A real-time control method for coal-fired power plants based on multi-source heterogeneous data

Abstract

This invention relates to the field of coal-fired power plant control technology and discloses a real-time control method for coal-fired power plants based on multi-source heterogeneous data. The method includes: synchronously acquiring multi-source heterogeneous data; performing format conversion through a unified data access protocol to generate a standardized real-time data pool; calling corresponding preprocessing rules according to the operating condition type identifier to perform noise reduction and normalization processing, generating a high-quality feature dataset; obtaining physical coupling relationship parameters and predicted trend values ​​through parallel computation of a mechanistic model and a recurrent neural network; performing control parameter optimization to generate optimal control parameters; and generating control commands after performing safety verification on the optimal control parameters and sending them to the DCS underlying controller to form closed-loop real-time control.

Description

Technical Field

[0001] This invention relates to the field of coal-fired power plant control technology, specifically to a real-time control method for coal-fired power plants based on multi-source heterogeneous data. Background Technology

[0002] Coal-fired power plants generate various types of data during operation, including time-series operational data of the main systems (furnace, turbine, and electrical), relational process data from the DCS controller, status measurement data from auxiliary equipment, and online monitoring data from the environmental protection island. These data differ in format, acquisition frequency, and semantics, constituting multi-source heterogeneous data.

[0003] Traditional control systems can only collect and process single types of data, and cannot achieve efficient fusion and real-time analysis of multi-source data. Under complex operating conditions such as variable load and deep peak shaving, coal-fired units have large inertia and large delay characteristics. Data lag and multivariate coupling lead to control response delays, requiring frequent manual intervention to adjust control parameters.

[0004] The aforementioned problems prevent the unit from meeting the requirements for fully autonomous real-time control. Key parameters are prone to fluctuations exceeding the allowable range, and the control precision is insufficient, which restricts the safe and economical operation of the unit. Summary of the Invention

[0005] This invention provides a real-time control method for coal-fired power plants based on multi-source heterogeneous data, which solves the technical problems in related technologies such as the inability to uniformly access multi-source heterogeneous data of coal-fired power plants, the inconsistent data quality leading to insufficient accuracy in the calculation of control parameters, and the lack of security verification guarantee for the issuance of control commands.

[0006] This invention discloses a real-time control method for multi-source heterogeneous data fusion in coal-fired power plants, comprising at least the following steps: S1. Synchronously collect multi-source heterogeneous data from coal-fired power plants, perform format conversion on various heterogeneous data through a unified data access protocol, and generate a standardized real-time data pool; S2. Based on the current operating conditions of the unit, the data in the standardized real-time data pool is filtered to remove outliers and redundant data. The filtered data is then subjected to noise reduction and normalization to extract core control features and generate a high-quality feature dataset. S3. Input the high-quality feature dataset into the intelligent computing engine, calculate the physical coupling relationship parameters between each core control feature through the mechanism model, calculate the predicted trend value of each core control feature through the recurrent neural network, fuse the physical coupling relationship parameters with the predicted trend value, perform control parameter optimization calculation, and generate the optimal control parameters. S4. Perform a security check on the optimal control parameters, generate control commands and send them to the DCS underlying controller through a secure data channel, and have the DCS underlying controller execute the corresponding control operation.

[0007] Furthermore, the multi-source heterogeneous data includes at least: time-series operation data of the main mechanical, boiler, and electrical systems; relational process data of the DCS controller; status measurement data of auxiliary equipment; and online monitoring data of the environmental protection island. The unified data access protocol refers to a protocol that adapts the communication interfaces of different data sources and maps various types of data to predefined standard data fields and timestamp formats. Specifically, the time-series operation data is timestamped according to the acquisition time; the relational process data is converted into a unified key-value structure according to the field mapping table; and the equipment status measurement data and the online monitoring data of the environmental protection island are respectively subjected to numerical unit normalization and encoding format conversion. All converted data is written into the standardized real-time data pool.

[0008] Furthermore, the step of filtering the data in the standardized real-time data pool based on the current operating condition of the unit also includes: obtaining the operating condition type identifier of the current operating condition of the unit, wherein the operating condition type identifier is used to characterize that the unit is currently in one of the following conditions: deep peak shaving, steady-state operation, or rapid load change; invoking the corresponding preprocessing rules according to the operating condition type identifier to perform outlier removal and redundant data filtering on the data in the standardized real-time data pool; wherein, under the deep peak shaving condition, the outlier judgment threshold in the preprocessing rules is relaxed; under the steady-state operation condition, the outlier judgment threshold is tightened; under the rapid load change condition, the data acquisition window is shortened, and feature data with high correlation to the load change rate are preferentially retained.

[0009] Furthermore, the noise reduction process employs a Kalman filter algorithm, with the input being the time-series sampled value sequence of each measurement point and the output being a smoothed estimated value sequence after removing random noise. The normalization process employs a mean normalization algorithm based on the range, with the input being the original values ​​of each core control feature and the output being normalized feature values ​​mapped to a unified numerical range. After the noise reduction and normalization processes, feature correlation analysis is performed on the extracted core control features to calculate the correlation coefficient matrix between each core control feature. When the absolute value of the correlation coefficient between two features exceeds a preset correlation threshold, the feature with a higher correlation to the control target is retained, and the other redundant feature is removed.

