A control method of a water turbine speed regulation optimization system based on hybrid intelligent feedforward compensation

Abstract

A control method for a turbine speed regulation optimization system based on hybrid intelligent feedforward compensation belongs to the field of turbine speed regulation system control technology. This method simultaneously triggers two feedforward channels—data-driven and knowledge-driven—when a change in head or load is detected: the data-driven channel outputs a first opening compensation amount using a neural network model trained on historical operating data, while the knowledge-driven channel outputs a second opening compensation amount using a fuzzy logic controller built based on the turbine's physical characteristics and expert experience. The severity index of the operating condition change is calculated based on the changes in head and load, and the two compensation amounts are adaptively weighted and fused to generate the final feedforward opening compensation amount. This compensation amount is then superimposed on the PID controller output command to form the main guide vane opening control signal. This invention overcomes the lag of traditional feedback control through a hybrid feedforward mechanism, improving the dynamic response speed and regulation stability of the unit under conditions of large head fluctuations and frequent load changes.

Description

Technical Field

[0001] This invention belongs to the field of turbine speed regulation system control technology, and specifically relates to a turbine speed regulation optimization system control method based on hybrid intelligent feedforward compensation. Background Technology

[0002] The turbine speed control system is the core control unit for ensuring frequency stability and load regulation in hydropower stations, and its performance directly affects grid security and unit operating efficiency. Currently, most mainstream speed governors employ a PID feedback control structure, adjusting the guide vane opening by comparing speed deviations in real time. However, this control method is essentially "reactive": the controller only begins to act after the speed deviation has occurred, exhibiting inherent lag.

[0003] In actual operation, hydro-generator units often face two typical disturbances: head changes caused by upstream water level fluctuations and power demand changes caused by sudden changes in grid dispatch or user load. Both types of disturbances directly alter the turbine torque balance, leading to speed fluctuations. Because PID control cannot predict the impact of disturbances and can only respond passively, the overshoot during regulation is large and the recovery time is long. Especially under conditions of high head fluctuations or frequent peak shaving, the unit speed stability decreases significantly, and may even trigger protection actions.

[0004] To improve dynamic performance, some systems have introduced feedforward control. However, existing feedforward strategies mostly rely on a single model: feedforward based on physical equations requires precise parameters and has poor adaptability under off-design conditions; while data-driven feedforward (such as neural networks) can fit complex nonlinear relationships, it lacks interpretability and has weak generalization ability in small sample areas. Therefore, how to construct a feedforward mechanism that can compensate for the impact of sudden changes in head and load on speed in advance, accurately, and robustly, in order to overcome the lag of pure feedback control, is a key technical problem for improving the dynamic quality of turbine speed control systems. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a control method for a turbine speed regulation optimization system based on hybrid intelligent feedforward compensation. The present invention overcomes the lag of traditional feedback control through a hybrid feedforward mechanism, and improves the dynamic response speed and regulation stability of the unit under conditions of large head fluctuation and frequent load changes.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A control method for a turbine speed regulation optimization system based on hybrid intelligent feedforward compensation, comprising the following steps: S1: Real-time monitoring of turbine unit operation status. When a change in head or load condition is detected, the change in head, the change in load condition, the current head value, the current load value, and the current guide vane opening value are collected to trigger feedforward compensation control. S2: The head change, load condition change, current head value, current load value, and current guide vane opening value are simultaneously input to the data-driven feedforward channel and the knowledge-driven feedforward channel, which operate in parallel. The data-driven feedforward channel sends the input variables to a neural network model trained with historical operating data and outputs the first opening compensation value. The knowledge-driven feedforward channel sends the same input variables to a fuzzy logic controller built based on the physical characteristics of the turbine and expert experience and outputs the second opening compensation value. S3: Calculate the working condition change intensity index based on the input variables to characterize the degree of change in working conditions, and based on this index, perform adaptive weighted fusion of the first opening compensation amount and the second opening compensation amount through an intelligent arbitration strategy to generate the final feedforward opening compensation amount. S4: The final feedforward opening compensation amount is superimposed with the opening command output by the PID controller to form the total opening control command of the turbine guide vanes. The guide vane relay is then driven by the electro-hydraulic servo system to adjust the turbine inflow and achieve fast and stable load tracking.

[0007] As a preferred embodiment, the construction of the neural network model includes: A1: Select sample data containing steady-state switching processes from the historical operation database of the turbine units, and construct a training dataset containing load disturbance, head disturbance and corresponding guide vane opening changes; A2: Perform data cleaning, outlier removal, and missing value handling on the training dataset to ensure data quality; A3: Normalize the cleaned data to eliminate dimensional differences; A4: A neural network is constructed using a multilayer perceptron structure, with load disturbance, head disturbance, current load, current head and current opening as input nodes, and guide vane opening change as output node. Two to three hidden layers are set, and the ReLU activation function is used. A5: The network is trained using the backpropagation algorithm combined with the Adam optimizer, and L2 regularization is introduced to prevent overfitting. A6: Evaluate the model performance on an independent test set and optimize hyperparameters or adjust the network structure based on the validation results.

