Hydroelectric generating set parameter dynamic multi-objective optimization method and system oriented to variable working conditions

Abstract

The invention discloses a hydroelectric generating set parameter dynamic multi-objective optimization method and system oriented to variable working conditions. The method comprises the following steps: establishing a nonlinear simulation model; establishing a multi-dimensional performance evaluation system; constructing an off-line / on-line fusion optimization system architecture; for each new working condition point, generating a high-quality initial population based on the Pareto solution set of the nearest neighbor optimized working condition point; solving a Pareto optimal solution set of the working condition points; performing incremental training on the machine learning model to form an offline knowledge base; real-time working condition data are collected in a fixed period, and future water head and power are predicted; based on a prediction result, adopting a dual warm start mechanism to solve a dynamic multi-objective optimization problem at the current moment and the future moment in parallel; and monitoring a consistency index between the online optimization solution set and the prediction solution set in real time, and triggering incremental updating of the model when the index continuously exceeds the limit. According to the method, all-working-condition continuous self-adaptive optimization and prospective regulation and control of the hydroelectric generating set can be realized.

Description

Technical Field

[0001] This invention mainly relates to the field of hydropower unit technology, specifically to a dynamic multi-objective optimization method and system for hydropower unit parameters under varying operating conditions. Background Technology

[0002] As the core regulating unit of the power system, the performance of the speed control system of hydropower units directly determines the frequency stability of the power grid, power quality, and the safe operation level of the unit itself. The speed governor achieves precise control of the unit's output power and speed by adjusting the turbine guide vane opening in real time. In this process, the tuning of PID controller parameters and the optimization of key parameters such as compensation coefficients are the core issues for improving control quality and have always been a research hotspot in this field.

[0003] However, a hydropower unit is a typical nonlinear, time-varying dynamic system, and its dynamic characteristics vary significantly with changes in operating conditions (mainly reflected in the operating head and guide vane opening). A fixed set of controller parameters is unlikely to maintain optimal performance throughout the entire operating range, often leading to problems such as slow response, large regulation overshoot, and even power oscillations under certain operating conditions.

[0004] To improve the adaptive capability of controller parameters to operating conditions, some current methods focus on the adaptive characteristics of the governor. Among these, the operating point division and static optimization method is the most common. This method discretizes the head-opening operating space of the unit into several typical operating points and uses intelligent optimization algorithms to perform offline optimization at each point, ultimately constructing a parameter lookup table for online switching. This method is simple and effective, but in essence, it is a simple set of multiple static optimal solutions. When operating conditions change continuously, parameter jumps can easily lead to an unsmooth control process, and it is difficult to cover all non-preset operating conditions.

[0005] Furthermore, advanced algorithms such as sliding mode control and predictive control, as well as research on introducing BP neural networks into parameter adaptive optimization, have demonstrated potential at the theoretical level. While these methods may improve some performance indicators, they often come at the cost of increased system complexity and sacrificed engineering interpretability. Existing methods still face core challenges in solving the adaptive control problem of hydropower units under all operating conditions, including a single optimization objective, difficulty in handling continuous dynamic changes, and insufficient foresight.

[0006] Although the aforementioned existing technologies have achieved certain results in optimizing the parameters of hydropower units, they still have some inherent limitations or areas for improvement, as follows: 1. The fixed-parameter method lacks adaptability. Although this method is simple in structure and easy to implement, its core flaw lies in attempting to use a fixed set of controller parameters to cope with all operating conditions. A hydropower unit is a nonlinear system whose dynamic characteristics vary significantly with head and load. The fixed-parameter strategy cannot achieve dynamic matching across the entire operating range, inevitably leading to a decline in regulation quality when the unit deviates from its design conditions, such as slow response, increased overshoot, or deteriorated stability, making it difficult to meet the requirements of high-quality operation.

[0007] 2. The operating point segmentation and static optimization method lack continuity. While this method improves performance under typical operating conditions by discretizing operating points and optimizing independently at each point, the offline optimization process fails to reflect the continuous dynamic process during actual operation. Secondly, in online applications, parameter switching between operating points can lead to abrupt changes in control commands, causing system fluctuations. Finally, this method cannot cover all possible operating conditions, and its performance is not guaranteed for atypical operating points that have not been pre-optimized.

[0008] 3. Data-driven adaptive methods face application bottlenecks. Advanced methods such as fuzzy PID control and BP neural networks, while enabling continuous parameter adjustment, have significant limitations. Fuzzy controllers heavily rely on expert-defined rule bases and membership functions, resulting in a highly subjective design process and a lack of systematic global optimization guidance. Neural network methods typically require extensive supervised training based on static optimization results, essentially fitting and extrapolating historical data without introducing new optimization information. This not only increases workload but also makes the controller a "black box" model, reducing interpretability. Furthermore, these methods generally focus on a single control objective, failing to meet the decision-making needs of operators who require flexible trade-offs between multiple performance indicators in different scenarios. Summary of the Invention

[0009] To address the technical problems existing in the prior art, this invention provides a dynamic multi-objective optimization method and system for hydropower unit parameters under varying operating conditions, enabling continuous adaptive optimization and forward-looking control of hydropower units across all operating conditions.