[0010] Furthermore, before inputting the high-quality feature dataset into the intelligent computing engine, the method further includes: aligning the time-series operational features and relational process features in the high-quality feature dataset along the time dimension to generate a fused dataset; wherein, hysteresis compensation processing is performed on the time-series operational features, and the timestamps of the time-series operational features are shifted forward by the corresponding delay amount according to the known transmission delay time of each measurement point, so that the time-series operational features and the relational process features are aligned on the same time reference.

[0011] Furthermore, the calculation of the physical coupling relationship parameters between the core control features obtained through the mechanism model and the calculation of the predicted trend values ​​of the core control features obtained through the recurrent neural network are performed in parallel. The mechanism model is a physical calculation model pre-established based on the thermodynamic and fluid dynamic principles of coal-fired power units. The input of the mechanism model is the numerical values ​​of each core control feature in the fused dataset, and the output is the physical coupling relationship parameters characterizing the coupling strength between the features. The input layer of the recurrent neural network receives the temporal feature sequence in the fused dataset, the hidden layer consists of multiple long short-term memory units, and the output layer is a fully connected layer that outputs the predicted trend values ​​of each core control feature within a preset future time step.

[0012] Furthermore, the optimization calculation of the execution control parameters further includes: the objective function of the optimization calculation is defined as the sum of the products of the weight coefficients corresponding to each core control feature and the absolute difference between the normalized predicted trend value and the normalized target set value of the core control feature at the current control cycle time; the optimization process uses the feasible region defined by the physical coupling relationship parameters as the constraint boundary, and the predicted trend value as the optimization direction, to search for the combination of control parameters that minimizes the objective function within the feasible region as the optimal control parameters; wherein, the weight coefficients are dynamically adjusted according to the operating condition type identifier of the current operating condition of the unit, increasing the weight of load tracking-related features under deep peak shaving conditions, increasing the weight of economic-related features under steady-state operating conditions, and increasing the weight of safety-related features under rapid load change conditions; after the optimization is completed, the optimal control parameters are subjected to inverse normalization transformation to restore the physical dimension values ​​of each core control feature.

[0013] Furthermore, the safety verification of the optimal control parameters further includes: comparing the optimal control parameters after inverse normalization transformation with pre-set safety boundary parameters, wherein the safety boundary parameters are a set of upper and lower limits pre-set according to the safe operating range of various operating parameters of the coal-fired unit; when all parameter values ​​in the optimal control parameters are within the allowable range defined by the safety boundary parameters, the verification is deemed successful; when any parameter value exceeds the allowable range, the parameter value is truncated to the boundary value of the allowable range, and an alarm flag is generated; based on the verified control parameters or the truncated control parameters, a control instruction set is generated, wherein the control instruction set includes at least one of the following: boiler-machine coordination control instruction, main steam temperature regulation instruction, reheat steam temperature regulation instruction, environmental protection system control instruction, and auxiliary equipment regulation instruction.

[0014] Furthermore, after the control command set is sent to the DCS underlying controller, the process returns to the step of synchronously collecting multi-source heterogeneous data from the coal-fired power plant, and re-collects the latest multi-source heterogeneous data, forming a continuous loop of data acquisition, preprocessing, fusion calculation, and control execution. Within a preset confirmation time after the control command set is sent, the execution confirmation signal returned by the DCS underlying controller is obtained. If the execution confirmation signal is not received within the preset confirmation time, the control command set is re-sent and the number of re-sentences is accumulated. When the number of re-sentences reaches a preset upper limit, the issuance of the control command set is paused and a fault notification is generated.

[0015] This invention discloses a real-time control system for multi-source heterogeneous data fusion in a coal-fired power plant, used to execute the aforementioned real-time control method for multi-source heterogeneous data fusion in a coal-fired power plant, comprising: The data standardization acquisition module is used to synchronously collect multi-source heterogeneous data from coal-fired power plants. It performs format conversion on various heterogeneous data through a unified data access protocol to generate a standardized real-time data pool. The operating condition-driven preprocessing module is used to filter the data in the standardized real-time data pool based on the current operating condition of the unit, remove outliers and redundant data, perform noise reduction and normalization processing on the filtered data, extract core control features, and generate a high-quality feature dataset. The multi-dimensional fusion computing module is used to input the high-quality feature dataset into the intelligent computing engine, calculate and obtain the physical coupling relationship parameters between each core control feature through the mechanism model, calculate and obtain the predicted trend value of each core control feature through the recurrent neural network, fuse the physical coupling relationship parameters and the predicted trend value, perform control parameter optimization calculation, and generate the optimal control parameters. The closed-loop control command module is used to perform safety verification on the optimal control parameters, generate control commands and send them to the DCS underlying controller through a secure data channel, and the DCS underlying controller executes the corresponding control operation.

[0016] This invention performs standardized format conversion on multi-source heterogeneous data through a unified data access protocol, enabling time-series operation data, relational process data, equipment measurement data, and environmental monitoring data to be merged into the same standardized real-time data pool and achieve unified timestamps, overcoming the limitation of traditional control systems that can only collect a single type of data.

