A data-driven model and a method and system for improving the controlled management of operationally constrained assets.
Through a systematic quality verification and continuous round counting mechanism, the candidate assets of the thermal power unit optimization system can be stably operated when the on-site operating conditions change. This solves the problem of decreased optimization system hit rate caused by improper model updates in the existing technology, and realizes the stability and security of the optimization system.
Patent Information
- Application Number
- CN202611277268.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-21
- Publication Date
- 2026-11-17
Smart Images

Figure CN123115918A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of cold-end optimization control technology for thermal power units, specifically relating to a frequency optimization control method and system for air-cooled fans in thermal power units. Background Technology
[0002] Thermal power unit optimization systems typically rely on assets such as back pressure prediction models, heat consumption models, wind turbine power curves, back pressure boundaries, and frequency boundaries. These assets determine the optimization algorithm's understanding of the current operating conditions, its evaluation of candidate frequency combinations, and its adherence to field constraints.
[0003] Field operating conditions vary with seasons, ambient temperature, equipment status, vacuum pump status, and operating strategies. If the model and boundary conditions are not updated for a long period, the optimization system may gradually deviate from the actual field characteristics, manifesting as increased prediction errors, a decrease in the proportion of positive-return suggestions, or a recommendation frequency inconsistent with operating habits.
[0004] On the other hand, directly deploying a newly trained model or boundary-life model online also carries risks. There might be too few training samples, insufficient effective days, uneven load coverage, an excessively high proportion of low-load samples, or long-term frozen turbine frequencies leading to a lack of adjustment information in the samples. In these cases, even if a candidate model performs well on local samples, it may not be suitable for replacing operational assets.
[0005] In existing thermal power unit optimization systems, the updating methods for models and operational constraint assets lack systematic quality verification, continuous confirmation, pre-upgrade observation, and post-upgrade rollback mechanisms. This results in candidate assets being deployed immediately upon generation. Under continuously changing on-site operating conditions, the hit rate of the optimization system is easily reduced due to poor training data quality or insufficient model accuracy. Furthermore, problems that occur after deployment cannot be automatically recovered, affecting the stability and safety of on-site closed-loop control. Summary of the Invention
[0006] This invention provides a frequency optimization control method and system for air-cooled fans in thermal power units. It can at least solve the technical problems in existing thermal power unit optimization systems, such as the lack of systematic quality verification, continuous confirmation, pre-upgrade observation, and post-upgrade rollback mechanisms for updating models and operational constraint assets. This leads to the immediate deployment of candidate assets, which, under continuously changing field conditions, can easily reduce the hit rate of the optimization system due to poor training data quality or insufficient model accuracy. Furthermore, problems that occur after deployment cannot be automatically recovered, affecting the stability and safety of the field closed-loop control.
[0007] Includes the following steps: Step S1: Obtain and clean the running samples to form a candidate training dataset; Step S2: Train based on the candidate training dataset to generate candidate models and candidate runtime constraint assets; Step S3: Perform a model acceptance step on the candidate model. The model acceptance includes at least: accuracy acceptance and improvement ratio acceptance relative to the current official model. Step S4: Perform a data quality acceptance step on the candidate training dataset. The data quality acceptance includes at least five indicators: number of valid samples, number of valid days, load segment coverage, low load ratio, and frequency freeze ratio. Step S5: Count the consecutive rounds of model acceptance and data quality acceptance. When the number of consecutive passes reaches the preset threshold, enter the state to be improved. Step S6: After entering the upgrade state, read the closed-loop feedback data on site and calculate the optimization hit rate. If the feedback data is insufficient or the hit rate is invalid, skip this upgrade. Step S7: Perform backup and replacement on the candidate assets that have passed the enhancement threshold; Step S8: After replacement, continuously monitor and optimize the hit rate. If the hit rate drops below the preset threshold, automatically roll back. If no rollback is triggered within the preset observation period, the controlled improvement is completed.
[0008] In one optional implementation, the accuracy acceptance in the model acceptance is to determine that the average absolute error of the candidate model on the validation sample set is not greater than a preset absolute error threshold. The improvement ratio acceptance is determined by judging that the error reduction ratio of the candidate model relative to the current official model is not less than a preset minimum improvement ratio.
[0009] In an optional implementation, the model acceptance in step S3 further includes a stratified sample degradation verification step: Construct historical baseline datasets and recent datasets respectively, calculate the error of the candidate model on the baseline dataset, and determine that the error of the candidate model on the baseline dataset is not greater than a preset multiple of the error of the current official model on the same baseline dataset. The preset multiple is greater than 1, which is used to prevent the candidate model from improving on recent samples but degrading too much on historical baseline samples.
[0010] In one optional implementation, the frequency freeze ratio is specifically calculated by calculating the average frequency of all fans at each sampling time point; Then calculate the absolute value of the average frequency change at adjacent sampling time points; The percentage of sampling points whose absolute value of the average frequency change is less than the preset freezing threshold is calculated.
[0011] Suppose there are K wind turbines in total, and there are M consecutive sampling time points in a certain time period. Obtain the average frequency of all wind turbines at each sampling time point, and obtain the absolute value of the average frequency change between adjacent sampling time points. The number of sampling points whose absolute value of the statistical average frequency change is less than the preset freezing threshold is used to divide the number of sampling points obtained by statistics by (M-1) to obtain the frequency freezing ratio.
