A data-driven self-learning-based optimal pulse width feedforward control method for hydropower units

By employing a data-driven self-learning optimal pulse width feedforward control method, an independent sample library is constructed and combined with feedforward and feedback control. This solves the problems of full-condition adaptation, response lag, and manual dependence in the regulation and control of hydropower units, achieving efficient, safe, and consistent unit regulation, adapting to unit differences, and possessing continuous optimization capabilities.

CN123073994APending Publication Date: 2026-11-10NANJING HEHAI NANZI HYDROPOWER AUTOMATION
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611585568.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-14
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

The existing hydropower unit regulation and control has problems such as fixed parameter PID which cannot cover the full operating conditions, response lag, unutilized data resources, inconsistent control effects due to differences between units, and reliance on manual experience for parameter tuning. It is difficult to meet the needs of rapid load regulation and the decline in control effect caused by equipment aging.

Method used

An optimal pulse width feedforward control method based on data self-learning is adopted. Through offline database construction, online matching and closed-loop update, an independent historical optimal operating condition sample library is built. The optimal adjustment pulse width is calculated in real time. Combined with feedforward and feedback control, autonomous optimization and safety constraints are achieved to avoid control abrupt changes.

Benefits of technology

It achieves efficient regulation under all operating conditions, shortens regulation time, reduces overshoot, reduces maintenance workload, improves control consistency and safety, adapts to unit differences, has continuous optimization capabilities, and is suitable for rapid grid response.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123073994A_ABST
    Figure CN123073994A_ABST
Patent Text Reader

Abstract

This invention discloses a data-driven self-learning-based optimal pulse width feedforward control method for hydropower units, belonging to the field of active power regulation and control technology for hydropower units. The method includes: collecting historical operating data of hydropower units and establishing an independent historical optimal operating condition sample library for each unit; generating operating condition feature vectors in real time upon receiving load regulation commands; matching K similar historical samples using dynamic weighted Euclidean distance based on the region to which the operating condition belongs; calculating a recommended pulse width using the inverse of the distance weight and performing nonlinear correction based on the current head; sequentially performing amplitude limiting verification, rate of change constraint, vibration zone avoidance, and inertial smoothing on the corrected pulse width to generate the final feedforward pulse width output to the speed control actuator; evaluating the effect after regulation and automatically updating the sample library, forming a closed-loop self-learning process. This invention is applicable to scenarios involving improved load response speed of hydropower units, adaptation to all operating conditions, and differentiated control of multiple units.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of active power regulation and control technology for hydropower units, and in particular to an optimal pulse width feedforward control method for hydropower units based on data self-learning. Background Technology

[0002] Hydropower generator units are the core frequency regulation and peak shaving power sources in the power system. Their regulation response speed and control accuracy directly affect the frequency stability and power balance of the power grid. With the continuous increase in the proportion of new energy sources such as wind power and photovoltaic power connected to the grid, the amplitude of active power fluctuations and the frequency of regulation in the system have increased significantly, which has placed more stringent requirements on the rapid load regulation capability and the stable operation capability under all operating conditions of hydropower units.

[0003] Currently, active power regulation of hydropower units generally adopts a proportional-integral-derivative (PID) closed-loop control architecture: the speed control system collects the actual output power of the unit in real time, calculates the difference between it and the target power issued by the dispatcher to obtain the deviation signal, and then calculates the control quantity based on the pre-tuned PID parameters, outputting the corresponding pulse width adjustment command to drive the guide vane actuator to act, ultimately achieving power tracking. This scheme is mature in principle and reliable in operation, and can meet basic control requirements in stable operating conditions. However, it still has many technical limitations when facing the nonlinear and time-varying operating characteristics of hydropower units.

[0004] First, fixed-parameter PID controllers cannot cover the full range of operating conditions, and parameter tuning is highly dependent on manual experience. The operating head of hydropower units varies significantly with the seasons and reservoir water levels. The head difference between the flood season and the dry season can reach tens of meters for some power stations. The hydraulic inertia and mechanical response characteristics of the units differ significantly under different operating conditions. Existing PID parameters are usually manually tuned through on-site tests for only a few operating conditions such as rated head and typical load. Once the operating conditions deviate from the design conditions, the control performance will deteriorate significantly. Maintaining the control effect across all operating conditions requires maintenance personnel to frequently test and adjust parameters on-site, which is labor-intensive and the tuning effect varies from person to person.

[0005] Secondly, pure feedback control has an inherent lag in response, making it difficult to adapt to high-frequency adjustment scenarios. In scenarios where automatic generation control (AGC) commands are frequently issued and load steps are large, the initial control output is too small and the power response is slow. After the deviation is fully manifested, increasing the adjustment intensity is prone to overshoot oscillation due to hydraulic and mechanical inertia. This not only fails to meet the requirements of rapid response of the power grid, but also aggravates the mechanical wear of the guide vane actuator.

[0006] Third, the massive historical operating data accumulated by the hydropower station monitoring system over a long period of time is only used for post-operation analysis, fault diagnosis and performance statistics. It has not been refined into a control knowledge base that can be directly accessed. Historical experience cannot assist in real-time control decisions, resulting in a waste of data resources.

[0007] Fourth, due to differences in manufacturing and installation errors, hydraulic channels, years of operation, and maintenance, the same type of generating units within the same power station exhibit different regulation response characteristics. Using uniform control parameters will inevitably lead to insufficient regulation performance of some units, making it impossible to achieve optimal control for each unit.

[0008] Fifth, once the parameters of a traditional PID controller are tuned, they are usually kept fixed for a long time. They cannot actively iterate and optimize the control law based on new operating data. As the equipment ages and the components wear out, the control effect will gradually decline, and the parameters can only be manually readjusted periodically.

[0009] In addition, two technical biases have long been prevalent in this field: first, that high-precision load regulation must rely on a precise hydraulic-mechanical coupling mathematical model, and that purely data-driven methods cannot directly support real-time feedforward control; and second, that the tuning of control parameters must rely on manual field tests, and that parameter optimization affected by operating condition fluctuations cannot be completed autonomously without human intervention.

[0010] This common perception has hindered the practical application of self-learning feedforward control technology in the field of hydropower speed regulation. Summary of the Invention

[0011] The purpose of this invention is to provide an optimal pulse width feedforward control method for hydropower units based on data self-learning, so as to solve the above-mentioned technical problems existing in the prior art.

[0012] To achieve the above objectives, the technical solution adopted by this invention is as follows: an optimal pulse width feedforward control method for hydropower units based on data self-learning. The overall technical route unfolds according to the logic of offline database construction, online matching, and closed-loop update, including the following steps:

[0013] Step S1: Collect historical operating data of the hydropower unit. The historical operating data includes at least the unit's basic operating status data, control process data, regulation response result data, and auxiliary marker data.

