Power plant steam turbine valve sequence detection and adjustment system

Abstract

The invention relates to the technical field of steam turbine operation control and vibration suppression, in particular to a power plant steam turbine valve sequence detecting and adjusting system. Comprising a multi-modal data acquisition unit which is used for acquiring the valve position opening degree, the valve rod movement rate and the unit real-time load of a steam turbine speed regulation steam valve in real time, an original vibration vector signal of a rotor vibration sensor, and temperature and pressure parameters of a key position; the vibration feedforward prediction unit is used for solving a predicted vibration vector at a next valve switching point; the valve sequence time sequence optimizing unit is used for optimizing an optimal time sequence shaping parameter; the control instruction generation unit is used for compiling the optimal time sequence shaping parameter into a control instruction for an actuator of the steam turbine speed regulation system and issuing and executing the control instruction; and the model self-adaptive correction unit is used for acquiring the actual vibration vector measured by the vibration sensor and calculating the prediction deviation between the actual vibration vector and the predicted vibration vector. According to the system, the logic of vibration control is changed, and the stability and safety of unit operation are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of steam turbine operation control and vibration suppression, in particular to a power plant steam turbine valve sequence detection and adjustment system. BACKGROUND

[0002] In the operation of large power plants, the stability and safety of the steam turbine unit are of great importance. During the transient process of speed regulating valve switching, the rotor vibration is easily triggered by steam excitation, which threatens the safety of the equipment. At present, the control method for such vibration mainly relies on the traditional passive response mechanism. This mechanism monitors the vibration value, and only when the vibration amplitude exceeds the preset safety threshold does it trigger the corresponding protection action. Although this method can provide post-protection, it is essentially a lagging and passive coping strategy. The core defect of this passive strategy is that it cannot intervene at the source of vibration, but only reacts after the vibration has threatened the stability of the unit. This not only affects the smooth operation of the unit during the transient process, but also may shorten the service life of key components due to frequent over-limit vibration. Therefore, how to change from passive response based on vibration threshold to a new control paradigm that can predict vibration trends and actively intervene in advance, actively suppress excitation by optimizing valve action before vibration occurs, has become a technical problem to be solved to ensure the long-term safe and stable operation of the steam turbine unit.

[0003] The above information disclosed in the above background section is only used to strengthen the understanding of the background of the present disclosure, and therefore it can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0004] To solve the above technical problems, the present application discloses a power plant steam turbine valve sequence detection and adjustment system. Specifically, the technical solution of the present application comprises: A multi-modal data acquisition unit is used to acquire in real time the valve position opening, valve rod movement rate, unit real-time load, original vibration vector signal of the rotor vibration sensor, and temperature and pressure parameters at key positions of the steam turbine speed regulating valve, and to synchronize the timestamp processing of the acquired multi-dimensional data to construct a multi-modal data set; A vibration feedforward prediction unit is used to construct and run a vibration prediction model based on the multi-modal data set to calculate the predicted vibration vector at the next valve switching point; A valve sequence timing optimization unit is used to take the amplitude of the predicted vibration vector as the objective function to perform iterative search within the dynamic reshaping window to optimize the optimal timing shaping parameter; A control instruction generation unit is used to compile the optimal timing shaping parameter into a control instruction for the steam turbine speed regulating system actuator and issue it for execution. The model adaptive correction unit is used to acquire the actual vibration vector measured by the vibration sensor after the control command is executed, calculate the prediction deviation between the actual vibration vector and the predicted vibration vector, and then correct the internal parameters of the vibration prediction model online based on the prediction deviation.

[0005] Furthermore, the original vibration vector signal includes vibration amplitude and vibration phase.

[0006] Furthermore, the dynamic shaping window is composed of a multi-dimensional parameter space consisting of the opening rate of the speed regulating steam valve, the overlap time of the opening of adjacent steam valves, and the phase difference of the opening of multiple valve combinations.

[0007] Furthermore, the boundary of the dynamic shaping window is set based on the physical response limit of the valve and the preset turbine operation safety regulations.

[0008] Furthermore, the optimal timing shaping parameters are a combination of parameters—the opening rate, overlap time, and phase difference—that minimize the amplitude of the predicted vibration vector within the dynamic shaping window.

[0009] Furthermore, the prediction deviation is a prediction deviation vector, which is the value obtained by subtracting the predicted vibration vector output by the vibration prediction model under the same input parameters from the actual vibration vector collected by the sensor after the control command is executed.

[0010] Furthermore, the model adaptive correction unit is used to construct a loss function based on the prediction bias, and to iteratively update the internal parameters of the vibration prediction model using a gradient descent algorithm based on the gradient of the loss function with respect to the current model parameters.

[0011] Furthermore, the gradient descent algorithm employs a preset learning rate to control the magnitude of the internal parameter updates in each iteration.

