Product quality prediction method and system based on methanol rectification system

Abstract

The invention relates to a product quality prediction and optimization method and system based on a methanol rectification system, and belongs to the field of product prediction, and the method comprises the following steps: (1) collecting process operation parameters and historical data in a methanol rectification production process; (2) carrying out supersampling and preprocessing on the collected data to obtain a training data set; (3) carrying out dimensionless processing on the preprocessed training data set; (4) training the dimensionless data by using a machine learning model to obtain a basic prediction model; and (5) predicting the ethanol content in the refined methanol by using the prediction model. The method has the advantages that flexibility is achieved, the basic models in the prediction model can be conveniently added or deleted, and various basic model combination schemes can be matched for deployment and use according to selection and requirements of a user; the method has the advantages of high accuracy and generalization, different basic models and fusion models are constructed by collecting historical data of the methanol rectification production process, and the evaluation result is more comprehensive and accurate.

Description

Technical Field

[0001] The invention relates to a product quality prediction method and system based on a methanol distillation system, belonging to the field of product prediction. Background Art

[0002] The methanol distillation system is a complex system with time-varying, multi-parameter, nonlinear, large hysteresis and multi-variable coupling. The ethanol content in refined methanol is an important indicator affecting the quality of methanol products. How to reduce energy consumption while ensuring product quality is a problem that current manufacturers are very concerned about.

[0003] The ethanol content in refined methanol is affected by many factors. In the system production process, the crude methanol feed amount, feed components, tower bottom liquid level, pressure and steam flow of the methanol distillation system will change with the change of the previous stage production situation, and the product quality of refined methanol mostly depends on manual testing, which is seriously delayed in time. When the production conditions change, the corresponding adjustments are not made in time, which easily leads to the risk of excessive energy consumption or substandard product quality. Therefore, it is necessary to comprehensively consider the impact of different working conditions on product quality. For example, CN103488085A constructs a multi-objective optimization control method for a methanol four-tower distillation system, selects 7 parameters such as crude methanol flow rate, methanol content, ethanol content, and extraction flow rate, and constructs a BP neural network for model training. The output result is the unit energy consumption of refined methanol, the refined methanol product flow rate of the pressure tower and the atmospheric tower, and the ethanol content of the pressure tower and the atmospheric tower. The invention only considers 7 process parameters as input, and does not consider the influence of factors such as tower bottom liquid level, tower top pressure, tower top temperature, and reflux flow on the ethanol content in refined methanol. Therefore, for the prediction of methanol product quality, not only the input and output flow rates of the product should be considered, but also factors such as temperature, pressure, and liquid level in the cascade production of the equipment.

[0004] In the methanol distillation process, energy consumption mainly involves steam and electricity. Steam is the main heat source, used to heat and evaporate methanol and its impurities. Steam consumption accounts for a significant proportion of total energy consumption. In the context of increasingly fierce market competition, reducing production costs and improving product quality are the key to maintaining competitiveness of enterprises. The methanol distillation system mainly adjusts the reflux ratio on the premise of ensuring product quality to reduce steam consumption and achieve the goal of energy saving and consumption reduction. At present, the quality of methanol products is mainly judged by quality inspection and testing data, which has a certain lag in time and cannot respond in time to changes in upstream production conditions. In addition, the process parameter adjustment of the distillation system basically relies on manual experience for regulation and control, which is subjective and cannot guarantee the optimal production state of product quality and energy consumption.

[0005] like Figure 1The figure shows the five-tower three-effect process flow of methanol distillation, which removes light components such as dimethyl ether, water, ethanol and other components in crude methanol to produce refined methanol. The refined methanol storage tank in the intermediate tank area is sent to the methanol finished product tank area for storage, and a certain amount of fusel alcohol is produced as a by-product. The entire process equipment is relatively large, and the heat of the three main distillation towers is coupled with each other and the materials are connected in series. At present, there are some technologies that build optimization control models based on some historical production data, but the collected data dimensions are not comprehensive enough to accurately guide production. Therefore, it is necessary to comprehensively consider the impact of the entire production process on the quality of the final product. Summary of the invention

[0006] In order to overcome the defects of the prior art, the present invention provides a product quality prediction and optimization method based on a methanol distillation system. The technical solution of the present invention is:

[0007] A method for predicting product quality based on a methanol distillation system comprises the following steps:

[0008] (1) Collecting process operation parameters and historical data during methanol distillation production;

[0009] (2) Supersample and preprocess the collected data to obtain a training data set; ensure the integrity, validity, timeliness, consistency, accuracy and uniqueness of the data;

[0010] (3) Perform dimensionless processing on the preprocessed training data set to eliminate the dimensional influence between different parameters;

[0011] (4) Use the machine learning model to train the dimensionless data to obtain the basic prediction model. The output results of several basic prediction models are stacked and fused to form a meta-model. The meta-models obtained by each branch are used to learn the relationship between each meta-model through the graph neural network to obtain the final prediction model.

