Key production index prediction method based on online learning integrated random configuration network
By integrating randomized networks through online learning, the problem of real-time and accurate prediction of key production indicators in complex industrial processes was solved, achieving efficient and stable prediction results.
Patent Information
- Application Number
- CN202511690309.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-18
- Publication Date
- 2026-02-13
AI Technical Summary
Key production indicators in complex industrial processes are difficult to predict accurately in real time. Existing mechanistic models are inaccurate and costly, and traditional ensemble learning methods cannot be updated in real time and have high computational costs.
An ensemble stochastic configuration network based on online learning is adopted. The model is constructed and updated through negative correlation learning and SOR iteration. Combined with the online learning strategy, the ensemble stochastic configuration network is trained using sensor data for prediction.
It improves the prediction accuracy of key production indicators and the stability of the model, reduces computing costs, and enables real-time response to changes in operating conditions.
Smart Images
Figure CN121525967A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of production indicator prediction technology, and specifically to a method for predicting key production indicators based on an online learning ensemble randomized network. Background Technology
[0002] Complex industrial production systems (including industries such as steel, non-ferrous metallurgy, and petrochemicals) are important pillars of social development, but they are constrained by high energy consumption, severe environmental pollution, and resource scarcity. Production scheduling optimization and process optimization are key to solving these problems, and timely detection of key production indicators (important variables characterizing production status) in industrial processes is crucial for controlling industrial processes, production scheduling, and optimizing equipment operation. However, complex industrial processes exhibit time lags, and the measurement of some key production indicators is time-consuming or even impossible to detect directly. Therefore, establishing predictive models and making scientific predictions is essential.
[0003] Traditional industrial process key production indicator prediction often employs mechanistic models based on kinetics, material and energy balance. However, complex industrial production processes involve the material transformation of primary raw materials into finished products, and these transformations are complex and their mechanisms are often unclear. Therefore, mechanistic models for industrial process key production indicators are frequently simplified, resulting in low model accuracy and parameter mismatch. Establishing rigorous mechanistic models also suffers from drawbacks such as high cost, long development cycles, and significant difficulty, and their reliability is difficult to guarantee.
[0004] With the development of computer and information technology, a large amount of process variable data related to key production indicators in industrial processes can now be collected at industrial sites. Therefore, data-driven methods for predicting key production indicators in industrial processes have been widely applied. Based on the number of individual models included in the prediction model, key production indicator prediction models can be divided into single models and ensemble models. Single prediction models for key production indicators in industrial processes typically suffer from parameter sensitivity and performance instability; ensemble models, on the other hand, can improve prediction accuracy and stability and are frequently used. However, classic ensemble learning methods (such as bagging and boosting) only obtain model diversity through different training data, without considering the interaction between the basic models; furthermore, updating ensemble models requires retraining with historical and new data, resulting in high computational costs, poor real-time performance, and an inability to update in real time according to changes in the working environment. Summary of the Invention
[0005] In view of this, the present invention provides a method for predicting key production indicators based on an online learning ensemble of randomized networks.
[0006] The technical solution of this invention is:
[0007] A method for predicting key production indicators based on online learning ensemble randomized networks includes the following steps:
[0008] S1. Collect data related to key production indicators in complex industrial production processes;
[0009] S2. Preprocess the collected data and construct the training dataset;
[0010] S3. Train the ensemble randomized network using the training dataset and update the ensemble randomized network online to obtain a key production indicator prediction model based on the ensemble randomized network.
[0011] S4. Deploy the obtained integrated stochastic configuration network prediction model for predicting key production indicators in complex industrial processes.
[0012] Furthermore, according to the key production indicator prediction method, the data related to key production indicators in the complex industrial production process described in S1 includes the measurement data of each sensor and key production indicator information throughout the entire complex industrial production process.
[0013] Furthermore, based on the aforementioned key production indicator prediction method, the process of preprocessing the collected data to construct the training dataset in S2 is as follows:
[0014] First, samples with NaN values are removed from the data. Then, a linear interpolation algorithm is used to pad the missing data, ensuring that the processed data has the same length. Next, the preprocessed data is aligned and normalized to obtain the industrial production process data. Accordingly, the training dataset is finally obtained. ,in, For the sample size, For the sample Key production metrics labels.