[0017] During the data preprocessing stage, the preprocessing rules are dynamically adjusted according to different operating conditions such as deep peak shaving, steady-state operation, and rapid load change, so that outlier identification and feature extraction can be adapted to the characteristics of the current operating conditions, reducing the erroneous removal of valid data or the retention of noisy data due to fixed rules.

[0018] In the fusion calculation stage, the mechanistic model provides the constraint boundary of the physical coupling relationship, and the recurrent neural network compensates for the nonlinear dynamic characteristics. After the two are fused, the control parameters are optimized. All values ​​in the objective function are based on normalized numerical calculations, which eliminates the influence of the difference in the dimensions of different physical quantities on the optimization calculation. The hysteresis compensation process reduces the data time deviation in large inertia and large delay systems, thereby reducing the control response delay.

[0019] During the control execution phase, the optimal control parameters are directly sent to the DCS underlying controller after safety verification, forming a continuous closed-loop control. This enables the unit to autonomously adjust control parameters under complex operating conditions, reducing manual intervention and improving the control accuracy of key parameters. Attached Figure Description

[0020] Figure 1 This is a flowchart of the real-time control method for multi-source heterogeneous data fusion in a coal-fired power plant provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the distribution of normalized feature values ​​of the core control feature provided in the embodiments of the present invention; Figure 3 This is a schematic diagram of the distribution of normalized eigenvalues ​​of the control period provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the distribution of normalized ratio values ​​of control parameters provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the distribution of normalized control parameter values ​​provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the distribution of control feature parameter values ​​provided in an embodiment of the present invention; Figure 7 This is a schematic diagram illustrating the number of effective features in the processing stage provided in this embodiment of the invention; Figure 8 This is a schematic diagram of the weighted deviation contribution value of the control features provided in the embodiments of the present invention; Figure 9 This is a schematic diagram illustrating the number of data source fields provided in the embodiments of the present invention. Detailed Implementation

[0021] Coal-fired power plants generate various types of data during operation, including time-series operational data of the main boiler, turbine, and electrical systems; relational process data from the DCS controller; status measurement data from auxiliary equipment; and online monitoring data from desulfurization, denitrification, and dust removal in the environmental protection island. These data differ in format, acquisition frequency, and semantics, constituting multi-source heterogeneous data. Traditional control systems can only collect and process single-type data, failing to achieve efficient fusion and real-time analysis of multi-source data. Under complex operating conditions such as variable loads and deep peak shaving, coal-fired units exhibit high inertia and significant latency. Data lag and multi-variable coupling lead to delayed control response, requiring frequent manual intervention to adjust control parameters, which cannot meet the needs of fully autonomous real-time control of the unit. Furthermore, fluctuations in key parameters exceeding allowable ranges and insufficient control precision restrict the safe and economical operation of the unit.

[0022] The method in this embodiment runs on the computing platform where the real-time data middleware and intelligent computing engine of the coal-fired power plant are located. This computing platform is connected to the DCS underlying controller through a secure data channel.

[0023] Reference Figure 1 According to an embodiment of the present invention, a real-time control method for multi-source heterogeneous data fusion in a coal-fired power plant includes the following steps: Step 1: Standardized collection of multi-source heterogeneous data.

[0024] Multi-source heterogeneous data from coal-fired power plants are collected synchronously. A unified data access protocol is used to perform format conversion on various types of heterogeneous data, generating a standardized real-time data pool. This multi-source heterogeneous data includes at least: time-series operational data of the main boiler, turbine, and electrical systems; relational process data from the DCS controller; status measurement data from auxiliary equipment; and online monitoring data from the environmental protection island. The standardized real-time data pool contains data fields covering pressure, temperature, flow rate, load, pollutant concentration, and equipment start-up / shutdown status.

[0025] It should be noted that the aforementioned unified data access protocol refers to a protocol that adapts the communication interfaces of different data sources and maps various types of data to predefined standard data fields and timestamp formats. Specifically, time-series operational data is timestamped according to the acquisition time; relational process data is converted into a unified key-value structure according to a field mapping table; and equipment status measurement point data and environmental monitoring data undergo numerical unit normalization and encoding format conversion, respectively. All converted data is then written to a standardized real-time data pool.

[0026] In this embodiment of the application, to ensure the real-time performance and integrity of the data, concurrent acquisition of tens of thousands of measurement points is supported. When the acquired data of a certain measurement point is missing in multiple consecutive acquisition cycles, the data of that measurement point is marked as missing, and the most recent valid acquisition value is retained in the standardized real-time data pool as a temporary filler value for subsequent preprocessing steps.

[0027] Step 2, condition-driven data preprocessing.

[0028] Based on the current operating conditions of the unit, data in the standardized real-time data pool is filtered to remove outliers and redundant data. The filtered data is then subjected to noise reduction and normalization to extract core control features, generating a high-quality feature dataset. (Refer to...) Figure 2 The core control characteristics include at least load, main steam pressure, main steam temperature, reheat steam temperature, SO2 concentration, and NO. x Concentration and coal mill output.

