf nodes in the k-th decision tree (more leaf nodes result in higher output complexity), ε represents the L2 regularization coefficient, θ represents the number of leaves, and w j This represents the weight of each leaf node;

[0131] A multilayer perceptron (MLP) is a feedforward neural network that can model nonlinear relationships. The output of its l-th layer is expressed as follows:

[0132] h (l) =σ(A (l) h (l-1) +b (l) )

[0133] In the formula, h (l) Let A represent the output vector of the l-th layer, σ represent the non-linear activation function, and A represent the output vector of the l-th layer. (l) Let h represent the weight matrix. (l-1) This represents the output vector of layer l-1, which is the input vector of layer l. (l) The deviation vector is represented by the following formula:

[0134]

[0135] In the formula, This represents the model's predicted output, where L represents the total number of layers in the network, and A represents the predicted output. (L) h represents the weight matrix of the output layer. (L-1) This represents the output vector of layer L-1, and is the input vector of layer L. (L) Represents the bias vector of the output layer;

[0136] Training a multilayer perceptron by minimizing the mean squared error (MSE) involves iteratively optimizing the network's weights and biases. First, the input data is processed layer by layer through linear transformations and non-linear activations in the hidden layers. Finally, the output layer performs linear calculations to obtain the prediction result. The MSE between the predicted and true values ​​is then calculated as the loss. Next, the chain rule is used to derive the loss from the output layer backwards, calculating the gradient of the loss with respect to the weights and biases of each layer, thus determining the direction and magnitude of parameter adjustments. Then, the learning rate controls the step size, updating the parameters of all layers along the gradient descent direction to reduce the loss. This process is repeated until the MSE converges to a stable minimum or the preset number of training iterations is reached, completing model training. The loss function (…) The following formula represents:

[0137]

[0138] In the formula, Represents the loss function. y represents the model's predicted output. i This represents the actual value.

[0139] S4. Based on the prediction results of the dynamic weighted feature fusion model, with the yield and quality of pyrolysis products and the energy consumption of the pyrolysis process as optimization objectives, and the key process parameters of the pyrolysis system such as reaction temperature, heating time, and material size as the parameters to be optimized, a multi-objective optimization model for the pyrolysis system of decommissioned wind turbine blades is established. The particle swarm optimization algorithm is used to solve the model to obtain the solution set of process parameters. The multi-objective optimization model is constructed using the prediction results of the fusion model and the parameters are solved.

[0140] The specific training steps for constructing the process parameter optimization model for the pyrolysis system of decommissioned wind turbine blades are as follows:

[0141] S41, Optimization objective and constraints:

[0142] The yield prediction result Y(x) is obtained from the dynamic weighted feature fusion model. Other objectives can be obtained through physical relationships. Let the optimization variables be expressed as:

[0143] x = [T, r(T), t]

[0144] In the formula, T represents temperature, r(T) represents heating rate, and t represents reaction time;

[0145] The other objectives refer to multiple optimization objectives that need to be considered simultaneously in the process parameter optimization model, in addition to the yield prediction results, including product quality, energy consumption, and process safety constraints.

[0146] The product quality Q(x) is related to the yield and temperature conditions, and can be estimated using an empirical function, as shown in the following formula:

[0147] Q(x) = θ1Y(x) - θ2(T - T0) 2

[0148] In the formula, θ1 and θ2 are fitting coefficients, and T0 is the optimal temperature range;

[0149] Energy consumption E(x) is expressed as the product of heat input power and time, and its formula is as follows:

[0150]

[0151] Where c is the system correlation coefficient;

[0152] The objective function can then be expressed as:

[0153]

[0154] The constraints (including process parameter range constraints and system safety constraints) are as follows:

[0155]

[0156] In the formula, g j (x) represents a safety constraint;

[0157] S42, Particle Swarm Optimization Solution:

[0158] Each particle represents a set of process parameters x = [T, r(T), t]. Each particle updates its position based on its own historical best solution and the global best solution, gradually approaching the optimal process conditions. The update formula is as follows:

[0159]

[0160] Where ω represents the inertia weight, which controls the balance of particle exploration or development; c1 and c2 represent learning factors, which control the influence of individuals and groups; and r1 and r2 represent random numbers, which increase randomness. Finally, the particle swarm optimization outputs a Pareto optimal solution set P, which reflects different trade-offs between high yield, good quality, and low energy consumption, and can then be selected from the solution set according to actual needs.

