Industrial process control method and system based on expected setpoint error neural network

Abstract

Disclosed in the present invention are an industrial process control method and system based on an expected setpoint error neural network. The method comprises: establishing a control output network model and a target prediction network model, determining specific control targets and corresponding controlled objects in an industrial process, selecting control variables on the basis of the control targets, and setting corresponding expected values; collecting time-series data from the industrial process, and performing data processing on the collected time-series data; using the time-series data subjected to data processing to train a target prediction network, obtaining prediction results, and calculating and integrating a plurality of error types to optimize the target prediction network; on the basis of the prediction results provided by the target prediction network and the time-series data, training a control output network, and calculating and integrating a plurality of error types to optimize the control output network; and applying trained network models to industrial process control, and on the basis of environmental changes, performing continuous incremental training. The present invention can improve the control effect for industrial processes.

Description

An industrial process control method and system based on a neural network for desired setting error. Technical Field

[0001] This invention relates to the field of industrial process control technology, and specifically to an industrial process control method and system based on a neural network for desired setting error. Background Technology

[0002] Process manufacturing, as a crucial component of the modern industrial system, encompasses multiple sectors including chemicals, petroleum, pharmaceuticals, and food. In process manufacturing, numerous pieces of equipment work collaboratively through automated control, and the quality of this control directly determines the quality of the products produced. Especially with increasing globalization and market competition, process manufacturing faces multiple challenges, including improving production efficiency, reducing production costs, ensuring product quality, and protecting the environment. Against this backdrop, the introduction and application of intelligent control technology has become key to the transformation and upgrading of the process manufacturing industry.

[0003] Currently, most industrial production processes still rely on traditional PID control for automatic control. With the emergence of the concept of intelligent manufacturing in recent years, and the advancement of machine learning technology, more and more companies are using machine learning algorithms to achieve automatic control of production processes. However, the currently available machine learning control schemes generally fall into two categories: one is to use neural networks and other algorithms to fit historical control data that has performed well, for use in subsequent production control outputs; the other is to predict the future trends of control variables and make rule-based compensation adjustments to the traditional controller.

[0004] The first approach is flawed because the control factors in the production process are constantly changing, so relying on excellent historical data and experience is insufficient to handle abnormal scenarios. The second approach is flawed because the factors influencing the control objective are diverse and constitute a very complex process. Simply using future target predictions to modify traditional controllers such as PID controllers is one-sided. Summary of the Invention

[0005] Technical objective: To address the shortcomings of existing industrial process control technologies, this invention discloses an industrial process control method and system based on a neural network for expected setting error, which can improve control performance.

[0006] Technical solution: To achieve the above technical objectives, the present invention adopts the following technical solution:

[0007] An industrial process control method based on a neural network for desired setting error includes:

[0008] Establish a control output network model and a target prediction network model, determine the specific control target and corresponding control object in the industrial process, select control variables according to the control target, and set the corresponding expected values. The control output network includes the state-time network of the control output network, the expected target neural network, and the integrated neural network of the control output network. The target prediction network includes the state-time network of the target prediction network, the next step output neural network, and the integrated neural network of the target prediction network.

[0009] Collect time-series data from industrial processes and perform data processing on the collected time-series data;

[0010] The target prediction network is trained using the processed time-series data, the prediction results are obtained, and multiple error types are calculated and integrated to optimize the target prediction network;

[0011] The control output network is trained based on the prediction results and time series data provided by the target prediction network, and the control output network is optimized by integrating multiple error types.

[0012] The trained control output network and target prediction network are applied to industrial process control, and continuous additional training is conducted based on environmental changes.

[0013] This invention also provides an industrial process control system based on a desired setpoint error neural network, used to implement the above-described industrial process control method based on a desired setpoint error neural network, comprising:

[0014] The control output network is used to output subsequent control actions based on the current control state and the desired target. The control output network includes the state-temporal network of the control output network, the desired target neural network, and the integrated neural network of the control output network.

[0015] The state-time network of the control output network is used to learn and extract features from the time-series data of the current control state;

[0016] A target neural network is used to compute the features of the desired control target.