[0014] Step S2: Clean the historical operating data, classify the cleaned historical regulation processes according to the unit number and regulation direction, select high-quality regulation processes according to unified evaluation indicators, and establish an independent historical optimal operating condition sample library for each unit. Each sample includes at least the operating head, load change deviation, initial guide vane opening, unit operating status evaluation parameters and the corresponding optimal regulation pulse width.

[0015] Step S3: When the dispatching system issues a load adjustment command, the current operating status is collected in real time and a working condition feature vector is generated. ,in To operate the water head, For load variation deviation, The initial guide vane opening, These are parameters for evaluating the operating status of the generating unit.

[0016] Step S4: Based on the current adjustment direction and the region to which the working condition belongs, calculate the weighted distance between the current working condition features and the features of each historical sample using dynamic feature weights, and select K similar historical samples.

[0017] Step S5: Calculate the recommended pulse width from the optimal adjustment pulse width of the K similar historical samples using a distance-inverse weighted method, and perform nonlinear correction by combining the current head with the average head of similar samples.

[0018] Step S6: The corrected pulse width is sequentially subjected to amplitude limiting verification, rate of change constraint, vibration zone avoidance and inertial smoothing processing to generate the final feedforward pulse width and output it to the speed control actuator.

[0019] Step S7: Record the data of the entire adjustment process and calculate the adjustment performance evaluation index; when the evaluation index meets the preset threshold, add the new sample to the sample library, or replace the similar sample with a worse evaluation in the sample library to complete the sample library update.

[0020] By organically linking the above steps, a closed-loop self-learning process is formed, realizing the transformation from error-driven passive adjustment to experience-based predictive active control.

[0021] Furthermore, the unified evaluation index mentioned in step S2 is:

[0022]

[0023] in To adjust performance evaluation indicators, To adjust the time, For the maximum overshoot, To adjust the number of oscillations during the process, For steady-state error, , , , The corresponding weighting coefficients can be adjusted according to the power plant's emphasis on regulation speed and stability. Typical values ​​are 0.4, 0.3, 0.2, and 0.1. A smaller comprehensive evaluation index indicates a better regulation process. Adjustment processes that fall below a preset evaluation threshold are considered high-quality adjustment processes; the sample library is stored separately in two directions, increasing and decreasing load, to adapt to the asymmetric characteristics of the guide vane response of the hydropower unit.

[0024] Preferably, the unit operating status evaluation parameter S is calculated by weighting the equipment operating time, the operating time after maintenance, and the recent average regulation performance degradation rate, with a value ranging from 0 to 1. The closer the value is to 1, the better the equipment status. The typical weights of the three indicators are 0.3, 0.3, and 0.4, respectively. Each unit corresponds to an independent historical sample database.

[0025]

[0026] During online control, the control model of the corresponding unit is invoked according to the unit number, and the control logic of different units is completely decoupled. The architecture of unit-by-unit modeling and independent invocation avoids data interference between different units, enabling the control strategy to accurately match the actual characteristics of a single unit.

[0027] Specifically, the load change deviation mentioned in step S3 is:

[0028]

[0029] in For the target power, The current actual power is used; the operating condition feature vector may also include at least one of the following according to the specific needs of the power station: unit speed, head change rate, grid frequency deviation, and guide vane opening change rate, in order to further improve the accuracy of the operating condition description.

[0030] Feature selection follows the principles of "strong correlation, easy acquisition, and low redundancy," with the core features being the operating head H (reflecting the unit's hydraulic conditions) and load variation deviation. (Reflects the adjustment range), initial guide vane opening G (reflects the mechanical actuation position), and unit operating status evaluation parameter S (reflects the equipment health status).

[0031] Furthermore, in step S4, the entire operating condition is divided into a core operating area, a transition operating area, and a boundary operating area based on the distribution density of historical sample data;

[0032] The core operating area corresponds to the head and load range with the highest daily operating frequency of the unit. The weights of the head and load deviations in this area are 0.4 and 0.35, respectively, and K is 1~2.

[0033] The transition operation zone adopts a balanced weight configuration, with K ranging from 3 to 4; the boundary operation zone corresponds to less frequent operating conditions such as extreme head and ultimate load. In this zone, the weights of the unit status parameters and guide vane opening are 0.25 and 0.3, respectively, with K set to 5. The risk caused by a single sample anomaly is reduced by weighted averaging of more samples.

[0034] When screening samples, samples that are consistent with the current adjustment direction are prioritized. If the number of samples in the same direction is less than half of the K value, samples in the opposite direction are added and the direction coefficient is corrected. The correction coefficient is calibrated according to the difference in response characteristics of the increase and decrease of load, and the typical value is 0.9 to 1.1.

[0035] The division of operating areas is determined by clustering algorithms combined with operational statistics. The boundaries can be dynamically adjusted according to the power plant's operation plan, without the need for manually fixing the division threshold.

[0036] Preferably, the weighted distance in step S4 is calculated using the following formula:

[0037]

[0038] in The number of input features, For the j-th feature of the current working condition, Features corresponding to historical samples The feature weights are calculated as follows: before calculation, each feature is normalized to map its value to the range of 0 to 1, thus eliminating the influence of dimensional differences on distance calculation; when a real-time feature exceeds the historical extreme value range, it is mapped to the interval boundary and marked as a boundary condition, triggering the corresponding matching strategy.

[0039] Further, the recommended pulse width mentioned in step S5 is calculated using the following formula:

[0040]

[0041] in The optimal pulse width corresponds to the i-th historical sample. The distance between the current working condition and the i-th sample. To prevent the correction parameter from being zero in the denominator, a typical value of 0.001 is used. Samples that are closer together have a higher proportion in the calculation, which fully reflects the principle of "nearest neighbor priority". This ensures that the output results match the current working conditions and avoids sudden changes in output caused by a single abnormal sample.

[0042] The nonlinear correction is:

[0043]

[0044] in The revised recommended pulse width. Let K be the average head of similar samples. The current operating head, It is a nonlinear correction index, which can be calibrated according to the actual hydraulic characteristics of the unit. The typical value is 0.4 to 0.6 for mixed-flow units and 0.3 to 0.5 for axial-flow units.

[0045] This correction step is used to compensate for the nonlinear relationship between the turbine flow characteristics and the head, effectively improving the accuracy of pulse width calculation in scenarios with wide head fluctuations.