[0161] S5, based on the optimized process parameter solution set, monitors key process parameters such as temperature, pressure, heating rate, and carrier gas flow rate in real time during the operation of the pyrolysis system, and compares them with the prediction results of the dynamic weighted feature fusion model. When the monitored values ​​deviate from the optimal process range, the control system automatically executes corresponding adjustment operations based on the prediction deviation, including: adjusting the heating power to correct temperature changes, adjusting the heating rate to control the reaction rate, adjusting the carrier gas flow rate to change the volatile matter transport rate, and adjusting the reaction time if necessary to ensure complete material conversion. The above adjustment actions are completed by controlling the pyrolysis furnace temperature control module, heating rate adjustment module, and gas flow control valve, thereby keeping the pyrolysis process within the optimal process parameter range and achieving stable yield and high-quality recovery of pyrolysis products.

[0162] Example

[0163] The decommissioned wind turbine blades in this embodiment are mainly composed of 60-70% glass fiber and 30-40% epoxy resin matrix. The pyrolysis experiment uses high-purity nitrogen as an inert carrier gas to purge the reaction system, ensuring the pyrolysis process is carried out under oxygen-deficient conditions. Key input parameters include temperature (range 300℃ to 600℃) and heating rate (set to 10℃ / min, 20℃ / min, and 40℃ / min), used to study the influence of these parameters on the thermodynamic characteristics of the pyrolysis process. The conversion rate is obtained by thermogravimetric analysis, specifically expressed as:

[0164]

[0165] In the formula, m0 represents the initial sample block, m t m represents the residual block at time t. ash The final ash block represents the pyrolysis process. The main output parameters include the yields of the solid, liquid, and gas phases. The solid phase yield includes the composition of pyrolytic carbon and glass fiber residue, the liquid phase yield includes the condensate of water and organic liquids, and the gas phase yield includes non-condensable gases such as hydrogen, methane, and carbon monoxide. These input and output parameters are used to evaluate the pyrolysis reaction characteristics under different experimental conditions.

[0166] The performance metrics used in this embodiment include: root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and coefficient of determination (R²). 2Furthermore, the Digital Micrograph (DM) test, based on the asymptotic standard normal distribution statistic, can be used to verify whether different candidate models have the same predictive ability. The DM test is used to evaluate the superiority of the prediction results of the method of this invention (dynamic weighted feature fusion model, abbreviated as DWFFM) compared with other prediction methods, such as support vector machine (SVM), random forest (RF), and artificial neural network (ANN).

[0167] Figure 3 This is a scatter plot showing the predicted carbon yield of each method in the embodiments of the present invention. As can be seen from the figure...

[0168] Table 1 Comparison of tar yield performance indicators of different methods

[0169]

[0170] Figure 4 This is a scatter plot showing the predicted tar yield of each method in the embodiments of the present invention.

[0171] Table 2 Comparison of carbon yield performance indicators of each method

[0172]

[0173] Figure 5 This is a scatter plot showing the predicted syngas yield for each method in the embodiments of the present invention.

[0174] Table 3 Comparison of Syngas Yield Performance Indicators of Different Methods

[0175]

[0176] Figure 6 This is a bar chart showing the DM test results of each method in the embodiments of the present invention.

[0177] Table 4 Comparison of DM test results by different methods

[0178]

[0179] In this embodiment, by Figures 4 to 6It can be seen that the method of the present invention (DWFFM) more accurately characterizes the kinetic characteristics of the rapid pyrolysis process of decommissioned wind turbine blades. Furthermore, the statistical evaluation indicators in Tables 1, 2, and 3 further confirm the superior performance of DWFFM in capturing the nonlinear dynamics of decommissioned wind turbine blade pyrolysis. Specifically, the SVM, RF, and ANN models all exhibited large prediction errors in liquid phase product prediction. The RMSE and MAE of the SVM model were 1.675 and 1.256, respectively; the RF model showed lower accuracy in predicting carbon yield, with RMSE and MAE reaching 2.094 and 1.105, respectively, and the corresponding R... 2 The values ​​are also relatively low. In contrast, the DWFFM model achieved the lowest RMSE and MAE across all product predictions, demonstrating its ability to significantly reduce prediction errors. For liquid-phase product prediction, DWFFM's RMSE was 0.021, a 98.75% reduction compared to SVM, while MAE and MAPE were reduced by 96.84% and 95.92%, respectively; compared to ANN, the RMSE improvement was 5.64%. In carbon yield prediction, DWFFM's RMSE decreased from 0.352 in RF to 0.042, a reduction of 85.37% in error. In syngas prediction, compared to SVM, the RMSE decreased by 95.57% to 0.040, and the corresponding MAE significantly decreased from 0.775 to 0.040. The R-value is as high as 0.993. 2 This further demonstrates the superior performance of DWFFM in terms of fitting accuracy, enabling it to more closely approximate actual experimental data. In summary, compared to existing methods, the dynamic weighted feature fusion algorithm can more accurately predict the product distribution of rapid pyrolysis of decommissioned wind turbine blades, and it is superior in simulating the nonlinear dynamic characteristics of the rapid pyrolysis process of decommissioned wind turbine blades. In the DM test, the method of this invention also exhibits the lowest statistical value of 2.734, further reflecting the smallest difference in loss from the actual value, demonstrating more stable and consistent predictive performance.