[0008] As a preferred embodiment, the design of the fuzzy logic controller includes: B1: Define the universe of discourse for load condition changes and head changes, and divide it into multiple fuzzy subsets. The linguistic variables include negative large, negative small, zero, positive small, and positive large. B2: Define the output universe of discourse and the corresponding fuzzy subset of the second opening compensation quantity, with the linguistic variables being consistent with the input variables; B3: Based on the dynamic balance relationship of "head-opening-output" of the water turbine and the safety operation constraints, establish a rule base containing multiple fuzzy rules. The rule form is "If the load change is A and the head change is B, then the opening compensation is C". B4: The Mamdani inference method is used for fuzzy inference, and the fuzzy output is defuzzified into a precise second opening compensation amount by the centroid method. B5: Verify the performance of the fuzzy controller through simulation or actual operating data, and optimize the membership function or rule base based on response speed, overshoot, and steady-state error indicators.

[0009] As a preferred embodiment, the calculation of the operating condition change severity index includes: C1: Based on the actual operating range of the unit, define the typical variation range of load conditions and head as the normalization benchmark. C2: Divide the change in head and the change in load condition by their corresponding typical ranges to obtain the dimensionless normalized change in head and the normalized change in load. C3: The sum of the absolute values ​​of the normalized head change and the normalized load change yields the index of the severity of the change in operating conditions.

[0010] As a preferred embodiment, the intelligent arbitration strategy employs a continuously monotonically decreasing weight function, which causes the weight of the data-driven feedforward channel to decrease as the intensity of changes in operating conditions increases, while the weight of the knowledge-driven feedforward channel increases accordingly. When the intensity of changes in operating conditions approaches zero, the data-driven channel is fully trusted; when the intensity of changes in operating conditions is extremely high, the knowledge-driven channel is fully relied upon.

[0011] As a preferred embodiment, the weighting function includes an adjustable parameter to adjust the aggressiveness or conservatism of the control strategy of the system under different power plants or units; this parameter can be determined through offline simulation or online experiment optimization.

[0012] As a preferred embodiment, a first-order inertial element and a dead-zone element are set in the feedforward compensation control loop; the first-order inertial element is used to filter out high-frequency noise and equivalently delay the system sampling time, while the dead-zone element is used to suppress frequent feedforward actions caused by small disturbances and improve system stability.

[0013] As a preferred option, the method described herein is applicable to the rapid load response control of hydro-generator units under power grid peak shaving and frequency regulation conditions, and is particularly suitable for operating scenarios with large head fluctuations and frequent load changes.

[0014] As a preferred option, the hybrid intelligent feedforward compensation control module is integrated into the existing turbine governor control system in the form of a software functional module, without changing the main hardware structure of the governor system, and has good engineering compatibility and portability.

[0015] As a preferred option, the method is applicable to turbine speed control in both opening mode and power mode, and can be adapted to different control objectives by switching the type of input variable.

[0016] In A1, the dataset construction process is as follows: Historical operating data of the hydro turbine unit Selected from and with , ; i represents the i-th time, where the baseline load condition, head, and guide vane opening begin to change, and a dataset is constructed. ; in, These represent the load condition, head, and guide vane opening of the unit during its current stable operation, respectively. These represent the unit's state points. During stable operation, load and head disturbances cause the unit to shift from the current state point to the next stable state point. During the transfer process, the changes in load conditions, the changes in head, and the resulting changes in guide vane opening are considered.

[0017] In A3, the normalization process uses the Min-Max normalization method, which independently normalizes each feature to... The interval is calculated using the following formula: ; Where X represents the original data. and These are the minimum and maximum values ​​of the feature in the entire dataset, respectively.

[0018] In A4, neural networks use a set of variables. As input, the change in guide vane opening As output, the network input layer contains 5 nodes and the output layer contains 1 node; the hidden layers are set to 2 to 3 layers, and the number of neurons in each layer is selected from 64, 128, and 256; the ReLU function is used as the activation function for the hidden layers.

[0019] In B1, the change in load condition The domain of discourse is defined as Change in water head The domain of discourse is defined as Both are divided into 5 fuzzy subsets, with linguistic variables being negative large NB, negative small NS, zero ZO, positive small PS, and positive large PB, respectively; the membership function of the fuzzy subsets adopts a trigonometric function, and its parameters are determined based on actual operating data.

[0020] In B2, the second opening compensation amount The domain of discourse is defined as It is divided into 5 fuzzy subsets, and the linguistic variables include negative large NB, negative small NS, zero ZO, positive small PS, and positive large PB; its membership function adopts trigonometric function or Gaussian function, and the specific parameters are determined according to the actual running data.

[0021] In B4, the centroid method is used for defuzzification, and its calculation formula is as follows:

[0022] Where y is the value on the output universe of discourse. This represents the total membership degree of the corresponding output value.

[0023] A turbine speed regulation optimization control system based on hybrid intelligent feedforward compensation, employing the aforementioned control method, comprises: The data acquisition module is used to monitor the operating status of the turbine unit in real time. When a change in head or load is detected, it collects the change in head, the change in load, the current head value, the current load value, and the current guide vane opening value, and triggers feedforward compensation control. The hybrid feedforward compensation module includes a parallel data-driven feedforward channel and a knowledge-driven feedforward channel. The data-driven feedforward channel contains a neural network model trained on historical operating data, which is used to output the first opening compensation amount. The knowledge-driven feedforward channel contains a fuzzy logic controller built based on the physical characteristics of the turbine and expert experience, which is used to output the second opening compensation amount. The intelligent arbitration module is used to adaptively weight and fuse the first opening compensation amount and the second opening compensation amount according to the index of the severity of changes in working conditions, and generate the final feedforward opening compensation amount. The command output module is used to superimpose the final feedforward opening compensation amount with the opening command output by the PID controller to form the total opening control command of the turbine guide vanes, and drive the guide vane relay to adjust the turbine inflow through the electro-hydraulic servo system.