[0046] Furthermore, the amplitude limiting verification in step S6 is as follows:

[0047] When the recommended pulse width is less than the minimum allowable pulse width, the minimum allowable pulse width is used; when the recommended pulse width is greater than the maximum allowable pulse width, the maximum allowable pulse width is used. This avoids control commands exceeding the safe execution range of the speed regulation system due to abnormal data matching. The rate of change constraint is:

[0048] and

[0049] in The constraint, representing the maximum allowable pulse width variation in a single control cycle, ensures that changes in control commands smoothly align with the dynamic response capability of the guide vane actuator, avoiding adjustment oscillations caused by sudden command changes; the inertial smoothing process is as follows:

[0050]

[0051] in The final output pulse width for the k-th control cycle. The pulse width after amplitude limiting. The output pulse width of the previous cycle. The smoothing coefficient and Larger It can improve response speed and reduce the size It can enhance output smoothness.

[0052] Preferably, the vibration zone avoidance in step S6 is: predicting and adjusting the unit power after the adjustment is completed based on the current power and the recommended pulse width;

[0053] When the predicted power falls within the preset vibration range, if the target power is higher than the upper limit of the vibration range, the pulse width is increased to allow the unit to quickly pass through the vibration range; if the target power is lower than the lower limit of the vibration range, the pulse width is decreased to avoid entering the vibration range.

[0054] When a single adjustment cannot completely avoid the vibration zone, the pulse width is split into two or more steps, with a short transition phase in between, to allow the unit to quickly pass through the vibration zone and reduce the dwell time.

[0055] The vibration range can be set based on the actual test data of the unit, and usually includes upper and lower boundaries and a transition margin. The margin value can be taken as 2% of the rated power. This vibration zone active avoidance design reduces the time the unit stays in the vibration zone from the source in the feedforward stage, which is different from the traditional feedback control method that can only be passively adjusted after entering the vibration zone.

[0056] In some embodiments, the final feedforward pulse width and the original PID feedback loop constitute a composite control with feedforward as the main component and feedback as the auxiliary component. The feedforward control undertakes 80% to 90% of the total adjustment, while the PID feedback loop undertakes the remaining fine-tuning. The feedforward weight is highest in the initial stage of adjustment, and decreases linearly as the power approaches the target value. After entering the steady-state error band, the feedforward output returns to zero, and the steady state is maintained entirely by the feedback loop. When an abnormality in the sample library, sensor data failure, or communication interruption is detected, the feedforward output gradually decays according to the control cycle and switches to pure PID control without disturbance. The switching process is shock-free, ensuring safe operation.

[0057] Furthermore, the sample library update in step S7 includes:

[0058] When there are no historical samples in the sample library that are less than the similarity threshold to the current working condition, a new sample is added directly.

[0059] When similar samples exist, their comprehensive evaluation indicators are compared and the better sample is retained; the matching weight of expired samples gradually decreases over time, and the typical effective period of the sample is 2 years.

[0060] After the unit completes a Class A overhaul or guide vane system modification, the samples before the overhaul are marked as historical reference samples and the matching weight is reduced, while the collection speed of new samples after the overhaul is accelerated.

[0061] When a special update is triggered, the sample collection frequency will be increased. The triggering conditions for a special update include: the operating head crossing the preset partition threshold and entering the head range of infrequent operation; the first load operation stage after the unit has been overhauled; and the comprehensive evaluation index of three consecutive adjustments being lower than the quality threshold.

[0062] As a preferred approach, for units lacking their own historical data, an initial control model is generated by scaling up a sample library of similar units according to rated parameters, or by generating virtual high-quality samples covering all operating conditions using turbine digital simulation data. During the small sample phase, the feedforward output weight is reduced, and PID feedback control is prioritized. As the number of high-quality samples increases, the feedforward weight is gradually increased, and the normal update frequency is restored after the sample size reaches a preset threshold. For new units, the sample update frequency is appropriately increased in the first three months, and the normal update frequency is restored after six months of operation and a sample size exceeding 500. This effectively solves the problem of no usable data for new units and significantly shortens the model adaptation cycle.

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] 1. By constructing a control knowledge base based on historical full-condition data and combining it with an adaptive matching strategy for different operating conditions, the system automatically matches the optimal control pulse width according to the current head, load, and equipment status. This eliminates the dependence on typical operating conditions and significantly improves the adaptability to all operating conditions. In low-head and high-head scenarios that deviate from the rated operating conditions, the settling time is reduced by more than 30%, the overshoot is reduced by more than 50%, and the consistency of performance across all operating conditions is greatly improved.

[0065] 2. By using feedforward predictive control, the control pulse width, which is close to the optimal value, is directly output in the initial stage of regulation, driving the actuator to act in advance. This eliminates the process of feedback control from the generation of deviation to gradual adjustment, greatly speeding up the regulation response and better meeting the assessment requirements of the power grid's rapid response.

[0066] 3. Control parameters are learned autonomously by the system from operational data, eliminating the need for repeated on-site testing and tuning, as well as periodic manual parameter calibration, thus reducing the workload of manual operation and maintenance. For a single unit, this can reduce parameter tuning and maintenance workload by more than 80% annually, while also avoiding the experience-based discrepancies of manual tuning, resulting in more consistent control performance.

[0067] 4. Through multiple safeguard mechanisms including similar sample weighting, amplitude limiting verification, rate constraint, and inertial smoothing, control abrupt changes caused by abnormal data are effectively avoided, resulting in continuous and smooth control output and a stable and safe control process. The newly added active vibration zone avoidance function can reduce the time the unit spends traversing the vibration zone by more than 40%, significantly reducing the unit's dwell time under adverse operating conditions, reducing equipment vibration and wear, and improving operational safety.

[0068] 5. The independent modeling architecture for each unit allows each unit to form a control strategy based on its own operating rules, perfectly adapting to the differences in manufacturing, aging, and hydraulic characteristics of different units. This ensures that all units in the power station can achieve their optimal regulation effect, realizing precise control with one strategy per unit, and avoiding the problem of "some units meeting the standards while others deteriorate" under unified parameters.

[0069] 6. The self-learning closed-loop mechanism, combined with sample aging and elimination and special update functions, allows the control model to continuously iterate and optimize as equipment status changes and operational data accumulates. Control performance will not decline with equipment aging, and it has continuous autonomous optimization capabilities. After unit maintenance, the model can be autonomously adapted within one month without manual intervention.

[0070] 7. This invention does not rely on a precise hydraulic mathematical model of the turbine, but achieves high-precision feedforward control entirely based on the unit's own operating data. It breaks the traditional understanding that "high-precision control of hydropower units must rely on a precise mechanism model" and overcomes the technical prejudice in this field, providing a brand-new data-driven technical path for hydropower speed regulation control.