[0180] Under constraints such as temperature, heating rate, and reaction time, yield weights are allocated preferentially to high-value-added products. Particle swarm optimization with 30 particles and 50 iterations is used, and five Pareto optimal solutions are output after convergence. Engineers can choose the solution according to downstream needs. The prediction results of the DWFFM model are used to deal with parameter deviations, so that fluctuations in parameters such as temperature and pressure are controlled within the warning threshold, and product indicators are stable. This is conducive to improving the efficiency of pyrolysis equipment and processes and reducing experimental research costs.

[0181] This invention's physical information neural network can improve the prediction accuracy of pyrolysis product distribution in decommissioned wind turbine blades by revealing the intrinsic mechanism by which process parameters affect product distribution. The dynamic weighted feature fusion model based on an attention mechanism fully leverages the advantages of the three sub-modules—the extreme gradient boosting algorithm, the multilayer perceptron, and the physical information neural network—to dynamically and adaptively adapt input data features, providing accurate and physically consistent predictions of pyrolysis product distribution in decommissioned wind turbine blades and optimized process parameters. Multi-objective optimization focuses on yield, quality, and energy consumption as core objectives, quantifying objective priorities using empirical functions and physical relationships, defining the feasible solution space based on constraints, and solving for the optimal solution set using a particle swarm optimization algorithm. This meets the personalized needs of high yield, high quality, and low energy consumption in different scenarios, effectively improving the prediction accuracy of pyrolysis products and enabling more precise adjustment of pyrolysis system process parameters.

Claims

1. A method for optimizing parameters of a decommissioned wind turbine blade pyrolysis system based on a dynamic weighted feature fusion algorithm, characterized in that, The following steps are used: S1. Construct a pyrolysis product distribution model based on a physical information neural network to provide a physical constraint model for the pyrolysis process; S2. Determine the objective function and constraints, construct a dynamic weighted feature fusion model for the pyrolysis system of decommissioned wind turbine blades based on the attention mechanism, and build a fusion framework to achieve multi-source model collaboration. Multi-source model collaboration includes three sub-modules: extreme gradient boosting algorithm, multilayer perceptron, and physical information neural network. The algorithms complement each other from the dimensions of statistical fitting, nonlinear modeling, and physical constraints. Then, the sub-module outputs are normalized and the comprehensive dataset is spliced ​​to achieve cross-module data linkage. Finally, the attention mechanism dynamically allocates the weights of each module and performs weighted summation to output a prediction result that has both fitting accuracy and physical consistency. S3. The normalized prediction results of the three sub-modules, namely the extreme gradient boosting algorithm, the multilayer perceptron and the physical information neural network, are weighted and summed according to the weights output by the attention mechanism to output the fused prediction result, thus completing the input, training and output of the dynamic weighted feature fusion model. S4. Based on the prediction results of the dynamic weighted feature fusion model, with the yield and quality of pyrolysis products and the energy consumption of the pyrolysis process as optimization objectives, the key process parameters of the pyrolysis system are used as the parameters to be optimized. A multi-objective optimization model for the pyrolysis system of decommissioned wind turbine blades is established, and the particle swarm optimization algorithm is used to solve the model to obtain the solution set of process parameters. The multi-objective optimization model is constructed and the parameters are solved using the prediction results of the fusion model. S5. Based on the optimized process parameter solution set, the pyrolysis system is dynamically adjusted and controlled by real-time monitoring of key parameters of the pyrolysis process and prediction results based on the dynamic weighted feature fusion model.