[0010] To solve the above-mentioned technical problems, the technical solution proposed by this invention is as follows: A dynamic multi-objective optimization method for hydropower unit parameters under varying operating conditions includes the following steps: S1. A nonlinear simulation model is established based on the comprehensive characteristic curve of the turbine, and the turbine torque M and flow rate Q are expressed as functions of guide vane opening α, rotational speed n and head H; a multi-dimensional performance evaluation system including frequency regulation performance, mechanical stability and damping characteristics is established; the controller parameter optimization problem is formalized into a dynamic multi-objective optimization problem, and the optimization objective is defined as minimizing the performance index vector; S2. Construct an offline / online integrated optimization system architecture, including an offline global optimization layer, an online rolling optimization layer, and an online self-learning layer; S3. For each new working point, a high-quality initial population is generated using a machine learning model based on the Pareto solution set of the nearest neighbor optimized working points; the Pareto optimal solution set of the working point is solved using the NSGA-II algorithm; after optimization, the machine learning model is incrementally trained to update its mapping capability, and this process is repeated until all working points are covered to form an offline knowledge base. S4. Collect real-time operating data at fixed intervals and call LSTM and LightGBM models to predict future head and power; based on the prediction results, adopt a dual hot-start mechanism to solve the dynamic multi-objective optimization problem at the current time and the future time in parallel; send the optimal solution selected from the optimization results at the current time to the speed control system for execution, and push the Pareto solution set at the future time to the human-machine interface for decision-making reference; S5. Monitor the consistency index between the online optimized solution set and the predicted solution set in real time. Trigger incremental model updates when the index continuously exceeds the limit. At the same time, compare the differences between the online optimized solution set and the offline interpolation solution set, adaptively adjust the threshold according to the distance, and issue a knowledge base update warning when regional inconsistency is detected.

[0011] Preferably, in step S1, establishing a nonlinear simulation model based on the comprehensive characteristic curve of the turbine specifically includes: Calculate the unit speed based on the actual rotational speed n, the turbine runner diameter D, and the working head H. : ; per unit rotational speed and current guide vane opening Using the input as input, query the comprehensive characteristic curve of the turbine to obtain the corresponding unit torque. and unit flow Finally, the actual torque is calculated using the following formula. and traffic :

[0012]

[0013] Preferably, in step S3, for each new operating point, the specific process of generating a high-quality initial population using a machine learning model based on the Pareto solution set of the nearest neighbor optimized operating points is as follows: First, select k operating conditions from the already optimized operating conditions that are compatible with the new operating conditions. The closest operating point is denoted as ; To ensure the comparability of head H and guide vane opening Y in distance calculations, the degree of closeness is determined by the normalized Euclidean distance. measure:

[0014] in, , , , These are the minimum and maximum values ​​of head and guide vane opening, respectively, in the entire discrete operating space; Let the initial population size be N; each nearest neighbor condition point Contribution solution of allocation It is inversely proportional to its distance; calculate the contribution solution data of each nearest working point; first, according to the calculation weights :

[0015] Then normalize the weights. ; Finally, each neighboring operating point The number of solutions to the assignment is Adjustments to ensure ; For each neighboring operating point From its Pareto solution set Random selection There are one solution; for each selected solution A BP neural network model M is used to predict its... The corresponding solution below :

[0016] Model M is trained synchronously during each optimization, and its input is... The output is the predicted solution. All predicted solutions To form a high-quality initial population .

[0017] Preferably, in step S3, the initial population is used. Starting from the NSGA-II algorithm, solve for the new operating condition. Pareto optimal solution set ; After optimization, the machine learning model M is incrementally trained to incorporate new knowledge; training data is obtained by constructing solution pair mappings; for each nearest neighbor operating point... its Pareto solution set and Pairing is performed to minimize the sum of the overall Euclidean distances, which is the solution:

[0018] Where Z is a binary allocation matrix, satisfying and The Hungarian algorithm is used for matching; after constructing the matching matrix, the model M is incrementally updated using the newly constructed training samples; the training loss function is the mean squared error. ; Finally, the model parameters are updated through backpropagation and optimization algorithms, enabling the model to better capture the patterns of solution changes with operating conditions.

[0019] Preferably, in step S4, the specific process of collecting real-time operating data at fixed intervals and calling the LSTM and LightGBM models to predict future head and power is as follows: The head prediction module employs a Long Short-Term Memory (LSTM) network model to capture long-term temporal dependencies in hydrological processes. This module uses historical hydrological time-series data as input to predict head values ​​at multiple future moments. Its core calculation process is as follows:

[0020] The model input consists of sequence data from the past n time points, including historical water head values ​​H, watershed rainfall R, and meteorological data W; the output is future data. Predicted head sequence at each time point ; The power forecasting module employs the Lightweight Gradient Boost Machine (LightGBM) model to handle short-term load demand fluctuations. This module uses historical power, temporal characteristics, and meteorological factors as input features to predict future short-term load demand. The forecasting process is described as follows:

[0021] The model input includes three feature categories: historical load data. Time characteristics and weather forecast information LightGBM efficiently learns the complex nonlinear relationship between features and loads using a histogram-based decision tree algorithm and gradient boosting framework, outputting future... Load forecast sequence at each time point .

[0022] Preferably, in step S4, the specific process of the dual hot-start mechanism is as follows: For optimization at the current moment, its initial population The results are directly generated by the historical Pareto front predictor M based on the current measured operating conditions. :

[0023] Regarding the optimization of the look-ahead time, its initial population... It is composed of two parts, which utilize both historical prior information and real-time recursive information:

[0024] in, It is based on historical prior populations that predict operating conditions. It is the Pareto solution set obtained at the current moment, which serves as the real-time recursive population.

[0025] Preferably, in step S5, the specific process of incremental model update is as follows: All high-quality "operating condition-online optimization solution set" paired data collected within the corresponding time period of the sliding window W. Construct as a fine-tuning dataset ; Subsequently, with the original The model serves as pre-trained weights to fine-tune the dataset. Using the new training samples, a finite number of iterations are performed with the objective of minimizing the following loss function:

[0026] in The loss function; The loss is calculated on the original large-scale offline dataset to preserve historical knowledge and prevent catastrophic forgetting. In the incremental dataset The loss is calculated above; λ is a hyperparameter that weighs the importance of new and old knowledge; after training, the updated model is obtained. .