[0024] The present invention can achieve the following beneficial effects: 1. By introducing a hybrid intelligent feedforward compensation mechanism, the lag of traditional feedback control is effectively overcome, making the speed regulation system respond more quickly and timely to changes in head and load, while retaining the stability advantage of feedback control in "eliminating steady-state error". This significantly improves the regulation quality during transient processes such as load changes and helps to enhance the frequency stability of the power grid.

[0025] 2. By integrating the advantages of data-driven and knowledge-driven methods, the system can maintain high-precision control under normal operating conditions and rely on expert knowledge to maintain reliable operation under extreme conditions where training data is not covered, significantly expanding the stable operating range of the system.

[0026] 3. By optimizing the control process, the unit's adjustment actions are smoothed, the mechanical and electrical stress of the equipment is reduced, thereby reducing maintenance requirements, extending the service life of key equipment, and improving the overall operating economy of the unit. Attached Figure Description

[0027] The present invention will be further described below with reference to the accompanying drawings and embodiments: Fig. 1 This is an optimized block diagram of a hybrid intelligent feedforward compensation control speed regulation system for a water turbine under the opening mode of the present invention; Fig. 2 This invention provides a nonlinear control structure diagram for a turbine speed regulation system. Detailed Implementation

[0028] Preferred solutions include Figs. 1-2 As shown, a control method for a turbine speed regulation optimization system based on hybrid intelligent feedforward compensation includes the following steps: S1: Real-time monitoring of turbine unit operation status. When a change in head or load condition is detected, the change in head, the change in load condition, the current head value, the current load value, and the current guide vane opening value are collected to trigger feedforward compensation control. In this embodiment, the operating status of the turbine unit is monitored in real time. When changes in head and load conditions are detected, the head change is collected. Load variation Current head value and current load condition value Current guide vane opening value This triggers feedforward compensation control.

[0029] S2: The head change, load condition change, current head value, current load value, and current guide vane opening value are simultaneously input to the data-driven feedforward channel and the knowledge-driven feedforward channel, which operate in parallel. The data-driven feedforward channel sends the input variables to a neural network model trained with historical operating data and outputs the first opening compensation value. The knowledge-driven feedforward channel sends the same input variables to a fuzzy logic controller built based on the physical characteristics of the turbine and expert experience and outputs the second opening compensation value. In this embodiment, the... , , , Simultaneously input to the data-driven feedforward channel and the knowledge-driven feedforward channel, which operate in parallel; the data-driven feedforward channel is used to process the... , , , The input is fed into a neural network model that has been trained using historical data to obtain the first opening compensation amount. The knowledge-driven feedforward channel is used to... , , , The input is fed into a fuzzy logic controller based on expert experience to obtain the second opening compensation amount. .

[0030] The construction of the neural network model includes: (1) Construction of the dataset Historical operating data of the hydro turbine unit 、 、 、...、 、 、 Selected from and with ( x The numerators are 0, 1, 2, ..., n; i For the first i A dataset is constructed based on the load conditions, head, and guide vane opening that begin to change at a given time. .

[0031] in, These represent the load condition, opening degree, and head of the unit during its current stable operation, respectively. These represent the unit's state points. During stable operation, load and head disturbances cause the unit to shift from the current state point to the next stable state point. During the transfer process, the changes in load conditions, the changes in head, and the resulting changes in guide vane opening are considered.

[0032] (2) Data preprocessing For the constructed dataset Preprocessing is performed to improve data quality, neural network training efficiency, and model generalization performance, taking into account the characteristics of turbine operation data. Preprocessing mainly includes the following steps: 1) Data cleaning: Remove outliers and missing data to ensure the quality of the dataset.

[0033] In terms of outlier handling, a physical threshold method is adopted to address the potential instantaneous jump errors or communication interference in the turbine sensor signals. Based on the physical operating limits and safety regulations of the turbine unit, an absolute threshold range is set for each variable, and data points exceeding this range are directly removed.

[0034] Regarding missing value handling, for samples with missing data records, if the missing rate is low (such as a single missing sample point), linear interpolation is used to fill the missing data, which is consistent with the characteristics of the turbine operating conditions when they change smoothly. If multiple consecutive data points are missing or the missing rate is high, the sample record is directly deleted to ensure the reliability of the data.

[0035] 2) Normalization: Hydropower turbine data exhibits characteristics of varying dimensions and orders of magnitude. Directly using it for training can lead to an imbalance in network weights, hindering convergence. Therefore, normalization of the dataset is necessary. Normalization is performed to bring the features within the range of [-1, 1], thereby improving the training efficiency and model performance of the neural network. This invention employs the Min-Max normalization method, independently normalizing each feature to the interval [-1, 1]. The calculation formula is as follows:

[0036] in, This is the original data. and These are the minimum and maximum values ​​of the feature in the entire dataset, respectively. This method effectively preserves the distribution relationship between data points and accelerates the convergence speed of gradient descent.