[0015] Furthermore, according to the aforementioned key production indicator prediction method, S3 specifically includes the following steps:
[0016] S31. Construct an ensemble randomized network by ensemble the K basic randomized networks with average weighting, and input the data from the training dataset into the ensemble randomized network.
[0017] S32. Constructing an ensemble stochastically configured network based on a cost function for negative correlation learning. ;
[0018] S33. Solve for the output weights of the ensemble stochastic network using the SOR iterative method;
[0019] S34. When new collected data is added to the training set, the integrated randomized network model is updated using an online learning strategy.
[0020] Furthermore, according to the aforementioned key production indicator prediction method, the integrated random configuration network is as follows:
[0021]
[0022] in, Data for industrial production processes; This represents the k-th basic randomized network; , and Let represent the output weight, input weight, and threshold of the l-th node in the k-th basic randomized network, respectively. This represents the number of hidden layer nodes in a basic randomized network. Indicates transpose; This indicates the integration of a randomly configured network; and These are the output weights of the k-th basic random configuration network and the output weights of the ensemble random configuration network, respectively.
[0023] Furthermore, according to the key production indicator prediction method described above, the cost function of the integrated randomized network based on regularized negative correlation learning in S32 is:
[0024]
[0025] in, This represents the coefficient of the negative correlation penalty term.
[0026] Furthermore, according to the aforementioned key production indicator prediction method, the output weights of the integrated stochastic network obtained by using the SOR iterative method in S33 are:
[0027]
[0028] Where q is the number of iterations of SOR; Let be the relaxation factor and 0 < ω < 2; defined To integrate the hidden layer correlation matrix of a randomly configured network, for diagonal matrix, and They are respectively Strict lower triangular and strict upper triangular matrices.
[0029] Furthermore, according to the key production indicator prediction method, updating the ensemble stochastic network using the online learning strategy in S34 is as follows: it only requires using the newly collected data added to the training dataset to obtain the corresponding hidden layer correlation matrix of the ensemble stochastic network. and target matrix Then, correlated with the historical hidden layer matrix. and target matrix Summing yields the hidden layer correlation matrix for all data. and target matrix Then, the global output weights of the integrated random configuration network are obtained by using the SOR iterative method in S33, thereby realizing the update of the integrated random configuration network.
[0030] Compared with the prior art, the present invention has the following beneficial effects:
[0031] This invention addresses the challenges of diversity and collaboration among individual stochastic configuration networks (SCNs) in ensemble stochastic configuration networks (SCNs). It employs a negative correlation learning method to integrate these individual SCNs for processing sequence data in complex industrial production processes. Furthermore, to overcome the limitations of offline learning in handling industrial sequence data, an online learning strategy is used to learn from new data while simultaneously updating the ensemble SCN, improving the model's adaptability. Additionally, the successive overrelaxation (SOR) iterative method is used instead of the least squares method to calculate the output weights of the ensemble SCN, reducing computational costs and improving model stability, thus enhancing the predictive model's practical application value. Attached Figure Description
[0032] Figure 1 This is a flowchart illustrating the key production indicator prediction method based on online learning integrated random configuration networks in this embodiment.
[0033] Figure 2 This is a block diagram of the algorithm structure for the key production indicator prediction method based on online learning integrated random configuration networks in this embodiment. Detailed Implementation
[0034] To facilitate understanding of this application, a more comprehensive description of this application will be provided below with reference to the accompanying drawings.
[0035] Figure 1 This is a flowchart illustrating the key production indicator prediction method based on online learning ensemble randomized network in this embodiment. Figure 2 This is a block diagram of the algorithm structure for the key production indicator prediction method based on an online learning ensemble randomized network in this embodiment. Figure 1 and Figure 2 As shown, the key production indicator prediction method based on online learning ensemble stochastic configuration network (ACM) uses collected process variable data to train an ACM to predict key production indicators. Specifically, it includes the following steps:
[0036] S1. Utilize sensors deployed in actual industrial production processes to collect data related to key production indicators in complex industrial production processes, including measurement data and key production indicator information from various sensors throughout the entire complex industrial production process.
[0037] S2. Preprocess the collected data and construct the training dataset;
[0038] This implementation preprocesses the collected data as follows: First, samples with NaN values are removed; then, a linear interpolation algorithm is used to interpolate and pad the missing data, ensuring that the processed data have the same length; finally, the preprocessed data is aligned, and the aligned data is normalized to obtain industrial production process data. Accordingly, the training dataset is finally obtained. ,in, For the sample size, For the sample Key production metrics labels.
