Multi-objective intelligent optimization method for coal gasification system production process based on knee region point guiding mechanism

Abstract

The invention discloses a coal gasification system production process multi-objective intelligent optimization method based on a knee region point guiding mechanism, and aims at the multi-objective optimization problem of synthesis gas product energy quality and system operation stability, and the production efficiency and the operation stability of a coal gasification system are improved at the same time. According to the method, a heterogeneous integrated optimization algorithm framework combining a Kriging model and a self-organizing feedforward neural network is constructed, and the model is constructed and trained through real-time data collected by sensor equipment of an industrial coal gasification system so as to accurately predict target parameters. And performing multi-target optimization on set values such as reaction temperature, pressure and flow in the coal gasification process. By introducing a knee region point guiding mechanism, an optimal tradeoff solution on a Pareto leading edge is identified in an algorithm evolution process, and an optimization scheme is dynamically adjusted by using a reinforced sampling strategy so as to cope with nonlinear changes in a complex industrial environment. The method shows good convergence, diversity and practical applicability in multi-objective optimization of the coal gasification system.

Description

Technical Field

[0001] The present invention belongs to the technical field of coal gasification system optimization. By using data modeling and intelligent optimization algorithms, it accurately predicts the key parameters of the coal gasification system, conducts multi-objective optimization on the set values of reaction temperature, pressure, flow rate, etc., so as to improve the energy quality of the syngas product, enhance the operation stability of the system, and achieve the improvement of the production efficiency of the coal gasification system and the efficient utilization of energy. Background Art

[0002] The coal gasification system is the core link of the coal chemical industry, and its operation status is directly related to product quality, production efficiency, and energy consumption. In actual production, the coal gasification process involves many complex physical and chemical reactions. The energy quality of the syngas product is affected by various factors, and at the same time, the operation stability of the system also faces many challenges. Traditional optimization control means for coal gasification systems have many deficiencies and are difficult to meet the requirements of modern coal chemical industries for high-efficiency and stable production.

[0003] The deep neural network combustion stability monitoring method based on stacked sparse autoencoders proposed by Han et al. has improved the prediction accuracy to a certain extent, but when dealing with the complex and changeable working conditions of the coal gasification system, the model generalization ability is insufficient. The one-dimensional convolutional autoencoder evaluated by Theunissen et al. has certain advantages in predicting positive pressure events in industrial furnaces, but when applied to the coal gasification system, the processing effect of multi-variable coupling and large time lag problems in the system is not good.

[0004] The RML machine learning stacking model proposed by Singh et al. shows high accuracy in predicting temperatures such as basic oxygen furnaces, but in the multi-objective optimization of the coal gasification system, it is difficult to balance the relationship between various objectives. The key variable prediction and interpretation model established by Lughofer et al. for the ironmaking blast furnace process cannot be directly applied to the energy quality prediction and optimization of the coal gasification system. The deep weighted joint distribution adaptive network fault diagnosis method proposed by Gao et al. has good effects in the abnormal state diagnosis of the blast furnace ironmaking process, but in the coal gasification system, the accuracy of fault diagnosis under complex working conditions needs to be improved.

[0005] The inventor has found that there are many problems in the current research on coal gasification systems. On the one hand, there are time lags and fluctuations in the detection of key parameters of the coal gasification system, the detection instruments are expensive and not easy to maintain, and it is difficult to obtain the operation status of the system in real time and accurately. On the other hand, existing optimization algorithms and models have defects in dealing with multi-objective optimization and complex non-linear relationships of the coal gasification system, resulting in the inability to achieve precise optimization and effective control of the coal gasification system, which seriously restricts the development of the coal chemical industry. Summary of the Invention