[0027] Preferably, in step S5, the specific process of issuing a knowledge base update warning is as follows: The online system completes each assessment of the current actual working conditions. After dynamic multi-objective optimization, a real-time Pareto solution set will be obtained. Simultaneously, the system queries the Pareto solution set of adjacent operating conditions in the offline knowledge base and uses an interpolation algorithm to generate a corresponding Pareto solution set. interpolation Pareto solution set Subsequently, a set similarity metric was used to calculate the distributional difference between the set and the online optimization solution set. : Then, calculate the normalized Euclidean distance from the current actual operating condition to the nearest offline operating condition point. : Based on the calculated normalized Euclidean distance Calculate dynamic threshold ; When a significant inconsistency is detected, i.e.:

[0028] Automatically perform regional verification on N offline reference points surrounding the operating point; for each neighboring operating point The same consistency check is performed: by calculating the proportion of inconsistencies within a region, when the proportion of inconsistencies exceeds a set threshold, the system generates a diagnostic report and issues an update warning, indicating that there may be a systematic deviation in the static knowledge base, and prompting the operators to decide whether to initiate a global update.

[0029] The present invention also discloses a computer-readable storage medium having a computer program stored thereon, the computer program performing the steps of the method described above when run by a processor.

[0030] This invention further discloses a dynamic multi-objective optimization system for hydropower unit parameters under varying operating conditions, comprising an interconnected memory and a processor. The memory stores a computer program, which, when run by the processor, executes the steps of the method described above.

[0031] Compared with the prior art, the advantages of the present invention are as follows: This invention breaks through the limitations of traditional static optimization methods, and for the first time constructs the optimization of hydropower unit parameters under varying operating conditions as a dynamic multi-objective optimization problem. Operating parameters such as head and load are defined as time-varying environmental variables, enabling the algorithm to actively track the Pareto front during the continuous change of operating conditions.

[0032] This invention innovatively proposes a hot-start mechanism based on nearest-neighbor mapping. By learning a continuous mapping relationship from the operating condition space to the decision space through a machine learning model, a high-quality initial population is provided for optimization at each new operating condition. This method not only significantly accelerates convergence but also ensures a smooth transition of the solution set across the entire operating condition range, avoiding parameter jumps.

[0033] This invention deeply embeds water and power prediction into the optimization process, constructing a unique parallel look-ahead optimization mechanism. The system simultaneously solves for the optimal parameter set for both the current moment and the predicted future moment, achieving a leap from passive response to proactive decision-making. This design provides operators with forward-looking decision support, enhancing the intelligence level of unit operation.

[0034] This invention addresses the long-term performance drift problem of generator units by establishing a complete online self-learning closed loop. The system automatically triggers an incremental learning process by monitoring the deviation between the optimization effect and the model prediction in real time, enabling the core prediction model to adaptively update in response to changes in generator unit state, ensuring the system maintains optimal performance throughout its entire lifecycle. Attached Figure Description

[0035] Figure 1 This is a diagram illustrating an embodiment of the multi-objective optimization method of the present invention in a specific application.

[0036] Figure 2 This is a flowchart of an embodiment of the multi-objective optimization method of the present invention. Detailed Implementation

[0037] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0038] like Figure 1 and Figure 2 As shown in the figure, the dynamic multi-objective optimization method for hydropower unit parameters under varying operating conditions according to an embodiment of the present invention includes the following steps: S1. Establish a multi-objective optimization model for hydropower units under varying operating conditions. Simulation models of hydropower units are fundamental for parameter optimization research. Their accuracy and applicability directly determine the reliability of the optimization results. The core components include four main parts: the governor, the electro-hydraulic servo system, the turbine and water intake system, and the generator. Traditional linear models struggle to accurately capture the complex nonlinear characteristics of turbines operating under varying conditions and fail to reflect the essential laws governing their dynamic characteristics as a function of guide vane opening and head. Therefore, a nonlinear model is used for simulation. The method is based on the turbine's comprehensive characteristic curves, expressing the turbine torque M and flow rate Q as functions of guide vane opening α, rotational speed n, and head H.

[0039]

[0040] The dynamic characteristics of a water turbine are determined by its torque and flow characteristics, as shown in the equation above. These two core parameters have a strong nonlinear coupling relationship with the guide vane opening, rotational speed, and operating head. During grid connection, the rotational speed remains relatively stable, with the main variations originating from the guide vane opening and the operating head. Therefore, to accurately describe this characteristic, a unit parameter is introduced for standardized calculations. (Unit rotational speed) The calculation formula is as follows:

[0041] Where n is the actual rotational speed, D is the turbine runner diameter, and H is the operating head. In the actual simulation calculation, a two-dimensional lookup table method is used to accurately solve the dynamic characteristics of the turbine.

[0042] First, calculate the unit rotational speed based on the current rotational speed, water head, and impeller diameter. Then, using the unit rotational speed and the current guide vane opening... Using the input as input, query the comprehensive characteristic curve of the turbine to obtain the corresponding unit torque. and unit flow Finally, the actual torque and flow rate are calculated using the following formula:

[0043]

[0044] This characteristic curve-based modeling method can effectively reflect the real characteristics of a turbine under non-design conditions, including complex phenomena such as efficiency changes, cavitation characteristics, and pressure pulsations.

[0045] To comprehensively evaluate the unit's overall performance under wide load and variable head operating conditions, a multi-dimensional performance evaluation system was established. Frequency regulation performance primarily examines the dynamic quality of the unit's participation in grid frequency regulation, including indicators such as settling time, overshoot, and settling time, ensuring compliance with the grid's "two detailed rules" assessment requirements. Mechanical stability focuses on the operating status of key unit components, quantitatively assessed through parameters such as frame vibration, main shaft runout, and tailrace pressure pulsation, providing assurance for safe unit operation. Damping characteristics are characterized by the system's oscillation attenuation after disturbances, reflecting the unit's ability to resist interference and maintain stable operation.

