A data-driven modeling method based on DCS logic autonomous exploration
By building a dictionary library in the DCS system, the machine learning model is seamlessly integrated with the DCS algorithm block, solving the problem of model deployment in the DCS system and realizing direct model execution and efficient prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANDONG DAOHE IOT TECH CO LTD
- Filing Date
- 2026-02-11
- Publication Date
- 2026-06-02
AI Technical Summary
Existing technologies make it difficult to directly convert neural network prediction models into DCS logic execution structures, and traditional mathematical modeling methods require manual selection of function structures, which cannot adapt to complex systems, making them difficult to deploy and use in DCS systems.
By constructing a dictionary library, the mathematical logic of the DCS algorithm block is encapsulated into basic units. Machine learning algorithms are used to train and explore the mathematical model structure of influencing factors and temperature changes, and then deployed directly in the DCS system to achieve seamless integration of the model and the system.
It enables the direct deployment and execution of machine learning models in DCS systems, improving the flexibility of model adaptation to operating conditions and predictive capabilities, while reducing system integration complexity and failure risks.
Smart Images
Figure CN122133472A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the technical field of engineering control, and more specifically, relates to a data-driven modeling method based on DCS logic autonomous exploration. Background Technology
[0002] In large industrial equipment (such as thermal power units, chemical reaction plants, and smelting plants), process control systems typically rely on DCS (Distributed Control System) to perform functions such as measurement, calculation, logical judgment, and sequential control logic execution. Operators construct logic using logic blocks, arithmetic blocks, limiting blocks, and comparison blocks to achieve complex control objectives.
[0003] Chinese patent document CN118859888A discloses a DCS artificial intelligence module and its optimization method. This model can output target variable parameters as an applied machine learning model, and can also perform machine learning training. During the training process, the input timestamp and the feature parameters and target variable parameters corresponding to the specified time after it are input, enabling the incremental learning model to have predictive capabilities on the time axis. Weight coefficients are set for the target variable, and an optimized variance algorithm is used to evaluate the model and calculate the Gaussian distribution shift of the target variable parameter error, providing a precise basis for model fine-tuning. The DCS distributed control system is combined with artificial intelligence to perform collaborative control, improving the control optimization capability of the distributed control system. The optimized training / application integrated incremental learning model is integrated into the DCS.
[0004] Industrial systems generally have strong coupling characteristics. For example, boiler combustion simultaneously affects steam temperature, flue gas composition, and thermal deviation. Operating conditions change frequently, making it difficult for traditional static logic to adapt correctly. The ever-increasing amount of data already contains a large amount of information that can be used for "self-learning" logic patterns, but traditional DCS cannot use data to automatically generate logic.
[0005] Therefore, people began to explore data-driven modeling methods, such as deep learning-based prediction methods, data regression models, and black-box models. However, these methods have problems: the model structure is inconsistent with the semantics of DCS engineering, making it difficult to deploy directly in the field; most models cannot be decomposed into executable logical block expressions; and the models lack interpretability, making them difficult to use in industrial control systems.
[0006] Therefore, there is an urgent need for a method that enables machine learning models to "autonomously explore" mathematical expressions that conform to the DCS pattern within the DCS logical architecture and can be directly deployed and executed on the DCS. Summary of the Invention
[0007] The present invention aims to overcome at least one of the defects of the prior art and provide a data-driven modeling method based on DCS logic autonomous exploration. This method addresses the problems that although neural network prediction models can make predictions, they cannot be directly converted into DCS logic execution structures; and that current mathematical modeling methods require manual selection of function structures, which cannot adapt to complex systems and are often disconnected from DCS structures, thus hindering engineering implementation.
[0008] The detailed technical solution of this invention is as follows: A data-driven modeling method based on DCS logic autonomous exploration, the method comprising: S1. Collect real-time data from the DCS system or historical database at the industrial site and preprocess it. The real-time data includes various influencing factors and temperature changes. Then calculate the time delay relationship of each influencing factor under different operating conditions.
[0009] S2. Construct a dictionary library, encapsulating the mathematical logic of the basic algorithm blocks of the DCS algorithm block function library into basic units in the dictionary library, which serve as the basic units for building the required model. Addressing the industry pain point of poor compatibility between traditional machine learning models and DCS systems, requiring additional interface development, this solution designs a configurable dictionary library module based on DCS algorithm blocks. This enables native, seamless integration of models with the DCS system, eliminating the need to store and access models as external files. Instead, models can be directly built and deployed using the basic units in the dictionary library through the DCS system's algorithm block configuration function. This achieves deep integration between the model and the DCS system, significantly reducing system integration complexity and failure risks. Furthermore, the dictionary library supports configurability, allowing users to adjust the parameter mapping rules of the basic algorithm blocks according to different operating conditions through the DCS visual interface, enhancing the model's flexibility in adapting to various operating conditions.