[0006] To solve the above problems, the present invention proposes a multi-objective intelligent optimization method for the production process of a coal gasification system based on a knee point guiding mechanism. A multi-objective optimization algorithm framework based on the Pareto-dominated evolutionary algorithm, combined with heterogeneous models, knee point guiding mechanisms, and enhanced sampling strategies, uses real-time data collected by sensors in the coal gasification system for data modeling and optimization algorithm processing to accurately optimize the operating setpoints of the coal gasification system. This method comprehensively applies the Kriging model and the Self-Organizing Feedforward Neural Network (SO-FNN) to construct heterogeneous models. By improving the energy quality of the syngas product and enhancing the operational stability of the system, while ensuring the production efficiency of the coal gasification system, the stable and reliable operation of the system is achieved. The invention provides an innovative solution for the efficient and stable operation of the coal gasification system, significantly improving the overall performance and economy of the system, and at the same time providing a new technical path for multi-objective optimization problems in complex industrial environments.

[0007] A multi-objective intelligent optimization method for the production process of a coal gasification system based on a knee point guiding mechanism comprises the following steps:

[0008] Step 1: First, perform population initialization and parameter setting. Collect real-time operation data through sensor devices in the coal gasification system, including key parameters such as reaction temperature, pressure, and flow rate, construct an initial data set, and perform data preprocessing;

[0009] Step 2: Construct heterogeneous models based on the Kriging model and the self-organizing feedforward neural network to describe the non-linear mapping relationship between the operating setpoints of the coal gasification system and the energy quality of the syngas, and at the same time use mathematical expressions to characterize the relationship between the operating setpoints of the system and the operational stability;

[0010] Step 3: According to the production objectives of the coal gasification system, set a multi-objective optimization problem with the energy quality of the syngas product and the operational stability of the system as the optimization objectives;

[0011] Step 4: Introduce a knee point guiding mechanism to identify the knee points in the Pareto front during the algorithm evolution process to guide the population evolution direction, and combine an enhanced sampling strategy to generate new sample points to improve the quality and diversity of the solution set.

[0012] Step 5: Adopt a multi-performance comprehensive evaluation method during the optimization solution process, use the finally obtained solution set to guide the regulation of the key operating parameters of the coal gasification system, and verify the practical feasibility and superiority of the optimization scheme.

[0013] Compared with the prior art, the beneficial effects of the present invention are:

[0014] The method of the present invention solves the problems of complex data modeling and low computational efficiency in the multi-objective optimization problem of the coal gasification system by introducing a heterogeneous integrated multi-objective evolutionary algorithm; constructs a heterogeneous model using the Kriging model and a self-organizing feedforward neural network to solve the problem of insufficient accuracy of traditional single models in non-linear mapping relationships; and dynamically identifies the optimal trade-off solutions on the Pareto front through the knee point guidance mechanism and the reinforcement sampling strategy, solving the problem of difficult balance between the diversity and convergence of the solution set in the multi-objective optimization process. The present invention can achieve rapid modeling and optimization of the operating state of the coal gasification system while obtaining the best balance point between the energy quality of the syngas product and the operating stability of the system, effectively improving the production efficiency and operating stability of the coal gasification system, and providing an innovative solution for the efficient and environmentally friendly operation of industrial coal gasification systems. Description of the Drawings

[0015] Figure 1 : Schematic diagram of the knee point on the Pareto front.

[0016] Figure 2 : Algorithm flow framework. Detailed Embodiment

[0017] The following provides a detailed description of the embodiments of the present invention. These embodiments are implemented on the premise of the technical solution of the present invention, and detailed implementation methods and specific operation processes are given. However, the protection scope of the present invention is not limited to the following embodiments.

[0018] First step: Collect real-time operation data through the sensor devices in the coal gasification system, construct an initial data set and perform data preprocessing. Use the sensor devices of the coal gasification system to obtain key operation parameters including reaction temperature, reaction pressure, gasifier agent flow rate, raw coal feeding rate, coal to gasifier agent mixing ratio, air / fuel ratio, waste gas emission flow rate, etc., and construct the original data set D0.

[0019] In order to screen out the characteristic variables highly correlated with the two optimization objectives of the energy quality of the syngas and the operating stability of the system, the Pearson correlation coefficient (PCC) is used to perform correlation analysis on each characteristic. The Pearson correlation coefficient γ is used to measure the linear correlation between two variables, and its calculation formula is as follows:

[0020] Where x n represents the observed value of the nth operating variable, Y n represents the energy quality of the syngas at the nth moment, represents the mean of the operating set value, μ Y represents the mean of the syngas energy quality index.