[0029] Step 2 may specifically include: Step 201: Obtain the operating condition type identifier of the unit's current operating condition. The operating condition type identifier is used to characterize whether the unit is currently in a deep peak shaving condition, a steady-state operating condition, or a rapid load change condition.

[0030] Step 202: Based on the working condition type identifier, call the corresponding preprocessing rules to perform outlier removal and redundant data filtering on the data in the standardized real-time data pool to generate a valid dataset.

[0031] Step 203: Perform denoising and normalization processing on the effective dataset to extract the values ​​of each core control feature and generate a high-quality feature dataset. The denoising process uses a Kalman filter algorithm, with the input being the time-series sampled values ​​of each measurement point and the output being a smoothed estimate sequence after removing random noise. The normalization process uses a range-based mean normalization algorithm, with the input being the original values ​​of each core control feature and the output being normalized feature values ​​mapped to a unified numerical range.

[0032] It should be noted that the above preprocessing rules are dynamically adjusted according to different operating conditions. Under deep peak shaving conditions, the load change is large, and the outlier judgment threshold in the preprocessing rules is relaxed accordingly to avoid misjudging normal large load change data as outliers. Under steady-state operating conditions, the fluctuation range of each parameter is small, and the outlier judgment threshold is tightened to improve the accuracy of data screening. Under rapid load change conditions, the data acquisition window is shortened, and feature data with high correlation to the load change rate are retained first.

[0033] In this embodiment of the application, in order to further improve the quality of the high-quality feature dataset, based on step 203, feature correlation analysis is also performed on the extracted core control features. Specifically, the correlation coefficient matrix between each core control feature is calculated. When the absolute value of the correlation coefficient between two features exceeds a preset correlation threshold, the feature with a higher correlation to the control target is retained, and the other redundant feature is removed, thereby reducing the redundant input dimensions in subsequent fusion calculations.

[0034] Step 3: Multi-dimensional data fusion calculation.

[0035] High-quality feature datasets are input into the intelligent computing engine to perform multivariate coupling relationship analysis and control parameter optimization, generating optimal control parameters.

[0036] Step 3 may specifically include: Step 301: Align the time-series running features and relational process features in the high-quality feature dataset with each other along the time dimension to generate an aligned fused dataset.

[0037] Step 302: Based on the fused dataset, calculate the physical coupling relationship parameters between each core control feature using a mechanistic model. The mechanistic model is a pre-established physical calculation model based on the thermodynamic and fluid dynamic principles of a coal-fired power unit. The input to the mechanistic model is the numerical values ​​of each core control feature in the fused dataset, and the output is the physical coupling relationship parameters characterizing the coupling strength between each feature.

[0038] Reference Figure 6 Step 303: Based on the fused dataset, the predicted trend values ​​of each core control feature are calculated using a recurrent neural network (RNN). The input layer of the RNN receives the temporal feature sequence from the fused dataset, the hidden layer consists of multiple long short-term memory units, and the output layer is a fully connected layer that outputs the predicted trend values ​​of each core control feature within a preset future time step. The RNN is labeled with the measured values ​​of each core control feature in historical operating data, uses the mean squared error loss function, and is trained using the time backpropagation algorithm and the Adam optimization algorithm.

[0039] It should be noted that the input to the recurrent neural network is the normalized feature value after normalization processing in step 203, and the output predicted trend value is in the same numerical range as the normalized feature value.

[0040] Step 304: The physical coupling parameters and predicted trend values ​​are fused, and control parameter optimization calculation is performed to generate the optimal control parameters. The objective function for the optimization calculation is defined as:

[0041] in, To control the value of the objective function, The number of core control features, This refers to the serial number of the core control feature. This is the moment of the current control cycle. For the first The weight coefficients corresponding to the core control features of each item For the first The core control feature at time Normalized predicted trend value, For the first The core control feature at time The normalized target setting value, and All values ​​must be within the same numerical range to ensure consistency of dimensions. The optimization process uses the feasible region defined by the physical coupling parameters as the constraint boundary, and the predicted trend value as the optimization direction, searching within the feasible region to make... The combination of control parameters that minimizes this value is the optimal control parameter. After optimization, the optimal control parameters are subjected to inverse normalization to restore them to the physical dimension values ​​of each core control characteristic.

[0042] It should be noted that steps 302 and 303 above employ a parallel computation method combining the mechanistic model and the recurrent neural network. The mechanistic model provides steady-state coupling relationships based on physical laws, while the recurrent neural network compensates for nonlinear dynamic characteristics that are difficult for the mechanistic model to describe. The fusion method in step 304 is as follows: using the physical coupling relationship parameters output by the mechanistic model as the constraint boundary, and the predicted trend value output by the recurrent neural network as the optimization direction, a search is performed within the constraint boundary range to make the... The combination of control parameters that takes the minimum value.

[0043] In this embodiment, to address control deviations caused by data lag in systems with high inertia and high latency, lag compensation processing is performed on the timing operation features in step 301. Specifically, based on the known transmission delay time of each measurement point, the timestamp of the timing operation features is shifted forward by the corresponding delay amount, aligning the timing operation features with the relational process features on the same time reference. Through this processing, the timing data and relational data in the fused dataset are kept consistent in the time dimension, reducing fusion errors caused by data lag.