[0071] 8. This invention adopts a feedforward-feedback composite architecture, which can be directly superimposed on the original speed regulation system of the power plant without large-scale hardware modification, and has strong engineering applicability; the supporting cold start and small sample solutions solve the problem of no data available for new units, and the full-scenario fault tolerance mechanism ensures the safety of operation under extreme conditions, and has high promotion and application value. Attached Figure Description

[0072] Figure 1 This is a schematic diagram of the overall architecture of the optimal pulse width feedforward control method for hydropower units based on data self-learning, as proposed in this invention.

[0073] Figure 2 This is a schematic diagram of the pulse width calculation process based on historical optimal operating condition matching in this invention.

[0074] Figure 3 This is a schematic diagram of the multi-unit independent self-learning model structure of the present invention.

[0075] Figure 4 This is a schematic diagram of the data self-learning update closed loop of the present invention. Detailed Implementation

[0076] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the following embodiments are used to illustrate the technical solutions of the present application, and are not intended to limit the scope of protection of the present application.

[0077] like Figure 1 As shown, the core idea of ​​the data self-learning-based optimal pulse width feedforward control method for hydropower units provided in this application is to break away from the traditional control framework that relies on precise mathematical models, and adopt a data-driven approach to extract high-quality regulation patterns from the historical operating data accumulated by the unit over a long period of time, and construct a mapping relationship between operating state characteristics and optimal regulation pulse width.

[0078] When a new load adjustment command is received, the system matches the current operating status with similar historical operating conditions, and generates the optimal feedforward adjustment pulse width that is adapted to the current operating conditions through weighted calculation. This pulse width is then directly applied to the speed control actuator, realizing the transformation from error-driven passive adjustment to experience-predictive active control.

[0079] The overall technical approach unfolds according to the logic of offline database construction, online matching, and closed-loop updates, thereby forming a closed-loop self-learning process of "operational data collection, experience learning, control output, effect evaluation, and model update".

[0080] Step S1 is to collect historical running data.

[0081] All historical data is sourced from the hydropower station's existing computer monitoring and speed control systems, eliminating the need for extensive new data acquisition equipment. The acquired data covers the entire data chain, from operating condition inputs and control actions to response outputs, specifically including four categories:

[0082] The first category is basic operating status data of the unit, which includes unit number, real-time active power, target active power, working head, guide vane opening, unit speed, grid frequency, cumulative running time, and time since last maintenance, etc., which are used to describe the unit's operating conditions before regulation occurs.

[0083] The second category is control process data, which includes adjustment start time, adjustment end time, output control pulse width, adjustment direction, and the sequence of changes in control commands during adjustment. It is used to record the complete process of control input.

[0084] The third category is the regulation response result data, which includes the entire power change curve, maximum overshoot, steady-state error, regulation time, number of oscillations during regulation, etc., and is used to evaluate the regulation effect;

[0085] The fourth category is auxiliary labeling data, which includes operating condition labels such as whether the area is in a vibration zone, whether there is a limited load command, and whether an equipment malfunction has occurred. These are used for subsequent sample filtering and anomaly removal.

[0086] Based on the above data, each complete load adjustment process can be organized into a correlation sample that includes operating status characteristics, control pulse width, and response effect, providing a data foundation for subsequent screening and modeling. The samples are stored separately in two directions: load increase and load decrease, to accommodate the asymmetric characteristics of the guide vane response of the hydropower unit.

[0087] Step S2 involves constructing a historical best operating condition sample library.

[0088] The historical operating data collected in step S1 is cleaned by deleting missing data, obviously abnormal data, and abnormal operating status data. Then, the cleaned historical regulation processes are classified according to the unit number and regulation direction.

[0089] To eliminate low-quality and abnormal operating conditions in the regulation data and ensure the accuracy of the sample database, it is necessary to establish a unified regulation performance evaluation index to quantitatively score and screen historical regulation processes.

[0090] This method uses comprehensive evaluation indicators. The evaluation of a single adjustment process is calculated using the following formula:

[0091]

[0092] in To adjust the time, For the maximum overshoot, To adjust the number of oscillations during the process, For steady-state error, , , , The corresponding weighting coefficients can be adjusted according to the power plant's emphasis on regulation speed and stability. Typical values ​​are 0.4, 0.3, 0.2, and 0.1, respectively. A smaller comprehensive evaluation index indicates a better regulation process.

[0093] The system filters based on the set evaluation threshold. Adjustment processes with values ​​below a preset threshold are considered high-quality adjustment processes, and their corresponding operating state characteristics and optimal control pulse widths are saved as samples. The feature vector for each historical sample is:

[0094]

[0095] The corresponding control output is ,in To operate the water head, For load variation deviation, The initial guide vane opening, These are parameters for evaluating the unit's operating status. This is the optimal adjustable pulse width obtained under this operating condition.

[0096] The unit operation status evaluation parameter S is calculated by weighting three indicators: equipment running time, running time after maintenance, and recent average regulation performance degradation rate. The value ranges from 0 to 1. The closer the value is to 1, the better the equipment status. The typical weights of the three indicators are 0.3, 0.3, and 0.4, respectively, and can be adjusted according to the power plant equipment management rules.

[0097] Considering the inherent differences in the regulation characteristics of different generating units, this method adopts an architecture of unit-specific modeling and independent invocation, establishing an independent historical sample database for each unit. During online control, the control model of the corresponding unit is called according to the unit number. The control logic of different units is completely decoupled to avoid mutual interference between data from different units.

[0098] Step S3 is to extract real-time running status features.

[0099] When the dispatching system issues a new load adjustment command, the control system first collects the current operating parameters of the unit in real time and extracts the feature vector as the input for operating condition matching.

[0100] Feature selection follows the principles of "strong correlation, easy acquisition, and low redundancy," and the core features include:

[0101] Operating head It directly reflects the hydraulic conditions of the unit;

[0102] Load variation deviation ,in For the target power, This represents the current actual power, reflecting the magnitude of load change during this adjustment;

[0103] Current guide vane opening This reflects the current mechanical operating position of the unit;

[0104] Unit operating status evaluation value It is generated by taking into account factors such as the equipment's operating time, maintenance cycle, and recent performance degradation, and is used to characterize the equipment's current operating status.

[0105] The final feature vector of the current working condition is formed. Then input the subsequent working condition matching module.

[0106] In practical applications, additional feature dimensions such as unit speed, head change rate, grid frequency deviation, and guide vane opening change rate can be added according to the specific needs of the power plant to further improve the accuracy of the operating condition description.