[0021] By calculating the absolute value of γ, features with relatively high correlation are selected as input variables for optimizing the model, namely reaction temperature, gasifier agent flow rate, reaction pressure, raw coal feeding rate, and air / fuel ratio. Then, a normalization method is adopted. The minimum value in the dataset is set as d min , and the maximum value is set as d max . Using the min-max scaling method, the data is linearly scaled to the specified minimum and maximum value intervals. The normalized dataset is defined as D, and the formula is expressed as:

[0022] where d0 is the original data, d max and d min are the maximum and minimum values in the dataset, norm min and norm max are the minimum and maximum boundary values of the normalized data range.

[0023] Step 2: Construct a heterogeneous model based on the Kriging model and the self-organizing feedforward neural network to describe the non-linear mapping relationship between the operating set values of the coal gasification system and the energy quality of the syngas. Select the five key parameters screened out, namely reaction temperature, gasifier agent flow rate, reaction pressure, air / fuel ratio, and raw coal feeding rate, as input features. First, construct the Kriging model, and use a statistical model to fit the relationship between the operating set values and the energy quality of the syngas. The Kriging model adopts the following form:

[0024] Z(x) = μ(x) + ε(x), ε(x) ~ N(0, σ 2 )(3) where Z(x) represents the predicted value of the objective function, μ(x) is the deterministic trend function, ε(x) is the local random perturbation term, which conforms to the normal distribution N(0, σ 2 ), σ 2 is a parameter obtained by dynamic calculation and calculated based on the maximum likelihood estimation, depending on the sample data and the correlation matrix calculation during model training.

[0025] Secondly, construct another component of the heterogeneous model. During the construction process of the self-organizing feedforward neural network model, the attributes of each neuron in the input layer are defined as: reaction temperature, gasifier agent flow rate, reaction pressure, air / fuel ratio, and raw coal feeding rate. The neuron in the output layer is defined as the energy quality of the syngas. The non-linear activation function selected between the input layer and the hidden layer is the Sigmoid function, set as g(x), and the expression is:

[0026]

[0027] The mathematical expectation of the mean square error is used as the objective function to test the performance of the neural network. Then, the mean square error at the l-th iteration is:

[0028] Among them, L is the total number of iterations of network training, t(l) is the expected output at the l-th iteration, and y(l) represents the actual output at the l-th iteration.

[0029] The weight between the i-th neuron in the hidden layer and the j-th neuron in the input layer is defined as w ij , and the adjustment amount at the l-th iteration is Δw ij (l), and the calculation formulas are as follows:

[0030] Among them, α is the learning rate during the neural network training process, and its value range is [10 -3 , 10 -2 , and E(l) is the mean square error at the l-th iteration.

[0031] To optimize the network structure, the number of hidden layer neurons is dynamically adjusted using a self-organizing growth mechanism. The upper limit of the hidden layer neurons is set to N max , which is set according to the number of input features and the trend of network error decline. Generally, a value within 5 times the number of input layer neurons is taken, that is, N max ≤25. In addition, during the network training process, when the number of hidden layer neurons N increases, the error E declines and tends to balance. It is stipulated that during the network training process, when N > 25 or N is in the state of |E(N) - E(N - 1)<εE, εE = 10-3, N no longer increases, and let N max = N, then the growth mechanism formula can be as follows:

[0032] N new = min(N + g * ·SGI, N max ), g * ∈[0.01, 0.1](7) Among them, g * is the self-growth coefficient, which reflects the growth rate of new neurons, and SGI is the neuron adaptive growth index, which reflects the contribution degree of hidden layer neurons to input data. The corresponding formula is:

[0033] Among them, h tj out is the output of the j-th hidden layer neuron for the t-th sample, and T is the total number of samples.