[0017] An integrated neural network for controlling the output network is used to integrate the outputs of the state-temporal network of the control output network and the desired target neural network to generate the next control action;

[0018] The target prediction network is used to predict the control target based on the current control state and the next control action. The target prediction network includes a state-sequential network, a next output neural network, and an ensemble neural network of the target prediction network.

[0019] The target prediction network is a state-series network used to learn and extract features from the time-series data of the current control state;

[0020] The next step is to output a neural network, which is used to calculate the features of the next control action;

[0021] An ensemble neural network for target prediction is used to integrate the state-sequence network of the target prediction network with the output of the next-step output neural network to generate the predicted target.

[0022] Beneficial Effects: The industrial process control method and system based on a neural network for desired setting error provided by the present invention have the following beneficial effects:

[0023] 1. This invention evaluates and optimizes system performance at multiple levels by introducing factual data error, adversarial error, and empirical error. In particular, adversarial error can enhance the system's adaptability to abnormal and extreme situations. By integrating state-time network and desired target neural network, the accuracy of control output is improved. The integrated neural network can effectively fuse the output features of different networks, ensuring that the generated control commands are more accurate and efficient.

[0024] 2. By continuously collecting new data and performing real-time incremental learning and model retraining, the system can dynamically adapt to changes in the production environment, maintain long-term stability and efficient operation, and continuously conduct additional training on the network in actual production to ensure that the system can adapt to changes in the production environment and improve the robustness and adaptability of the control system.

[0025] 3. Compared with the PID control in the prior art, the present invention, by combining the control output network and the target prediction network, can more accurately predict and adjust the control variables, thereby improving production efficiency and product quality; moreover, the present invention can process and optimize multiple control variables simultaneously, making complex industrial control processes more efficient and stable. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0027] Figure 1 is a flowchart of the industrial process control method based on the desired setting error neural network of the present invention;

[0028] Figure 2 is a structural framework diagram of the industrial process control system based on the desired setting error neural network of the present invention;

[0029] Figure 3 is a schematic diagram of the temperature error control results in the actual application of the present invention.

[0030] In the diagram: 100, Control output network; 101, State-temporal network of control output network; 102, Expected target neural network; 103, Integrated neural network of control output network; 110, Target prediction network; 111, State-temporal network of target prediction network; 112, Next step output neural network; 113, Integrated neural network of target prediction network. Detailed Implementation

[0031] The present invention will now be described more clearly and completely by way of a preferred embodiment in conjunction with the accompanying drawings, but this does not limit the invention to the scope of the described embodiment.

[0032] In one embodiment, as shown in Figure 1, an industrial process control method based on a desired setting error neural network is provided, comprising the following steps:

[0033] S101. Establish a control output network model and a target prediction network model, determine the specific control target and corresponding control object in the industrial process, select control variables according to the control target, and set the corresponding expected values. The control output network includes the state-time network of the control output network, the expected target neural network, and the integrated neural network of the control output network. The target prediction network includes the state-time network of the target prediction network, the next output neural network, and the integrated neural network of the target prediction network.

[0034] In one embodiment, the control target refers to the process parameters that need to be controlled in the industrial process, such as temperature and pressure, while the control object refers to the equipment or certain processes associated with the process parameters.

[0035] S102. Collect time-series data from industrial processes and process the collected time-series data.

[0036] In one embodiment, the time-series data includes historical operational data and real-time monitoring data to ensure the comprehensiveness and accuracy of the data, covering various operational states and abnormal situations.

[0037] In one embodiment, data processing of the collected time-series data includes feature expansion, feature selection, and normalization. The purpose of feature expansion is to introduce derived features to enhance the predictive ability of the system model. The purpose of feature selection is to select the most representative data features and extract redundant information to simplify the system model. The purpose of data normalization is to unify the size and reduce the impact of data differences on the training of the system model.

[0038] S103. Train the target prediction network using the processed time-series data, obtain the prediction results, and calculate and integrate multiple error types to optimize the target prediction network.

[0039] The purpose of training the target prediction network is to predict future control targets and provide a basis for the system's automated decision-making. This involves the following steps:

[0040] S301. Input the time series data corresponding to the Kth time series node into the state time series network of the target prediction network. Input the next output C calculated by the control output network corresponding to the time series data of the K+1 time series nodes into the next output neural network. After the ensemble neural network of the target prediction network calculates, output the predicted target P1 of the K+Nth time series node.