[0012] (5) Use the prediction model to predict the ethanol content in refined methanol.

[0013] The process operating parameters and historical data in step (1) mainly include the process operating parameters and historical data of the distillation tower temperature, pressure, feed flow, steam flow, tower bottom liquid level, and quality inspection and testing data.

[0014] The step (3) is specifically as follows: dimensionless transformation is performed by normalization, and for each factor x i The numerical value is calculated as follows:

[0015] The step (4) is specifically as follows:

[0016] 4.1 Select a variety of machine learning models including but not limited to RF, SVM, XGBoost, LightGBM, and LSTM to perform parallel model training on these sample sets. The hyperparameter search method of each basic model is implemented with the help of the Hyperopt package. This method is based on the Bayesian method heuristic parameter search, which can accelerate the search process and obtain better parameter values;

[0017] 4.2 For the trained basic model, use the stacking fusion method, including but not limited to, to integrate the output results of the basic model as sample features, and use the original sample labels as new data sample labels to generate a new training set;

[0018] 4.3 Then train a new model based on the new training set, and finally use the new model to predict the samples.

[0019] The step 4.2 is specifically as follows:

[0020] 4.2.1 Since the basic model does not show structural information after the stacked fusion network, initialize their spatial structure: Among them, E1 and E2 are the embedded representations of the initialized nodes, which are subsequently optimized during training;

[0021] 4.2.2 Calculate the structural dependency weights between different basic models through matrix multiplication. Since the above step 4.2.1 obtains a fully connected graph, which is often unnecessary in graph representation learning and will increase the computational complexity of the model, the ReLU (·) activation function is used to eliminate weak connections and obtain the implicit structural relationship of self-learning.

[0022] In the step (1),

[0023] Each process parameter has an inherent change regularity and correspondence. For example, when other parameters remain unchanged, a decrease in reflux flow will lead to an increase in the ethanol content in refined methanol. Therefore, the single change of the process parameter is also crucial. The change of a process parameter over time is regarded as a set of sequences. The method of void causal convolution is used to capture the dependency on the time axis. It slides on the input sequence with a preset step jump value. It can process long-distance sequences in a non-recursive way, parallelize calculations and alleviate the gradient explosion problem. The convolution formula for each layer is: Among them, f(t) represents the representation before time t, x represents the current information, * is the convolution symbol, d is the sliding step size, and S is the total length of the sequence;

[0024] The gating structure is used to control the information flow of the temporal convolutional network so that only part of the information is retained after each convolution: h = g(W1×X+b1)☉σ(W2×X+b2), where h is the node feature in the new feature space, w is the transformation matrix, x is the original node feature, b is the bias term, and g and σ are activation functions.

[0025] The obtained representation h is considered as a subset of E1 and E2. The graph structure obtained in this way shows the influence of the process parameters in each basic model assigned on the ethanol content in refined methanol. Since the graph structure A is optimized according to the final structure, that is, the predicted value and the true value of the ethanol content are used as the loss optimization model, different importance can be assigned to different process parameters;

[0026] For spatial relationship capture, it is described as: Among them, A represents the adjacency matrix, k represents the network order. The larger k is, the more high-order relationships are captured. H is the information matrix of this layer, and W is the weight to be learned. Z is the node representation captured under the spatial relationship. The final Z representation is used as the predicted value. The predicted value is compared with the true value to calculate the MAE loss, so as to optimize the loss function: in, represents the predicted value, X represents the true value, k represents the network order, t represents the time step, T represents the predicted value length, and N represents the true value length.

[0027] A system for realizing a product quality prediction method based on a methanol distillation system, comprising:

[0028] Data acquisition module, used to collect process operation parameters and historical data in the methanol distillation production process;

[0029] A training data set generation module is used to perform supersampling and preprocessing on the collected data to obtain a training data set;

[0030] The dimensionless processing module is used to perform dimensionless processing on the preprocessed training data set to eliminate the dimensional influence between different parameters;

[0031] The prediction model generation module is used to use the machine learning model to train the dimensionless data to obtain the basic prediction model, and the output results of several basic prediction models are stacked and fused to form a meta-model. The meta-models obtained by each branch are learned through the graph neural network to learn the relationship between each meta-model to obtain the final prediction model;

[0032] A prediction module is used to predict the ethanol content in refined methanol using the prediction model.