[0012] In one alternative implementation, the improvement of the wind turbine power curve needs to meet the following condition: Each wind turbine participating in the fitting has at least a preset number of valid samples per turbine, covers at least a preset number of frequency sub-segments, and the correlation coefficient between the turbine frequency and the turbine current is not lower than a preset threshold.
[0013] In one alternative implementation, the back pressure upper limit table and the wind turbine frequency band table are slow-updating assets; The candidate generation of the slow-updating assets is controlled by an independent minimum candidate generation interval, and their formal promotion is controlled by an independent minimum promotion interval. The method for determining the upper limit of back pressure is as follows: historical operating data are grouped according to unit load range and ambient temperature range, and the preset quantile of the back pressure value in each group is taken as the upper limit of back pressure under that operating condition.
[0014] In one alternative implementation, after the automatic rollback is triggered in step S8, a cooling lock period is entered, during which any automatic promotion operation is prohibited. After the cooling lock period ends, the system will automatically reset to the initial state and clear the consecutive pass count to zero. If a rollback is triggered a preset number of times within a preset time window, the circuit breaker will be locked and can only be released after manual confirmation.
[0015] In one optional implementation, the candidate back pressure prediction model is established using a BP neural network or a least squares support vector machine, and its input variables include at least the unit load, ambient temperature, fan average frequency and condenser temperature, and the output variable is the condenser back pressure. The candidate heat consumption model is established using extreme learning machine or support vector regression. Its input variables include at least unit load, main steam parameters, reheat steam temperature and condenser back pressure, and the output variable is the turbine heat consumption rate.
[0016] In one optional implementation, the field closed-loop feedback data is the actual execution result data returned by the data acquisition system after the optimization system sends the recommended control scheme to the field, including at least the actual wind turbine frequency, actual back pressure value, and actual unit load.
[0017] In one optional implementation, after the post-improvement observation is passed, the recently cleaned data in the candidate training dataset is merged into the historical baseline of the training pool as a long-term data basis for subsequent candidate training. The upgrade report includes at least the backup path, a list of upgraded assets, and a snapshot of observations before the upgrade, which is used to trace the reason for each automatic update and the scope of assets.
[0018] In another aspect, a frequency optimization control system for air-cooled fans in thermal power units is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described in any of the above.
[0019] The method provided in this embodiment of the invention has at least the following beneficial effects: This invention's embodiments ensure that candidate models achieve usable accuracy levels while offering substantial improvements over current models through model acceptance, preventing performance degradation. Data quality acceptance ensures training data possesses sufficient statistical significance, representativeness of operating conditions, and sufficient stimulus information, reducing the risk of deployment due to poor-quality samples from the outset. A continuous round-counting mechanism filters out accidental passes; only candidate assets that pass all validations multiple times consecutively enter the improvement process, reducing the probability of erroneous improvements due to data fluctuations. A pre-improvement observation mechanism prevents blind replacement when on-site feedback data is insufficient, ensuring sufficient information for improvement decisions. A backup replacement and post-improvement observation / rollback mechanism forms a closed-loop operation and maintenance system. If the optimization hit rate significantly decreases after improvement, the system automatically rolls back to the pre-improvement asset, eliminating the need for manual investigation and recovery, ensuring on-site closed-loop stability. By merging recently cleaned data into the training pool baseline after observation, continuous evolution of the training pool is achieved, enabling subsequent model training to incorporate new on-site operating conditions. Attached Figure Description
[0020] Figure 1 A schematic diagram of the data-driven model and the controlled improvement method for operational constraints of assets provided by this invention. Detailed Implementation
[0021] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.
[0022] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this invention is for describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.