2. The method for optimizing parameters of a decommissioned wind turbine blade pyrolysis system based on a dynamic weighted feature fusion algorithm as described in claim 1, characterized in that, The specific steps for constructing the pyrolysis product distribution model based on a physical information neural network in S1 are as follows: The physical information neural network embeds the differential equation controlling the reaction rate into the loss function, making the network output not only dependent on data fitting but also constrained by the pyrolysis reaction kinetics. Therefore, the Arrhenius law representing the pyrolysis mechanism and the nth-order reaction rate equation are substituted into the physical information neural network as physical constraint terms; the resulting equation is shown below: Among them, E a Let r(T) represent the activation energy, r(T) represent the reaction rate, T represent the temperature, R represent the gas constant, A represent the pre-exponential factor, β represent the reaction exponent that controls the nonlinear behavior of the conversion rate, and α(t) represent the conversion rate of the decommissioned wind turbine blade material at time t. It represents the rate of change of material conversion rate per unit time and is a quantitative representation of the pyrolysis reaction rate; The total loss function of the pyrolysis product distribution model based on the physical information neural network is composed of the data loss L. data and physical constraint loss L phys composition; Data loss L data The difference between the predicted yield and the experimental yield is calculated based on the mean squared error quantification formula, as shown below: in, y represents the predicted value of the physical information neural network. data Represents the true value of the data; Physical constraint loss L phys Used to measure the pyrolysis reaction rate of decommissioned wind turbine blades predicted by neural networks. Does it satisfy Arrhenius's law? If the network prediction deviates from Arrhenius's law, the loss increases. Through training, the network can be made to approximate a solution that satisfies the physical law. The calculation formula is as follows: By weighting and summing the data loss and physical constraint loss according to a certain proportion λ, the total loss function is obtained as follows: THE total =λL data +(1-λ)L phys Here, λ represents an adjustable weight parameter used to balance the accuracy of data fitting with the conformity to physical laws. The value of λ is determined through cross-validation. When the experimental data is noisy, λ is increased to improve the data fitting ability. When the sample is insufficient or physical constraints need to be strengthened, λ is decreased to improve the interpretability of the model.

3. The method for optimizing parameters of a decommissioned wind turbine blade pyrolysis system based on a dynamic weighted feature fusion algorithm as described in claim 1, characterized in that, The objective function in S2 includes the training loss and a regularization term, as shown in the following formula: in, is the regularization loss function, which is the optimization objective function for model training, integrating prediction error and regularization constraints. n is the number of samples in the training dataset, λ is a coefficient used to control the strength of regularization, and W... j b is the weight matrix of the j-th sub-network layer. j It is the bias vector of the j-th sub-network layer. It is the model prediction value, y i It is the actual value of the data; Constraints include process parameter range constraints and system safety constraints; The constraints on the range of process parameters include pyrolysis temperature, heating rate, and reaction time, which are preset based on the capacity of the pyrolysis equipment, the thermal stability of the material, and experimental conditions. System safety constraints are designed to ensure the safe and stable operation of the reaction system, including the maximum allowable temperature of the reactor, the maximum operating pressure of the system, the upper limit of the tail gas emission concentration, and the safe value of the heating power. To construct a dynamic weighted feature fusion model of the pyrolysis system of decommissioned wind turbine blades based on the attention mechanism, the allowable range of pyrolysis process parameters is first used as a constraint condition, including the upper and lower limits of pyrolysis temperature, heating rate, and reaction time. Furthermore, safety constraints such as safe reactor temperature, safe pressure, and tail gas emission concentration are set to ensure that the model training and optimization process is carried out only within a feasible and safe process range. Based on this, three sub-models—the extreme gradient boosting algorithm, the multilayer perceptron, and the physical information neural network—are used to train the constrained sample data, respectively obtaining the predicted results of the product distribution. Subsequently, the process parameters, the predicted outputs of the three sub-models, and the actual experimental yield are used as comprehensive input features and introduced into a fully connected attention network for learning. The attention network analyzes the comprehensive features and adaptively determines the contribution weights among the three sub-models, enabling the weights to be dynamically adjusted according to different operating conditions. The fused predicted value is obtained by weighting the prediction results of the three sub-models according to the attention weights. The training objective of the model consists of a prediction error term and a regularization term, where the prediction error term is used to measure the difference between the fused prediction and the experimental yield, and the regularization term is used to limit the network parameter size and prevent overfitting. During the optimization process of the objective function, all training samples must meet the constraints of the process parameter range and the system safety constraints, and the output of the physical information neural network must also conform to the pyrolysis reaction kinetics. By limiting the feasible solution space through constraints, the objective function drives the attention network to converge to the optimal weight distribution, thereby realizing the dynamic fusion of the advantages of the three sub-models and finally obtaining a pyrolysis product prediction model that combines high accuracy, physical consistency, and operational safety.