[0044] Reference Figure 6 In this embodiment of the application, in order to improve the analytical accuracy of multivariate coupling relationships under complex operating conditions, an adaptive weight adjustment based on the operating condition is introduced in step 302. Specifically, based on the operating condition type identifier obtained in step 201, the weight coefficients of each core control feature in the control objective function are dynamically adjusted. Under deep peak-shaving conditions, the weight of load tracking-related characteristics is increased; under steady-state operation conditions, the weight of economic-related characteristics is increased; and under rapid load-changing conditions, the weight of safety-related characteristics is increased. Through adaptive weight adjustment based on operating conditions, the results of the optimization calculation can be adapted to the emphasis requirements of the current operating conditions on the control objectives.

[0045] Step 4: Generation and issuance of closed-loop real-time control commands.

[0046] Reference Figure 4 and Figure 5 The system performs a safety check on the optimal control parameters, generates control commands, and sends them to the DCS underlying controller via a secure data channel. The DCS underlying controller then executes the corresponding control operations.

[0047] Step 4 may specifically include: Step 401: Compare and verify the optimal control parameters with the pre-set safety boundary parameters. If all parameter values ​​in the optimal control parameters are within the allowable range defined by the safety boundary parameters, the verification is deemed successful; if any parameter value exceeds the allowable range, the parameter value is truncated to the boundary value of the allowable range, and an alarm flag is generated.

[0048] It should be noted that the above-mentioned optimal control parameters have undergone inverse normalization transformation before the safety verification, restoring them to the physical dimension values ​​of each core control feature. The safety boundary parameters are also expressed in physical dimensions, and the two are compared under the same dimension. The safety boundary parameters are a set of upper and lower limits preset according to the safe operating range of various operating parameters of the coal-fired unit. The purpose of the safety boundary parameters is to ensure that the control commands issued to the DCS lower-level controller will not cause the unit's operating parameters to exceed the safe range.

[0049] Step 402: Based on the verified control parameters or truncated control parameters, generate a control instruction set. The control instruction set includes at least one of the following: boiler-turbine coordination control instructions, main steam temperature regulation instructions, reheat steam temperature regulation instructions, environmental protection system control instructions, and auxiliary equipment regulation instructions.

[0050] Step 403: The control instruction set is sent to the DCS underlying controller through the secure data channel, and the DCS underlying controller executes the corresponding control operation according to each control instruction.

[0051] In this embodiment, to achieve fully closed-loop control, after the control command is issued in step 403, the system returns to step 1 to re-collect the latest multi-source heterogeneous data, forming a continuous loop of data acquisition, preprocessing, fusion calculation, and control execution. In each loop, the intelligent computing engine recalculates the optimal control parameters based on the latest collected multi-source heterogeneous data, enabling the control command to continuously track changes in the unit's operating status and adapt to operating condition switching without manual intervention.

[0052] In this embodiment, to improve the reliability of control command issuance, in addition to step 403, an execution confirmation signal returned by the DCS underlying controller is obtained within a preset confirmation time after the control command set is issued. If no execution confirmation signal is received within the preset confirmation time, the control command set is reissued, and the number of reissues is accumulated. When the number of reissues reaches a preset upper limit, the issuance of the control command set is paused, and a fault notification is generated.

[0053] Reference Figure 7 and Figure 8 This implementation method uses a unified data access protocol to perform standardized format conversion on time-series operational data, relational process data, equipment measurement point data, and environmental monitoring data. This allows multi-source heterogeneous data to be merged into the same standardized real-time data pool and achieves unified timestamps, thus overcoming the limitation of traditional control systems that can only collect a single type of data. In the data preprocessing stage, preprocessing rules are dynamically adjusted according to different operating conditions such as deep peak shaving, steady-state operation, and rapid load changes. This ensures that outlier detection and feature extraction are adapted to the characteristics of the current operating condition, reducing the erroneous removal of valid data or the retention of noisy data due to fixed rules. In the fusion calculation stage, a parallel calculation and fusion approach using a mechanistic model and a recurrent neural network is adopted. The mechanistic model provides the constraint boundary of the physical coupling relationship, while the recurrent neural network compensates for nonlinear dynamic characteristics. After fusion, control parameter optimization is performed. The predicted trend values ​​and normalized target setpoints in the objective function are all based on the current control cycle time. Calculations are performed based on normalized values, eliminating time inconsistencies caused by mixing data from different acquisition times. All values ​​are calculated using normalized values, eliminating the impact of differences in the dimensions of different physical quantities on the optimization calculation. After optimization, the values ​​are restored to their physical dimensions through inverse normalization before safety verification and instruction generation, ensuring the executability of the optimal control parameters. Furthermore, hysteresis compensation reduces data time deviations in systems with large inertia and large delays, thereby reducing control response latency. During the control execution phase, the optimal control parameters, after safety verification, are directly sent to the DCS lower-level controller to form closed-loop control. This allows the unit to autonomously adjust control parameters under complex operating conditions, reducing manual intervention, improving the control accuracy of key parameters, and ensuring the safe and economical operation of the unit.