[0046] Based on the aforementioned refined simulation model and comprehensive evaluation system, the optimization problem of hydropower unit control parameters is formally defined as a dynamic multi-objective optimization problem. The optimization objective is to minimize the performance index vector:

[0047] in The set of control parameters to be optimized. The operating conditions are time-varying, including head and guide vane opening. Objective function. These are quantitative indicators corresponding to frequency modulation performance, mechanical stability, and damping characteristics, respectively.

[0048] The problem definition clearly defines the core task of seeking a Pareto optimal solution set that achieves the best balance among multiple conflicting objectives in a dynamically changing operating environment, providing a rigorous mathematical framework and a clear optimization direction for the design of subsequent dynamic multi-objective optimization algorithms.

[0049] S2. Offline / Online Fusion Optimization Architecture To systematically achieve the aforementioned optimization objectives, this invention proposes a hydropower unit parameter optimization system that integrates offline and online methods. This system follows a progressive technical path from static presets to dynamic decision-making and can self-evolve, systematically addressing the multiple challenges faced by hydropower units during variable operating conditions. Its core architecture and workflow are as follows: Figure 1 As shown, the core architecture includes an offline global optimization layer, an online rolling optimization layer, and an online self-learning layer, which are described in detail below: The offline global optimization layer forms the foundation of the entire system, corresponding to Figure 1 The upper-level part of the system. In this stage, the system first discretizes the entire operating space of the unit, constructing a comprehensive operating condition grid. Then, using an innovative dynamic multi-objective optimization algorithm, combined with the continuity of operating conditions, a set of Pareto optimal parameters is calculated for each operating point in the grid. The output of this process consists of two core assets: firstly, a static parameter matrix that can be directly accessed by the speed control system to support basic operation, forming the system's offline knowledge base; secondly, a Pareto front predictor capable of characterizing the intrinsic mapping between operating conditions and parameters. This establishes the global optimization benchmark for the entire framework.

[0050] The online rolling optimization layer is the core of achieving intelligent decision-making, corresponding to Figure 1 The lower layer of the system incorporates real-time operating data to predict short-term head and opening, overcoming the limitations of static grid points based on the prediction model provided by the offline layer. The system can dynamically generate Pareto optimal solution sets for the current and future moments. This overcomes the accuracy limitations of static matrices relying on interpolation at non-grid operating points and provides operators with a flexible interactive decision-making space, allowing them to weigh and choose among multiple optimal solutions based on real-time scheduling needs.

[0051] The online self-learning layer is key to ensuring the long-term effectiveness of the system, corresponding to... Figure 1 The middle layer of the system addresses the performance drift of hydropower units due to long-term operation by incorporating a closed-loop monitoring and updating mechanism. The system continuously compares the expected performance based on a static matrix with the actual operating performance of the unit. When the performance deviation consistently exceeds a set threshold, a global re-optimization is automatically triggered to update the static parameter matrix and simultaneously train and update the prediction model. This mechanism ensures that the optimization system can continuously improve itself as the unit's state changes, thus achieving a significant leap from one-time design to lifelong adaptation.

[0052] S3. Offline Global Optimization Based on Knowledge Transfer The core objective of the offline optimization phase is to systematically explore the Pareto front variation patterns of hydropower units across the entire operating range and to construct a knowledge-rich pre-trained model and solution set database. This allows for both direct practical operation using offline data and the laying of a foundation for online applications. This process achieves efficient and continuous global exploration through iterative cycles of operating space discretization, intelligent population initialization based on nearest neighbor prediction, multi-objective optimization, and incremental model learning, while ensuring the smoothness of the solution set across the operating space.

[0053] This process begins with the rational discretization of the continuous operating space. The head H is then set within its operating range. With fixed step size Discretize the image; similarly, consider the guide vane opening Y in... Within the interval, by step size Discretize the data to form a two-dimensional grid of operating points covering the entire operating domain:

[0054] Where i and j are indices for head and guide vane opening, respectively. This grid ensures comprehensive coverage of operating points. For each new operating point to be optimized... It obtains information from its nearest-neighbor optimized operating points to generate a high-quality initial population.

[0055] Specifically, the method first selects k operating conditions from the already optimized operating points that are compatible with... The closest operating point is denoted as To ensure the comparability of the head H and guide vane opening Y in distance calculations, the degree of proximity is measured by a normalized Euclidean distance:

[0056] in, , , , These represent the minimum and maximum values ​​of the head and guide vane opening, respectively, across the entire discrete operating space. This formula unifies the two operating parameters to the dimensionless interval [0,1], thus ensuring their balanced contribution to the distance metric. To exclude excessively distant operating points, a distance threshold is set here. Only considering The working points are determined. If there are fewer than k available working points, all of them are selected; if there are no available working points, the population is still initialized randomly.

[0057] Let the initial population size be N. For each nearest neighbor condition point... Contribution solution of allocation It is inversely proportional to its distance. The contribution of each nearest working point to the solution is calculated as follows. First, the weights are calculated using the following formula:

[0058] Then normalize the weights according to the following formula:

[0059] Finally, each neighbor The number of solutions to the assignment is Adjustments to ensure .

[0060] For each neighboring operating point From its Pareto solution set Random selection There are several solutions. For each selected solution... A BP neural network model M is used to predict its... The corresponding solution below:

[0061] Model M is trained synchronously during each optimization, and its input is... The output is the predicted solution. All predicted solutions To form a high-quality initial population .

[0062] With the initial population Starting from the NSGA-II algorithm, solve for the operating point. Pareto optimal solution set During the optimization process, the algorithm iterates based on multiple objective functions, ultimately obtaining a convergent and diverse solution set.