[0041] The predicted target P1 is the future target predicted by the target prediction network under normal conditions based on the current time series data and control output. The specific process is as follows: the current time series data, such as temperature and pressure, and other control states are input into the state time series network 111 of the target prediction network. The state time series network 111 of the target prediction network learns the time series characteristics of the current state and outputs a feature vector. The current control output is used as input and input into the next output neural network 112. The ensemble neural network 113 of the target prediction network integrates the outputs of the state time series network 111 and the next output neural network 112 to generate the predicted target P1.

[0042] S302. Calculate the training error E1 based on the actual output Q of K+N time series nodes and the prediction target P1. The calculation formula is as follows: E1 = MSE(0, Sigmoid(Abs(P1-Q)-0.5)*2.0))

[0043] Where MSE represents mean squared error, Sigmoid represents the S-curve, and Abs represents the absolute value.

[0044] The training error E1 based on the factual data is used to evaluate the difference between the predicted target and the actual output. It is processed using the mean square error (MSE) and the sigmoid function to make the error between 0 and 1.

[0045] S303. Input the next output C2 calculated by the control output network corresponding to the time series data of K+1 time series nodes into the next output neural network. After the target prediction network is processed by the integrated neural network, the predicted target P2 of the K+N time series node is output.

[0046] The predicted target P2 is the next control action predicted by the control output network under normal conditions based on the current state. The specific process is as follows: The desired target, such as ideal temperature or pressure, is input into the desired target neural network 102. The current time series data is input into the state time series network 101 of the control output network. The integrated neural network 103 of the control output network integrates the outputs of the state time series network 101 and the desired target neural network 102 to generate the control action C2. C2 is input into the next output neural network 112 in the target prediction network. The predicted target P2 is calculated again through the state time series network 111, the next output neural network 112, and the integrated neural network 113 of the target prediction network.

[0047] S304. Calculate the adversarial error E2 of the control output network based on the actual output Q of K+N time-series nodes and the predicted target P2. The adversarial error E2 is used to evaluate the robustness of the system under abnormal conditions. The calculation formula is as follows: E2 = MSE(0, Sigmoid(Abs(P2-Q)-0.5)*2.0)).

[0048] This calculation formula uses adversarial examples to evaluate the system's ability to cope with extreme conditions.

[0049] S305. Select a value C3 in the interval [C, x1] as the next output corresponding to the time series data of K+1 time series nodes, and input it into the next output neural network. After the integrated neural network of the target prediction network calculates, the predicted target P3 of the K+N time series node is output, where x1 represents the upper limit of the value of C3. Generally, after normalization, the value of x1 is 1.

[0050] The predicted target P3 is obtained by randomly selecting a value within a reasonable range based on the current control output and then calculating it through the target prediction network. The specific process is as follows: Based on the current control output, a random value within a reasonable range is selected and used as the new control output input into the target prediction network. The target prediction network obtains the predicted target P3 through the joint action of the target prediction network's state-sequential network 111, the next step output neural network 112, and the target prediction network's ensemble neural network 113.

[0051] S306. Based on the actual output Q of K+N time series nodes and the predicted target P3, calculate the empirical error E3 of the positive correlation between the control variable and the expected target. The calculation formula is as follows: E3=MSE(0,(Sigmoid((Abs(Q-P3)-(Q-P3)) / 2.0)-0.5)*2.0)

[0052] If the actual system control objective and control input are positively correlated, then the theoretical P3 must be larger than the actual value Q. If P3 > Q, then the loss is 0; otherwise, the loss E3 > 0.

[0053] S307. Select a value C4 within the interval [y1, C] and input it into the next output neural network. The ensemble neural network of the target prediction network calculates the predicted target P4 of the K+Nth time node, where y1 represents the lower limit of the value of C4. After normalization, the value of y1 is -1.

[0054] The predicted target P4 is another value randomly selected within a reasonable range based on the current control output, and then calculated by the target prediction network. The specific process is as follows: Based on the current control output, a random value within a reasonable range is selected, and the selected random value is used as the new control output input into the target prediction network. The target prediction network obtains the predicted target P4 through the joint action of the target prediction network's state-sequential network 111, the next step output neural network 112, and the target prediction network's ensemble neural network 113.