[0034] The adaptive growth mechanism of hidden layer neurons determines whether to add new hidden layer neurons according to the neuron adaptive growth index. If the neuron adaptive growth index of the j-th hidden layer neuron exceeds the specified adaptive threshold G th , that is, SGIj >G th indicates that the contribution value of the neuron has exceeded the preset value of the growth mechanism, and a new neuron needs to be added to the neuron to enhance the expression ability of the network. Threshold G th The specific formula is as follows:

[0035]

[0036] The Kriging model and the self-organizing feedforward neural network model are trained by cross-validation, and the root mean square error (RMSE) of their respective test groups is updated after each training and construction. Then, the Kriging and neural network models respectively predict the input data to generate the syngas energy quality estimation output. For the output of the heterogeneous models, the output weights are allocated according to the proportion of the RMSE between the two models to share the population parameter information. Use ratio to represent the weight of the output of the neural network model, and define the output of the heterogeneous model as O H The formula is as follows:

[0037]

[0038] O H = ratio · O F + (1 - ratio) · O K (11) where R K and R F respectively represent the RMSE of each model, and O F and O K respectively represent the respective output values of the two models. The output value O H of the heterogeneous model is used as the fitness value of the offspring solution in the population evolution process and subsequent operations are carried out.

[0039] Step 3: Design a multi-objective optimization problem with the syngas product energy quality and the system operation stability as the optimization objectives according to the production process characteristics of the coal gasification system. Set the optimization objectives of the coal gasification system as F1(t) and F2(t) respectively. The syngas product energy quality is used as one of the objective functions F1(t), represented by SQ, and is described using the integrated model as follows:

[0040] SQ = τ1 · f Kriging (x) + τ2 · f SO-FNN (x) (12) where f Kriging and f SO-FNN respectively refer to the two models, and τ1 and τ2 respectively represent the weight of each model. When using the equal fusion method for model output calculation, τ1 = τ2 = 0.5.

[0041] Another objective function F2(t) is the system operation stability, denoted by ST, which is represented by minimizing the fluctuation range of operation parameters. The reference set values x of the key operation parameters of the system r are generally set according to the requirements of the gasification process and expert experience. Specifically, the reaction temperature is usually controlled at 1300°C - 1500°C, and the typical value is 1400°C; the flow rate of the gasifying agent is generally 2000 - 3000 Nm 3 / h, and the typical value is 2500 Nm 3 / h; the reaction pressure is usually controlled at 2.5 - 4.0 MPa, and the typical value is 3.0 MPa; the conveying rate of the raw coal is generally 20 - 40 t / h, and the typical value is 30 t / h; the air / fuel ratio is usually 1.2 - 1.5, and the typical value is 1.3. In actual industrial production, flexible adjustments are made according to the specific process specifications and on-site production requirements in different factories and under different working conditions. The specific formula for the system operation stability is as follows:

[0042] where x p represents the real-time value of the p-th key operation parameter, and P is the total number of key operation parameters, that is, P = 5. The smaller the fluctuation of the operation stability, the stronger the stability and the smoother the production process.

[0043] Based on the above objectives, the multi-objective optimization problem model can be described as: min{-F1(t), F2(t)}, and the multi-objective optimization problem F(x) is designed as shown in Equation (14):

[0044] Minimize F(x) = [SQ, ST] (14)

[0045] Fourthly, introduce the knee point guiding mechanism. By identifying the knee points on the Pareto front, guide the population evolution direction, and incorporate the enhanced sampling strategy to conduct intensive sampling near the knee points to generate new sample points to improve the quality and diversity of the solution set. On the Pareto front, the knee point is the point with the maximum curvature, and the corresponding knee interval is a region on the Pareto front around the knee point, representing the approximate optimal trade-off solution between the objective functions. The knee point and the knee interval can be represented as Figure 1 shown, where point C is the knee point on the Pareto front, and the corresponding [R1, R2] is the knee interval.