[0037] 3) Data partitioning: Divide the dataset into training, validation and test sets, usually using a ratio of 70% training set, 15% validation set and 15% test set to ensure the model's generalization ability.

[0038] 4) Neural network architecture design: To address the multivariable, strongly coupled, and nonlinear characteristics of hydro-turbine regulation systems and to accurately fit their complex dynamic processes, this invention employs a multilayer perceptron (MLP) structure to construct a data-driven feedforward model. This network consists of a set of variables characterizing the current operating state and disturbances of the unit. As input to the model, the amount of change in guide vane opening to be predicted. As the expected output, the network input layer is designed with 5 nodes, corresponding to the 5 input variables mentioned above; the output layer is designed with 1 node, directly outputting the predicted value. value.

[0039] To ensure the model possesses sufficient nonlinear mapping capability to capture the complex operating condition transitions of the hydro turbine system, the network has 2 to 3 hidden layers. The number of neurons in each layer is optimally selected from typical configurations such as 64, 128, and 256, based on the dataset size and complexity. This aims to fully learn the higher-order coupling relationships between input and output, while avoiding overfitting due to excessive model complexity or underfitting due to excessive simplicity. The ReLU (Rectified Linear Unit) activation function is used in the hidden layers to effectively alleviate the vanishing gradient problem, accelerate the model training process, and enhance its ability to express nonlinear characteristics. The neural network structure is shown in the figure.

[0040] (4) Training process The neural network model is trained using the backpropagation algorithm, and the specific steps are as follows: Initialization parameters: Given that neural networks use the ReLU activation function, in order to alleviate the gradient vanishing problem and promote stable gradient propagation in the early stage of training, this invention adopts the He initialization method to randomly initialize network weights and biases. This method is particularly suitable for ReLU family activation functions and can accelerate model convergence.

[0041] Define the loss function: Use the mean squared error (MSE) as the loss function to calculate the error between the predicted value and the true value.

[0042] Optimizer Selection: To address the noise and non-stationary optimization objective present in the turbine operating data, the Adaptive Moment Estimation (Adam) optimizer was selected. This optimizer combines momentum and adaptive learning rate mechanisms, enabling efficient handling of sparse gradients and non-stationary objective functions. The initial learning rate was set to... %, and introduce a learning rate decay strategy (such as per %). decay per training cycle (%), to refine the search step size in the later stages of training, improve convergence accuracy, and avoid getting trapped in local optima.

[0043] Regularization method: In order to prevent the model from overfitting the noise and specific working details that may exist in the training data and improve the generalization ability, L2 regularization (weight decay term) is introduced into the loss function to constrain the weight magnitude and promote the model to learn more robust features.

[0044] Training iteration and validation: Set the training cycle as follows The training process uses mini-batch gradient descent for each epoch, with forward and backward propagation performed on the training set data in each iteration to update the network parameters. After each epoch, an independent validation set is used to calculate the loss metric and monitor model performance. During training, the overfitting trend is dynamically assessed based on the validation set loss curve, and early stopping strategies or hyperparameter adjustments are employed to ensure that the final model parameters possess both excellent fitting ability and generalization performance.

[0045] (5) Model evaluation and optimization Performance metrics: Calculate the mean squared error (MSE), root mean squared error (RMSE), and coefficient of determination (R²) of the model on the test set to evaluate the model's prediction accuracy and generalization ability.

[0046] Hyperparameter tuning: Optimize the hyperparameters of the neural network, such as the number of hidden layer nodes, learning rate, and regularization coefficient, through grid search or random search methods to obtain the best model performance.

[0047] Model comparison: The neural network model is compared with traditional methods (such as linear regression and support vector regression) to verify its superiority.

[0048] (6) Validation of training results Real-world verification: The trained neural network model is applied to real-world operating data to verify its prediction accuracy and robustness under different operating conditions.

[0049] Practicality assessment: The practicality of the model is assessed by comparing the actual change in guide vane opening with the model's predicted value.

[0050] Continuous optimization mechanism: Based on the validation results, further optimize the model structure or training strategy to improve the model's performance.

[0051] The design of the fuzzy logic controller includes: (1) Define the input variable: load condition change and change in head The universe of discourse, fuzzy subsets, and their membership functions.

[0052] The load condition change The domain of discourse is defined as and divided into A fuzzy subset; water head change The domain of discourse is defined as and divided into A fuzzy subset; the linguistic variables of the fuzzy subset include negative large, negative small, zero, positive small, and positive large; the linguistic variables of the fuzzy subset include negative large (NB), negative small (NS), zero (ZO), positive small (PS), and positive large (PB). For example, load condition variation. The membership function of the fuzzy subset can be a trigonometric function or a Gaussian function. The specific parameters can be determined based on the actual operating data, such as NB: NS: ZO: PS: PB: .