[0055] S308. Based on the actual output Q of K+N time series nodes and the predicted target P4, calculate the empirical error E4 of the inverse correlation between the control variable and the expected target. The calculation formula is as follows: E4=MSE(0,Sigmoid((Abs(P4-Q)-(P4-Q)) / 2.0)-0.5)*2.0)

[0056] The empirical error E3, which is positively correlated with the expected target, and the empirical error E4, which is negatively correlated with the expected target, are used to evaluate the system's handling of the positive and negative correlations between the control variables. By adjusting the input variables, the relationship between the prediction and the actual situation is observed to ensure that the system responds reasonably to changes.

[0057] S309. Using E1+E2+E3+E4 as the total error, the weights of the target prediction network are adjusted. Here, the backpropagation algorithm is used to minimize the total error, and the network weights are updated by gradient descent to improve the prediction ability of the system.

[0058] S104. Train the control output network based on the prediction results and time series data provided by the target prediction network, and optimize the control output network by integrating multiple error types. The purpose of training the control output network is to ensure the accuracy and reliability of the system operation. Specifically, it includes the following steps:

[0059] S401. Input the actual outputs Q of K+N time-series nodes into the state-series network of the control input network, and obtain the next output C2 through the integrated neural network of the control output network.

[0060] By utilizing a state-temporal network with the same state-temporal network structure as the target prediction network, focusing on training the desired target neural network can reduce computational burden and ensure the accuracy of control output.

[0061] S402. Calculate the adversarial error E5 of the control output network based on the actual output Q of K+N time-series nodes and the predicted target P2. The adversarial error E5 is used to evaluate the difference between the output and the actual requirement. The calculation formula is as follows: E5 = MSE(1, Sigmoid(Abs(P2-Q)-0.5)*2.0))

[0062] S403. Input 0 into the desired target neural network of the control output network. The next output C6 is obtained through the integrated neural network of the control output network. Then, input the next output C6 into the next output neural network of the target prediction network. The target prediction P6 is obtained through the integrated neural network of the target prediction network.

[0063] The predicted target P6 is a future control target predicted by a target prediction network based on the expected target value 0 being input into the expected target neural network 102 of the control output network. Specifically, it includes the following steps: inputting the expected target value 0 into the expected target neural network 102, which processes the expected target value and generates an expected target feature vector; inputting current control state data, such as temperature and pressure, into the state temporal network 101 of the control output network, extracting the state feature vector; integrating the feature vectors of the state temporal network 101 and the expected target neural network 102 of the control output network into the integrated neural network 103 of the control output network to generate a control action C6; and then inputting the control action C6 into the target prediction network to calculate the target prediction P6.

[0064] S404. Calculate the ideal expected error E6 based on the target prediction P6. The ideal expected error E6 is used to ensure that the control output achieves the expected target. The calculation formula is as follows: E6 = MSE(0, Sigmoid(Abs(P6)-0.5)*2.0)

[0065] S405. Select a value Q1 from the interval [Q, x2] and input it into the desired target neural network. The next output C7 is obtained by the integrated neural network that controls the output network. x2 represents the upper limit of the value of Q1. The value of x2 after normalization is generally 1.

[0066] S406. Calculate the empirical error E7, which is negatively correlated between the control variable and the next output, based on the next output C and the next output C7. The calculation formula is as follows: E7 = MSE(0, (Sigmoid((Abs(C-C7)-(C-C7)) / 2.0)-0.5)*2.0)

[0067] S407. Select a value Q2 from the interval [y2, Q] and input it into the desired target neural network. The next output C8 is obtained by the integrated neural network that controls the output network. After normalization, the value of Y2 is generally -1.

[0068] S408. Calculate the empirical error E8, which is negatively correlated between the control variable and the next output, based on the next output C and C8. The calculation formula is as follows: E8 = MSE(0, (Sigmoid((Abs(C8-C)-(C8-C)) / 2.0)-0.5)*2.0)

[0069] The empirical error E7, which is positively correlated with the control variable and the next output, and the empirical error E8, which is negatively correlated with the control variable and the next output, are used to ensure that the control output achieves the expected goal, adjust the control output, observe the relationship between the control output and the actual system state, and adjust the error to optimize the control effect.