[0054] Reference Figure 9A coal-fired power plant unit (hereinafter referred to as Unit A) has a rated capacity of 600MW. On a certain operating day in 20XX, it faced a deep peak-shaving task, with the load command rapidly decreasing from the rated value to 50% of the rated value. Unit A is equipped with a DCS control system, with more than 10,000 field monitoring points. The data sources cover four categories: the main boiler, turbine, and electrical systems, the DCS process database, auxiliary equipment monitoring points, and online instruments in the environmental protection island. This case study focuses on the data flow of four consecutive control cycles during this peak-shaving process to illustrate the specific operation of each step.

[0055] Implementation process of step 1 Reference Figure 3 The four data sources for Unit A differ in acquisition frequency, communication protocol, and field format. The main boiler, turbine, and electrical system outputs time-series data at 1-second acquisition intervals; the DCS process database stores data in relational table format, with field names consisting of equipment codes and parameter abbreviations; auxiliary equipment measuring points use enumerated codes to represent start / stop status; and the concentration unit output by the environmental protection island instruments is mg / Nm³. After processing with a unified data access protocol, all the above data are mapped to standard fields, timestamps are standardized to whole-second format, equipment status codes are standardized to "Run / Stop" text, the concentration unit is retained as mg / Nm³, and each field is written to a standardized real-time data pool. The acquisition time for this control cycle is T1.

[0056] Table 1 Results of Multi-Source Heterogeneous Data Acquisition and Standardization

[0057] At time T1, the data of the coal mill B measuring point was missing for three consecutive acquisition cycles. The measuring point was marked as missing. The acquired value of the previous valid cycle (coal mill B output 138.50 t / h) was retained in the standardized real-time data pool as a temporary filling value for step 2 judgment and processing.

[0058] Implementation process of step 2 Step 201 identifies the current operating condition as "deep peak shaving condition" and calls the deep peak shaving preprocessing rules based on this identifier: the threshold for judging outliers is relaxed, the data acquisition window for load-related features is maintained at the standard length, the missing measurement point of coal mill B is judged as a data interruption, its temporary filling value is not included in the effective dataset, and is replaced by the average value estimation of similar operating coal mills.

[0059] Step 202 performs outlier removal and redundancy filtering on the standardized real-time data pool to generate a valid dataset. The reheat steam temperature measurement point reads 621.00℃ in this cycle, which exceeds the upper limit of the relaxed threshold (620.00℃) under deep peak shaving conditions. It is identified as an outlier and removed. The subsequent value is replaced by a Kalman filter smoothed estimate.

[0060] Step 203 performs Kalman filtering for noise reduction on the valid dataset, followed by mean normalization based on the range. The normalization formula is:

[0061] in, These are the original eigenvalues. and These are the minimum and maximum values ​​of this feature in the historical data. These are normalized characteristic values. Taking the main steam pressure as an example, MPa MPa MPa, normalized result .

[0062] Table 2. Noise Reduction and Normalization Results of Core Control Features

[0063] Feature correlation analysis showed that NO x The absolute value of the correlation coefficient between NO concentration and coal mill output is 0.81, exceeding the preset threshold of 0.80. Because coal mill output has a higher correlation with the control target, NO... x Concentration is reduced to an auxiliary reference dimension in the fusion calculation input and does not participate in the main optimization calculation to reduce redundant input.

[0064] Implementation process of step 3 Step 301 aligns the time-series operational features with the relational process features along the time dimension. The main steam pressure measurement point has a known transmission delay of 2 seconds; its timestamp is shifted forward by 2 seconds and aligned with the load features to the same time reference T1. After alignment, a fused dataset is formed, with each feature represented by a normalized feature value.

[0065] Step 302 calculates the physical coupling parameters using thermodynamic and fluid dynamic mechanism models. Under the current deep peak shaving conditions, the mechanism model outputs a coupling strength parameter of 0.88 between the main steam pressure and the load, and a coupling strength parameter of 0.76 between the main steam temperature and the coal mill output. These two parameters determine the feasible domain constraint boundary for the optimization calculation.

[0066] Step 303 involves receiving the time-series feature sequences from the fused dataset using a recurrent neural network (LSTM) and outputting the predicted trend values ​​of each core control feature over the next three control cycles.

[0067] Table 3. Trend values ​​predicted by recurrent neural networks (normalized)

[0068] Step 304 involves optimizing control parameters. The current operating condition is deep peak shaving; the adaptive weight adjustment will adjust the load characteristic weights. Set to 0.45, main steam pressure weight The weight is set to 0.30, and the total weight of the remaining features is 0.25. The objective function is:

[0069] in, The total number of control features participating in the optimization process. For the first The weights of each feature, For this feature at the target time The predicted trend value, Set the corresponding normalized target value. With time T2 as the optimization objective, the load normalized target setting value is... Substitute the predicted trend value The contribution value is After comprehensive calculations The combination of control parameters corresponding to the minimum value is the optimal control parameter. After optimization, the optimal control parameters are inversely normalized to restore them to their physical dimension values. The inverse normalization formula is:

[0070] in, To achieve the normalized optimal value, and These are the minimum and maximum values ​​of this feature in the historical data. These are the restored physical values. Taking the load setpoint as an example... , MW, MW, inverse normalization result MW.