[0023] Please see Figure 1 This invention provides a data-driven model and a method for improving the controlled management of operationally constrained assets, comprising the following steps: Step S1: Read the running samples from the field operation data or training pool, and perform cleaning operations on the running samples to remove samples with missing key fields and abnormal working conditions to form a candidate training dataset. Step S2: Train based on the candidate training dataset to generate candidate models and candidate operating constraint assets; the candidate models include candidate back pressure prediction models and candidate heat consumption models, and the candidate operating constraint assets include candidate wind turbine power curves, candidate back pressure upper limit tables, and candidate wind turbine frequency band tables. Step S3: Perform a model acceptance step on the candidate model to obtain model acceptance. The model acceptance step includes: calculating the mean absolute error of the candidate model on the validation sample set and determining that the mean absolute error is not greater than a preset absolute error threshold; and, when the current official model exists, calculating the error reduction ratio of the candidate model relative to the current official model and determining that the error reduction ratio is not less than a preset minimum improvement ratio. Step S4: Perform a data quality acceptance step on the candidate training dataset to obtain the data quality acceptance. The data quality acceptance step includes: determining whether the number of valid samples is not less than the preset minimum number of samples, whether the number of valid days is not less than the preset minimum number of days, whether the number of covered load segments is not less than the preset minimum number of load segments, whether the proportion of low load samples is not greater than the preset maximum low load proportion, and whether the frequency freeze ratio is not greater than the preset maximum freeze ratio. The frequency freeze ratio is calculated by the proportion of sample points where the average frequency change of the fan between consecutive sampled samples is less than the preset freeze threshold to the total number of sample points. Step S5: Count consecutive rounds of model acceptance and data quality acceptance. When both model acceptance and data quality acceptance pass in the same candidate generation round, the consecutive pass count is incremented by one. If either model acceptance or data quality acceptance fails in any candidate generation round, the consecutive pass count is reset to zero. When the consecutive pass count reaches the preset consecutive pass count threshold, the status is updated to pending improvement status and is regarded as the candidate model and candidate operational constraint assets passing the improvement threshold. Step S6: Before the formal upgrade, read the on-site closed-loop feedback data and calculate the optimization hit rate. If the number of samples of the on-site closed-loop feedback data is insufficient or the optimization hit rate is invalid, skip this upgrade and keep it in the upgrade-waiting state. If the upgrade-waiting state continues to exceed the preset observation timeout and no valid on-site closed-loop feedback data can be obtained, automatically revert the state to the initial state and clear the consecutive pass count to zero. Step S7: For candidate models and candidate operational constraint assets that have passed the threshold improvement, first copy the current official assets to the backup directory with timestamps, then copy the candidate assets to the official operation directory to replace them, and record the improvement report. Step S8: After the upgrade, continuously read the closed-loop feedback data on site. If the current optimized hit rate decreases by a percentage relative to the baseline hit rate before the upgrade, reaching or exceeding the preset rollback percentage threshold, then automatically roll back to the assets before the upgrade according to the backup directory in the upgrade report. If no rollback conditions are triggered within the preset observation period, then the post-upgrade observation is confirmed to be passed, and the controlled upgrade of the candidate model and the operational constraint assets of the current round is completed. The baseline hit rate before the upgrade is the statistical value of the optimized hit rate within the preset duration before the upgrade.
[0024] The method provided in this embodiment of the invention has at least the following beneficial effects: This invention's embodiments ensure that candidate models achieve usable accuracy levels while offering substantial improvements over current models through model acceptance, preventing performance degradation. Data quality acceptance ensures training data possesses sufficient statistical significance, representativeness of operating conditions, and sufficient stimulus information, reducing the risk of deployment due to poor-quality samples from the outset. A continuous round-counting mechanism filters out accidental passes; only candidate assets that pass all validations multiple times consecutively enter the improvement process, reducing the probability of erroneous improvements due to data fluctuations. A pre-improvement observation mechanism prevents blind replacement when on-site feedback data is insufficient, ensuring sufficient information for improvement decisions. A backup replacement and post-improvement observation / rollback mechanism forms a closed-loop operation and maintenance system. If the optimization hit rate significantly decreases after improvement, the system automatically rolls back to the pre-improvement asset, eliminating the need for manual investigation and recovery, ensuring on-site closed-loop stability. By merging recently cleaned data into the training pool baseline after observation, continuous evolution of the training pool is achieved, enabling subsequent model training to incorporate new on-site operating conditions.
[0025] Step S1: Run sample loading and cleaning.
[0026] In step S1, running samples are read from the field operation data or training pool, and a cleaning operation is performed on the running samples to remove samples with missing key fields and abnormal operating conditions, forming a candidate training dataset.
[0027] Specifically, data can be read from sample files, databases, or training pools. If a training pool is used, historical baseline samples are merged with recent running samples by timestamp, and duplicate time points are removed. The key fields must include at least the following essential fields for model training: unit load, back pressure, turbine frequency, turbine current, and ambient temperature; samples lacking any of these fields will be discarded. Abnormal operating condition samples include those corresponding to turbine frequencies exceeding normal operating ranges, abnormal turbine currents, abnormal back pressures, or operating states unsuitable for training. Anomaly detection can employ statistical methods, such as classifying values exceeding the mean ± 3 standard deviations as abnormal, or based on upper and lower limits set by equipment operating procedures. The cleaned candidate training dataset serves as the basis for generating all subsequent candidate assets.
[0028] In a preferred embodiment of the present invention, the data sampling time interval is 1 minute, and the training pool is stored in time partitions to facilitate the subsequent extraction of historical baselines and recent data by time window.
[0029] In step S2, training is performed based on the candidate training dataset to generate candidate models and candidate operational constraint assets. The candidate models include candidate back pressure prediction models and candidate heat dissipation models, and the candidate operational constraint assets include candidate wind turbine power curves, candidate back pressure upper limit tables, and candidate wind turbine frequency band tables.
[0030] Specifically, the generation methods for various candidate assets are as follows: The candidate back pressure prediction model is established using a BP neural network or a least squares support vector machine. Its input variables include at least unit load, ambient temperature, average fan frequency, and condenser temperature, with the output variable being the condenser back pressure. During the data preprocessing stage, the input and output data are normalized and mapped to the [0,1] interval. If a BP neural network is used, the number of hidden layer nodes is determined through trial and error, preferably twice the number of input nodes plus one. The activation function is the Sigmoid function, and the Levenberg-Marquardt algorithm is used for network training.