[0070] S409. Using E5+E6+E7+E8 as the total error, the weights of the control output network are adjusted. Here, the network parameters are adjusted through backpropagation to improve the control effect and ensure the decision-making ability of the control output network.

[0071] S105. Apply the trained control output network and target prediction network to industrial process control, and continuously supplement the training according to environmental changes to ensure the stability and adaptability of the system in the real environment. This includes the following steps:

[0072] S501. Apply the trained network to actual production, generate control commands in real time, and adjust and optimize the control strategy based on feedback from the production environment.

[0073] S502. Continuously collect new operational data, update the training dataset, and regularly retrain the system to adapt to new data and environmental changes, thereby maintaining continuous optimization of system performance.

[0074] S503. Based on real-time feedback during the production process, adjust the control strategy in a timely manner, and use the incremental learning mechanism to continuously update and optimize the model, thereby improving the robustness and stability of the system.

[0075] This invention also provides an industrial process control system based on a desired setpoint error neural network, used to implement the above-described industrial process control method based on a desired setpoint error neural network, comprising:

[0076] The control output network 100 is used to output subsequent control actions based on the current control state and the desired target, so as to ensure the accuracy and reliability of the system operation. The control output network 100 includes a state timing network 101 of the control output network, a desired target neural network 102, and an integrated neural network 103 of the control output network.

[0077] The state-series network 101 of the control output network is used to learn and extract features from the time-series data of the current control state. A Long Short-Term Memory (LSTM) network or a gated recurrent unit (GRU) is used to process the time-series data. The Tanh activation function is used. The number of input and output neurons is determined according to the actual industrial control requirements. The number of hidden layer neurons is determined by an automated parameter tuning algorithm, such as Bayesian hyperparameter tuning.

[0078] The desired target neural network 102 is used to calculate the features of the desired control target. It uses a fully connected neural network (FCNN) with the Tanh activation function. The number of input and output neurons is the same as the number of control targets. If temperature and pressure need to be controlled simultaneously, the number of input and output neurons is 2.

[0079] An integrated neural network 103 for the control output network is used to integrate the outputs of the state-temporal network of the control output network and the desired target neural network to generate the next control action. It adopts FCNN and uses the Tanh activation function. The number of input neurons is the same as the number of output neurons of the state-temporal network 101 of the control output network. The number of hidden layer neurons is finally determined according to the overall network automatic parameter tuning algorithm, such as Bayesian hyperparameter tuning. The number of output neurons is determined according to the controlled object. For example, if the temperature and pressure are controlled by controlling one cold water valve and one high-pressure gas valve, then the number of output neurons is 2.

[0080] The target prediction network 110 is used to predict the control target based on the current control state and the next control action. The target prediction network includes a state-sequential network 111, a next output neural network 112, and an integrated neural network 113.

[0081] The state-time network 111 of the target prediction network is used to learn and extract features from the time-series data of the current control state. It shares the same architecture as the state-time network 101 of the control output network, adopts LSTM or GRU, uses the Tanh activation function, and the number of input and output neurons is determined according to actual needs. The number of hidden layer neurons is determined by an automated parameter tuning algorithm.

[0082] The next step is to output neural network 112, which is used to calculate the features of the next control action; FCNN is used, with the Tanh activation function, and the number of input and output neurons is the same as the number of controlled objects.

[0083] The ensemble neural network 113 of the target prediction network integrates the output of the state-temporal network of the target prediction network with the output of the next output neural network to generate the predicted target. It adopts FCNN and uses the Tanh activation function. The number of input neurons is the same as the number of output neurons of the state-temporal network 111 of the target prediction network. The number of hidden layer neurons is finally determined according to the overall network automatic parameter tuning algorithm, such as Bayesian hyperparameter tuning. The number of output neurons is determined according to the control target. If temperature and pressure need to be controlled simultaneously, the number of input and output neurons is 2.