[0107] Step S4 is adaptive similarity matching based on working condition partitioning.

[0108] like Figure 2 As shown, optimal pulse width calculation is the core component of feedforward control. This method first divides the entire operating condition into three regions based on the distribution density of the unit's historical operating data:

[0109] The core operating area corresponds to the head and load range where the units operate most frequently on a daily basis. This area has a large sample size and high data density.

[0110] The transitional operation zone corresponds to the less frequently used operating range of the unit, and the sample size is moderate;

[0111] The boundary operating zone corresponds to rare operating conditions such as extreme head and ultimate load, and the sample size is small.

[0112] The division of regions can be determined by clustering algorithms combined with operational statistics, and the boundaries can be dynamically adjusted according to the power plant's operation plan, without the need for manually fixing the division threshold.

[0113] Before calculating similarity, all features are first normalized to map the feature values ​​to the range of 0 to 1, thus eliminating the influence of dimensional differences on distance calculation.

[0114] When the real-time feature value exceeds the historical extreme value range, it is automatically mapped to the interval boundary and marked as a boundary condition, triggering the corresponding matching strategy.

[0115] This method uses dynamically weighted Euclidean distance to calculate the similarity of working conditions. The distance calculation formula is as follows:

[0116]

[0117] in The number of input features, For the j-th feature of the current working condition, Features corresponding to historical samples These are the feature weights. The feature weights are dynamically adjusted based on the region to which the current operating condition belongs.

[0118] In the core operating area, the weights for head and load deviations are 0.4 and 0.35, respectively, to enhance the accuracy of operating condition matching;

[0119] In the boundary operation zone, the weights of the unit status parameters and guide vane opening are 0.25 and 0.3, respectively, to enhance the stability and safety of the output; while in the transition operation zone, a balanced weight configuration is adopted.

[0120] After calculating the distance between the current working condition and all historical samples, sort them in ascending order of distance, and dynamically select the number of similar samples K based on the region to which the working condition belongs:

[0121] The K value in the core operating area is set to 3 to 4 to ensure the accuracy of the matching results;

[0122] The K value in the transition operation zone is set to 5 to 6, balancing accuracy and smoothness;

[0123] The K value for the boundary operating zone is set to 7 to 8, and the risk of anomalies from a single sample is reduced by using a weighted average of more samples.

[0124] During the screening process, samples that are consistent with the current adjustment direction are matched first. Only when the number of samples in the same direction is less than half of the K value are samples in the opposite direction added and the direction coefficient is corrected. The correction coefficient is calibrated according to the difference in response characteristics of the increase and decrease of load, and the typical value is 0.9 to 1.1.

[0125] Step S5 involves calculating the optimal pulse width and correcting for nonlinear head.

[0126] After selecting K similar samples, the recommended adjustment pulse width for the current operating condition is calculated using a distance-inverse weighted method. The calculation formula is as follows:

[0127]

[0128] in The optimal pulse width corresponds to the i-th historical sample. The distance between the current working condition and the i-th sample. To prevent the correction parameter from being zero in the denominator, a typical value of 0.001 is used.

[0129] Samples that are closer in distance have a higher proportion in the calculation, which fully reflects the principle of "nearest neighbor priority". This ensures that the output results match the current working conditions and avoids sudden changes in output caused by a single abnormal sample.

[0130] The flow characteristics of a water turbine are nonlinearly related to the head. The power change corresponding to the same guide vane adjustment pulse width varies under different heads. Relying solely on weighted similar samples cannot fully cover the nonlinear effect of the head. Therefore, this method adds a head nonlinearity correction step to perform secondary calibration on the initial recommended pulse width.

[0131] The corrected formula is:

[0132]

[0133] in The revised recommended pulse width. Let K be the average head of similar samples. The current operating head, It is a nonlinear correction index, which can be calibrated according to the actual hydraulic characteristics of the unit. The typical value is 0.4 to 0.6 for mixed-flow units and 0.3 to 0.5 for axial-flow units.

[0134] This correction step effectively compensates for the nonlinear differences in hydraulic characteristics under different heads, improving the accuracy of pulse width calculation in wide head fluctuation scenarios.

[0135] Step S6 is the processing of safety constraints for feedforward control commands.

[0136] Directly outputting the original calculated pulse width poses a safety risk, as abnormal sample matching, measurement data fluctuations, or boundary condition extrapolation may cause the output pulse width to exceed the safe operating range of the unit. Therefore, this method sequentially performs amplitude limiting verification, rate of change constraint, vibration zone avoidance, and inertial smoothing on the corrected pulse width.

[0137] Pulse width rationality limit verification: Since historical samples are derived from actual operation processes, there may be some atypical operating conditions in different operation stages. Therefore, the rationality of the calculated recommended pulse width needs to be judged.

[0138] Let the original recommendation pulse width output by the model be... The system sets the allowable output range based on the unit's regulation capacity:

[0139]

[0140] in: The minimum allowable adjustment pulse width; The maximum allowable adjustment pulse width.

[0141] when: ,but: ;

[0142] when: ,but: ;

[0143] otherwise: By using the above-mentioned limiting process, we can prevent control commands from exceeding the safe execution range of the speed regulation system due to abnormal data matching.

[0144] Pulse width change rate constraint: The guide vane actuator of a hydroelectric generator is a mechanical device with an upper limit to its response speed. If the pulse width change is too large in adjacent control cycles, the actuator will not be able to keep up in time, and may even cause mechanical shock and power oscillation. Therefore, this method incorporates a pulse width change rate constraint.

[0145] Let the output pulse width of the current calculation cycle be... The output pulse width of the previous control cycle was: Then the amount of change is limited:

[0146]

[0147] satisfy:

[0148]

[0149] in: This represents the maximum pulse width variation allowed in a single control cycle. The output after rate limiting is:

[0150]

[0151] in: This represents the amplitude limiting function.

[0152] Rate constraints allow changes in control commands to be smoothly matched with the dynamic response capability of the actuator, avoiding adjustment oscillations caused by sudden changes in commands.

[0153] Active vibration zone avoidance and correction: Hydropower units have specific power vibration operating ranges. Prolonged operation within these ranges exacerbates equipment vibration and wear, impacting operational safety. Traditional feedback control can only passively adjust after entering the vibration zone. This method incorporates active vibration zone avoidance and correction in the feedforward stage, reducing the unit's dwell time in the vibration zone from the outset.

[0154] The specific execution logic is as follows: the system predicts the unit power value after adjustment based on the current power and the recommended pulse width. If the predicted power falls within the preset vibration range, the recommended pulse width is adjusted. When the target power is higher than the upper limit of the vibration range, the pulse width is appropriately increased to allow the unit to quickly pass through the vibration range and stabilize at the target load. When the target power is lower than the lower limit of the vibration range, the pulse width is appropriately decreased to prevent the power from surging into the vibration range.