[0046] The non - dominated sorting method (NDSort) is used to sort each objective value in the current Pareto solution set. The maximum value points of each objective function are found, which represent the points where the objective functions reach the optimal values on the Pareto front, and a candidate set of extreme value points for each objective function is obtained. For the objective of syngas exergy, the maximum value point is the solution with the highest calorific value in the current solution set; for the objective of system operation stability, the maximum value point is the solution with the smallest fluctuation. These maximum value points reflect the key characteristics of the Pareto front in the objective space and are used to construct hyperplanes to describe the overall trend of the Pareto front. Specifically, for each objective function f o (o = 1, 2,..., O), the formula for selecting the maximum value point is as follows:

[0047] where O represents the total number of objective functions in the multi - objective optimization problem. In this optimization problem, O = 2, that is, the optimization objectives include syngas exergy and system operation stability.

[0048] All these maximum values are gathered to form the set of maximum value points of the current front

[0049] According to the sorting results, the extreme value points of each objective function are sequentially selected as the basis for hyperplane calculation, and it is ensured that the selected extreme value points are not repeated and cover the extremes of each objective function. Then, hyperplane construction and distance calculation are carried out. The hyperplane is defined by the linear equation:

[0050] a1f1 + a2f2+...+a M f M +b = 0 (16) where a1, a2,..., a M are the normal vector coefficients of the hyperplane, corresponding to the normal vector coefficients of the hyperplane in the m - th dimension respectively, and are calculated based on the sensitivity analysis of each optimization objective to the operating variables; f1, f2,..., f M are independent variables, corresponding to the coordinate values in the M - th dimension respectively; b is the intercept term of the hyperplane, representing the distance between the hyperplane and the origin.

[0051] The coefficients and intercept of the hyperplane are determined by minimizing the sum of squared fitting errors E, and the calculation formula is:

[0052]

[0053] Subsequently, the perpendicular distance d(x) from each solution on the Pareto front to the hyperplane is calculated as a measure of its degree of deviation from the hyperplane. The formula for the perpendicular distance is:

[0054]

[0055] In the multi-objective optimization algorithm of the coal gasification system, the crossover operation is an important genetic operation for generating new solutions (offspring individuals). Its core idea is to generate new solutions with diversity and potential superiority, that is, offspring individuals, by combining the information of two parent individuals, thereby enhancing the exploration ability of the population.

[0056] During the algorithm evolution process, as the solution that achieves the optimal balance between energy quality and stability, the knee point has strong guiding ability. Therefore, in order to make the newly generated individuals approach the knee point, we need to generate offspring individuals by combining the position characteristics of the parent individuals and the flexibility of the crossover operation. The crossover operation formula is as follows:

[0057] x new = α c ·x p1 +(1 - α c )·x p2 (19) Among them, x new represents the newly generated individual, x p1 and x p2 are the parent individuals, respectively from the optimal solution set of the previous generation, α c is the crossover coefficient, which controls the mixing degree of the information of the two parents. Generally, α c = 0.5;

[0058] After that, the mutation operation is performed. Δx is the perturbation term, and its value is a random number in the range of [0, 1]. The corresponding formula is:

[0059] x new = x parent1 + △x(20)

[0060] To further improve the quality and diversity of the solution set, a reinforcement sampling strategy is introduced during the population evolution process, and intensive sampling is carried out near the knee point based on the uniform distribution method. Specifically, let the knee point be v represents the dimension of the decision variable. In this optimization problem, v = 5. For the q-th decision variable The scaling coefficient λ q is used to determine the neighborhood radius d q , and the calculation formula for the neighborhood radius d q is as follows:

[0061] d q = λ q ·(b q - a q )(21) Among them, a q and b q are respectively the lower bound and the upper bound of the q-th decision variable, and λ qis a constant with a value in the range (0, 0.5), and generally λ q = 0.2.

[0062] Within the neighborhood of the q-th decision variable, generate a random offset ε q ∈[-d q , d q , and the calculation formula is as follows:

[0063] ε q = d q ·(2·θ q - 1), where θ q ∈[0, 1](22) where θ q is a random number uniformly distributed on the interval [0, 1].