[0063] After optimization, the machine learning model M is immediately incrementally trained to incorporate new knowledge. Training data is obtained by constructing solution pair mappings. For each nearest neighbor operating point... its Pareto solution set and Pairing is performed to minimize the sum of the overall Euclidean distances. That is, solving for:

[0064] Where Z is a binary allocation matrix, satisfying and Here, the Hungarian algorithm is used for matching. After constructing the matching matrix, the model M is incrementally updated using the newly constructed training samples. The training loss function is the mean squared error, as shown in the following formula:

[0065] Finally, the model parameters are updated through backpropagation and optimization algorithms, enabling the model to better capture the patterns of solution changes with operating conditions.

[0066] Repeat the above process; the method applies to the working space. Each working condition point in the model is optimized. Initially, for the first working condition point, since there are no optimized points, a randomized initialization population and standard optimization are used. As optimization progresses, model M is continuously improved through incremental training, the quality of population initialization is improved, and overall convergence is accelerated.

[0067] After optimization of all operating conditions, an offline global knowledge base is obtained, containing the Pareto solution set and the trained model M for each operating condition. This knowledge base ensures the continuity and smoothness of the solution set in the operating condition space, providing a solid foundation for online dynamic optimization.

[0068] S4. Online rolling optimization integrating forward-looking predictions The global knowledge base built during the offline optimization phase lays a solid foundation for online applications. The online rolling optimization framework of this invention aims to transform offline global cognition into real-time closed-loop control capabilities, and through the integration of forward-looking predictive information, achieve a leap from passive response to proactive decision-making. The core of this framework lies in establishing a periodic "perception-prediction-optimization-execution" closed loop, driven primarily by a fixed-time rolling window.

[0069] Online systems operate on a fixed control cycle. Activated. At each activation moment. The system first senses and collects the current real-time operating data, mainly including the current working head. and actual power Subsequently, the system invokes a multi-timescale prediction model deployed in the background to obtain an estimate of future operating conditions.

[0070] The head prediction module employs a Long Short-Term Memory (LSTM) network model to capture the long-term temporal dependencies of hydrological processes. This module uses historical hydrological time-series data as input to predict head values ​​at multiple future points in time. Its core computational process can be represented as follows:

[0071] The model input consists of sequence data from the past n time points, including historical water head values ​​H, watershed rainfall R, and meteorological data W; the output is future data. Predicted head sequence at each time point LSTM selectively memorizes and forgets information through its unique gating mechanism, thereby effectively capturing the long-term trends and periodic characteristics of head changes.

[0072] The power forecasting module employs a Lightweight Gradient Boost Machine (LightGBM) model to handle short-term load demand fluctuations. This module uses historical power, temporal characteristics, and meteorological factors as input features to predict future load demand in the short term. The forecasting process can be described as follows:

[0073] The model input includes three feature categories: historical load data. Time characteristics and weather forecast information LightGBM efficiently learns the complex nonlinear relationship between features and loads using a histogram-based decision tree algorithm and gradient boosting framework, outputting the future... Load forecast sequence at each time point .

[0074] After calculating the predicted power sequence, and combining it with the predicted head information, the opening degree corresponding to each power is obtained through simple calculation. After acquiring the current and predicted operating conditions, the online optimizer is started. A look-ahead optimization strategy based on a double-warm-start approach is designed here. The core of this strategy is to fully utilize the historical Pareto front predictor M in the offline knowledge base, while combining it with real-time optimization results, to provide the highest quality initial population for both current and future optimization problems.

[0075] Online optimizer in each cycle We solve the following two dynamic multi-objective optimization problems in parallel. The optimization at the current time step can be expressed as:

[0076] The optimization of look-ahead time can be expressed as:

[0077] in, Let F be the feasible region of the decision variables, and F be the objective function vector. This formal definition clearly separates the current and future optimization tasks. To achieve efficient solution of the above problem, a dual warm-start mechanism is adopted. For the optimization at the current moment, the initial population is directly generated by the historical Pareto front predictor M based on the current measured conditions:

[0078] For the optimization of the look-ahead time, the initial population is formed by merging two parts, utilizing both historical prior information and real-time recursive information:

[0079] in, It is based on historical prior populations that predict operating conditions. It is the Pareto solution set obtained from optimization at the current moment, serving as the real-time recursive population. This fusion strategy provides a high-quality starting point for look-ahead optimization that combines a global perspective with local accuracy.

[0080] After the optimization calculation is complete, the system generates two outputs. For the current time... From its Pareto solution set In the process, a final solution is selected based on preset default preferences or weights specified in real time by the operator. After passing the safety verification logic, the command is sent to the unit speed control system for execution, completing the closed-loop control. Then, the next predicted time... Pareto solution set It is pushed to the human-computer interaction interface as advanced decision support information. This allows operators to anticipate potential performance trade-offs in the future.

[0081] This rolling optimization framework significantly improves the system's agility in responding to changes in operating conditions and the foresight of control quality by deeply embedding predictive information into the optimization loop and utilizing a time-progressive hot-start mechanism. It realizes a fundamental shift in the intelligent operation of hydropower units from static table lookup to dynamic foresight.

[0082] S5. Online self-learning mechanism to ensure long-term performance The aforementioned offline knowledge base, including the static parameter matrix, the Pareto front predictor M, and the online optimization framework, relies on historical data and simulation models for its initial performance. However, as a complex physical system, the operating characteristics of a hydropower unit slowly drift due to equipment aging, component wear, and changes in water flow characteristics. This can lead to a gradual deviation between the system built based on the initial model and the unit's actual optimal operating range. To address this issue, a hybrid online self-learning mechanism is introduced. Through real-time incremental updates to the predictor M and progressive updates to the static knowledge base, the entire optimization system possesses the ability to continuously evolve and adapt to long-term performance changes in the unit.

[0083] The mechanism first includes a closed loop for online monitoring and iterative optimization of the predictor model M's performance. During each online rolling optimization cycle, the system not only executes control tasks but also simultaneously collects data and evaluates performance. Specifically, the system uses the online optimizer to reflect the current real-world operating conditions. The Pareto approximate solution set obtained by calculation This serves as the gold standard for truly optimal performance in the current environment. Simultaneously, the system invokes the existing Pareto front predictor. For the same working condition Make predictions and obtain the predicted solution set. .