[0010] S3. Use machine learning algorithms to train and explore the mathematical model structure of each influencing factor and temperature change that matches the dictionary database; S4. Deployment is performed based on the DCS system and the generated mathematical model structure.
[0011] Furthermore, the preprocessing includes removing null values, filling in missing values, and filtering: Null value removal: The deletion method is used to remove null data caused by sensor failure or communication interruption during the acquisition process, ensuring the integrity of each time series data; Missing values are imputed: For consecutively missing data (i.e., missing duration ≤ 5 sampling periods), linear interpolation is used for imputation, expressed as: (1); In formula (1), For the k-th data point corresponding to the missing time, For the sampling time of the k-th data point, For the (k+1)th data, For the (k-1)th data, For the sampling time of the (k+1)th data point, This refers to the sampling time of the (k-1)th data point; Filtering: Median filtering algorithm is used to eliminate instantaneous pulse interference. The window size is set to m1~m2 sampling periods. Data fluctuations are smoothed by sorting and taking the median value to ensure the stability and reliability of the data.
[0012] The preprocessed data is time-series aligned to obtain the training dataset, ensuring that all parameters match on the same time dimension, laying the foundation for subsequent time delay relationship calculations.
[0013] Furthermore, the calculation of the time delay relationships corresponding to each influencing factor under different operating conditions specifically includes: ① Training dataset selection: From the preprocessed data, stratified sampling is performed according to different working conditions divided into different load ranges to ensure that the dataset includes time series data of each typical working condition, while removing extreme abnormal working condition data to avoid interfering with model training. ② Algorithm selection and model training: A combination of machine learning algorithms with strong time-series adaptability is adopted, with the dynamic time warping algorithm being the first choice for preliminary time delay matching to quickly locate the time delay range of each influencing factor on temperature changes; Then, the Long Short-Term Memory (LSTM) network is used for precise calculations. Specifically, the time lag range of each influencing factor on temperature change is used as the limiting interval of the LSTM network. The time memory capability of the LSTM network model is used to capture the dynamic change pattern of the time lag relationship. ③ Differentiated time delay relationship output: The model is trained using the training dataset, and the time delay relationship of each influencing factor under different operating conditions is output. Finally, it is presented in the form of time delay piecewise linear functions for different power ranges. This function can intuitively reflect the time delay duration change law of a certain influencing factor under different operating condition parameters, and provide core parameter support for the operating condition adaptation of the subsequent temperature prediction model.
[0014] Furthermore, S3 specifically includes: S31. Time Delay Data Adaptation Processing: Based on the time delay duration of each influencing factor under different load segments, the preprocessed time series data is shifted according to the corresponding time delay duration of different load segments, so that the data of each influencing factor and the corresponding temperature response data are accurately aligned in the time dimension. S32. Change Extraction: For the shifted data of each influencing factor, perform difference calculations based on the corresponding time lag duration to obtain the change of each influencing factor within a time period. (2); In formula (2), This refers to the change in time delay duration corresponding to the first load segment interval of the influencing factors. This represents the value of the influencing factor at time T. This refers to the time delay duration corresponding to this influencing factor in the first load segment interval; The changes in each influencing factor across different load segments are then multiplied by the weighting coefficients of each load interval constructed using the Gaussian function to obtain the reconstructed changes. ; (3); (4); In formulas (3)-(4), is the weighting coefficient, i.e. the model reconstruction coefficient; x is the current input load; a is the midpoint of the load divided into each load interval; and b is an adjustable parameter.
[0015] S33. Model Training and Construction: The reconstructed changes of each influencing factor are used as the model input, and the temperature change is used as the model output. The model is trained using Bayesian methods and the DCS system function library to finally obtain the model parameters of the mathematical model for predicting the temperature of desuperheated water.
[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention provides a data-driven modeling method based on DCS logic autonomous exploration. It uses machine learning methods to automatically convert input data into mathematical logic expressions, which are interpretable and convenient for review and engineering analysis. At the same time, when processing data, the time delay time under different working conditions can be calculated for the influencing factors of time delay relationship, which is convenient for adjusting different time axes to align data, thereby achieving stronger prediction capabilities.