4. The method for optimizing parameters of a decommissioned wind turbine blade pyrolysis system based on a dynamic weighted feature fusion algorithm as described in claim 1, characterized in that, The process by which the three sub-modules of S3—the extreme gradient boosting algorithm, the multilayer perceptron, and the physical information neural network—obtain the training results is shown below: The extreme gradient boosting algorithm is based on gradient boosting decision trees, and its prediction formula is expressed as follows: In the formula, α is the total number of trees, and g k (x i ) represents the prediction for the k-th tree. This is the predicted value from the Limit Gradient Boosting algorithm; the objective function L(θ) of the Limit Gradient Boosting algorithm includes training loss and regularization term, and can be expressed as follows: In the formula, This represents the loss function, used to measure the true value y of the i-th sample. i and model predictions The difference between them, Ω(g) k The regularization term () represents the penalty for model complexity, γ represents the tree complexity coefficient, T represents the number of leaf nodes in the k-th decision tree (more leaf nodes result in higher output complexity), and δ represents the regularization coefficient. θ w represents the number of leaves j This represents the weight of each leaf node; A multilayer perceptron is a feedforward neural network that can model nonlinear relationships. The output of its l-th layer is expressed by the following formula: h (l) =ρ(A (l) h (l-1) +b (l) ) In the formula, h (l) Let A represent the output vector of the l-th layer, σ represent the non-linear activation function, and A represent the output vector of the l-th layer. (l) Let h represent the weight matrix. (l-1) This represents the output vector of layer l-1, which is the input vector of layer l. (l) The deviation vector is represented by the following formula: In the formula, This represents the model's predicted output, where L represents the total number of layers in the network, and A represents the predicted output. (L) h represents the weight matrix of the output layer. (L-1) This represents the output vector of layer L-1, and is the input vector of layer L. (L) Represents the bias vector of the output layer; The multilayer perceptron is trained by minimizing the mean squared error (MSE). Specifically, the weights and biases of the network are iteratively optimized. First, the input data is passed through the input layer, undergoing linear transformations and non-linear activations in the hidden layers. Finally, the prediction result is obtained through linear calculation in the output layer. The MSE between the predicted and true values ​​is then calculated as the loss. Next, the chain rule is used to derive the loss from the output layer backwards, calculating the gradient of the loss with respect to the weights and biases of each layer, thus determining the direction and magnitude of parameter adjustments. Then, the step size is controlled according to the learning rate, and the parameters of all layers are updated along the gradient descent direction to reduce the loss. This process is repeated until the MSE converges to a stable minimum or the preset number of training iterations is reached, completing the model training. The loss function is expressed as follows: In the formula, Represents the loss function. y represents the model's predicted output. i This represents the actual value.