[0039] S3. Train the ensemble randomized network using the training dataset and update the ensemble randomized network online to obtain a key production indicator prediction model based on the ensemble randomized network.
[0040] S31. Construct an ensemble randomized network by ensemble the K basic randomized networks with average weighting, and input the data from the training dataset into the ensemble randomized network.
[0041]
[0042] in, Data for industrial production processes; This represents the k-th basic randomized network. , and Let the output weight, input weight, and threshold of the l-th node in the k-th basic randomized network be represented respectively. This represents the number of hidden layer nodes in a basic randomized network. Indicates transpose. This indicates the integration of a randomly configured network. and These are the output weights of the k-th basic random configuration network and the output weights of the ensemble random configuration network, respectively.
[0043] S32. To balance the diversity of individual randomized networks, the complexity of ensemble randomized networks, and the mutual cooperation among individual randomized networks, a cost function based on negative correlation learning is constructed for ensemble randomized networks. as follows:
[0044]
[0045] in, This represents the coefficient of the negative correlation penalty term.
[0046] S33. The output weights of the ensemble stochastic network are solved using a fast and stable successive overrelaxation (SOR) iterative method, as shown below:
[0047] Define the correlation matrix of the hidden layer of the integrated random configuration network as follows. and target matrix :
[0048] ,
[0049] in, , , , ( Let be the output matrix of the hidden layer of the k-th basic randomized network:
[0050]
[0051] The iterative solution for the output weights of the ensemble stochastically configured network obtained using the SOR method is shown below:
[0052]
[0053] Where q is the number of iterations of the SOR, and ω is the relaxation factor, where 0 < ω < 2. for diagonal matrix, and They are respectively The strictly lower triangular and strictly upper triangular matrices are shown below:
[0054]
[0055]
[0056]
[0057] S34. When new collected data is added to the training dataset, the integrated randomized network is updated using an online learning strategy.
[0058] For new data, the online learning strategy only needs to use that data to obtain the hidden layer correlation matrix of the corresponding ensemble stochastic network. and target matrix Then, correlated with the historical hidden layer matrix. and target matrix Summing yields the hidden layer correlation matrix for all data. and target matrix .
[0059] Integrated random configuration network hidden layer correlation matrix The specific update method is shown in the following formula:
[0060]
[0061] in, and represents the output matrix of the hidden layer of the k-th basic randomized network on historical data and new data, respectively.
[0062] Target matrix The specific update method is shown in the following formula:
[0063]
[0064] in, and These are label matrices representing the original data and the newly added data, respectively.
[0065] After adding new data, recalculate the matrix. Then Decompose into matrices , and Finally, the output weights of the integrated randomized network are updated using the SOR iterative method.
[0066] S4. Deploy the key production indicator prediction model based on the integrated random configuration network to the server of the actual complex industrial production process for the prediction of key production indicators in the complex industrial production process.
[0067] This implementation deploys a key production indicator prediction model based on an integrated randomized network on a server within the actual industrial production process. After actual production begins, firstly, step S1 is executed to collect data related to the key production indicators; then, step S2 is executed to preprocess the data; finally, the preprocessed data is... The input is fed into a key performance indicator (KPI) prediction model based on an ensemble randomized network to predict the final KPI values. As shown below:
[0068]
[0069] Example
[0070] This embodiment applies the method of the present invention to the prediction of crude oil residual carbon content during refinery production, including the following steps:
[0071] First, the nuclear magnetic resonance (NMR) system deployed in the refinery production process is used to collect NMR spectrum data of crude oil during the refinery production process. This includes converting the NMR spectrum into 700-dimensional data slices through a computer system, as well as crude oil residual carbon content labeling information obtained through laboratory testing.
[0072] Then, the collected NMR spectrum data was cleaned and preprocessed to construct a training dataset: the preprocessing of the collected NMR spectrum data involved removing samples with NaN values and interpolating missing data; the preprocessed NMR spectrum data was then aligned, and the aligned data was normalized to obtain the NMR spectrum data. Accordingly, the training dataset can eventually be obtained. ,in, For the sample size, Nuclear magnetic resonance data samples The label indicates the crude oil residual carbon content.
[0073] Then, the integrated randomized network was trained and updated online using the training dataset to obtain a crude oil residual carbon content prediction model based on the integrated randomized network.