[0155] If the adjustment range is too large to avoid the vibration zone in one go, a step-by-step adjustment strategy is adopted, which breaks down a single large pulse width into two small pulse widths, with a short transition phase in between, so that the unit can quickly pass through the vibration zone and reduce the dwell time.

[0156] The vibration range can be set based on the actual test data of the unit, and usually includes upper and lower boundaries and a transition margin. The margin value can be taken as 2% of the rated power.

[0157] For inertial smoothing, to further filter out high-frequency fluctuations in the pulse width output, a first-order inertial filter is added after the rate constraint. The calculation formula is as follows:

[0158]

[0159] in The final output pulse width for the k-th control cycle. The pulse width after amplitude limiting. The output pulse width of the previous cycle. The smoothing coefficient and Larger It can improve response speed and reduce the size It can enhance output smoothness and allows for parameter configuration based on unit regulation characteristics, head variation range, and control cycle.

[0160] After the above processing, the final result is As a feedforward control pulse width sent to the speed control system, it is output to the speed control actuator to drive the guide vanes and complete the unit power adjustment.

[0161] In some implementations, the final feedforward pulse width and the original PID feedback loop form a composite control architecture with feedforward as the main component and feedback as the auxiliary component. The two work together to balance response speed and control accuracy.

[0162] Feedforward control handles the main regulation, responsible for quickly bringing the unit power to near the target value in the initial stage of regulation, accounting for 80% to 90% of the total regulation.

[0163] The PID feedback loop handles the remaining fine-tuning, eliminating residual deviations and steady-state errors from the feedforward control, and ensuring final control accuracy.

[0164] The weights of both are dynamically adjusted during the adjustment process: in the initial stage of adjustment, when the power deviation is large, the feedforward weight is set to the highest to ensure the response speed; as the power gradually approaches the target value, the feedforward weight decreases linearly and the feedback weight increases linearly to avoid overshoot due to excessive feedforward; when the power enters the steady-state error band, the feedforward output returns to zero and the steady state is maintained entirely by the feedback loop.

[0165] Meanwhile, a seamless switching logic is set up. When faults such as abnormal sample library, failure of key sensor data, or communication interruption are detected, the feedforward output gradually decays according to the control cycle and smoothly switches to pure PID control mode. The switching process is shock-free and will not cause unit power fluctuations, thus ensuring safe operation.

[0166] Command Output and Execution: The final calculated regulating pulse width is sent to the speed regulating actuator through the hydropower station's automatic control system. The specific process is as follows:

[0167] The intelligent control module receives the target power command issued by the scheduling system.

[0168] Real-time acquisition of current unit operating status data generates operating condition feature vectors.

[0169] The historical best sample model of the corresponding unit is invoked based on the unit number.

[0170] Perform similar working condition matching and weighted calculation to generate the initial recommended pulse width.

[0171] The following steps are performed sequentially: head nonlinearity correction, rationality limit, rate constraint, vibration zone avoidance, and smoothing.

[0172] The final control pulse width is output to the unit speed governor.

[0173] The speed governor controls the guide vane movement based on the pulse width to complete the unit power adjustment.

[0174] Step S7 is a data self-learning update based on the evaluation of the running effect.

[0175] like Figure 4 As shown, after each complete load adjustment process, the system automatically records the entire link data of this adjustment, including the operating status characteristics before adjustment. Actual output feedforward pulse width and adjustment response results New running samples are generated and stored in the dataset to be evaluated. The data recording process is completed automatically without human intervention.

[0176] The system then performs a comprehensive performance evaluation of the newly completed adjustment process, with evaluation indicators including response speed. (Settlement time), overshoot index and stability indicators (Number of oscillations), the comprehensive evaluation index is:

[0177]

[0178] Weighting coefficient , , Typical values ​​are 0.5, 0.3, and 0.2.

[0179] when When the value is less than the preset threshold, it indicates that the adjustment effect meets the requirements, and the system performs a sample database update.

[0180] The system searches the sample library for historical samples that are highly similar to the current working condition (with a distance less than a set similarity threshold). If no historical samples with a distance less than the similarity threshold exist in the sample library, the new sample is added directly to the sample library to expand the coverage of working conditions.

[0181] If highly similar historical samples exist, compare the comprehensive evaluation indicators of the two and retain the sample with better evaluation to ensure that each working point in the sample library stores the current best adjustment experience.

[0182] Update the corresponding unit control knowledge set to form:

[0183]

[0184] in: For the original historical sample database, To add high-quality samples.

[0185] This mechanism enables the control model to be continuously optimized as operational data accumulates.

[0186] Building upon this, our method also incorporates a sample aging and elimination mechanism to accommodate the gradual degradation of unit equipment performance over time. Each sample in the sample library has a set validity period, typically two years. Expired samples have their matching weight gradually reduced over time and are eventually replaced by newly generated high-quality samples. When the unit completes operations that may alter its regulation characteristics, such as Class A overhauls or guide vane system modifications, the system automatically marks all samples from before the overhaul as historical reference samples, reducing their matching weight. Simultaneously, it accelerates the inclusion of new samples after the overhaul, enabling rapid model adaptation.

[0187] In addition to the conventional successive adjustment evaluation updates, this method also sets up a special update trigger mechanism: when three special scenarios occur, such as the operating head crossing the preset partition threshold and entering the head range of infrequent operation, the first load operation stage after the unit has been overhauled, and the comprehensive evaluation index of three consecutive adjustments being lower than the quality threshold, the system automatically triggers special sample collection and model updates, temporarily increases the sample collection frequency, accelerates the model iteration under the corresponding operating conditions, and quickly makes up for control shortcomings.

[0188] In one embodiment, the implementation process of the above method is illustrated by taking the optimal pulse width feedforward control in the load regulation process of a single hydropower unit as an example.

[0189] During the normal operation of a turbine generator unit at a hydropower station, the dispatch system issued a load adjustment command, requiring the unit to adjust its current active power to a new target power.

[0190] The system pre-collects historical operating data of the unit, covering load adjustment processes under different head conditions, response processes under different load change amplitudes, and adjustment processes under different guide vane initial positions. For each historical adjustment process, the system calculates the adjustment time, maximum power deviation, steady-state error, and number of oscillations, and selects the adjustment processes with faster response speed and better stability to form a historical sample set. .