[0084] To quantify the accuracy of the model's predictions, the system calculates a performance consistency metric. This metric measures the distributional difference between the predicted solution set and the actual online optimization solution set, and its form is as follows:

[0085] in It is a distance function that measures the difference in distribution between two solution sets. The smaller the value, the better the predictor. The better the performance, the larger the value, the further it deviates from reality.

[0086] The system maintains a fixed-length sliding window W to store the performance consistency metrics calculated over the most recent m control cycles. The model update is triggered based on the overall performance of the metrics within this window. Updates are triggered when the moving average of the metrics within the window continuously exceeds a preset threshold. At that time, that is:

[0087] Once the above conditions are met, the system determines the original predictor. Performance degradation has occurred, and the model update process has been automatically initiated. The update process does not involve rebuilding the model from scratch; instead, it employs an efficient incremental learning strategy. The system will use all high-quality "working condition-online optimization solution set" paired data collected within the time period corresponding to the sliding window W. Construct as a fine-tuning dataset .

[0088] Subsequently, with the original The model serves as pre-trained weights to fine-tune the dataset. Using the new training samples, a finite number of iterations are performed with the objective of minimizing the following loss function:

[0089] in The loss is calculated on the original large-scale offline dataset to preserve historical knowledge and prevent catastrophic forgetting. In the incremental dataset The loss is calculated above; λ is a hyperparameter that weighs the importance of new and old knowledge. After training, the updated model is obtained. .

[0090] Meanwhile, to further ensure the long-term effectiveness of the entire offline knowledge base, this module also designs a progressive update reminder mechanism for the static parameter matrix. The online system updates the current actual working conditions each time. After dynamic multi-objective optimization, a real-time Pareto solution set will be obtained. Simultaneously, the system queries the Pareto solution set of adjacent operating conditions in the offline knowledge base and uses an interpolation algorithm to generate a corresponding Pareto solution set. interpolation Pareto solution set Subsequently, a set similarity metric was used to calculate the distributional difference between the set and the online optimization solution set:

[0091] Then, the system calculates the normalized Euclidean distance from the current actual operating condition to the nearest offline operating condition point:

[0092] Based on the distance calculated using the above formula, a dynamic threshold mechanism is calculated. The relationship between the threshold and the distance d is defined as follows:

[0093] in Here, k is the base threshold, and k is the scaling factor. This means that the further away from the fixed operating point, the lower the confidence in the interpolation result naturally becomes, and therefore the larger the allowable deviation θ. When a significant inconsistency is detected, i.e.:

[0094] The system automatically performs regional verification on N offline reference points surrounding the operating point. For each neighboring operating point... Perform the same consistency check:

[0095] The proportion of inconsistencies within a region is calculated using the following formula:

[0096] When the inconsistency rate exceeds the set threshold, the system generates a diagnostic report and issues an update warning, indicating that there may be a systematic deviation in the static knowledge base, and prompting the operators to decide whether to initiate a global update.

[0097] This hybrid dynamic update mechanism enables the optimization system of this invention to maintain the real-time accuracy of the predictor M through high-frequency incremental learning, while ensuring the long-term reliability of the static knowledge base through low-frequency global updates. This forms a dual-cycle industrial intelligent body that balances agility and robustness, providing core technical support for the long-term intelligent optimization operation of hydropower units.

[0098] This invention breaks through the limitations of traditional static optimization methods, and for the first time constructs the optimization of hydropower unit parameters under varying operating conditions as a dynamic multi-objective optimization problem. Operating parameters such as head and load are defined as time-varying environmental variables, enabling the algorithm to actively track the Pareto front during the continuous change of operating conditions.

[0099] This invention innovatively proposes a hot-start mechanism based on nearest-neighbor mapping. By learning a continuous mapping relationship from the operating condition space to the decision space through a machine learning model, a high-quality initial population is provided for optimization at each new operating condition. This method not only significantly accelerates convergence but also ensures a smooth transition of the solution set across the entire operating condition range, avoiding parameter jumps.

[0100] This invention deeply embeds water and power prediction into the optimization process, constructing a unique parallel look-ahead optimization mechanism. The system simultaneously solves for the optimal parameter set for both the current moment and the predicted future moment, achieving a leap from passive response to proactive decision-making. This design provides operators with forward-looking decision support, enhancing the intelligence level of unit operation.

[0101] This invention addresses the long-term performance drift problem of generator units by establishing a complete online self-learning closed loop. The system automatically triggers an incremental learning process by monitoring the deviation between the optimization effect and the model prediction in real time, enabling the core prediction model to adaptively update in response to changes in generator unit state, ensuring the system maintains optimal performance throughout its entire lifecycle.

[0102] Compared with the prior art, the significant advantages and beneficial effects of the present invention are as follows: 1. Introducing the continuity of the optimization process: This invention aims to address the problems caused by the forced discretization of the continuous operating space in traditional "operating point partitioning and static optimization" methods. It treats operating parameters such as head and guide vane opening as a continuously changing dynamic environment, and directly seeks optimization within the continuous operating space using a dynamic multi-objective optimization algorithm. This method not only ensures that the Pareto optimal solution set evolves smoothly and continuously with the changing operating conditions, avoiding abrupt changes in control parameters, but also effectively accelerates the convergence of the optimization process by utilizing the correlation between operating conditions.

[0103] 2. Achieving Real-Time and Foresight in System Optimization: This invention aims to improve the responsiveness of parameter optimization to actual operating environments. It connects the dynamic optimization system with real-time operating data, enabling online adaptive adjustment of controller parameters. Simultaneously, by introducing water level and power prediction functions, the system is upgraded to a rolling optimization framework with a certain degree of predictive capability. This allows the system to calculate the Pareto optimal frontier in advance, providing operators with future decision-making space and achieving a leap from passive response to proactive control.