[0084] In practical applications, such as a project that requires controlling the temperature during a polymerization reaction by controlling the flow rate of cold water, the timing characteristics of this system are selected as liquid level, water flow rate, and temperature, and the control output is the opening degree of the cold water valve.

[0085] After data normalization, the time-series data corresponding to the first 100 time-series nodes are used as input to the time-series state network. A temperature of 130℃ is used as training data for the desired and predicted targets. The desired target is defined as the ideal state that the system wants to achieve, which guides the output of the control network during training. The opening degree of the cold water valve is set to 101% and used as training data for the next output neural network. Through this data, the system can learn how to adjust the opening degree of the cold water valve under a specific temperature target. The industrial process control method based on the desired setting error neural network provided by this invention is used for training and control.

[0086] The target prediction network is trained to simulate complex multivariate control processes in actual production. By learning the relationships between different state characteristics, the model can predict future temperature changes.

[0087] The goal of the control output network is to generate corresponding control commands (such as adjusting the opening of the cold water valve in the next step) based on the expected temperature changes in the future (e.g., the setpoint error is 0). The output of this network is the control command used directly in production.

[0088] After training, the target prediction network can ultimately reflect the complex model of multivariate control in actual production, while the control output network can output the control operation to be executed next based on the expected target after 30 steps, such as the set value error being 0.

[0089] In actual production, only the control output network is used. However, since the system changes gradually over time and with the environment during production, incremental training is required. The purpose of incremental training is to enable the system to adapt to new production conditions and ensure its stability and effectiveness in actual production. During incremental training, both networks need to be updated and trained together.

[0090] [Detailed Rules 91, 25.11.2025] As shown in Figure 3, the horizontal axis represents the index of the time series, and the vertical axis represents the deviation of the control result. By using historical data simulation, the control method of the present invention can effectively control the temperature error to around 0. The line with large fluctuations in the figure represents the temperature error of the production PID control, and the line with small fluctuations represents the temperature error controlled by the present invention.

[0091] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An industrial process control method based on a neural network for desired setting error, characterized in that, include: Establish a control output network model and a target prediction network model, determine the specific control target and corresponding control object in the industrial process, select control variables according to the control target, and set the corresponding expected values. The control output network includes the state-time network of the control output network, the expected target neural network, and the integrated neural network of the control output network. The target prediction network includes the state-time network of the target prediction network, the next step output neural network, and the integrated neural network of the target prediction network. Collect time-series data from industrial processes and perform data processing on the collected time-series data; The target prediction network is trained using the processed time-series data, the prediction results are obtained, and multiple error types are calculated and integrated to optimize the target prediction network; The control output network is trained based on the prediction results and time series data provided by the target prediction network, and the control output network is optimized by integrating multiple error types. The trained control output network and target prediction network are applied to industrial process control, and continuous additional training is conducted based on environmental changes.

2. The industrial process control method based on a neural network for desired setting error according to claim 1, characterized in that, The target prediction network is trained using time-series data, and the prediction results are obtained. Then, multiple error types are calculated and integrated to optimize the target prediction network, including: The time series data corresponding to the Kth time series node is input into the state time series network of the target prediction network. The next output C calculated by the control output network corresponding to the time series data of K+1 time series nodes is input into the next output neural network. After the ensemble neural network of the target prediction network is calculated, the predicted target P1 of the K+Nth time series node is output. The training error E1 based on the actual output Q of K+N time series nodes and the prediction target P1 is calculated using the following formula: E1=MSE(0,Sigmoid(Abs(P1-Q)-0.5)*2.0)) Where MSE represents mean squared error, Sigmoid represents the S-curve, and Abs represents the absolute value; The next output C2 calculated by the control output network corresponding to the time series data of K+1 time series nodes is input into the next output neural network. After the target prediction network is processed by the integrated neural network, the predicted target P2 of the K+N time series node is output. The adversarial error E2 of the control output network is calculated based on the actual output Q of K+N time-series nodes and the predicted target P2, using the following formula: E2=MSE(0,Sigmoid(Abs(P2-Q)-0.5)*2.0)) Select a value C3 in the interval [C, x1] as the next output corresponding to the time series data of K+1 time series nodes, and input it into the next output neural network. After the integrated neural network of the target prediction network is calculated, the predicted target P3 of the K+N time series node is output, where x1 represents the upper limit of the value of C3. Based on the actual output Q of K+N time series nodes and the predicted target P3, the empirical error E3, which is positively correlated with the expected target and the control variable, is calculated using the following formula: E3=MSE(0,(Sigmoid((Abs(Q-P3)-(Q-P3)) / 2.0)-0.5)*2.0) A value C4 is selected within the interval [y1, C] and input into the next output neural network. The ensemble neural network of the target prediction network calculates the predicted target P4 of the K+Nth time node, where y1 represents the lower limit of the value of C4. Based on the actual output Q of K+N time series nodes and the predicted target P4, the empirical error E4, which is inversely correlated with the expected target, is calculated using the following formula: E4=MSE(0,Sigmoid((Abs(P4-Q)-(P4-Q)) / 2.0)-0.5)*2.0) The weights of the target prediction network are adjusted using E1+E2+E3+E4 as the total error.