[0017] (2) This invention provides a data-driven modeling method based on DCS logic autonomous exploration. It can be built directly using the algorithm logic blocks in the DCS system and deployed directly in the DCS system. It does not need to access the model externally and can be retrained using the latest data to obtain new mathematical logic expressions and update the original model. Attached Figure Description
[0018] Figure 1 This is a flowchart of a data-driven modeling method based on DCS logic autonomous exploration, as described in this invention.
[0019] Figure 2 This is a schematic diagram of the construction structure of some mathematical models in Embodiment 1 of the present invention.
[0020] Figure 3 This is an error diagram before and after model deployment in Embodiment 1 of the present invention. Detailed Implementation
[0021] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.
[0022] It should be noted that the following detailed description is exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.
[0023] It should be noted that the terms used herein are merely for describing specific embodiments and are not intended to limit the exemplary embodiments of the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.
[0024] In the case of no conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.
[0025] Embodiment 1 Refer to Figure 1 , this embodiment provides a data-driven modeling method based on DCS logic autonomous exploration, and the method includes: The present invention proposes a data-driven mathematical modeling method based on the "DCS algorithm block function library", which automatically explores the combined form through a machine learning algorithm to obtain a mathematical expression that can be directly run on the DCS for predicting and optimizing industrial logic behavior. The method includes steps such as data collection, feature generation, autonomous search modeling, structured output of expressions, control system deployment and operation.
[0026] S1. Data collection and data preprocessing, collect real-time data from the DCS system or historical database in the industrial field and perform preprocessing. The real-time data includes various influencing factors and temperature change amounts, and then calculate the time-delay relationship corresponding to each influencing factor under different working conditions.
[0027] Taking the prediction of the desuperheating water temperature of a boiler as an example, collect real-time data related to the target logic, that is, data of various influencing factors, from the DCS or historical database in the industrial field, including but not limited to: Coal feeding amount, actual power generation, desuperheating water flow rate, main steam temperature, main steam pressure, drum pressure, flue gas flow rate, flue gas velocity, desuperheating water valve opening, etc. Alternative data can be selected according to different units.
[0028] The specific implementation process of the preprocessing operation is: Remove null values: Use the deletion method to剔除null value data caused by sensor failures and communication interruptions during the collection process to ensure the integrity of each time series data; Missing values are imputed: For a small number of consecutively missing data points (missing duration ≤ 5 sampling periods), linear interpolation is used to impute them. The formula is as follows: (1); In formula (1), For the k-th data, For the sampling time of the k-th data point, For the (k+1)th data, For the (k-1)th data, For the sampling time of the (k+1)th data point, This refers to the sampling time of the (k-1)th data point; Filtering: A median filter algorithm is used to eliminate transient impulse interference. The window size is set to 3-5 sampling periods. Data fluctuations are smoothed by sorting and taking the median value, ensuring data stability and reliability. The preprocessed dataset needs to be time-series aligned to ensure that all parameters match on the same time dimension, laying the foundation for subsequent time delay calculations.
[0029] The steps for calculating the time delay relationship are as follows: ① Training dataset selection: From the preprocessed historical data, stratified sampling is performed according to different operating conditions divided into different load ranges. Load ranges are divided as follows: [90-132], [133-165], [166-210], [211-250], [251-330]. This ensures that the dataset includes time series data for each typical operating condition, while removing data from extreme abnormal operating conditions (such as unit start-up and shutdown, and fault shutdown) to avoid interfering with model training. ② Algorithm selection and model training: A combination of machine learning algorithms with strong time adaptability is adopted, with the Dynamic Time Warping (DTW) algorithm being the first choice for initial time delay matching to quickly locate the time delay range of each influencing factor on temperature changes; Then, it is combined with the Long Short-Term Memory Network (LSTM) for accurate calculation. That is, the time lag range of each influencing factor on temperature change is used as the limiting interval of the Long Short-Term Memory Network (LSTM). The time memory ability of the LSTM model is used to capture the dynamic change pattern of the time lag relationship. ③ Differentiated time delay relationship output: The model is trained using the training dataset, and the time delay relationship of each influencing factor is output under different load segments (in this embodiment, the load segment of the 330MW unit is divided into five intervals [90-132], [133-165], [166-210], [211-250], and [251-330]). Finally, it is presented in the form of time delay piecewise linear functions for different power ranges. This function can intuitively reflect the time delay duration variation law of each influencing factor under different operating conditions, and provide core parameter support for the subsequent temperature prediction model to adapt to the operating conditions.
[0030] S2. Construct a dictionary library, encapsulating the mathematical logic of the basic algorithm blocks in the DCS algorithm block function library into basic units in the dictionary library, which will serve as the basic units for building the required model.