[0071] Table 4. Inverse normalization results of optimal control parameters

[0072] Implementation process of step 4 Step 401 compares the optimal control parameters after inverse normalization with the pre-set safety boundary parameters. The main steam pressure setpoint of 16.55 MPa is within the safety lower limit of 12.50 MPa to the safety upper limit of 17.50 MPa, and the verification passes; the other parameters also pass the verification, with no truncation operation and no alarm flag generated.

[0073] Table 5 Security Verification Comparison Results

[0074] Step 402 generates a control command set, including boiler-turbine coordination control commands (load target 297.60 MW), main steam temperature adjustment commands (536.42℃), and auxiliary equipment adjustment commands (total coal mill output 267.60 t / h). Step 403 sends the control command set to the DCS lower-level controller via a secure data channel. Upon receiving an execution confirmation signal within a preset confirmation time (500 milliseconds), the current control cycle is complete. The system then returns to step 1 to begin data acquisition for the next control cycle, forming a continuous closed loop.

[0075] The entire data flow process demonstrates a clear logical progression: the standardized real-time data pool output from step 1 provides a unified format input for step 2; the normalized feature dataset output from step 2 directly enters the fusion calculation in step 3; step 3 completes optimization with normalized values ​​and then inversely normalizes to restore the physical values, which are then passed to step 4 for security verification; after the verification in step 4 passes, an instruction is generated and issued, and the control execution result is reflected back through the data collected in the next round of step 1, forming a closed loop. This includes labeling and estimating missing data from coal mill B, filtering and replacing outlier reheat steam temperatures, and NO... x Correlation reduction of concentration and hysteresis compensation alignment of main steam pressure both maintain data coherence and physical consistency between steps.

[0076] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A real-time control method for multi-source heterogeneous data fusion in a coal-fired power plant, characterized in that, At least the following steps are included: S1. Synchronously collect multi-source heterogeneous data from coal-fired power plants, perform format conversion on various heterogeneous data through a unified data access protocol, and generate a standardized real-time data pool; S2. Based on the current operating conditions of the unit, the data in the standardized real-time data pool is filtered to remove outliers and redundant data. The filtered data is then subjected to noise reduction and normalization to extract core control features and generate a high-quality feature dataset. S3. Input the high-quality feature dataset into the intelligent computing engine, calculate the physical coupling relationship parameters between each core control feature through the mechanism model, calculate the predicted trend value of each core control feature through the recurrent neural network, fuse the physical coupling relationship parameters with the predicted trend value, perform control parameter optimization calculation, and generate the optimal control parameters. S4. Perform a security check on the optimal control parameters, generate control commands, and send them to the DCS underlying controller through a secure data channel. The DCS underlying controller then executes the corresponding control operation.

2. The real-time control method for multi-source heterogeneous data fusion in a coal-fired power plant according to claim 1, characterized in that, The multi-source heterogeneous data includes at least: time-series operation data of the main mechanical, boiler, and electrical systems; relational process data of the DCS controller; status measurement data of auxiliary equipment; and online monitoring data of the environmental protection island. The unified data access protocol refers to a protocol that adapts the communication interfaces of different data sources and maps various types of data to predefined standard data fields and timestamp formats. Specifically, the time-series operation data is timestamped according to the acquisition time; the relational process data is converted into a unified key-value structure according to the field mapping table; and the equipment status measurement data and the online monitoring data of the environmental protection island are respectively subjected to numerical unit normalization and encoding format conversion. All converted data is written into the standardized real-time data pool.

3. The real-time control method for multi-source heterogeneous data fusion in a coal-fired power plant according to claim 1, characterized in that, The process of filtering data in the standardized real-time data pool based on the current operating conditions of the unit also includes: Obtain the operating condition type identifier of the unit's current operating condition. The operating condition type identifier is used to characterize that the unit is currently in one of the following conditions: deep peak shaving, steady-state operation, or rapid load change. According to the operating condition type identifier, the corresponding preprocessing rules are invoked to perform outlier removal and redundant data filtering on the data in the standardized real-time data pool. Specifically, under deep peak shaving conditions, the outlier determination threshold in the preprocessing rules is relaxed; under steady-state operating conditions, the outlier determination threshold is tightened; and under rapid load change conditions, the data acquisition window is shortened, and feature data with high correlation to the load change rate are preferentially retained.

4. The real-time control method for multi-source heterogeneous data fusion in a coal-fired power plant according to claim 3, characterized in that, The noise reduction process employs a Kalman filter algorithm, with the input being the time-series sampled value sequence of each measurement point and the output being a smoothed estimated value sequence after removing random noise. The normalization process employs a mean normalization algorithm based on the range, with the input being the original values ​​of each core control feature and the output being normalized feature values ​​mapped to a unified numerical range. After the noise reduction and normalization processes, feature correlation analysis is performed on the extracted core control features to calculate the correlation coefficient matrix between each core control feature. When the absolute value of the correlation coefficient between two features exceeds a preset correlation threshold, the feature with a higher correlation to the control target is retained, and the other redundant feature is removed.