[0104] 3. Enhancing the Long-Term Adaptability and Self-Evolution Capability of the Optimization System: This invention aims to address the long-term performance drift of the unit caused by equipment aging, changes in water flow characteristics, etc. It establishes an online self-learning mechanism to monitor the deviation between the optimization effect and the model prediction in real time, and automatically triggers an incremental learning process when performance degradation is detected. This system can utilize freshly generated optimization data online to dynamically fine-tune the core prediction model, enabling the optimization system to continuously update itself as the unit's operating status changes, maintaining the long-term effectiveness of the optimization effect and reducing the need for later system maintenance.

[0105] This invention achieves a complete optimization process at each key stage through a systematic technical design. Specifically, the establishment of a dynamic multi-objective optimization framework, by defining operating parameters such as head and load as time-varying environmental variables, enables the algorithm to continuously track the Pareto front during the continuous change of operating conditions. This achieves a transformation from static discrete optimization to dynamic continuous optimization, resulting in continuous adaptive optimization across all operating conditions. Based on a near-neighbor mapping-based hot-start mechanism, a machine learning model learns the continuous mapping relationship from the operating condition space to the decision space. This provides a high-quality initial population for each new operating condition point, improving the computational efficiency of offline optimization while ensuring a smooth transition of the solution set within the operating condition space.

[0106] Online rolling optimization that integrates forecast information constructs a parallel look-ahead optimization mechanism by deeply embedding head prediction and power prediction into the optimization process. This enables the system to solve for the optimal parameter set for both the current moment and the future prediction moment, thereby enhancing the system's ability to respond to changes in operating conditions.

[0107] The introduction of an online self-learning mechanism, by monitoring the deviation between the optimization effect and the model prediction in real time and automatically triggering the incremental learning process, enables the core prediction model and the offline knowledge base to adaptively update in response to changes in the random group state, maintaining the effectiveness of the optimization system throughout its entire lifecycle. The implementation of a dual warm-start strategy, by combining the global prior of the historical knowledge base with the local recursive information of the real-time optimization results, provides a high-quality starting point for look-ahead optimization that combines a global perspective with local accuracy, improving the accuracy and reliability of the optimization results.

[0108] Through systematic design and coordination, these technical aspects have created a synergistic effect in terms of efficiency optimization, operational stability, timely response, and long-term adaptability, providing complete technical support for the intelligent operation of hydropower units.

[0109] This invention also provides a computer-readable storage medium having a computer program stored thereon, the computer program executing the steps of the method described above when run by a processor.

[0110] This invention further provides a dynamic multi-objective optimization system for hydropower unit parameters under varying operating conditions, including an interconnected memory and a processor. The memory stores a computer program, which executes the steps of the method described above when run by the processor.

[0111] The medium and system of the present invention, corresponding to the methods described above, also have the advantages described above.

[0112] The present invention can implement all or part of the processes in the methods of the above embodiments, or it can be implemented by hardware related to computer program instructions. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of the above method embodiments. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable storage medium includes: any entity or device capable of carrying computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. The memory is used to store computer programs and / or modules. The processor implements various functions by running or executing the computer programs and / or modules stored in the memory, and by calling data stored in the memory. The memory may include high-speed random access memory, as well as non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital (SD) cards, flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0113] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A dynamic multi-objective optimization method for parameters of hydropower units under varying operating conditions, characterized in that, Including the following steps: S1. A nonlinear simulation model is established based on the comprehensive characteristic curve of the turbine, and the turbine torque M and flow rate Q are expressed as functions of guide vane opening α, rotational speed n and head H; a multi-dimensional performance evaluation system including frequency regulation performance, mechanical stability and damping characteristics is established. The controller parameter optimization problem is formalized as a dynamic multi-objective optimization problem, and the optimization objective is defined as minimizing the performance index vector. S2. Construct an offline / online integrated optimization system architecture, including an offline global optimization layer, an online rolling optimization layer, and an online self-learning layer; S3. For each new operating point, a high-quality initial population is generated using a machine learning model based on the Pareto solution set of the nearest neighboring optimized operating points. The NSGA-II algorithm is used to solve the Pareto optimal solution set for the working conditions. After optimization, the machine learning model is incrementally trained to update its mapping capability. This process is repeated until all working conditions are covered, forming an offline knowledge base. S4. Collect real-time operating data at fixed intervals and call LSTM and LightGBM models to predict future head and power; based on the prediction results, adopt a dual hot-start mechanism to solve the dynamic multi-objective optimization problem at the current time and the future time in parallel; send the optimal solution selected from the optimization results at the current time to the speed control system for execution, and push the Pareto solution set at the future time to the human-machine interface for decision-making reference; S5. Monitor the consistency index between the online optimized solution set and the predicted solution set in real time. Trigger incremental model updates when the index continuously exceeds the limit. At the same time, compare the differences between the online optimized solution set and the offline interpolation solution set, adaptively adjust the threshold according to the distance, and issue a knowledge base update warning when regional inconsistency is detected.