[0031] To address the industry pain point of poor compatibility between traditional machine learning models and DCS systems, requiring additional interface development, this solution innovatively designs a configurable dictionary library module based on DCS algorithm blocks. This enables native and seamless integration of the model with the DCS system. The DCS algorithm block function library includes multiple basic algorithm blocks used to build the computational logic of the mathematical model for predicting the temperature of desuperheated water.
[0032] The core construction logic of this dictionary module is: to build upon the algorithm block logic natively contained in the DCS system, covering basic algorithm blocks such as addition blocks, subtraction blocks, multiplication blocks, exponential function blocks, proportional blocks, and integral blocks, and to encapsulate the mathematical logic of these basic algorithm blocks into the basic units in the dictionary.
[0033] The core function of the dictionary is to provide a "directly deployable DCS adaptation carrier" for temperature prediction models. Specifically, it uses the fundamental mathematical logic within the dictionary as the underlying foundation for model building. This allows the internal computational logic of the trained cooling water temperature prediction mathematical model to be completely decomposed into combinations of pre-packaged basic algorithm blocks from the dictionary. There's no need to store and call the model as an external file; instead, the model can be built and deployed directly using the algorithm block configuration function of the DCS system, leveraging the basic units in the dictionary. This achieves deep integration between the model and the DCS system, significantly reducing system integration complexity and failure risks. Furthermore, the dictionary supports configurable functions, allowing users to adjust the parameter mapping rules of the basic algorithm blocks through the DCS visual interface according to different operating conditions, enhancing the model's flexibility in adapting to various operating conditions.
[0034] S3. Autonomously explore data-driven mathematical modeling algorithms, using machine learning algorithms to train and explore mathematical model structures for various influencing factors and temperature changes that match the dictionary database.
[0035] Specifically, S3 includes: S31. Time-Lag Data Adaptation Processing: Based on the time lag durations of each influencing factor under different load segments obtained above, the preprocessed time series data is shifted five times according to the time lag durations corresponding to different load segments, so that the data of each influencing factor and the corresponding temperature response data are accurately aligned in the time dimension (if there is no time lag relationship between the data, the time alignment process is not required); for example, if the water spray volume lags behind the temperature by 100 seconds, then the water spray volume is shifted forward by 100 data points, and the data containing NAN values at the beginning and end are deleted. S32. Change Extraction: For the shifted data of each influencing factor, perform a difference operation (implemented by a subtraction block) with the corresponding time lag duration to obtain the change of each influencing factor within a time period; (2); In formula (2), This refers to the change in time delay duration corresponding to the first load segment interval of the influencing factors. This represents the value of the influencing factor at time T. This refers to the time delay duration corresponding to the influencing factor in the first load segment interval.
[0036] The changes in each influencing factor across different load segments are then multiplied by the weighting coefficients of each load interval constructed using the Gaussian function to obtain the reconstructed changes. ; (3); (4); In formulas (3)-(4), is the weighting coefficient, i.e. the model reconstruction coefficient; x is the current input load; a is the midpoint of the load divided into each load interval; b is an adjustable parameter that can be adjusted according to the model performance, and can be set to an initial value of 50.
[0037] S33. Model Training and Construction: The changes in each influencing factor are used as the model input, and the temperature change is used as the model output. The model is trained using Bayesian methods and the DCS system function library to obtain the model parameters of the mathematical model for predicting the temperature of desuperheated water.
[0038] The characteristic of the mathematical model for predicting the temperature of desuperheated water is that its mathematical logic expression is a custom-defined basic mathematical formula, which can be directly built using the algorithm blocks in the DCS system, thus enabling the model to be directly deployed in the DCS logic.
[0039] S4, Model Deployment and Control: Deployment is performed based on the DCS system and the generated mathematical model structure.
[0040] The time delay effect of the model input data can be implemented using the pure delay block in the DCS system, while the mathematical logic expression of the model can be directly constructed using the algorithm block of the DCS system based on the mathematical formula generated by the model.
[0041] Specifically, because the mathematical model for predicting the temperature of desuperheated water involved in the actual operation of this embodiment involves many influencing factors and is a large system, taking a portion of the mathematical model for predicting the temperature of desuperheated water as an example, such as... Figure 2As shown, the influencing factors include the steam drum pressure and the superheater first-stage desuperheating water flow. The algorithm blocks used in the construction include model parameter blocks, multiplication blocks, subtraction blocks, and pure delay blocks. Among them, the steam drum pressure pure delay block executes the corresponding time delay duration to offset the time delay effect.