[0031] Candidate heat consumption models were established using Extreme Learning Machine (ELM) or Support Vector Regression (SVR). The input variables included at least unit load, main steam temperature, main steam pressure, reheat steam temperature, and condenser back pressure; the output variable was the turbine heat rate. Data preprocessing employed the 3σ criterion to remove outliers, Z-score standardization for the remaining data, and 5-fold cross-validation to evaluate the model's generalization ability. Candidate turbine power curves were fitted using piecewise least squares. The independent variable was turbine frequency, and the dependent variable was turbine power or turbine current. Frequency segments were divided at 2Hz intervals from the lowest to the highest operating frequency. Linear or quadratic polynomial fitting was performed within each frequency segment, with the fitting criterion being minimizing the sum of squared residuals between the predicted and measured power. The candidate back pressure upper limit table grouped historical operating data according to unit load and ambient temperature segments. Within each group, the 90th or 95th quantile of back pressure was calculated as the upper limit value for that operating condition. The upper limit values for back pressure under each operating condition were then compiled into the upper limit table.
[0032] The candidate wind turbine frequency band table groups historical operating data according to unit load range and ambient temperature range. Within each load-temperature group, the distribution range of wind turbine frequency during historical operation is statistically analyzed. The 10th percentile of the frequency distribution is taken as the lower limit and the 90th percentile as the upper limit to form the recommended frequency band for that operating condition. After the candidate assets are generated, they are first saved in the candidate directory and do not directly overwrite the formal operating directory.
[0033] In step S3, a model acceptance step is performed on the candidate model to obtain model acceptance. The model acceptance step includes two judgments. The first judgment is an absolute accuracy threshold: the mean absolute error of the candidate model on the validation sample set is calculated, and it is determined that the mean absolute error is not greater than a preset absolute error threshold. The validation sample set is a hold-out set partitioned from the candidate training dataset, which does not overlap with the training samples, and is used to evaluate the model's generalization ability. The preferred ratio of the training set to the validation set is 7:3. This threshold ensures that the candidate model at least reaches a usable accuracy level.
[0034] The second criterion is a relative improvement threshold. When a current official model exists, the error reduction ratio of the candidate model relative to the current official model is calculated, and it is determined that this error reduction ratio is not less than a preset minimum improvement ratio. The error reduction ratio refers to the ratio obtained by subtracting the average absolute error of the candidate model from the average absolute error of the current official model on the validation sample set, and then dividing by the average absolute error of the current official model on the validation sample set. This condition is used to ensure that the candidate model does not merely reach a usable level, but rather has a substantial improvement relative to the current model. In a preferred embodiment of the present invention, the minimum improvement ratio is 5%, that is, the candidate model error must be reduced by at least 5% compared to the current model to pass acceptance. This threshold can be adjusted within the range of 3% to 10% according to the actual requirements for model accuracy on-site.
[0035] In step S4, a data quality acceptance step is performed on the candidate training dataset to obtain data quality acceptance results. The data quality acceptance step includes the following judgments: The system determines whether the number of valid samples is not less than a preset minimum number of samples. Valid samples refer to the number of rows of samples retained after cleaning in step S1 that can be used for model training. In a preferred embodiment of the invention, the minimum number of samples is 5000 rows, which can be adjusted within the range of 2000 to 10000 rows. This threshold ensures that there are sufficient statistical samples for model training.
[0036] The system determines whether the number of valid days is not less than a preset minimum number of days. The number of valid days refers to the number of natural days covered by the candidate training dataset. In a preferred embodiment of the invention, the minimum number of days is 7 days, which can be adjusted within the range of 5 to 15 days. This threshold ensures that the data covers multiple operating days, preventing the model's characteristics from being dominated by special operating conditions on a single day.
[0037] The system determines whether the number of covered load segments is not less than a preset minimum number of load segments. Load segments are divided into 10% intervals based on the rated load, for example, five load segments: 50%~60%, 60%~70%, 70%~80%, 80%~90%, and 90%~100%. In a preferred embodiment of the invention, the minimum number of load segments is three, requiring coverage of at least three consecutive load segments. This threshold ensures that the candidate model is representative under different load conditions.
[0038] The system determines whether the proportion of low-load samples is not greater than a preset maximum low-load proportion. Low load refers to less than 50% of the rated load. In a preferred embodiment of the invention, the maximum low-load proportion is 30%, which can be adjusted within the range of 20% to 40%. An excessively high proportion of low-load samples will cause the model to be biased towards low-load characteristics, affecting the model accuracy under medium- and high-load conditions.
[0039] The frequency freeze ratio is determined to be no greater than a preset maximum freeze ratio. The frequency freeze ratio is calculated as the proportion of sample points where the average frequency change of the fan between consecutive samples is less than a preset freeze threshold, relative to the total number of sample points. An excessively high frequency freeze ratio indicates that the fan has not been adjusted for a long time, and the mapping relationship between frequency and back pressure / power in the samples lacks effective excitation, making it unsuitable for model training. In a preferred embodiment of the invention, the maximum freeze ratio is 40%, which can be adjusted within the range of 30% to 50%.
[0040] The above five indicators together constitute the data quality threshold. Only when all five indicators are met can the data quality acceptance be considered passed in this round.