[0064] According to the generated random offset, calculate the new sample point of the q-th decision variable

[0065] At this time, if exceeds the value range [a q , b q of the q-th decision variable, then correct it to the boundary value, and the specific formula is as follows:

[0066]

[0067] Step 5: In the process of algorithm evolution, adopt a multi-performance comprehensive evaluation method, and use the finally obtained solution set to guide the regulation of the key operating parameters of the coal gasification system, and verify the practical feasibility and superiority of the optimization scheme. For the newly sampled individuals, use the convergence index Inverted Generational Distance (IGD) and the diversity index Spacing to comprehensively evaluate the quality of individuals. The smaller the IGD value, the closer the solution set is to the optimal front, and the better the convergence of the algorithm; the smaller the Spacing value, the more uniform the distribution of the solutions, and the better the diversity of the algorithm. The calculation results are taken in an equal fusion manner, and the individuals corresponding to the former are retained until the number of individuals is less than or equal to the preset value of 120. Then, according to the objective function, evaluate and update the solution archive of the above population evolution process again.

[0068] This process is continuously iterated, and the algorithm continuously optimizes the population, so that the solution set is continuously improved in terms of convergence and diversity. Until the maximum evaluation times (300 times) of the preset fitness function are reached, the algorithm terminates. At this time, the obtained solution set will provide scientific and effective guidance for the regulation of the key operating parameters of the coal gasification system, and fully demonstrate the feasibility and superiority of this optimization scheme in practical applications.

Claims

1. A multi-objective intelligent optimization method for the production process of a coal gasification system based on a knee region point guidance mechanism, characterized in that, It includes the following steps: Step 1: Collect real-time operation data through the sensor devices in the coal gasification system to construct an initial data set; Step 2: Construct a heterogeneous model for describing the complex non-linear relationship between the operation set values of the coal gasification system and the energy quality of syngas; Step 3: According to the characteristics of the coal gasification system, set a multi-objective optimization problem including the energy quality of syngas products and the operation stability of the system; Step 4: Introduce a knee point guiding mechanism to guide the optimization direction, and combine with a reinforcement sampling strategy to generate new sample points to optimize the quality and diversity of the solution set; Step 5: Use a multi-performance comprehensive evaluation method in the optimization solution process, and use the finally obtained solution set to guide the regulation of the key operation parameters of the coal gasification system.

2. The multi-objective intelligent optimization method for the production process of a coal gasification system based on a knee region point guiding mechanism according to claim 1, characterized in that, In Step 1, real-time operation data is collected through the sensor devices in the coal gasification system, an initial data set is constructed and data preprocessing is carried out; the selected data types include key operation parameters such as reaction temperature, reaction pressure, gasifier agent flow rate, raw coal feed rate, coal-gasifier agent mixing ratio, air / fuel ratio, waste gas emission flow rate, etc., and the original data set D0 is constructed based on this; the Pearson Correlation Coefficient (PCC) is used to analyze the correlation of each feature to screen the feature variables highly relevant to the optimization objective. The Pearson correlation coefficient γ is used to measure the linear correlation between two variables, and its calculation formula is: Among them, x n represents the nth observed value of the operating variable, and Y n represents the exergy of syngas at the nth moment, represents the mean value of the operating setpoint, and μ Y represents the mean value of the syngas exergy index; By calculating the absolute value of γ, the features with higher correlation are screened as the input variables of the optimization model, which are reaction temperature, gasifier agent flow rate, reaction pressure, raw coal conveying rate, air / fuel ratio respectively; After that, a normalization method is adopted, and the minimum value in the dataset is set to d min , and the maximum value is d max . The min-max scaling method is used to linearly scale the data to the specified minimum and maximum value intervals. The normalized dataset is defined as D, and the formula is expressed as: Among them, d0 is the original data, d max and d min are the maximum and minimum values in the dataset, norm min and norm max are the minimum and maximum boundary values of the data range after normalization.