[0033] A computer-readable storage medium is used to store a computer program for implementing the method.

[0034] A computer program product includes a computer-readable code or a computer-readable storage medium carrying the computer-readable code. When the computer-readable code is executed in an electronic device, a processor in the electronic device executes the method.

[0035] The advantages of the present invention are:

[0036] Data-driven accurate prediction: By collecting process operating parameters and historical data in the methanol distillation production process and training with machine learning models, product quality, such as the ethanol content in refined methanol, can be accurately predicted. This method can reduce reliance on experience and improve the accuracy and reliability of predictions.

[0037] Supersampling and preprocessing: Supersampling and preprocessing the collected data can increase the diversity and coverage of the data, improve the generalization ability of the model, and remove noise and outliers to ensure the quality of training data.

[0038] Dimensionless processing: Through dimensionless processing, the dimensional influence between different parameters is eliminated, making model training more efficient and improving the model's adaptability to parameters of different magnitudes.

[0039] Advanced model structure: Use machine learning models to train dimensionless data to obtain a basic prediction model, which is then fused into a meta-model through stacking networks. This structure can fully utilize the advantages of multiple basic models to improve prediction accuracy.

[0040] Application of graph neural networks: The meta-models obtained from each branch are learned through graph neural networks to learn the relationship between the meta-basic models, which can capture the complex interactions between the various parameters in the methanol distillation process and improve the expressiveness and accuracy of the prediction model.

[0041] Optimize the production process: Through accurate product quality prediction, the optimization of the production process can be better guided, such as adjusting operating parameters such as reflux ratio and load distribution ratio, thereby improving production efficiency and product quality and reducing production costs.

[0042] In summary, the present invention has:

[0043] 1) Flexibility: The basic models in the prediction model can be easily added or deleted, and can be deployed and used with a variety of basic model combinations according to user selection and needs.

[0044] 2) Generalization: This method collects historical data of the methanol distillation production process and combines a variety of equipment and process parameters to adaptively construct different basic models and fusion models to achieve higher accuracy and more comprehensive and accurate evaluation results.

[0045] 3) Reasonably set the parameter adjustment frequency: Since methanol distillation mainly relies on adjusting the reflux ratio and steam flow rate to control product quality, it takes a certain amount of time from the system being sent to the control system to achieve a stable operating state. If the reflux ratio and steam flow rate are adjusted too quickly, the system will not be able to reach a steady state. Therefore, it is necessary to reasonably set the parameter adjustment frequency according to different control systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a process flow chart of methanol distillation of the present invention.

[0047] Figure 2 It is a schematic diagram of the prediction method flow of the present invention.

[0048] Figure 3 It is a schematic diagram of a stacking network of the present invention.

[0049] Figure 4 It is a comparison chart of the predicted value and the actual value of the ethanol content in the refined methanol of the present invention.

[0050] Figure 5 It is a system structure block diagram of the present invention. DETAILED DESCRIPTION

[0051] The present invention will be further described below in conjunction with specific embodiments, and the advantages and features of the present invention will become clearer as the description proceeds. However, these embodiments are exemplary only and do not constitute any limitation to the scope of the present invention. It should be understood by those skilled in the art that the details and forms of the technical solution of the present invention may be modified or replaced without departing from the spirit and scope of the present invention, but these modifications and replacements all fall within the scope of protection of the present invention.

[0052] See also Figures 1 to 5 The present invention relates to a product quality prediction method based on a methanol distillation system, comprising the following steps:

[0053] (1) Collecting process operation parameters and historical data during methanol distillation production;

[0054] (2) Supersample and preprocess the collected data to obtain a training data set; ensure the integrity, validity, timeliness, consistency, accuracy and uniqueness of the data;

[0055] (3) Perform dimensionless processing on the preprocessed training data set to eliminate the dimensional influence between different parameters;

[0056] (4) Use the machine learning model to train the dimensionless data to obtain the basic prediction model. The output results of several basic prediction models are stacked and fused to form a meta-model. The meta-models obtained by each branch are used to learn the relationship between each meta-model through the graph neural network to obtain the final prediction model.