[0012] Compared with the prior art, the present invention has the following beneficial effects: 1. This system constructs a complete technical loop from data perception, trend prediction, optimization decision-making to closed-loop correction through multimodal data acquisition, feedforward vibration prediction, and model adaptive correction. This system innovates the traditional passive response mechanism based on vibration thresholds into an active intervention mechanism, capable of anticipating and actively suppressing potential steam flow excitation at the source of valve switching. This fundamentally changes the logic of vibration control, significantly improving the stability and safety of unit operation.

[0013] 2. This system incorporates the original vibration vector signal, which includes vibration amplitude and phase, into a multimodal dataset, enabling the prediction model to not only estimate vibration intensity but also accurately predict its direction. This vectorized prediction allows the valve sequence optimization unit to implement more refined control strategies. For example, by fine-tuning the phase difference of valve opening, oppositely oriented excitation forces can cancel each other out, thereby achieving a more efficient and precise vibration control effect than simply suppressing amplitude.

[0014] 3. This system transforms the complex valve sequence control problem into a mathematical optimization problem with clear boundaries by constructing a dynamic shaping window constrained by the valve's physical response limits and operational safety procedures. This method effectively integrates expert experience and physical constraints into the intelligent optimization algorithm, ensuring that the optimization process always proceeds within a safe and feasible range. It eliminates the generation of control commands that may endanger unit safety from the source, achieving a balance between efficient optimization and absolute safety.

[0015] 4. The adaptive correction unit designed in this system accurately calculates the deviation between the predicted vibration vector and the actual vibration vector, and uses this as a basis to continuously fine-tune the model parameters online using the gradient descent algorithm. This design constructs a stable online learning closed loop, enabling the system to autonomously track the gradual changes in unit characteristics caused by factors such as wear and scaling, ensuring that the prediction model maintains high accuracy throughout the equipment's entire life cycle, greatly enhancing the system's robustness and long-term application effectiveness. Attached Figure Description

[0016] The present invention will be further explained below with reference to the accompanying drawings and embodiments: Figure 1 This is a system structure diagram of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0018] Example 1:

[0019] Please see Figure 1 A power plant turbine valve sequence detection and adjustment system, comprising: The multimodal data acquisition unit is used to acquire in real time the valve position opening, valve stem movement rate, unit real-time load, original vibration vector signal of rotor vibration sensor, and temperature and pressure parameters at key locations of the turbine speed regulating valve, and to perform time stamp synchronization processing on the acquired data of each dimension to construct a multimodal dataset. The vibration feedforward prediction unit is used to build and run a vibration prediction model based on a multimodal dataset to calculate the predicted vibration vector at the next valve switching point. The valve sequence timing optimization unit is used to perform iterative search within the dynamic shaping window, taking the amplitude of the predicted vibration vector as the objective function, in order to find the optimal timing shaping parameters. The control command generation unit is used to compile the optimal timing shaping parameters into control commands for the turbine speed control system actuators and issue them for execution. The model adaptive correction unit is used to obtain the actual vibration vector measured by the vibration sensor after the control command is executed, calculate the prediction deviation between the actual vibration vector and the predicted vibration vector, and then correct the internal parameters of the vibration prediction model online based on the prediction deviation. This embodiment provides a power plant turbine valve sequence detection and adjustment system; it innovates the traditional passive response mechanism based on vibration threshold into an active, feedforward intervention mechanism based on vibration trend prediction, thereby actively suppressing potential steam flow excitation at the source of turbine valve switching, ensuring the stability and safety of unit operation; through the coordinated action of internal units, the system achieves a complete technical closed loop from data acquisition, vibration prediction, strategy optimization, command execution to model adaptive correction; The multimodal data acquisition unit provides a comprehensive and high-quality data foundation for subsequent accurate prediction and control. In this embodiment, the unit uses multiple sensors deployed at key measuring points of the turbine, such as valve position sensors, rate sensors, load sensors, eddy current sensors, and temperature and pressure transmitters, to collect in real time the valve opening degree of the turbine speed regulating valve, the valve stem movement rate, the real-time load of the unit, the raw vibration vector signal of the rotor vibration sensor, and the temperature and pressure parameters at key locations. The unit performs high-precision timestamp synchronization processing on the collected data streams from each dimension to ensure strict alignment of all data in the time dimension, thereby constructing a multimodal dataset that can truly reflect the overall state of the unit at a specific instant. The construction of this dataset is the fundamental prerequisite for overcoming the one-sidedness of traditional single-parameter monitoring information and achieving high-fidelity prediction. The vibration feedforward prediction unit quantifies and predicts the vibrations that may be caused by future valve switching behavior based on the current unit status. The core of this unit is a deep learning model, such as a recurrent neural network or a long short-term memory network. This vibration prediction model has been trained offline using a large amount of historical multimodal datasets during the design phase. In this embodiment, the model is specifically a sequence model containing two layers of LSTM units, with 128 neurons in each layer. The model's input is a time-series window containing multimodal feature vectors sampled every 100 milliseconds over the past 5 seconds. Each time point's multimodal feature vector is composed of physical quantities such as valve opening degree, valve stem movement rate, unit real-time load, original vibration vector signal, temperature, and pressure, all normalized and then concatenated along the feature dimension. The output is the predicted vibration vector at the next valve switching point. The vector consists of two floating-point values: predicted amplitude and predicted phase. When the system is running, this unit receives the real-time dataset provided by the multimodal data acquisition unit as input, and calculates the predicted vibration vector of the unit rotor at the next valve switching point through the nonlinear mapping relationship inside the model. This predicted vector provides an accurate and quantifiable target for subsequent optimization. The valve sequence optimization unit actively seeks a set of valve action sequences that can optimally suppress predicted vibrations; this unit calculates the amplitude of the predicted vibration vector output by the vibration feedforward prediction unit, i.e. As the objective function of the optimization algorithm, it aims to find the minimum value of the objective function; the unit defines a dynamic shaping window, which is a multi-dimensional feasible region composed of multiple valve action parameters, such as opening rate, overlap time, and phase difference; intelligent search algorithms, such as particle swarm optimization or genetic algorithm, are used to perform iterative search within the constraints of this window to find the optimal combination of time-series shaping parameters that minimizes the predicted vibration amplitude; Taking particle swarm optimization as an example, each particle represents a set of time-series shaping parameters, and its fitness value is determined by the objective function. The calculation shows that, among which This set of parameters; the algorithm searches for the optimal solution by iteratively updating the particle's velocity and position. Key parameters include a population size of 50 and inertia weights. The learning factor decreases linearly from 0.9 to 0.4. and All values ​​were set to 2.0, which was determined during the system debugging phase. This design transforms the vibration control problem into a constrained mathematical optimization problem, achieving a technological leap from passive acceptance to active shaping. The control command generation unit transforms the abstract parameters obtained through optimization into specific actions that can be executed in the physical world. As the interface between the decision-making layer and the execution layer, this unit receives the optimal timing shaping parameters output by the valve sequence timing optimization unit and compiles them into a precise sequence of control commands for specific turbine speed control system actuators, such as hydraulic servo valves or electric actuators, according to the communication protocol and instruction set of the turbine digital electro-hydraulic control system. These commands are then sent to the control system and executed, thereby ensuring that the actual valve actions are highly consistent with the optimization results. The model adaptive correction unit constructs a closed-loop feedback loop to ensure that the prediction model can continuously track the slow changes in the unit's state caused by factors such as wear and scaling, maintaining the system's high performance throughout its entire lifecycle. When a control command is executed, this unit obtains the actual vibration vector caused by the valve action through a vibration sensor and calculates the prediction deviation between the actual vibration vector and the previously predicted vibration vector output by the model. This deviation directly reflects the model's current performance. Based on this deviation, the unit uses online learning algorithms, such as gradient descent, to fine-tune the internal parameters of the vibration prediction model. This continuous parameter iterative optimization process enables the entire system to have adaptive learning capabilities, greatly enhancing its robustness and long-term application effectiveness. Compared to the passive strategy of triggering protection actions only after vibration exceeds the limit in existing technologies, this system can predict the trend of vibration before it occurs and actively suppress it by optimizing the valve action sequence, thereby significantly reducing the vibration level of the unit during transient processes, improving operational stability, and extending equipment life.

[0020] Example 2: The original vibration vector signal includes vibration amplitude and vibration phase; Based on Example 1, this embodiment provides a complete description of the rotor vibration state using the original vibration vector signal. Specifically, the signal includes vibration amplitude and vibration phase. The vibration amplitude is the maximum offset of the rotor's vibration trajectory in space, which is an indicator of the intensity of the vibration. The vibration phase is the angular position of the rotor relative to a fixed marker point on the rotor, such as a key phase marker, when the rotor reaches its maximum offset, which describes the directionality of the vibration. By incorporating the vibration phase into the multimodal dataset, the vibration prediction model can predict not only the vibration intensity but also the vibration direction. This vectorized prediction allows for more refined control in the subsequent valve sequence optimization unit. If it is predicted that the opening of two consecutive valves will generate excitation forces in opposite directions, the system can fine-tune their opening phase difference to make the two excitation forces cancel each other out in time, thereby achieving more efficient vibration control than simply suppressing the amplitude, and thus greatly improving the accuracy and effectiveness of vibration suppression.