[0191] When a new adjustment task arrives, the system collects the current head data in real time. Current load deviation Current guide vane opening and unit status parameters This forms the current operating condition vector. Calculate the weighted Euclidean distance between the current operating condition and each historical sample, and select the K historical samples with the smallest distances, for example, K=3, to obtain the optimal pulse width corresponding to the three closest historical adjustment cases. ;

[0192] According to the reciprocal weighted formula:

[0193]

[0194] The recommended pulse width for the current operating condition is obtained, and after head nonlinearity correction, maximum and minimum pulse width limits, rate of change limits, and first-order inertial smoothing are applied sequentially before the final control command is sent to the speed control system. The speed controller controls the guide vane movement based on this pulse width, enabling the unit to quickly approach the target power.

[0195] Through this implementation method, the system can utilize historical operating experience to directly provide control actions that match the current operating conditions at the beginning of the adjustment phase, reducing the search process in the initial adjustment phase and improving the matching degree between the control actions and the actual dynamic characteristics of the unit.

[0196] In another implementation, such as Figure 3 As shown, the application of the multi-unit independent self-learning model is illustrated using the coordinated application scenario of multiple hydropower units as an example.

[0197] Large hydropower stations are typically equipped with multiple turbine generator units. Due to differences in manufacturing, hydraulic parameters, operating time, and equipment status among different units, the same control parameters cannot guarantee that all units will achieve the best regulation effect.

[0198] This method establishes a historical sample model for each unit separately; for example, it establishes a control model for Unit 1. Establish a control model for Unit 2 And so on, among which This represents the historical high-quality regulation sample of the i-th generating unit.

[0199] When the power grid dispatching system sends regulation tasks to multiple generating units, the system identifies the unit number participating in the regulation, calls the corresponding unit control model, generates the corresponding feedforward pulse width according to the real-time operating status of each unit, and sends it to the speed control system of each unit.

[0200] By using a multi-unit independent modeling approach, interference between operating data from different units is avoided, improving the specificity of the control model. It also facilitates the expansion of new units and the maintenance and updating of individual unit models, ensuring that the sample libraries and control models of each unit do not interfere with each other.

[0201] In another implementation, the process of continuous iteration of the sample library is illustrated using a self-learning optimization scenario during long-term operation as an example.

[0202] During long-term operation, the regulation characteristics of hydropower units may change due to factors such as equipment aging, maintenance and adjustment, and changes in the operating environment. Therefore, the control strategy needs to have continuous optimization capabilities.

[0203] After each load adjustment is completed, the system automatically saves the pre-adjustment state, feedforward output pulse width, actual response curve, and adjustment performance indicators, forming a new operating sample. The system then calculates the performance indicators for that adjustment. If the new sample If a sample is superior to a historical sample for the corresponding working condition in the sample library, then the sample is added to the historical knowledge base; otherwise, the original high-quality sample is retained.

[0204] As the system's operating time accumulates and the number of historical samples continues to grow, the range of operating conditions covered by the model gradually expands. The system gradually forms a continuous learning loop of increasing operating data, enriching control experience, improving the accuracy of operating condition matching, and optimizing control effects, so that the model's control capability is continuously enhanced as the operating time increases.

[0205] In some implementations, for scenarios where historical data is insufficient, such as newly commissioned units or units whose characteristics have changed significantly after maintenance, this method adopts a cold start and small sample adaptation scheme.

[0206] During the cold start phase, when the unit has no historical operating data of its own, the sample library of similar units is first used as the initial migration model. The initial control model is generated by scaling the rated parameters of the unit. At the same time, the turbine digital simulation system can be connected to generate virtual high-quality samples covering all operating conditions based on the simulation data to supplement the initial sample library.

[0207] In the small sample stage, the system adopts a composite control strategy of feedforward and feedback. The weight of the feedforward output is set to a low value, and the main adjustment is still undertaken by traditional PID control, with the feedforward only serving as an auxiliary correction. As the operating data accumulates and the number of high-quality samples gradually increases, the weight of the feedforward control is gradually increased, smoothly transitioning to the full feedforward control mode.

[0208] For the first three months of a new unit, the sample update frequency should be appropriately increased to speed up model adaptation. After running for six months and having more than 500 samples, the normal update frequency should be restored. This can effectively solve the problem of no data available for the new unit and significantly shorten the model adaptation cycle.

[0209] To ensure the safe and reliable operation of the control method in all scenarios, this method is designed with complete fault-tolerant degradation processing logic for various extreme working conditions and abnormal states.

[0210] When critical sensor data becomes abnormal, such as a jump in head or active power data that exceeds the range allowed by the unit's physical characteristics, the system immediately freezes the feedforward output, keeps the control quantity of the previous cycle unchanged, and smoothly switches the control power to the PID feedback loop. After the sensor data returns to normal and stabilizes for 3 control cycles, the feedforward control is gradually restored. The entire switching process is free of power surges.

[0211] When the feature distance between all historical samples in the sample library and the current operating condition is greater than the set threshold, the current condition is determined to be a completely new operating condition without a match. At this time, the system automatically reduces the feedforward output ratio, outputting only 50% of the calculated value as the feedforward quantity. The remaining adjustment quantity is borne by the feedback loop to avoid overshoot caused by inaccurate feedforward. At the same time, this adjustment process is marked as a key observation sample. After the adjustment is completed and the effect evaluation is passed, it will be included in the sample library first to quickly supplement the operating condition coverage.

[0212] When the actuator of the speed control system malfunctions and the guide vane action response speed decreases, the system identifies the change in response characteristics through the evaluation of the adjustment effect, automatically reduces the pulse width change rate limit, and at the same time accelerates the sample update frequency to adapt to the performance changes of the actuator and avoid adjustment oscillation.

[0213] It is understood that the specific parameters in the above embodiments are exemplary descriptions and do not constitute a limitation on the scope of protection of this application. The parameters include: working condition region division method, feature weight values, K value range, and nonlinear correction index. Specific values ​​and smoothing coefficients All parameters can be adjusted according to the specific power plant's unit type, head variation range, operating characteristics, and control precision requirements. None of the above variations deviate from the core technical concept of this application.