[0041] In step S5, the model acceptance and the data quality acceptance are counted in consecutive rounds.
[0042] When both the model acceptance and the data quality acceptance are passed in the same candidate generation round, the consecutive pass count is incremented by one. If either the model acceptance or the data quality acceptance fails in any candidate generation round, the consecutive pass count is reset to zero. When the consecutive pass count reaches a preset consecutive pass count threshold, the status is updated to "pending improvement," and the candidate model and the candidate operational constraint asset are considered to have passed the improvement threshold.
[0043] This continuous pass mechanism can reduce errors caused by occasional data fluctuations. If a round of verification fails, the continuous pass count is immediately reset to zero, and the candidate asset needs to accumulate continuous pass counts again before entering the improvement process. In a preferred embodiment of the invention, the continuous pass count threshold is 3 times.
[0044] The system maintains an automatically updated status file that records the number of consecutive passes and the current system status.
[0045] In step S6, before the formal upgrade, the field closed-loop feedback data is read to calculate the optimization hit rate. Field closed-loop feedback data refers to the actual execution result data returned by the data acquisition system after the optimization system sends the recommended control scheme to the field. This data includes at least the actual wind turbine frequency, actual back pressure value, and actual unit load. This data is used to evaluate the actual optimization effect of the current formal model in the field.
[0046] The optimized hit rate is defined as follows: the number of control schemes recommended by the optimization system and adopted on-site, divided by the total number of control schemes recommended by the optimization system, and then multiplied by 100%. Here, the control scheme recommended by the optimization system refers to the optimal frequency combination suggestion calculated by the optimization system based on the current operating conditions at a certain point in time; adoption on-site means that the suggestion has been confirmed and implemented by operators, or automatically executed by the closed-loop control system. The preferred observation period is 1 hour, i.e., the hit rate is calculated once per hour.
[0047] If the number of samples of the on-site closed-loop feedback data is insufficient or the optimized hit rate is invalid, the current improvement is skipped and the system remains in the pending improvement state. If the pending improvement state continues for more than the preset observation timeout and no valid on-site closed-loop feedback data can be obtained, the state is automatically rolled back to the initial state and the consecutive pass count is reset to zero. This design avoids unnecessary delays caused by insufficient single observation data requiring the re-accumulation of consecutive pass counts, and also prevents the system from deadlocking when observation data is unavailable for an extended period. In a preferred embodiment of the invention, the observation timeout is 7 days.
[0048] In step S7, for the candidate models and candidate running constraint assets that have passed the improvement threshold, the current official assets are first copied to the backup directory with timestamps, and then the candidate assets are copied to the official running directory for replacement, and an improvement report is recorded.
[0049] Backup and replacement operations are performed in the order of backup first, then replacement. Replacement operations use a method of first writing to a temporary directory and then atomically renaming the files to ensure that asset files in the production directory are never incomplete. Detailed logs are recorded for each operation to facilitate manual troubleshooting.
[0050] The upgrade report should include at least the backup path, a list of upgraded assets, and a snapshot of observations before the upgrade, to trace the reason for each automatic update and the scope of assets.
[0051] In step S8, the on-site closed-loop feedback data is continuously read after the upgrade. If the decrease in the current optimized hit rate relative to the baseline hit rate before the upgrade reaches or exceeds a preset rollback ratio threshold, the system automatically rolls back to the assets before the upgrade according to the backup directory in the upgrade report. The baseline hit rate before the upgrade is the statistical value of the optimized hit rate within a preset time period before the upgrade.
[0052] If no rollback condition is triggered within the preset observation period, the post-improvement observation is confirmed as passed, and the controlled improvement of the candidate model and the operational constraint asset in the current round is completed. In a preferred embodiment of the present invention, the post-improvement observation period is 1 hour, and the preset observation period is 7 consecutive periods, that is, the post-improvement observation is determined to be passed only after 7 consecutive hours of observation without triggering the rollback condition.
[0053] After the improved data passes the observation, the recently cleaned data in the candidate training dataset is merged into the historical baseline of the training pool, serving as the long-term data foundation for subsequent candidate training. This step allows the training pool to incorporate new on-site conditions while preventing unconfirmed data from directly becoming the baseline.
[0054] In one optional implementation, the accuracy acceptance in the model acceptance is to determine that the average absolute error of the candidate model on the validation sample set is not greater than a preset absolute error threshold. The improvement ratio acceptance is determined by judging that the error reduction ratio of the candidate model relative to the current official model is not less than a preset minimum improvement ratio.
[0055] During candidate model training, the candidate training dataset needs to be divided into a training set and a validation set, preferably in a 7:3 ratio. The training set is used to train the model parameters, and the validation set is used to evaluate the model's generalization ability on unseen data. The mean absolute error (MAE) of the candidate model on the validation set is calculated; this metric directly reflects the prediction accuracy of the candidate model on new data. If this error is not greater than a preset absolute error threshold, it indicates that the candidate model has at least reached a usable accuracy level. The significance of this threshold is that regardless of the current state of the official model (e.g., it may be a manually initialized model or a severely outdated model), the candidate model must meet a minimum accuracy standard to be considered. For example, in a preferred embodiment of the present invention, the maximum permissible absolute error threshold for the backpressure prediction model is 0.5 kPa, and the maximum permissible absolute error threshold for the heat dissipation model is 50 kJ / kWh.