5. The real-time control method for multi-source heterogeneous data fusion in a coal-fired power plant according to claim 1, characterized in that, Before inputting the high-quality feature dataset into the intelligent computing engine, the following steps are also included: The time-series operation features and relational process features in the high-quality feature dataset are aligned in time dimension to generate a fused dataset; wherein, hysteresis compensation processing is performed on the time-series operation features, and the timestamps of the time-series operation features are shifted forward by the corresponding delay amount according to the known transmission delay time of each measurement point, so that the time-series operation features and the relational process features are aligned on the same time base.

6. The real-time control method for multi-source heterogeneous data fusion in a coal-fired power plant according to claim 5, characterized in that, The calculation of physical coupling parameters between core control features using a mechanistic model and the calculation of predicted trend values ​​of core control features using a recurrent neural network are performed in parallel. The mechanistic model is a pre-established physical calculation model based on the thermodynamics and fluid dynamics principles of a coal-fired power unit. The input of the mechanistic model is the numerical values ​​of each core control feature in the fused dataset, and the output is the physical coupling parameters characterizing the coupling strength between the features. The input layer of the recurrent neural network receives the temporal feature sequence from the fused dataset, the hidden layer consists of multiple long short-term memory units, and the output layer is a fully connected layer that outputs the predicted trend values ​​of each core control feature within a preset future time step.

7. The real-time control method for multi-source heterogeneous data fusion in a coal-fired power plant according to claim 6, characterized in that, The optimization calculation of the execution control parameters also includes: The objective function for the optimization calculation is defined as the sum of the products of the weight coefficients corresponding to each core control feature and the absolute difference between the normalized predicted trend value and the normalized target setpoint of that core control feature at the current control cycle time. The optimization process uses the feasible region defined by the physical coupling relationship parameters as the constraint boundary, and the predicted trend value as the optimization direction. It searches for the combination of control parameters that minimizes the objective function within the feasible region as the optimal control parameters. Among them, the weight coefficients are dynamically adjusted according to the operating condition type identifier of the current operating condition of the unit. Under deep peak shaving conditions, the weight of load tracking-related features is increased; under steady-state operating conditions, the weight of economic-related features is increased; and under rapid load change conditions, the weight of safety-related features is increased. After optimization, the optimal control parameters are subjected to inverse normalization transformation to restore the physical dimension values ​​of each core control feature.

8. The real-time control method for multi-source heterogeneous data fusion in a coal-fired power plant according to claim 1, characterized in that, The step of performing a security check on the optimal control parameters further includes: The optimal control parameters after inverse normalization transformation are compared and verified with pre-set safety boundary parameters, which are a set of upper and lower limits pre-set according to the safe operating range of various operating parameters of the coal-fired unit. When all parameter values ​​in the optimal control parameters are within the allowable range defined by the safety boundary parameters, the verification is deemed successful. When any parameter value exceeds the allowable range, the parameter value is truncated to the boundary value of the allowable range, and an alarm flag is generated. Based on the verified control parameters or the truncated control parameters, a control instruction set is generated. The control instruction set includes at least one of the following: boiler-generator coordination control instruction, main steam temperature regulation instruction, reheat steam temperature regulation instruction, environmental protection system control instruction, and auxiliary equipment regulation instruction.

9. The real-time control method for multi-source heterogeneous data fusion in a coal-fired power plant according to claim 8, characterized in that, After the control command set is sent to the DCS underlying controller, the process returns to the step of synchronously collecting multi-source heterogeneous data from the coal-fired power plant, and the latest multi-source heterogeneous data is collected again, forming a continuous loop of data acquisition, preprocessing, fusion calculation and control execution. Within a preset confirmation time after the control command set is issued, the execution confirmation signal returned by the DCS underlying controller is obtained. If the execution confirmation signal is not received within the preset confirmation time, the control command set is reissued and the number of reissues is accumulated. When the number of reissues reaches a preset upper limit, the issuance of the control command set is suspended and a fault notification is generated.

10. A real-time control system for multi-source heterogeneous data fusion in a coal-fired power plant, used to execute the real-time control method for multi-source heterogeneous data fusion in a coal-fired power plant as described in any one of claims 1 to 9, characterized in that, include: The data standardization acquisition module is used to synchronously collect multi-source heterogeneous data from coal-fired power plants. It performs format conversion on various heterogeneous data through a unified data access protocol to generate a standardized real-time data pool. The operating condition-driven preprocessing module is used to filter the data in the standardized real-time data pool based on the current operating condition of the unit, remove outliers and redundant data, perform noise reduction and normalization processing on the filtered data, extract core control features, and generate a high-quality feature dataset. The multi-dimensional fusion computing module is used to input the high-quality feature dataset into the intelligent computing engine, calculate and obtain the physical coupling relationship parameters between each core control feature through the mechanism model, calculate and obtain the predicted trend value of each core control feature through the recurrent neural network, fuse the physical coupling relationship parameters and the predicted trend value, perform control parameter optimization calculation, and generate the optimal control parameters. The closed-loop control command module is used to perform safety verification on the optimal control parameters, generate control commands and send them to the DCS underlying controller through a secure data channel, and the DCS underlying controller executes the corresponding control operation.

Images