[0057] (5) Use the prediction model to predict the ethanol content in refined methanol.

[0058] In addition to the above steps (1) to (5), the model relationship between the historical process parameters and the ethanol content in the refined methanol constructed in the above steps can be combined with the current process parameters and the ethanol content. Under the conditions of ensuring product quality and the limitation of each parameter, the parameter optimization algorithm can be combined with the prediction model to search for the optimal process parameter setting values, so as to minimize the deviation between the predicted value and the target value of the ethanol content in the refined methanol.

[0059] The results of the present invention are shown as Figure 4 As shown in the figure, by collecting methanol distillation production process data and quality inspection data from a chemical plant in Shaanxi, the final ethanol content prediction optimization model was obtained through data dimensionless conversion, basic prediction model, stacked network, graph neural network fusion training and reverse optimization algorithm, where the horizontal axis is the time axis, the vertical axis is the ethanol content, the blue line is the true value of the ethanol content, and the green line is the predicted value of the ethanol content.

[0060] The process operating parameters and historical data in step (1) mainly include the process operating parameters and historical data of the distillation tower temperature, pressure, feed flow, steam flow, tower bottom liquid level, and quality inspection and testing data.

[0061] The step (3) is specifically as follows: dimensionless transformation is performed by normalization, and for each factor x i The numerical value is calculated as follows:

[0062] The step (4) is specifically as follows:

[0063] 4.1 Select a variety of machine learning models including but not limited to RF, SVM, XGBoost, LightGBM, and LSTM to perform parallel model training on these sample sets. The hyperparameter search method of each basic model is implemented with the help of the Hyperopt package. This method is based on the Bayesian method heuristic parameter search, which can accelerate the search process and obtain better parameter values;

[0064] 4.2 For the trained basic model, use the stacking fusion method, including but not limited to, to integrate the output results of the basic model as sample features, and use the original sample labels as new data sample labels to generate a new training set;

[0065] 4.3 Then train a new model based on the new training set, and finally use the new model to predict the samples.

[0066] The step 4.2 is specifically as follows:

[0067] 4.2.1 Since the basic model does not show structural information after the stacked fusion network, initialize their spatial structure: Among them, E1 and E2 are the embedded representations of the initialized nodes, which are subsequently optimized during training;

[0068] 4.2.2 Calculate the structural dependency weights between different basic models through matrix multiplication. Since the above step 4.2.1 obtains a fully connected graph, which is often unnecessary in graph representation learning and will increase the computational complexity of the model, the ReLU (·) activation function is used to eliminate weak connections and obtain the implicit structural relationship of self-learning.

[0069] In the step (1),

[0070] Each process parameter has an inherent change regularity and correspondence. For example, when other parameters remain unchanged, a decrease in reflux flow will lead to an increase in the ethanol content in refined methanol. Therefore, the single change of the process parameter is also crucial. The change of a process parameter over time is regarded as a set of sequences. The method of void causal convolution is used to capture the dependency on the time axis. It slides on the input sequence with a preset step jump value. It can process long-distance sequences in a non-recursive way, parallelize calculations and alleviate the gradient explosion problem. The convolution formula for each layer is: Among them, f(t) represents the representation before time t, x represents the current information, * is the convolution symbol, d is the sliding step size, and S is the total length of the sequence;

[0071] The gating structure is used to control the information flow of the temporal convolutional network so that only part of the information is retained after each convolution: h = g(W1×X+b1)☉σ(W2×X+b2); where h is the node feature in the new feature space, w is the transformation matrix, x is the original node feature, b is the bias term, and g and σ are activation functions.

[0072] The obtained representation h is considered as a subset of E1 and E2. The graph structure obtained in this way shows the influence of the process parameters in each basic model assigned on the ethanol content in refined methanol. Since the graph structure A is optimized according to the final structure, that is, the predicted value and the true value of the ethanol content are used as the loss optimization model, different importance can be assigned to different process parameters;

[0073] For spatial relationship capture, it is described as: Among them, A represents the adjacency matrix, k represents the network order. The larger k is, the more high-order relationships are captured. H is the information matrix of this layer, and W is the weight to be learned. Z is the node representation captured under the spatial relationship. The final Z representation is used as the predicted value. The predicted value is compared with the true value to calculate the MAE loss, so as to optimize the loss function: in, represents the predicted value, X represents the true value, k represents the network order, t represents the time step, T represents the predicted value length, and N represents the true value length.