3. The multi-objective intelligent optimization method for the production process of a coal gasification system based on a knee-point guiding mechanism according to claim 1, characterized in that, Step 2 constructs a heterogeneous model based on the integration of the Kriging model and the self-organizing feedforward neural network SO-FNN to describe the complex non-linear relationship between the operation set values of the coal gasification system and the energy quality of syngas; among them, in the training process of the neural network model, each neuron in the input layer is defined as: reaction temperature, gasifier agent flow rate, reaction pressure, air / fuel ratio, raw coal feed rate, and the neuron in the output layer is defined as the energy quality of syngas. The non-linear activation function selected between the input layer and the hidden layer is the Sigmoid function, which is set as g(x), and the expression is: The mathematical expectation of the mean square error is used as the objective function to test the performance of the neural network, and the mean square error at the l-th iteration is: Among them, L is the total number of iterations of network training, t(l) is the expected output at the l-th iteration, and y(l) represents the actual output at the l-th iteration; The weight between the $i$-th neuron in the hidden layer and the $j$-th neuron in the input layer is defined as $w$ ij , and the adjustment amount at the $l$-th iteration is $\Delta w$ ij (l), and the calculation formula is as follows: Among them, α is the learning rate during the neural network training process, and its value range is [10 -3 , 10 -2 , and E(l) is the mean square error at the l-th iteration; The number of hidden layer neurons is dynamically adjusted using a self-organizing growth mechanism, and the upper limit of the hidden layer neurons is set to N max , taking a value within 5 times the number of input layer neurons, that is, N max ≤25. In addition, during the network training process, when the number of hidden layer neurons N increases, the error E decreases and tends to balance. It is stipulated that during the network training process, when N > 25 or N is in the state of |E(N) - E(N - 1)<ε E , ε E = 10 -3 |, N will no longer increase, and let N max = N, then the growth mechanism formula can be shown as follows: N new = min(N + g * ·SGI, N max ), g * ∈ [0.01, 0.1](7) Among them, g * is the self-growth coefficient, reflecting the growth rate of new neurons, and SGI is the neuron adaptive growth index, reflecting the contribution degree of hidden layer neurons to input data. The corresponding formula is: where h tj out is the output of the j-th hidden layer neuron for the t-th sample, and T is the total number of samples; The growth mechanism determines whether new hidden layer neurons need to be added according to the neuron's adaptive growth index; when the adaptive growth index of the j-th hidden layer neuron exceeds the specified adaptive threshold G th , namely SGI j > G th , it means that the contribution value of this neuron has exceeded the preset value of the growth mechanism, and a new neuron needs to be added to this neuron to enhance the expression ability of the network; the adaptive threshold G th The corresponding formula is as follows: After that, the heterogeneous model is trained in a cross-validation manner. The included Kriging model and neural network model update their respective test group mean square error RMSE after each training is completed; then the Kriging and neural network models respectively predict the input data to generate the estimated output of the energy quality of syngas; for the output of the heterogeneous model, the output weights are allocated according to the proportion of the RMSE size between the two models to share the population parameter information; the output value of the heterogeneous model is used as the fitness value of the offspring solution in the population evolution process and subsequent operations are carried out.

4. A multi-objective intelligent optimization method for the production process of a coal gasification system based on a knee point guidance mechanism according to claim 1, characterized in that, Step 3 designs a multi-objective optimization problem with the energy quality of syngas products and the system operation stability as the optimization objectives according to the production process characteristics of the coal gasification system; Set the optimization objectives of the coal gasification system as F1(t) and F2(t). The energy quality of syngas products is used as one of the objective functions F1(t), denoted by SQ, and is described using an integrated model as follows: SQ = τ·f Kriging (x) + υ·f SO-FNN (x) (10) where, f Kriging and f SO-FNN respectively refer to two models, τ1 and τ2 respectively represent the weight of each model. When using the equal fusion method for model output calculation, τ1 = τ2 = 0.5; Another objective function F2(t) is the system operation stability, denoted by ST, which is represented by minimizing the fluctuation range of operation parameters. The reference set value of the key operation parameters of the system is x r ; The key operation parameters specifically include: the reaction temperature is controlled at 1300°C - 1500°C; the gasification agent flow rate is 2000 - 3000 Nm 3 / h; the reaction pressure is controlled at 2.5 - 4.0 MPa, and the typical value is 3.0 MPa; the raw coal conveying rate is 20 - 40 t / h; the air / fuel ratio is 1.2 - 1.5; The specific formula for the system operation stability is as follows: where x p represents the real-time value of the p-th key operating parameter, P is the total number of key operating parameters, that is, P = 5; the smaller the fluctuation of the operating stability, the stronger the stability and the smoother the production process; Based on the above objectives, design the multi-objective optimization problem F(x) as follows: Minimize F(x) = [SQ, ST] (12) Finally, determine that the multi-objective optimization problem model is min{-F1(t), F2(t)}.