[0053] (2) Define the output variable, the second opening compensation amount. Fuzzy subsets and membership functions; Second opening compensation amount The domain of discourse is defined as and divided into A fuzzy subset; the linguistic variables of the fuzzy subset include negative large, negative small, zero, positive small, and positive large; the linguistic variables of the fuzzy subset include negative large (NB), negative small (NS), zero (ZO), positive small (PS), and positive large (PB). Second opening compensation amount. The membership function of the fuzzy subset can be a trigonometric function or a Gaussian function, and the specific parameters can be determined based on the actual running data.

[0054] (3) Establish a fuzzy rule base based on the experience of experts in turbine unit control.

[0055] The fuzzy rule base is based on the experience of turbine unit control experts and encapsulates the decision-making knowledge of operators under different operating conditions.

[0056] Therefore, based on the experience of experts in turbine unit control, multiple fuzzy control rules were established, with the rule form being: "If". It is A and (and) If it is B, then (Then) "It is C", where A, B, and C are fuzzy subset linguistic variables. The specific content of the fuzzy rule base can be designed as shown in Table 1.

[0057] Table 1. Based on the regulating characteristics of the water turbine Fuzzy control rule base

[0058] The design of the aforementioned rule base strictly adheres to the dynamic balance relationship of the turbine's "head-opening-output" and safety operation constraints. Its design logic is as follows: Load-dominated rules, i.e., when the water head remains constant ( When ZO = 0, the control strategy fully responds to load changes. This set of rules embodies the core power-aperture regulation relationship of the speed control system. For example, the rule "If ZO = 0, the control strategy fully responds to load changes." is NB and is ZO, Then The basis for "is NB" is: when the water head remains constant ( = ZO) while the load decreased significantly ( When = NB), to prevent the unit speed from being too high, a large negative guide vane opening must be output ( =NB) compensation amount to quickly reduce power generation.

[0059] Head-dominated rule, that is, when the load remains constant ( When ZO = 0, the control strategy responds to changes in head. This set of rules reflects the energy conversion characteristics of the turbine. For example, the rule "If ZO = 0, the control strategy responds to changes in head". is ZO and is NB, Then The basis for "is PS" is: when the load remains constant ( = ZO) while the head decreased significantly ( When the water flow potential energy is reduced significantly (= NB), a large positive second opening compensation amount is required. = PB) to compensate for the decrease in unit water flow energy in order to maintain rated output.

[0060] Composite operating condition rules (other areas): Rule Example 1 (Synergistic Effect): A significant increase in load ( = PB (needs to increase guide vane opening) and a slight decrease in head ( = NS (requiring increased guide vane opening) acts in the same direction as the other, producing a synergistic effect, thus outputting a large positive second opening compensation amount ( = PB).

[0061] Rule Example 2 (Conflict Effect - Conservative Strategy): Load significantly reduced ( =NB (guide vane opening needs to be reduced) and the head is significantly reduced ( =NB (requiring an increased guide vane opening) strongly conflict with the effect. In this case, a conservative strategy is adopted, outputting a second opening compensation amount with zero compensation ( = ZO).

[0062] Rule Example 3 (Conflict Effect - Load Dominance): Under this condition, the load increases slightly ( = PS (needs to increase guide vane opening) and a significant increase in head ( = PB (requiring a reduction in guide vane opening) has the opposite effect. Considering all factors, prioritizing safety to prevent overload, the output should have a smaller second opening compensation amount ( = NS), appropriately reduce the opening.

[0063] (4) Fuzzy reasoning and defuzzification Fuzzy inference employs the Mamdani inference method, based on input variables. and The fuzzy value is used to derive the second opening compensation amount of the output variable by combining it with the fuzzy rule base. The fuzzy value.

[0064] Defuzzification then uses the centroid method to transform the fuzzy inference result into a precise second opening compensation quantity. This is used for subsequent control. The centroid method calculates the centroid position of the output fuzzy set as the precise value. The preferred embodiment uses the centroid method for defuzzification calculation. This method uses the abscissa value corresponding to the centroid of the area enclosed by the membership functions and abscissas of all fuzzy output sets as the precise output value. The calculation formula is as follows:

[0065] in, To output the value on the universe of discourse, This represents the total membership degree of the corresponding output value.

[0066] (5) Verification and optimization of fuzzy logic controller The performance of the fuzzy logic controller is verified through simulation or actual operating data, mainly evaluating its response speed, overshoot and steady-state error under different operating conditions.

[0067] Based on the verification results, controller performance can be optimized by adjusting the partitioning of fuzzy subsets, optimizing the parameters of the membership function, or improving the fuzzy rule base. For example, if performance is poor under extreme conditions, corresponding fuzzy rules can be added or the partitioning range of fuzzy subsets can be adjusted; if the response speed is slow, the shape of the membership function can be optimized or the weights of the fuzzy rules can be adjusted.

[0068] S3: Calculate the working condition change intensity index based on the input variables to characterize the degree of change in working conditions, and based on this index, perform adaptive weighted fusion of the first opening compensation amount and the second opening compensation amount through an intelligent arbitration strategy to generate the final feedforward opening compensation amount. In this embodiment, according to the , , , The value is used to compensate for the first opening degree through an intelligent arbitration strategy. and the second opening compensation amount Make decisions or merge data to generate the final feedforward opening compensation amount. .