[0074] A system for realizing a product quality prediction method based on a methanol distillation system comprises: a data acquisition module 1 for collecting process operation parameters and historical data in a methanol distillation production process;

[0075] The training data set generation module 2 is used to perform super sampling and preprocessing on the collected data to obtain a training data set;

[0076] The dimensionless processing module 3 is used to perform dimensionless processing on the preprocessed training data set to eliminate the dimensional influence between different parameters;

[0077] Prediction model generation module 4 is used to train dimensionless data using a machine learning model to obtain a basic prediction model, and to stack the output results of several basic prediction models to form a meta-model. The meta-models obtained by each branch are subjected to a graph neural network to learn the relationship between each meta-model to obtain the final prediction model.

[0078] The prediction module 5 is used to predict the ethanol content in the refined methanol using the prediction model.

[0079] The present invention also relates to a computer-readable storage medium for storing a computer program for implementing the method.

[0080] A computer program product includes a computer-readable code or a computer-readable storage medium carrying the computer-readable code. When the computer-readable code is executed in an electronic device, a processor in the electronic device executes the method.

[0081] The working principle of the present invention is:

[0082] The present invention provides a product quality prediction and optimization method based on a methanol distillation system. The present invention combines the actual production process of methanol distillation, uses equipment sensors and quality inspection and testing data, and constructs a product quality prediction and process parameter optimization model.

[0083] First, the original data is preprocessed, oversampled, and feature-screened, and then dimensionless. Machine learning models including but not limited to RF, SVM, XGBoost, LightGBM, LSTM, etc. are used for model training to obtain multiple basic prediction models.

[0084] These basic models are then fused using a stacked network, and each machine learning model is combined into a new model by assigning different weights, or called a Meta model; the meta-models obtained from each branch learn the relationship between the base models through a graph neural network, and then re-weight each base model to achieve model aggregation; then the information of each parameter at each time point in the past is used as input for initialization, and the graph network structure is constructed according to the importance of the weights of each base model to the ethanol content.

[0085] Then, according to the trend of each factor changing over time, the past time information of a single factor is captured to predict the future change trend of a single factor. At the same time, the graph network structure is optimized, that is, different importance is assigned to each factor according to the strength of its impact on the ethanol content.

[0086] Finally, a prediction model for ethanol content is constructed based on the prediction results of each factor and their influence on ethanol content.

[0087] In combination with the deviation between the current ethanol content and the target content, the parameter optimization method is used to calculate the optimal distillation tower setting parameters while meeting the parameter limitation conditions. The selection of parameter optimization methods includes but is not limited to PSO, artificial bee colony algorithm, sparrow search algorithm and other optimization methods, thereby realizing the construction of the entire prediction model and parameter optimization model.

[0088] The main technical innovations of the present invention are as follows:

[0089] 1. Most existing methods only consider the input and output flow rates in the production process, without considering the impact of process parameters in the intermediate links on product quality. This method is based on process data from each link, using statistical analysis, machine learning or deep learning algorithms to construct a quality prediction and process optimization model for a methanol distillation unit. This model can comprehensively consider the interactions between various factors, predict and optimize the production process.

[0090] 2. By analyzing historical data, the model can quickly provide the optimal reflux ratio and steam flow rate and other operating parameters under different feed conditions and raw material compositions. These parameters help to achieve refined control of the methanol distillation process.

[0091] 3. Through model prediction, employees can be guided to adjust the steam flow in time to adapt to changes in feed flow and composition, thereby reducing unnecessary steam consumption and improving energy utilization efficiency.

[0092] 4. The optimized reflux ratio and steam flow rate can ensure that the purity of the methanol product is stable at a high level, reduce the impurity content and improve product quality.

[0093] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A product quality prediction method based on a methanol distillation system, characterized in that: The following steps are involved: (1) Collecting process operation parameters and historical data during methanol distillation production; (2) supersampling and preprocessing the collected data to obtain a training data set; (3) Perform dimensionless processing on the preprocessed training data set to eliminate the dimensional influence between different parameters; (4) Use the machine learning model to train the dimensionless data to obtain the basic prediction model. The output results of several basic prediction models are stacked and fused to form a meta-model. The meta-models obtained by each branch are used to learn the relationship between each meta-model through the graph neural network to obtain the final prediction model. (5) Use the prediction model to predict the ethanol content in refined methanol.