[0042] This embodiment operates in an industrial field and performs real-time predictions. Based on the predictions, the industrial logic behavior can be optimized, such as... Figure 3 As shown, the horizontal axis represents time points, and the vertical axis represents the deviation from the set value. Figure 3 The error control of the main steam temperature after de-cooling on the morning of January 3, 2025, is shown in red. The red curve represents the error control situation without the deployment of this prediction model, and the blue curve represents the error control situation after the deployment of this prediction model. The comparison of the two curves shows that the error after deploying this prediction model is smaller in peak, trough and fluctuation ranges than when deploying this prediction model.
[0043] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solutions of the present invention, and are not intended to limit the specific implementation of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the claims of the present invention should be included within the protection scope of the claims of the present invention.
Claims
1. A data-driven modeling method based on DCS logic autonomous exploration, characterized in that, The method includes: S1. Collect real-time data from the DCS system or historical database at the industrial site and preprocess it. The real-time data includes various influencing factors and temperature changes. Then calculate the time delay relationship of each influencing factor under different operating conditions. S2. Construct a dictionary library, encapsulating the mathematical logic of the basic algorithm blocks of the DCS algorithm block function library into basic units in the dictionary library, which serve as the basic units for building the required model. S3. Use machine learning algorithms to train and explore the mathematical model structure of each influencing factor and temperature change that matches the dictionary database; S4. Deployment is performed based on the DCS system and the generated mathematical model structure.
2. The data-driven modeling method based on DCS logic autonomous exploration according to claim 1, characterized in that, The preprocessing includes removing null values, filling in missing values, and filtering: Remove null values: Use deletion to remove null data caused by sensor failure or communication interruption during the data acquisition process; Missing values are imputed: For consecutively missing data (i.e., missing duration ≤ 5 sampling periods), linear interpolation is used for imputation, expressed as: （1）； In formula (1), For the k-th data point corresponding to the missing time, For the sampling time of the k-th data point, For the (k+1)th data, For the (k-1)th data, For the sampling time of the (k+1)th data point, This refers to the sampling time of the (k-1)th data point; Filtering: Median filtering algorithm is used to eliminate instantaneous impulse interference. The window size is set to m1~m2 sampling periods. Data fluctuations are smoothed by sorting and taking the median value. Finally, the preprocessed data is time-aligned to obtain the training dataset.
3. The data-driven modeling method based on DCS logic autonomous exploration according to claim 2, characterized in that, The calculation of the time delay relationships of various influencing factors under different operating conditions specifically includes: ① Training dataset selection: From the preprocessed data, stratified sampling is performed according to different working conditions divided into different load ranges to ensure that the dataset includes time series data of each typical working condition; ② Algorithm selection and model training: A combination of machine learning algorithms with strong time adaptability is adopted, namely the dynamic time warping algorithm, to perform preliminary time delay matching and quickly locate the time delay range of each influencing factor on temperature changes; Then, the Long Short-Term Memory (LSTM) network is used for precise calculations. Specifically, the time lag range of each influencing factor on temperature change is used as the limiting interval of the LSTM network. The time memory capability of the LSTM network model is used to capture the dynamic change pattern of the time lag relationship. ③ Differentiated time delay relationship output: The model is trained using the training dataset, and the time delay relationship of each influencing factor under different operating conditions is output, presented in the form of time delay piecewise linear functions in different power ranges.
4. The data-driven modeling method based on DCS logic autonomous exploration according to claim 2, characterized in that, S3 specifically includes: S31. Time Delay Data Adaptation Processing: Based on the time delay duration of each influencing factor under different load segments, the preprocessed time series data is shifted according to the corresponding time delay duration of different load segments, so that the data of each influencing factor and the corresponding temperature response data are accurately aligned in the time dimension. S32. Change Extraction: For the shifted data of each influencing factor, perform difference calculations based on the corresponding time lag duration to obtain the change of each influencing factor within a time period. （2）； In formula (2), This refers to the change in time delay duration corresponding to the first load segment interval of the influencing factors. This represents the value of the influencing factor at time T. This refers to the time delay duration corresponding to this influencing factor in the first load segment interval; The changes in each influencing factor across different load segments are then multiplied by the weighting coefficients of each load interval constructed using the Gaussian function to obtain the reconstructed changes. ; （3）； （4）； In formulas (3)-(4), Here, x represents the current load input, a represents the midpoint of the load divided into each load interval, and b represents an adjustable parameter. S33. Model Training and Construction: The reconstructed changes of each influencing factor are used as the model input, and the temperature change is used as the model output. The model is trained using Bayesian methods and the DCS system function library to finally obtain the model parameters of the mathematical model for predicting the temperature of desuperheated water.