[0069] The intelligent arbitration strategy is as follows: To construct an index reflecting the drastic changes in the unit's operating conditions, i.e., the drasticness of operating condition changes. To ensure the change in water head and load change To ensure that both factors contribute fairly to the assessment of the severity of changes in operating conditions, they must be normalized to eliminate the influence of dimensions.

[0070] This invention employs a normalization method based on the percentage of operating range, followed by weighted fusion and function transformation to obtain the drasticness of operating condition changes. Specifically, it includes the following steps: a: Define the normalization baseline value Based on the actual operating range of the unit, the typical range of load variations is defined. Typical range of water head variation .

[0071] b: Calculate the normalized change Normalized load change:

[0072] Normalized change in head:

[0073] After normalization and They all become dimensionless scalars, with values ​​fluctuating roughly in the range of [-1, 1], thus possessing additivity and comparability.

[0074] c: Calculate the index of the severity of normalized chemical condition changes Based on the normalized load change and normalized head change obtained in step b, the severity of the operating condition change is... Designed as follows:

[0075] d: Design the weighting function and perform arbitration. When designing a weighting function, its continuity and smoothness, as well as its monotonicity, boundary characteristics, and adjustability, should be considered. In this invention, [the following was selected]. As the weight function for the data-driven channel, the weight of the knowledge-driven feedforward channel is: , where 0 ≤ ≤ 1.

[0076] Will As a data-driven channel weighting function, it can achieve the following functions: 1) Avoid arbitration output A sudden change occurs, causing unnecessary jitter in the system.

[0077] 2) With increasing intensity The increase in weights of the neural network path The weight of the knowledge-driven feedforward channel continues to decrease. The core concept is to continuously increase the reliance on robust expert experience as the operating conditions become more extreme.

[0078] 3) To adapt to the characteristics of different power plants and generating units, the function has a specific parameter. This allows for flexible adjustment of the aggressiveness or conservatism of the control strategy. (Larger) A smaller value makes the system more inclined towards a knowledge-driven feedforward path, suitable for scenarios with high stability requirements. Values ​​that make the system more data-driven are suitable for scenarios with high response speed requirements. Additionally, parameters... It can be determined through experiments or simulation optimization.

[0079] 4) Satisfies the ideal boundary conditions, that is, when → 0 (nearly no disturbance), → 1 (Fully Trust Data-Driven Channel); when → ∞ (maximum perturbation), → 0 (Fully knowledge-driven feedforward channel).

[0080] By designing the weighting function and combining data-driven and knowledge-driven feedforward compensation, the final feedforward opening compensation amount is obtained:

[0081] Therefore, the present invention has the following advantages over the prior art: Improve system stability and response speed: Through intelligent arbitration strategy, the system can quickly adjust the control strategy when the operating conditions change drastically.

[0082] Reduce system jitter and overshoot: Avoid jumps in the arbitration output, improving the smoothness and reliability of the system.

[0083] Enhance system adaptability: It can flexibly switch between data-driven and knowledge-driven strategies to adapt to different working conditions.

[0084] Adjustable parameters are provided: via parameters The introduction of this technology enables the system to adapt to the characteristics of different power plants and generating units.

[0085] In this embodiment, the final feedforward opening compensation amount The opening command output by the PID controller is superimposed to generate the total opening control command for the turbine guide vanes.

[0086] S4. The final feedforward opening compensation amount is superimposed with the opening command output by the PID controller to form the total opening control command for the turbine guide vanes. This command is then used to drive the guide vane servo system to adjust the turbine inflow, achieving rapid and stable load tracking. The total opening control command for the turbine guide vanes is the final feedforward opening compensation amount. This is superimposed on the opening command output by the PID controller:

[0087] in, This is the opening command output by the PID controller. This completes the hybrid intelligent feedforward compensation control.

[0088] As a preferred embodiment, the intelligent arbitration strategy employs a continuously monotonically decreasing weight function, which causes the weight of the data-driven feedforward channel to decrease as the intensity of changes in operating conditions increases, while the weight of the knowledge-driven feedforward channel increases accordingly. When the intensity of changes in operating conditions approaches zero, the data-driven channel is fully trusted; when the intensity of changes in operating conditions is extremely high, the knowledge-driven channel is fully relied upon.

[0089] As a preferred embodiment, the weighting function includes an adjustable parameter to adjust the aggressiveness or conservatism of the control strategy of the system under different power plants or units; this parameter can be determined through offline simulation or online experiment optimization.

[0090] As a preferred embodiment, a first-order inertial element and a dead-zone element are set in the feedforward compensation control loop; the first-order inertial element is used to filter out high-frequency noise and equivalently delay the system sampling time, while the dead-zone element is used to suppress frequent feedforward actions caused by small disturbances and improve system stability.

[0091] As a preferred option, the method described herein is applicable to the rapid load response control of hydro-generator units under power grid peak shaving and frequency regulation conditions, and is particularly suitable for operating scenarios with large head fluctuations and frequent load changes.

[0092] As a preferred option, the hybrid intelligent feedforward compensation control module is integrated into the existing turbine governor control system in the form of a software functional module, without changing the main hardware structure of the governor system, and has good engineering compatibility and portability.

[0093] As a preferred option, the method is applicable to turbine speed control in both opening mode and power mode, and can be adapted to different control objectives by switching the type of input variable.