[0056] This invention ensures the basic usability of candidate models. The absolute accuracy threshold mandates that candidate models meet a minimum prediction accuracy standard. Regardless of the system's state, no model with insufficient accuracy will be deployed, thus guaranteeing that the optimization system will not cause recommendations to fail under any circumstances due to the model's inherently low accuracy.
[0057] In an optional implementation, the model acceptance in step S3 further includes a stratified sample degradation verification step: Construct historical baseline datasets and recent datasets separately, calculate the error of the candidate model on the baseline dataset, and determine that the error of the candidate model on the baseline dataset is not greater than a preset multiple of the error of the current official model on the same baseline dataset. The preset multiple is greater than 1 to prevent the candidate model from improving on recent samples but degrading too much on historical baseline samples.
[0058] Specifically, a historical baseline dataset and a recent layer dataset are constructed. The historical baseline dataset consists of historical stable operating condition samples accumulated over a long period in the training pool, while the recent layer dataset consists of running samples from the most recent period (e.g., the last 30 days). The error of the candidate model on the baseline dataset is calculated, and it is determined that the error of the candidate model on the baseline dataset is not greater than a preset multiple of the error of the current official model on the same baseline dataset, where the preset multiple is greater than 1.
[0059] This design allows for slight degradation of the candidate model at the baseline level, but the extent of degradation must be controlled within a preset factor. For example, a preset factor of 1.03 means that the candidate model's error at the baseline level must not exceed 103% of the current model's error. This verification is used to prevent the candidate model from overfitting to recent data and compromising the accuracy of the model under historical stable conditions.
[0060] While pursuing the improvement of candidate models in fitting recent operating conditions, this embodiment of the invention also constrains the extent of performance degradation under historical stable operating conditions, preventing the model from neglecting one aspect for another. It absorbs recent changes in the field while maintaining stability under historical operating conditions.
[0061] In one optional implementation, the frequency freezing ratio R_f is calculated as follows: Assume there are K wind turbines in total, and M consecutive sampling time points in a certain time period. First, calculate the average frequency of all wind turbines at each sampling time point, then calculate the absolute value of the average frequency change at adjacent sampling time points, count the number of sampling points whose absolute value of the average frequency change is less than the preset freezing threshold, and divide the counted number of sampling points by (M-1) to obtain the frequency freezing ratio; the sampling time interval is 1 minute, and the freezing threshold is 0.1 Hz.
[0062] The frequency freeze ratio of this invention reflects the proportion of time a wind turbine is in an unregulated state within a dataset. When the wind turbine frequency remains constant for a long period, the mapping relationship between frequency and back pressure / power lacks effective excitation, and such samples have low information value for model training. By quantifying this ratio and incorporating it into the data quality acceptance threshold, training data with sufficient regulation information can be effectively screened, avoiding overfitting or underfitting of the model on frozen data.
[0063] In one alternative implementation, the improvement of the wind turbine power curve needs to meet the following conditions: each wind turbine participating in the fitting has a number of valid samples that are no less than the preset number of single-unit samples, the number of frequency sub-segments covered is no less than the preset number of sub-segments, and the correlation coefficient between the wind turbine frequency and the wind turbine current is no less than the preset threshold.
[0064] Specifically, the power curve fitting object is a single wind turbine. Since the performance of each wind turbine may differ, a power curve needs to be fitted independently for each wind turbine. The effective sample size threshold for each wind turbine ensures that the amount of data for a single wind turbine is sufficient (preset data threshold); the frequency bin coverage ensures that the wind turbine has sample coverage across its entire operating frequency range to avoid the risk of extrapolation (preset frequency threshold); the correlation coefficient threshold between frequency and current ensures that there is a sufficient linear or non-linear correlation between frequency and power / current. If the correlation coefficient is too low, it indicates that the data noise is too high or the wind turbine is operating abnormally, and curve fitting is not suitable.
[0065] The wind turbine power curves of this invention are the fundamental assets for optimizing the energy consumption of different frequency combinations in the system evaluation, and their accuracy directly affects the reliability of the optimization recommendations. Through triple constraints—the number of single-unit samples, the number of frequency coverage sub-boxes, and the correlation coefficient—it is ensured that the power curve of each wind turbine is fitted based on sufficient, comprehensive, and statistically correlated data, avoiding curve distortion due to insufficient or poor data quality.
[0066] In one alternative implementation, the back pressure upper limit table and the wind turbine frequency band table are slow-updating assets; The generation of candidates for slow-updating assets is controlled by an independent minimum candidate generation interval, and their formal promotion is controlled by an independent minimum promotion interval. The method for determining the back pressure upper limit table is as follows: historical operating data are grouped according to unit load range and ambient temperature range, and the preset quantile of the back pressure value in each group is taken as the back pressure upper limit value under that operating condition.
[0067] In this invention, operating assets are categorized into fast-update assets and slow-update assets. Models, as fast-update assets, can be updated more frequently; wind turbine power curves are medium-update assets; and back pressure upper limit tables and frequency band tables are slow-update assets. Slow-update assets require additional time thresholds for candidate generation intervals and improvement intervals. This hierarchical design avoids frequent changes in boundary-type assets due to short-cycle sample fluctuations.