Claims

1. A method for optimal pulse width feedforward control of hydropower units based on data self-learning, characterized in that, Includes the following steps: S1: Collect historical operating data of hydropower units. The historical operating data shall include at least the basic operating status data of the units, control process data, regulation response result data, and auxiliary marker data. S2: Clean the historical operating data, classify it according to unit number and regulation direction, select high-quality regulation processes according to unified evaluation indicators, and establish an independent historical best operating condition sample library for each unit. S3: When the dispatching system issues a load adjustment command, it collects the current operating status in real time and generates a working condition feature vector; S4: Based on the current adjustment direction and the region to which the operating condition belongs, use dynamic feature weights to calculate the weighted distance between the current operating condition features and the features of each historical sample, and select K similar historical samples; S5: The recommended pulse width is calculated from the optimal adjustment pulse width of the K similar historical samples according to the distance reciprocal weighting method, and nonlinear correction is performed by combining the current water head and the average water head of similar samples; S6: Perform amplitude limiting verification, rate of change constraint, vibration zone avoidance and inertial smoothing on the corrected pulse width in sequence to generate the final feedforward pulse width and output it to the speed control actuator. S7: Record the data of the entire adjustment process and calculate the adjustment performance evaluation index; when the evaluation index meets the preset threshold, add the new sample to the sample library, or replace the similar sample with a worse evaluation in the sample library, and complete the sample library update.

2. The optimal pulse width feedforward control method for hydropower units based on data self-learning as described in claim 1, characterized in that, In S2, each sample includes at least the operating head, load change deviation, initial guide vane opening, unit operating status evaluation parameters, and the corresponding optimal adjustment pulse width; The unified evaluation index is: ; in To adjust performance evaluation indicators, To adjust the time, For the maximum overshoot, To adjust the number of oscillations during the process, For steady-state error, , , , For the corresponding weight coefficients; filter Adjustment processes that fall below a preset evaluation threshold are considered high-quality adjustment processes; the sample library is stored separately in two directions: increasing load and decreasing load.

3. The optimal pulse width feedforward control method for hydropower units based on data self-learning according to claim 2, characterized in that, The unit operating status evaluation parameter S is calculated by weighting the equipment operating time, the operating time after maintenance, and the recent average regulation performance degradation rate, and its value ranges from 0 to 1; each unit corresponds to an independent historical sample database. During online control, the control model of the corresponding unit is called according to the unit number.

4. The optimal pulse width feedforward control method for hydropower units based on data self-learning according to claim 2, characterized in that, The load variation deviation in S2 is: ; in For the target power, The current actual power; the operating condition feature vector also includes at least one of the following: unit speed, head change rate, grid frequency deviation, and guide vane opening change rate.

5. The optimal pulse width feedforward control method for hydropower units based on data self-learning according to claim 1, characterized in that: In S4, the entire operating condition is divided into a core operating zone, a transition operating zone, and a boundary operating zone based on the distribution density of historical sample data. In the core operating zone, the weights of head and load deviation are 0.4 and 0.35, respectively, and K is 1 to 2. In the transition operating zone, a balanced weight configuration is adopted, and K is 3 to 4. In the boundary operating zone, the weights of unit status parameters and guide vane opening are 0.25 and 0.3, respectively, and K is 5. When screening samples, prioritize matching samples that are consistent with the current adjustment direction. If the number of samples in the same direction is less than half of the K value, supplement the samples in the opposite direction and correct the direction coefficient.

6. The optimal pulse width feedforward control method for hydropower units based on data self-learning according to claim 1, characterized in that, The weighted distance in S4 is calculated using the following formula: ; in The number of input features, For the j-th feature of the current working condition, Features corresponding to historical samples The feature weights are calculated by normalizing the range of each feature before calculation. When a real-time feature exceeds the historical extreme value range, it is mapped to the interval boundary and marked as the boundary condition.

7. The optimal pulse width feedforward control method for hydropower units based on data self-learning according to claim 1, characterized in that, The recommended pulse width in S5 is calculated using the following formula: ; in The optimal pulse width corresponds to the i-th historical sample. The distance between the current working condition and the i-th sample. To prevent correction parameters with a denominator of zero; the nonlinear correction is as follows: ; in The revised recommended pulse width. Let K be the average head of similar samples. The current operating head, It is a non-linear correction exponent.

8. The optimal pulse width feedforward control method for hydropower units based on data self-learning according to claim 1, characterized in that, The amplitude limiting verification in S6 is as follows: When the recommended pulse width is less than the minimum allowed pulse width, the minimum allowed pulse width is used; when the recommended pulse width is greater than the maximum allowed pulse width, the maximum allowed pulse width is used. The rate of change constraint is: ; and: ; in The maximum allowable pulse width variation in a single control cycle; the inertial smoothing process is as follows: ; in The final output pulse width for the k-th control cycle. The pulse width after amplitude limiting. The output pulse width of the previous cycle. Smoothing coefficient and .

9. The optimal pulse width feedforward control method for hydropower units based on data self-learning according to claim 1, characterized in that, The vibration zone avoidance in S6 is as follows: predict the unit power after adjustment based on the current power and the recommended pulse width; when the predicted power falls into the preset vibration zone, if the target power is higher than the upper limit of the vibration zone, increase the pulse width to make the unit quickly pass through the vibration zone; if the target power is lower than the lower limit of the vibration zone, decrease the pulse width to avoid entering the vibration zone; when a single adjustment cannot completely avoid the vibration zone, the pulse width is split into two or more steps.

10. The optimal pulse width feedforward control method for hydropower units based on data self-learning according to claim 1, characterized in that, The final feedforward pulse width and the original PID feedback loop constitute a composite control with feedforward as the main function and feedback as the auxiliary function. Feedforward control handles 80% to 90% of the total adjustment, while the PID feedback loop handles the remaining fine-tuning. The feedforward weight is highest in the initial stage of adjustment. As the power approaches the target value, the feedforward weight decreases linearly. After entering the steady-state error band, the feedforward output returns to zero. When an anomaly is detected in the sample library, sensor data failure, or communication interruption, the feedforward output gradually decays according to the control cycle and switches to pure PID control without disturbance.

11. The optimal pulse width feedforward control method for hydropower units based on data self-learning according to claim 1, characterized in that, The sample library update in S7 includes: When there are no historical samples in the sample library that are less than the similarity threshold to the current working condition, a new sample is added directly. When similar samples exist, compare their comprehensive evaluation indicators and retain the better sample; The matching weight of expired samples gradually decreases over time. After the unit completes a Class A overhaul or guide vane system modification, the samples before the overhaul are marked as historical reference samples and their matching weight is reduced. Increase the frequency of sample collection when a special update is triggered.

12. The optimal pulse width feedforward control method for hydropower units based on data self-learning according to claim 1, characterized in that: For units without their own historical data, an initial control model is generated by scaling up a sample library of similar units according to the rated parameters, or by generating virtual high-quality samples covering all operating conditions using turbine digital simulation data. During the small sample stage, the feedforward output weight is reduced and PID feedback control is used as the main method. As the number of high-quality samples increases, the feedforward weight is gradually increased. After the sample size reaches the preset threshold, the normal update frequency is restored.