3. The industrial process control method based on a neural network for desired setting error according to claim 2, characterized in that, The control output network is trained based on the prediction results provided by the target prediction network and time series data, and then optimized by integrating multiple error types, including: Set the state-time network of the control output network to be the same as the state-time network of the target prediction network. Input the actual output Q of K+N time nodes into the state-time network of the control input network. The next output C2 is obtained by the integrated neural network of the control output network. The adversarial error E5 of the control output network is calculated based on the actual output Q of K+N time-series nodes and the predicted target P2, using the following formula: E5=MSE(1,Sigmoid(Abs(P2-Q)-0.5)*2.0)) 0 is input into the desired target neural network of the control output network. The next output C6 is calculated by the integrated neural network of the control output network. The next output C6 is then input into the next output neural network of the target prediction network. The target prediction P6 is calculated by the integrated neural network of the target prediction network. The ideal expected error E6 is calculated based on the target prediction P6, using the following formula: E6=MSE(0,Sigmoid(Abs(P6)-0.5)*2.0) A value Q1 is selected from the interval [Q, x2] and input into the desired target neural network. The next output C7 is obtained by the integrated neural network that controls the output network, where x2 represents the upper limit of the value of Q1. Calculate the empirical error E7, which is positively correlated with the next output, based on the next output C and the next output C7. The calculation formula is as follows: E7=MSE(0,(Sigmoid((Abs(C-C7)-(C-C7))) / 2.0)-0.5)*2.0) A value Q2 is selected from the interval [y2, Q] and input into the desired target neural network. The next output C8 is obtained by the integrated neural network that controls the output network, where y2 represents the upper limit of the value of Q2. Calculate the empirical error E8, which is negatively correlated between the control variable and the next output, based on the next output C and the next output C8. The calculation formula is as follows: E8=MSE(0,(Sigmoid((Abs(C8-C)-(C8-C)) / 2.0)-0.5)*2.0) The weights of the control output network are adjusted using E5+E6+E7+E8 as the total error.

4. The industrial process control method based on a neural network for desired setting error according to claim 1, characterized in that, Time series data includes historical data and real-time data. Data processing of collected time series data includes feature expansion, feature selection and normalization.

5. An industrial process control system based on a desired setpoint error neural network, used to implement the industrial process control method based on a desired setpoint error neural network as described in any one of claims 1-4, characterized in that, include: The control output network is used to output subsequent control actions based on the current control state and the desired target. The control output network includes the state-temporal network of the control output network, the desired target neural network, and the integrated neural network of the control output network. The state-time network of the control output network is used to learn and extract features from the time-series data of the current control state; A target neural network is used to compute the features of the desired control target. An integrated neural network for controlling the output network is used to integrate the outputs of the state-temporal network of the control output network and the desired target neural network to generate the next control action; The target prediction network is used to predict the control target based on the current control state and the next control action. The target prediction network includes a state-sequential network, a next output neural network, and an ensemble neural network of the target prediction network. The target prediction network is a state-series network used to learn and extract features from the time-series data of the current control state; The next step is to output a neural network, which is used to calculate the features of the next control action; An ensemble neural network for target prediction is used to integrate the state-sequence network of the target prediction network with the output of the next-step output neural network to generate the predicted target.

Images