[0094] A turbine speed regulation optimization control system based on hybrid intelligent feedforward compensation, employing the aforementioned control method, comprises: The data acquisition module is used to monitor the operating status of the turbine unit in real time. When a change in head or load is detected, it collects the change in head, the change in load, the current head value, the current load value, and the current guide vane opening value, and triggers feedforward compensation control. The hybrid feedforward compensation module includes a parallel data-driven feedforward channel and a knowledge-driven feedforward channel. The data-driven feedforward channel contains a neural network model trained on historical operating data, which is used to output the first opening compensation amount. The knowledge-driven feedforward channel contains a fuzzy logic controller built based on the physical characteristics of the turbine and expert experience, which is used to output the second opening compensation amount. The intelligent arbitration module is used to adaptively weight and fuse the first opening compensation amount and the second opening compensation amount according to the index of the severity of changes in working conditions, and generate the final feedforward opening compensation amount. The command output module is used to superimpose the final feedforward opening compensation amount with the opening command output by the PID controller to form the total opening control command of the turbine guide vanes, and drive the guide vane relay to adjust the turbine inflow through the electro-hydraulic servo system.

[0095] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A method for optimizing the speed control of a hydro turbine based on hybrid intelligent feedforward compensation, characterized in that, Includes the following steps: S1: Real-time monitoring of turbine unit operation status. When a change in head or load condition is detected, the change in head, the change in load condition, the current head value, the current load value, and the current guide vane opening value are collected to trigger feedforward compensation control. S2: The head change, load condition change, current head value, current load value, and current guide vane opening value are simultaneously input to the data-driven feedforward channel and the knowledge-driven feedforward channel, which operate in parallel. The data-driven feedforward channel sends the input variables to a neural network model trained with historical operating data and outputs the first opening compensation value. The knowledge-driven feedforward channel sends the same input variables to a fuzzy logic controller built based on the physical characteristics of the turbine and expert experience and outputs the second opening compensation value. S3: Calculate the working condition change intensity index based on the input variables to characterize the degree of change in working conditions, and based on this index, perform adaptive weighted fusion of the first opening compensation amount and the second opening compensation amount through an intelligent arbitration strategy to generate the final feedforward opening compensation amount. S4: The final feedforward opening compensation amount is superimposed with the opening command output by the PID controller to form the total opening control command of the turbine guide vanes. The guide vane relay is then driven by the electro-hydraulic servo system to adjust the turbine inflow and achieve fast and stable load tracking.

2. The turbine speed regulation optimization control method based on hybrid intelligent feedforward compensation according to claim 1, characterized in that, The construction of a neural network model includes: Sample data containing steady-state switching processes were selected from the historical operation database of the turbine units to construct a training dataset containing load disturbance, head disturbance and corresponding guide vane opening changes; Perform data cleaning, outlier removal, and missing value handling on the training dataset to ensure data quality; The cleaned data is normalized to eliminate dimensional differences; A neural network is constructed using a multilayer perceptron structure, with load disturbance, head disturbance, current load, current head, and current opening as input nodes, and guide vane opening change as output node. Two to three hidden layers are set, and the ReLU activation function is used. The network is trained using the backpropagation algorithm combined with the Adam optimizer, and L2 regularization is introduced to prevent overfitting. Evaluate model performance on independent test sets and optimize hyperparameters or adjust network structure based on validation results.

3. The turbine speed regulation optimization control method based on hybrid intelligent feedforward compensation according to claim 1, characterized in that, The design of the fuzzy logic controller includes: Define the universe of discourse for load condition variation and head variation, and divide it into multiple fuzzy subsets. The linguistic variables include negative large, negative small, zero, positive small, and positive large. Define the output universe of discourse and the corresponding fuzzy subset of the second opening compensation quantity, with the linguistic variables being consistent with the input variables; Based on the dynamic balance relationship of "head-opening-output" of the water turbine and the safety operation constraints, a rule base containing multiple fuzzy rules is established. The rule form is "If the load change is A and the head change is B, then the opening compensation is C". The Mamdani inference method is used for fuzzy inference, and the fuzzy output is defuzzified into a precise second opening compensation amount by the centroid method. The performance of the fuzzy controller is verified by simulation or actual operating data, and the membership function or rule base is optimized based on response speed, overshoot, and steady-state error indicators.

4. The turbine speed regulation optimization control method based on hybrid intelligent feedforward compensation according to claim 1, characterized in that, The calculation of the intensity of operating condition changes includes: Based on the actual operating range of the unit, the typical variation range of load conditions and head is defined as the normalization benchmark. Divide the head change and load change by their corresponding typical ranges to obtain the dimensionless normalized head change and load change. The summation of the absolute values ​​of the normalized head change and the normalized load change yields the index of the severity of the operating condition change.

5. The turbine speed regulation optimization control method based on hybrid intelligent feedforward compensation according to claim 1 or 4, characterized in that, The intelligent arbitration strategy employs a continuously monotonically decreasing weight function, which causes the weight of the data-driven feedforward channel to decrease as the intensity of changes in operating conditions increases, while the weight of the knowledge-driven feedforward channel increases accordingly. When the intensity of changes in operating conditions approaches zero, the data-driven channel is fully trusted; when the intensity of changes in operating conditions is extremely high, the knowledge-driven channel is fully relied upon.