[0074] Finally, the crude oil residual carbon content prediction model based on an integrated random configuration network is deployed on the refinery's production process server for predicting crude oil residual carbon content during the refinery's production process.
[0075] In this example, the root mean square error (RMSE) and mean absolute percentage error (MAPE) are used to evaluate the performance of the crude oil residual carbon content prediction model based on an ensemble randomized network. Furthermore, to more fully illustrate the effectiveness of the method of this invention, the prediction results of this method are compared with the prediction results of other common ensemble neural network algorithms, and the comparison results are shown in Table 1.
[0076] Table 1. Comparison of prediction results between the method of this invention and common algorithms.
[0077] As can be seen from Table 1, the method of the present invention has the lowest RMSE and MAPE, indicating that the method of the present invention has higher accuracy.
[0078] It should be understood that, inspired by the inventive concept, those skilled in the art can make various improvements or modifications based on the above without departing from the scope of the invention, and these modifications still fall within the protection scope of the invention.
Claims
1. A method for predicting key production indicators based on online learning ensemble stochastic configuration networks, characterized in that: Includes the following steps: S1. Collect data related to key production indicators in complex industrial production processes; S2. Preprocess the collected data and construct the training dataset; S3. Train the ensemble randomized network using the training dataset and update the ensemble randomized network online to obtain a key production indicator prediction model based on the ensemble randomized network. S4. Deploy the obtained integrated stochastic configuration network prediction model for predicting key production indicators in complex industrial processes.
2. The key production indicator prediction method according to claim 1, characterized in that: The data related to key production indicators in the complex industrial production process described in S1 include measurement data from various sensors and key production indicator information throughout the entire complex industrial production process.
3. The key production indicator prediction method according to claim 1, characterized in that: The process of preprocessing the collected data to construct the training dataset, as described in S2, is as follows: First, samples with NaN values are removed from the data. Then, a linear interpolation algorithm is used to pad the missing data, ensuring that the processed data has the same length. Next, the preprocessed data is aligned and normalized to obtain the industrial production process data. ; Accordingly, the training dataset is finally obtained. ,in, For the sample size, For the sample Key production metrics labels.
4. The key production indicator prediction method according to claim 1, characterized in that: S3 specifically includes the following steps: S31. Construct an ensemble randomized network by ensemble the K basic randomized networks with average weighting, and input the data from the training dataset into the ensemble randomized network. S32. Constructing an ensemble stochastically configured network based on a cost function for negative correlation learning. ; S33. Solve for the output weights of the ensemble stochastic network using the SOR iterative method; S34. When new collected data is added to the training set, the integrated randomized network model is updated using an online learning strategy.
5. The key production indicator prediction method according to claim 4, characterized in that: The integrated random configuration network is: ； in, Data for industrial production processes; This represents the k-th basic randomized network; , and Let represent the output weight, input weight, and threshold of the l-th node in the k-th basic randomized network, respectively. This represents the number of hidden layer nodes in a basic randomized network. Indicates transpose; This indicates the integration of a randomly configured network; and These are the output weights of the k-th basic random configuration network and the output weights of the ensemble random configuration network, respectively.
6. The key production indicator prediction method according to claim 5, characterized in that: The cost function of the integrated randomized network described in S32 based on regularized negative correlation learning is: ； in, This represents the coefficient of the negative correlation penalty term.
7. The key production indicator prediction method according to claim 6, characterized in that: The output weights of the ensemble stochastically configured network obtained using the SOR iterative method described in S33 are: ； Where q is the number of iterations of SOR; Let be the relaxation factor and 0 < ω < 2; defined To integrate the hidden layer correlation matrix of a randomly configured network, for diagonal matrix, and They are respectively Strict lower triangular and strict upper triangular matrices.
8. The key production indicator prediction method according to claim 7, characterized in that: The online learning strategy described in S34 for updating the ensemble stochastic network is as follows: it only requires using the newly acquired data added to the training dataset to obtain the corresponding hidden layer correlation matrix of the ensemble stochastic network. and target matrix Then, correlated with the historical hidden layer matrix. and target matrix Summing yields the hidden layer correlation matrix for all data. and target matrix Then, the global output weights of the integrated random configuration network are obtained by using the SOR iterative method in S33, thereby realizing the update of the integrated random configuration network.