5. The method for optimizing parameters of a decommissioned wind turbine blade pyrolysis system based on a dynamic weighted feature fusion algorithm as described in claim 1, characterized in that, The construction of the dynamic weighted feature fusion model for the pyrolysis system of decommissioned wind turbine blades based on the attention mechanism is as follows: The min-max normalization method is used to normalize the prediction outputs of the three sub-modules: the extreme gradient boosting algorithm, the multilayer perceptron, and the physical information neural network, mapping the output values ​​to the [0,1] interval. The min-max normalization formula used in the normalization process is as follows: In the formula, y norm This represents the normalization result, where y represents the original data value that needs to be standardized, specifically the predicted output value of the submodule. min y represents the minimum value in the original dataset. max The maximum value in the original dataset is represented by the sum of the predictions from the normallyd sub-models trained separately, the original input data, and the actual experimental yield, which are then concatenated to construct a comprehensive training dataset. The comprehensive dataset is represented as follows: Data fusion =[X,y XGB|norm ,and MLP|norm ,and PINN|norm ,and data ] In the formula, X represents the original input parameter, y XGB|norm y represents the predicted value after normalization by the extreme gradient boosting algorithm. MLP|norm y represents the normalized prediction value of the multilayer perceptron. PINN|norm y represents the normalized predicted value of physical information from the neural network. data Indicates the original output value; The comprehensive dataset is used to train a fully connected attention network. The attention network outputs three weights ω1, ω2, and ω3, where ω1, ω2, and ω3 represent the contributions of the extreme gradient boosting algorithm, the multilayer perceptron, and the physical information neural network, respectively. The attention network evaluates the correlation between the sub-models based on input features and performance. Its original output a j Normalization is performed using the softmax function, a j and ω j The following formula can be used to calculate: a j =W j X+b j In the formula, X represents the input parameter, W j and b j This represents the trainable weight matrix and bias vector; Based on this, the Adam optimizer is used to update the parameter W. j and b j The learning rate was adjusted through cross-validation. Cross-validation divides the dataset into multiple subsets, and different subsets are used as validation sets in turn to evaluate the model's performance. When adjusting the learning rate, different learning rate values ​​were tested in multiple rounds of cross-validation experiments to observe the model's convergence speed and generalization effect on the validation set. Finally, the optimal learning rate was selected that allows the Adam optimizer to update W. j and b j At this time, the optimal learning rate is found to ensure balanced convergence of the model, preventing parameter oscillations and non-convergence due to an excessively large learning rate, and avoiding slow training due to an excessively small learning rate. Fusion prediction y fusion Represented as: the fusion =ω1y XGB|norm +ω2y MLP|norm +ω3y PINN|norm 。 6. The method for optimizing parameters of a decommissioned wind turbine blade pyrolysis system based on a dynamic weighted feature fusion algorithm as described in claim 1, characterized in that, The specific training steps for constructing the process parameter optimization model of the decomposition system for retired wind turbine blades in S4 are as follows: S41, Optimization objective and constraints: The yield prediction result Y(x) is obtained from the dynamic weighted feature fusion model. Other objectives can be obtained through physical relationships. Let the optimization variables be expressed as: x = [T, r(T), t] In the formula, T represents temperature, r(T) represents heating rate, and t represents reaction time; The other objectives include constraints on product quality, energy consumption, and process safety. The product quality Q(x) is related to the yield and temperature conditions, and can be estimated using an empirical function, as shown in the following formula: Q(x)=θ1Y(x)-θ2(T-T0) 2 In the formula, θ1 and θ2 are fitting coefficients, and T0 is the optimal temperature range; Energy consumption E(x) is expressed as the product of heat input power and time, and its formula is as follows: Where c is the system correlation coefficient; The objective function can then be expressed as: The constraints are In the formula, g j (x) represents a safety constraint; S42, Particle Swarm Optimization Solution: Each particle represents a set of process parameters x = [T, r(T), t]. Each particle updates its position based on its own historical best solution and the global best solution, gradually approaching the optimal process conditions. The update formula is as follows: Where ω represents the inertia weight, which controls the balance of particle exploration or development; c1 and c2 represent learning factors, which control the influence of individuals and groups; and r1 and r2 represent random numbers, which increase randomness. Finally, the particle swarm optimization outputs a Pareto optimal solution set P, which reflects different trade-offs between high yield, good quality, and low energy consumption, and can then be selected from the solution set according to actual needs.

7. A pyrolysis system for decommissioned wind turbine blades controlled by a dynamic weighted feature fusion algorithm, characterized in that, It includes a pyrolysis reactor (2), a carbon storage tank (3) connected to the outlet of the pyrolysis reactor (2), a condenser tower group, a gas collecting bottle (7) connected to the gas outlet at the top of the condenser tower group, a dry packed tower (8) connected to the side of the condenser tower group, an adsorption tower (9) connected to the outlet of the dry packed tower (8), and an exhaust fan (6) connected to the outlet of the adsorption tower (9). The bottom condensate collection section of both the condenser tower group and the dry packed tower (8) is connected to the oil collection tank (5).

8. The pyrolysis system for decommissioned wind turbine blades based on a dynamic weighted feature fusion algorithm control according to claim 7, characterized in that, The condenser tower group includes a pre-condenser tower (41), a middle condenser tower (42), and a post-condenser tower (43); the feed inlet of the pre-condenser tower (41) is connected to the discharge outlet of the pyrolysis reactor (2), the discharge outlet of the pre-condenser tower (41) is connected to the feed inlet of the middle condenser tower (42), the discharge outlet of the middle condenser tower (42) is connected to the feed inlet of the post-condenser tower (43), forming a staged condensation flow path; the gas outlet of the post-condenser tower (43) is connected to the gas inlet of the gas collecting bottle (7), and the inlet of the dry packed tower (8) is connected to the side of the post-condenser tower (43); The bottom condensate collection sections of the pre-condenser (41), the intermediate condenser (42), the post-condenser (43), and the dry packed tower (8) are all connected to an oil collection tank (5).

Images