[0068] This invention achieves hierarchical management of assets with different update frequencies by classifying the backpressure upper limit table and frequency band table as slow-update assets and implementing independent time interval control. The backpressure upper limit and frequency band are operational constraint boundaries, and their changes should reflect long-term trends rather than short-term fluctuations. Independent time interval control prevents frequent changes in boundary assets due to short-cycle sample fluctuations, maintaining the stability of the operational boundaries.
[0069] In one optional implementation, the minimum candidate generation interval for slow-updating assets is 15 days, the minimum promotion interval is 30 days, and the threshold for the number of valid data days used to generate slow-updating assets is 15 days; when the candidate generation or formal promotion of a slow-updating asset is skipped because the minimum interval is not met, the number of consecutive passes is not affected.
[0070] Specifically, the minimum candidate generation interval of 15 days means that there must be at least 15 days between the generation of candidate assets for two consecutive slow update assets; the minimum promotion interval of 30 days means that there must be at least 30 days between the promotion of official assets for two consecutive slow update assets; and the effective data days threshold of 15 days means that the training data used to generate slow update assets must cover at least 15 effective running days. When the generation of candidate assets for slow update assets is skipped due to a 15-day interval or the promotion of official assets is skipped due to a 30-day interval, the skipping does not affect the count of consecutive passes, and the consecutive passes of fast update assets remain valid.
[0071] This invention quantifies the update rhythm of slow-updating assets by providing specific time parameters. First, the specific time parameters make the update rhythm of slow-updating assets predictable and traceable. Second, the dual interval design of 15 days and 30 days takes into account the response speed and stability of boundary assets to long-term trends. Third, the design of skipping without affecting the number of consecutive passes avoids the time rhythm of slow assets dragging down the update efficiency of fast assets, thus achieving decoupled management of assets with different update rhythms.
[0072] In one optional implementation, the preset threshold for consecutive passes is 3 times; the preset observation timeout is 7 days; the improved observation period is 1 hour, and the preset observation period is 7 consecutive periods.
[0073] Specifically, the consecutive pass threshold of 3 times means that candidate assets must pass model acceptance and data quality acceptance for 3 consecutive rounds before entering the upgrade waiting state. The observation timeout of 7 days means that if no valid on-site closed-loop feedback data is obtained for 7 consecutive days after entering the upgrade waiting state, the system will automatically roll back to the initial state. The post-upgrade observation period of 1 hour and 7 consecutive periods means that on-site closed-loop feedback data is read every 1 hour after the upgrade to calculate the hit rate, and the upgrade is confirmed as passed only after 7 consecutive times when the rollback condition is not triggered.
[0074] First, the three consecutive passes strike a balance between filtering out random fluctuations and avoiding excessive delays; second, the 7-day observation timeout prevents the system from deadlocking when data is unavailable for an extended period; third, the 1-hour / 7-cycle observation scheme strikes a balance between timely problem detection and sufficient observation.
[0075] In one alternative implementation, after the automatic rollback is triggered in step S8, a cooling lock period is entered, during which any automatic promotion operation is prohibited. After the cooldown lock period ends, it will automatically reset to the initial state and the consecutive pass count will be cleared to zero; If a rollback is triggered a preset number of times within a preset time window, the circuit breaker will be locked and can only be released after manual confirmation.
[0076] Specifically, the cooling-off period is a grace period during which all automatic promotion operations are suspended to prevent repeated rollbacks caused by attempting to promote again immediately after a rollback. After the cooling-off period, the system automatically resets and resumes the normal automatic update process without manual intervention. However, if there are multiple consecutive rollbacks within a short period, it indicates a potential systemic problem. In this case, the system enters a circuit breaker lockout state, and it can only be reactivated after maintenance personnel have investigated the problem and manually confirmed it.
[0077] This invention establishes a tiered fault handling mechanism, ranging from single rollback, cooling lock, and automatic recovery to multiple rollbacks, circuit breaker lock, and manual intervention. The cooling lock period avoids jittery repeated boosting and rollbacks after rollback, ensuring system stability; the circuit breaker lock mechanism promptly brakes during continuous multiple rollbacks, preventing the system from endlessly attempting to repeat the process when there are fundamental problems, thus protecting on-site operational safety; the tiered handling balances automation efficiency and safety.
[0078] In one alternative implementation, the cooling lock-in period lasts for 24 hours; the preset time window is 30 days, and the preset number of times is 3.
[0079] In one optional implementation, the candidate back pressure prediction model is established using a BP neural network or a least squares support vector machine. Its input variables include at least the unit load, ambient temperature, average fan frequency, and condenser temperature, and the output variable is the condenser back pressure. Candidate heat consumption models are established using extreme learning machine or support vector regression. Their input variables include at least unit load, main steam parameters, reheat steam temperature, and condenser back pressure, and the output variable is the turbine heat consumption rate.
[0080] This invention employs BP neural network / LSSVM and ELM / SVR, both of which are mature data-driven modeling methods with good nonlinear fitting ability and generalization performance, making them suitable for modeling complex thermal systems of thermal power units. The clear definitions of input and output variables enable those skilled in the art to implement this invention without creative effort. The input variables cover the main factors affecting back pressure and heat consumption, ensuring that the model can accurately reflect the real characteristics on site.