2. The dynamic multi-objective optimization method for hydropower unit parameters under varying operating conditions according to claim 1, characterized in that, In step S1, establishing a nonlinear simulation model based on the comprehensive characteristic curve of the turbine specifically includes: Calculate the unit speed based on the actual rotational speed n, the turbine runner diameter D, and the working head H. : ; per unit rotational speed and current guide vane opening Using the input as input, query the comprehensive characteristic curve of the turbine to obtain the corresponding unit torque. and unit flow Finally, the actual torque is calculated using the following formula. and traffic : 。 3. The dynamic multi-objective optimization method for hydropower unit parameters under varying operating conditions according to claim 1 or 2, characterized in that, In step S3, for each new operating condition, the specific process of generating a high-quality initial population using a machine learning model based on the Pareto solution set of the nearest neighbor optimized operating conditions is as follows: First, select k operating conditions from the already optimized operating conditions that are compatible with the new operating conditions. The closest operating point is denoted as ; To ensure the comparability of head H and guide vane opening Y in distance calculations, the degree of closeness is determined by the normalized Euclidean distance. measure: in, , , , These are the minimum and maximum values ​​of head and guide vane opening, respectively, in the entire discrete operating space; Let the initial population size be N; each nearest neighbor condition point Contribution solution of allocation It is inversely proportional to its distance; calculate the contribution data of each nearest working point; first calculate the weight. : Then normalize the weights. ; Finally, each neighboring operating point The number of solutions to the assignment is Adjustments to ensure ; For each neighboring operating point From its Pareto solution set Random selection There are one solution; for each selected solution A BP neural network model M is used to predict its... The corresponding solution below : Model M is trained synchronously during each optimization, and its input is... The output is the predicted solution. All predicted solutions To form a high-quality initial population .

4. The dynamic multi-objective optimization method for hydropower unit parameters under varying operating conditions according to claim 3, characterized in that, In step S3, the initial population is used. Starting from the NSGA-II algorithm, solve for the new operating condition. Pareto optimal solution set ; After optimization, the machine learning model M is incrementally trained to incorporate new knowledge; Training data is obtained by constructing solution pair mappings; for each nearest neighbor operating point... its Pareto solution set and Pairing is performed to minimize the sum of the overall Euclidean distances, which is the solution: Where Z is a binary allocation matrix, satisfying and The Hungarian algorithm is used for matching; for The j-th solution in the model; after constructing the matching matrix, the model M is incrementally updated using the newly constructed training samples; the training loss function is the mean squared error. ; Finally, the model parameters are updated through backpropagation and optimization algorithms, enabling the model to better capture the patterns of solution changes with operating conditions.

5. The dynamic multi-objective optimization method for hydropower unit parameters under varying operating conditions according to claim 1 or 2, characterized in that, In step S4, the specific process of collecting real-time operating data at fixed intervals and calling the LSTM and LightGBM models to predict future head and power is as follows: The head prediction module employs a Long Short-Term Memory (LSTM) network model to capture long-term temporal dependencies in hydrological processes. This module uses historical hydrological time-series data as input to predict head values ​​at multiple future moments. Its core calculation process is as follows: The model input consists of sequence data from the past n time points, including historical water head values ​​H, watershed rainfall R, and meteorological data W; the output is future data. Predicted head sequence at each time point ; The power forecasting module employs the Lightweight Gradient Boost Machine (LightGBM) model to handle short-term load demand fluctuations. This module uses historical power, temporal characteristics, and meteorological factors as input features to predict future short-term load demand. The forecasting process is described as follows: The model input includes three feature categories: historical load data. Time characteristics and weather forecast information LightGBM efficiently learns the complex nonlinear relationship between features and loads using a histogram-based decision tree algorithm and gradient boosting framework, outputting future... Load forecast sequence at each time point .

6. The dynamic multi-objective optimization method for hydropower unit parameters under varying operating conditions according to claim 5, characterized in that, In step S4, the specific process of the dual hot-start mechanism is as follows: For optimization at the current moment, its initial population The results are directly generated by the historical Pareto front predictor M based on the current measured operating conditions. : Regarding the optimization of the look-ahead time, its initial population... It is composed of two parts, which simultaneously utilize historical prior information and real-time recursive information: in, It is based on historical prior populations that predict operating conditions. It is the Pareto solution set obtained at the current moment, which serves as the real-time recursive population.

7. The dynamic multi-objective optimization method for hydropower unit parameters under varying operating conditions according to claim 1 or 2, characterized in that, In step S5, the specific process of incremental model update is as follows: All high-quality "operating condition-online optimization solution set" paired data collected within the corresponding time period of the sliding window W. Construct as a fine-tuning dataset ; Subsequently, with the original The model serves as pre-trained weights to fine-tune the dataset. Using the new training samples, a finite number of iterations are performed with the objective of minimizing the following loss function: in The loss function; The loss is calculated on the original large-scale offline dataset to preserve historical knowledge and prevent catastrophic forgetting. In the incremental dataset The loss calculated above; λ is a hyperparameter that weighs the importance of new and old knowledge; After training, the updated model is obtained. .

8. The dynamic multi-objective optimization method for hydropower unit parameters under varying operating conditions according to claim 7, characterized in that, In step S5, the specific process of issuing a knowledge base update warning is as follows: The online system completes each assessment of the current actual working conditions. After dynamic multi-objective optimization, a real-time Pareto solution set will be obtained. Simultaneously, the system queries the Pareto solution set of adjacent operating conditions in the offline knowledge base and uses an interpolation algorithm to generate a corresponding Pareto solution set. interpolation Pareto solution set ; Subsequently, a set similarity metric was used to calculate the distributional difference between the set and the online optimization solution set. : Then, calculate the normalized Euclidean distance from the current actual operating condition to the nearest offline operating condition point. : Based on the calculated normalized Euclidean distance Calculate dynamic threshold ; When a significant inconsistency is detected, i.e.: Automatically perform regional verification on N offline reference points surrounding the operating point; for each neighboring operating point The same consistency check is performed: by calculating the proportion of inconsistencies within a region, when the proportion of inconsistencies exceeds a set threshold, the system generates a diagnostic report and issues an update warning, indicating that there may be a systematic deviation in the static knowledge base, and prompting the operators to decide whether to initiate a global update.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program, when run by a processor, performs the steps of the method as described in any one of claims 1-7.

10. A dynamic multi-objective optimization system for parameters of a hydropower unit under varying operating conditions, comprising an interconnected memory and a processor, wherein the memory stores a computer program, characterized in that... The computer program, when run by a processor, performs the steps of the method as described in any one of claims 1-7.

Images