5. The multi-objective intelligent optimization method for the production process of a coal gasification system based on the knee region point guidance mechanism according to claim 1, characterized in that The evolutionary direction of the knee region point-guided mechanism optimization algorithm dynamically guides the optimization direction, incorporates a reinforcement sampling strategy to densely sample near the knee region points, generates new sample points to improve the quality and diversity of the solution set; uses the non-dominated sorting method (NDSort) to sort each objective value in the current Pareto solution set, finds the maximum value points of each objective function to represent the points where the objective function reaches the optimal value in the Pareto front, and obtains the extreme value point set of each objective function; for each objective function f o (o = 1, 2,..., O), the selection formula for the maximum value point is as follows: Among them, O represents the total number of objective functions in the multi-objective optimization problem; in this optimization problem, O = 2, that is, the optimization objectives include the energy quality of syngas and the system operation stability; Pool all these maximum values to form the set of maximum points of the current front According to the sorting results, successively select the extreme points of each objective function as the basis for hyperplane calculation. Then, construct the hyperplane and calculate the distance. The hyperplane is defined by the linear equation as: a1f1 + a2f2 +... + a M f M + b = 0(14) Among them, a1, a2,..., a M are the coefficients of the normal vector of the hyperplane, corresponding to the coefficients of the normal vector of the hyperplane in the m-th dimension respectively; f1, f2,..., f M are independent variables, corresponding to the coordinate values in the M-th dimension respectively; b is the intercept term of the hyperplane, representing the distance between the hyperplane and the origin; by minimizing the sum of squared fitting errors E * to determine the coefficients and intercept of the hyperplane, the calculation formula is: Subsequently, calculate the vertical distance d(x) from each solution on the Pareto front to the hyperplane as a measure of its degree of deviation from the hyperplane; the specific calculation formula is: After that, a reinforcement sampling strategy is introduced during the population evolution process, and intensive sampling is carried out in the vicinity of the knee point based on a uniform distribution. Let the knee point be v represents the dimension of the decision variable. In this optimization problem, v = 5. For the q-th decision variable Use the scaling factor λ q to determine the neighborhood radius d q , the neighborhood radius d q The calculation formula is as follows: d q = λ q ·(b q - a q )(17) where a q and b q are the lower and upper bounds of the q-th decision variable respectively, and λ q is a constant with a value between (0, 0.5); Generate a random offset ε within the neighborhood of the q-th decision variable q ∈[-d q ,d q , and the calculation formula is as follows: ε q = d q ·(2·θ q −1), θ q ∈ [0, 1](18) where θ q is a random number uniformly distributed on the interval [0, 1]; Calculate the new sample point of the q-th decision variable according to the generated random offset At this time, if exceeds the value range [a q , b q of the q-th decision variable, it is corrected to the boundary value, and the specific formula is as follows:

6. The multi-objective intelligent optimization method for the production process of a coal gasification system based on a knee point guidance mechanism according to claim 1, wherein Step 5 adopts a multi-performance comprehensive evaluation method in the optimization solution process, uses the finally obtained solution set to guide the regulation of the key operation parameters of the coal gasification system, and comprehensively evaluates the quality of the solution set by using the convergence performance index and the diversity index for the newly sampled individuals.

Images