[0081] In one optional implementation, the field closed-loop feedback data is the actual execution result data returned by the data acquisition system after the optimization system sends the recommended control scheme to the field, including at least the actual wind turbine frequency, actual back pressure value, and actual unit load.
[0082] This invention distinguishes on-site closed-loop feedback data from ordinary operational data. It specifically refers to feedback data after the implementation of optimization suggestions, which is the direct basis for evaluating the optimization effect. The clear data content enables those skilled in the art to clearly know which data needs to be collected to support pre-improvement and post-improvement observations. The term "closed-loop" emphasizes that the data comes from the actual implementation of optimization suggestions, rather than open-loop simulation calculations.
[0083] In one alternative implementation, after the post-improvement observation passes, the recently cleaned data in the candidate training dataset is merged into the historical baseline of the training pool as a long-term data basis for subsequent candidate training. The upgrade report should include at least the backup path, a list of upgraded assets, and a snapshot of observations before the upgrade, to trace the reason for each automatic update and the scope of assets.
[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific embodiments of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.
Claims
1. A data-driven model and a method for controlling the improvement of operationally constrained assets, characterized in that, Includes the following steps: Step S1: Obtain and clean the running samples to form a candidate training dataset; Step S2: Train based on the candidate training dataset to generate candidate models and candidate runtime constraint assets; Step S3: Perform a model acceptance step on the candidate model. The model acceptance includes at least: accuracy acceptance and improvement ratio acceptance relative to the current official model. Step S4: Perform a data quality acceptance step on the candidate training dataset. The data quality acceptance includes at least five indicators: number of valid samples, number of valid days, load segment coverage, low load ratio, and frequency freeze ratio. Step S5: Count the consecutive rounds of model acceptance and data quality acceptance. When the number of consecutive passes reaches the preset threshold, enter the state to be improved. Step S6: After entering the upgrade state, read the closed-loop feedback data on site and calculate the optimization hit rate. If the feedback data is insufficient or the hit rate is invalid, skip this upgrade. Step S7: Perform backup and replacement on the candidate assets that have passed the enhancement threshold; Step S8: After replacement, continuously monitor and optimize the hit rate. If the hit rate drops below the preset threshold, automatically roll back. If no rollback is triggered within the preset observation period, the controlled improvement is completed.
2. The method of claim 1, wherein, The candidate models include candidate back pressure prediction models and candidate heat consumption models; The candidate operating constraint assets include candidate wind turbine power curves, candidate back pressure upper limit tables, and candidate wind turbine frequency band tables.
3. The method of claim 1, wherein, In the model acceptance process, the accuracy acceptance is to determine that the average absolute error of the candidate model on the validation sample set is not greater than a preset absolute error threshold. The improvement ratio acceptance is determined by judging that the error reduction ratio of the candidate model relative to the current official model is not less than a preset minimum improvement ratio.
4. The method of claim 1, wherein, The model acceptance also includes stratified sample degradation verification: constructing historical baseline layer datasets and recent layer datasets respectively; The error of the candidate model on the baseline layer dataset is determined to be no greater than a preset multiple of the error of the current official model on the same baseline layer dataset.
5. The method of claim 1, wherein, The specific calculation method for the frequency freeze ratio is as follows: calculate the average frequency of all fans at each sampling time point; Then calculate the absolute value of the average frequency change at adjacent sampling time points; The percentage of sampling points whose absolute value of the average frequency change is less than the preset freezing threshold is calculated.
6. The method of claim 1, wherein, The improvement of the wind turbine power curve must meet the following requirements: the number of effective samples, the number of frequency sub-boxes covered, and the frequency-current correlation coefficient of each wind turbine participating in the fitting must not be lower than their respective preset thresholds.
7. The method of claim 2, wherein, The candidate back pressure prediction model is established using a BP neural network or a least squares support vector machine. Its input variables include at least the unit load, ambient temperature, average fan frequency and condenser temperature, and the output variable is the condenser back pressure. The candidate heat consumption model is established using extreme learning machine or support vector regression. Its input variables include at least unit load, main steam parameters, reheat steam temperature and condenser back pressure, and the output variable is turbine heat consumption rate. The on-site closed-loop feedback data includes at least the actual wind turbine frequency, actual back pressure value, and actual unit load.
8. The method of claim 7, wherein, After the improvement observation is passed, the recent cleaned data in the candidate training dataset is merged into the historical baseline of the training pool as the long-term data basis for subsequent candidate training. The upgrade report includes at least the backup path, a list of upgraded assets, and a snapshot of observations before the upgrade, which is used to trace the reason for each automatic update and the scope of assets.
9. The method according to claim 6, characterized in that, The improvement of the wind turbine power curve must meet the following requirements: the number of effective samples, the number of frequency sub-boxes covered, and the frequency-current correlation coefficient of each wind turbine participating in the fitting must not be lower than their respective preset thresholds.
10. A frequency optimization control system for air-cooled fans in thermal power units, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method as described in any one of claims 1 to 9.