2. The product quality prediction method based on the methanol distillation system according to claim 1, characterized in that: The process operating parameters and historical data in step (1) mainly include the process operating parameters and historical data of the distillation tower temperature, pressure, feed flow, steam flow, tower bottom liquid level, and quality inspection and testing data.

3. The method for predicting and optimizing product quality based on a methanol distillation system according to claim 1 or 2, characterized in that: The step (3) is specifically as follows: dimensionless transformation is performed by normalization, and for each factor x i The numerical value is calculated as follows:

4. The method for predicting and optimizing product quality based on a methanol distillation system according to claim 1 or 2, characterized in that: The step (4) is specifically as follows: 4.1 Select a variety of machine learning models including but not limited to RF, SVM, XGBoost, LightGBM, and LSTM to perform parallel model training on these sample sets. The hyperparameter search method of each basic model is implemented with the help of the Hyperopt package; 4.2 For the trained basic model, use the stacking fusion method, including but not limited to, to integrate the output results of the basic model as sample features, and use the original sample labels as new data sample labels to generate a new training set; 4.3 Then train a new model based on the new training set, and finally use the new model to predict the samples.

5. The method for predicting and optimizing product quality based on a methanol distillation system according to claim 4, characterized in that: The step 4.2 is specifically as follows: 4.2.1 Since the basic model does not show structural information after the stacked fusion network, initialize their spatial structure: Among them, E1 and E2 are the embedded representations of the initialized nodes, which are subsequently optimized during training; 4.2.2 Calculate the structural dependency weights between different basic models through matrix multiplication. Since the above step 4.2.1 obtains a fully connected graph, the ReLU (·) activation function is used to eliminate weak connections and obtain the implicit structural relationship of self-learning.

6. The method for predicting and optimizing product quality based on a methanol distillation system according to claim 5, characterized in that: In the step (1), The change of each process parameter over time is regarded as a set of sequences. The method of void causal convolution is used to capture the dependency on the time axis. The input sequence is slid with a preset step jump value. It can process long-distance sequences in a non-recursive way, parallelize calculations and alleviate the gradient explosion problem. The convolution formula for each layer is: Among them, f(t) represents the representation before time t, x represents the current information, * is the convolution symbol, d is the sliding step size, and S is the total length of the sequence; The gating structure is used to control the information flow of the temporal convolutional network so that only part of the information is retained after each convolution: h = g(W1×X+b1)⊙σ(W2×X+b2); where h is the node feature in the new feature space, w is the transformation matrix, x is the original node feature, b is the bias term, and g and σ are activation functions.

7. The method for predicting and optimizing product quality based on a methanol distillation system according to claim 6, characterized in that: The obtained representation h is considered as a subset of E1 and E2. The graph structure obtained in this way shows the influence of the process parameters in each basic model assigned on the ethanol content in refined methanol. Since the graph structure A is optimized according to the final structure, that is, the predicted value and the true value of the ethanol content are used as the loss optimization model, different importance can be assigned to different process parameters; For spatial relationship capture, it is described as: Among them, A represents the adjacency matrix, k represents the network order. The larger k is, the more high-order relationships are captured. H is the information matrix of this layer, and W is the weight to be learned. Z is the node representation captured under the spatial relationship. The final Z representation is used as the predicted value. The predicted value is compared with the true value to calculate the MAE loss, so as to optimize the loss function: in, represents the predicted value, X represents the true value, k represents the network order, t represents the time step, T represents the predicted value length, and N represents the true value length.

8. A system for implementing the product quality prediction method based on a methanol distillation system according to any one of claims 1 to 7, characterized in that: include: Data acquisition module, used to collect process operation parameters and historical data in the methanol distillation production process; A training data set generation module is used to perform supersampling and preprocessing on the collected data to obtain a training data set; The dimensionless processing module is used to perform dimensionless processing on the preprocessed training data set to eliminate the dimensional influence between different parameters; The prediction model generation module is used to use the machine learning model to train the dimensionless data to obtain the basic prediction model, and to stack the output results of several basic prediction models to form a meta-model. The meta-models obtained by each branch are used to learn the relationship between each meta-model through the graph neural network to obtain the final prediction model.

9. A computer-readable storage medium, characterized in that: A computer program for storing a computer program for implementing the method according to any one of claims 1 to 7.

10. A computer program product, comprising a computer-readable code, or a computer-readable storage medium carrying the computer-readable code, wherein when the computer-readable code is executed in an electronic device, a processor in the electronic device executes the method according to any one of claims 1 to 7.

Images