6. The turbine speed regulation optimization control method based on hybrid intelligent feedforward compensation according to claim 5, characterized in that, The weighting function includes an adjustable parameter to adjust the aggressiveness or conservatism of the control strategy for different power plants or units; this parameter can be determined through offline simulation or online experiment optimization.

7. The turbine speed regulation optimization control method based on hybrid intelligent feedforward compensation according to claim 1, characterized in that, A first-order inertial element and a dead-zone element are set in the feedforward compensation control loop. The first-order inertial element is used to filter out high-frequency noise and equivalently delay the system sampling time, while the dead-zone element is used to suppress frequent feedforward actions caused by small disturbances and improve system stability.

8. The turbine speed regulation optimization control method based on hybrid intelligent feedforward compensation according to claim 1, characterized in that, The method is applicable to the rapid load response control of hydro-generator units under power grid peak shaving and frequency regulation conditions, and is especially suitable for operating scenarios with large head fluctuations and frequent load changes.

9. The turbine speed regulation optimization control method based on hybrid intelligent feedforward compensation according to claim 1, characterized in that, The hybrid intelligent feedforward compensation control module is integrated into the existing turbine governor control system as a software functional module, without requiring changes to the main hardware structure of the governor system, and has good engineering compatibility and portability.

10. The turbine speed regulation optimization control method based on hybrid intelligent feedforward compensation according to claim 1, characterized in that, The method is applicable to turbine speed control in both opening and power modes, and can be adapted to different control objectives by switching the type of input variable.

11. The turbine speed regulation optimization control method based on hybrid intelligent feedforward compensation according to claim 2, characterized in that, The process of constructing the dataset is as follows: Historical operating data of the hydro turbine unit Selected from and with , ; i represents the i-th time, where the baseline load condition, head, and guide vane opening begin to change, and a dataset is constructed. ; in, These represent the load condition, head, and guide vane opening of the unit during its current stable operation, respectively. These represent the unit's state points. During stable operation, load and head disturbances cause the unit to shift from the current state point to the next stable state point. During the transfer process, the changes in load conditions, the changes in head, and the resulting changes in guide vane opening are considered.

12. The turbine speed regulation optimization control method based on hybrid intelligent feedforward compensation according to claim 2, characterized in that, The normalization process uses the Min-Max normalization method to independently normalize each feature to... The interval is calculated using the following formula: ； Where X represents the original data. and These are the minimum and maximum values ​​of the feature in the entire dataset, respectively.

13. The turbine speed regulation optimization control method based on hybrid intelligent feedforward compensation according to claim 2, characterized in that, Neural networks use a set of variables As input, the change in guide vane opening As output, the network input layer contains 5 nodes and the output layer contains 1 node; the hidden layers are set to 2 to 3 layers, and the number of neurons in each layer is selected from 64, 128, and 256; the ReLU function is used as the activation function for the hidden layers.

14. The turbine speed regulation optimization control method based on hybrid intelligent feedforward compensation according to claim 3, characterized in that, Load variation The domain of discourse is defined as Change in water head The domain of discourse is defined as Both are divided into 5 fuzzy subsets, with linguistic variables being negative large NB, negative small NS, zero ZO, positive small PS, and positive large PB, respectively; the membership function of the fuzzy subsets adopts a trigonometric function, and its parameters are determined based on actual operating data.

15. The turbine speed regulation optimization control method based on hybrid intelligent feedforward compensation according to claim 3, characterized in that, Second opening compensation amount The domain of discourse is defined as It is divided into 5 fuzzy subsets, and the linguistic variables include negative large NB, negative small NS, zero ZO, positive small PS, and positive large PB; its membership function adopts trigonometric function or Gaussian function, and the specific parameters are determined according to the actual running data.

16. The turbine speed regulation optimization control method based on hybrid intelligent feedforward compensation according to claim 3, characterized in that, Defuzzification uses the centroid method, and its calculation formula is as follows: ； Where y is the value on the output universe of discourse. This represents the total membership degree of the corresponding output value.

17. A turbine speed regulation optimization control system based on hybrid intelligent feedforward compensation, characterized in that, The system employs the turbine speed regulation optimization control method based on hybrid intelligent feedforward compensation according to any one of claims 1 to 16, comprising: The data acquisition module is used to monitor the operating status of the turbine unit in real time. When a change in head or load is detected, it collects the change in head, the change in load, the current head value, the current load value, and the current guide vane opening value, and triggers feedforward compensation control. The hybrid feedforward compensation module includes a parallel data-driven feedforward channel and a knowledge-driven feedforward channel. The data-driven feedforward channel contains a neural network model trained on historical operating data, which is used to output the first opening compensation amount. The knowledge-driven feedforward channel contains a fuzzy logic controller built based on the physical characteristics of the turbine and expert experience, which is used to output the second opening compensation amount. The intelligent arbitration module is used to adaptively weight and fuse the first opening compensation amount and the second opening compensation amount according to the index of the severity of changes in working conditions, and generate the final feedforward opening compensation amount. The command output module is used to superimpose the final feedforward opening compensation amount with the opening command output by the PID controller to form the total opening control command of the turbine guide vanes, and drive the guide vane relay to adjust the turbine inflow through the electro-hydraulic servo system.

Images