[0021] Example 3: The dynamic shaping window consists of a multi-dimensional parameter space composed of the opening rate of the speed regulating steam valve, the overlap time of the opening of adjacent steam valves, and the phase difference of the opening of multiple valves in combination. The boundaries of the dynamic shaping window are set based on the physical response limits of the valve and the preset turbine operating safety regulations; The optimal timing shaping parameters are the combination of on-state rate, overlap time, and phase difference that minimize the amplitude of the predicted vibration vector within the dynamic shaping window. Based on Example 1, this embodiment further elaborates on the internal working mechanism of the valve sequence timing optimization unit, whose synergistic effect ensures the efficiency, safety and accuracy of the optimization process; The dynamic shaping window defines a clear search range for the optimization algorithm. The window is defined as a multi-dimensional parameter space consisting of the opening rate of the speed regulating valve, the overlap time of the opening of adjacent valves, and the phase difference of the opening of multiple valves. The boundary of the dynamic shaping window ensures that the output of the optimization process is always within the range of physical feasibility and operational safety. This boundary is set based on two core criteria: the physical response limit of the valve and the preset turbine operation safety regulations. For example, according to the equipment manual provided by the turbine manufacturer, the physical maximum opening rate of a specific model of speed-regulating steam valve is 50% / s, and the operating safety regulations require that the overlap time between the opening of adjacent steam valves must not exceed 80 milliseconds. These specific, quantifiable values ​​are directly used as the constraint boundaries of the corresponding dimensions of the dynamic shaping window, thereby integrating expert experience and physical constraints into the optimization algorithm and avoiding the generation of control commands that are unexecutable or endanger safety. The optimal timing shaping parameter provides an explicit solution to the optimization problem, which is defined as: within the above bounded dynamic shaping window, the parameter that makes the objective function... The optimal framework is a combination of parameters, including a specific opening rate, overlap time, and phase difference, that achieves the minimum value. It can systematically search for the optimal solution while ensuring safety, thus achieving optimal and adaptive suppression of time-varying vibration characteristics under different operating conditions.

[0022] Example 4: The prediction deviation is the prediction deviation vector, which is the value obtained by subtracting the predicted vibration vector output by the vibration prediction model under the same input parameters from the actual vibration vector collected by the sensor after the control command is executed. The model adaptive correction unit is used to construct a loss function based on the prediction bias, and to iteratively update the internal parameters of the vibration prediction model using the gradient descent algorithm based on the gradient of the loss function with respect to the current model parameters. The gradient descent algorithm uses a preset learning rate to control the magnitude of the internal parameter updates in each iteration; Based on Example 1, this embodiment further explains the internal algorithm logic of the model adaptive correction unit. Their synergistic effect constitutes a stable and efficient online learning closed loop. The underlying logic is that the prediction deviation is precisely defined as a prediction deviation vector to simultaneously quantify the magnitude and direction of the prediction error; this prediction deviation vector is calculated by taking the actual vibration vector collected by the sensor after the control command is executed. Subtract the predicted vibration vector output by the vibration prediction model under the same input parameters. The value obtained later, that is The vectorized error definition provides richer gradient information for subsequent model parameter adjustments. Based on the aforementioned bias vector, the specific correction mechanism of the model adaptive correction unit becomes clear; this unit constructs a loss function based on the prediction bias vector. ,For example Based on this loss function Regarding the current model parameters gradient The gradient descent algorithm is used to analyze the internal parameters of the vibration prediction model. Iterative updates are performed; to ensure convergence, this gradient descent algorithm uses a fixed learning rate determined experimentally during the system debugging phase. For example, a typical value can be set to 0.001 to control internal parameters. The magnitude of each iteration update is given by the update formula. A key convergence control mechanism was introduced for this correction process to ensure that the model does not oscillate violently due to a single large error during continuous learning, thus guaranteeing the stability of the adaptive process; the online model correction closed loop achieves a combination of robustness and adaptability.

[0023] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A valve sequence detection and adjustment system for a power plant turbine, characterized in that, include: The multimodal data acquisition unit is used to acquire in real time the valve position opening, valve stem movement rate, unit real-time load, original vibration vector signal of rotor vibration sensor, and temperature and pressure parameters at key locations of the turbine speed regulating valve, and to perform time stamp synchronization processing on the acquired data of each dimension to construct a multimodal dataset. The vibration feedforward prediction unit is used to build and run a vibration prediction model based on a multimodal dataset to calculate the predicted vibration vector at the next valve switching point. The valve sequence timing optimization unit is used to perform iterative search within the dynamic shaping window, taking the amplitude of the predicted vibration vector as the objective function, in order to find the optimal timing shaping parameters. The control command generation unit is used to compile the optimal timing shaping parameters into control commands for the turbine speed control system actuators and issue them for execution. The model adaptive correction unit is used to acquire the actual vibration vector measured by the vibration sensor after the control command is executed, calculate the prediction deviation between the actual vibration vector and the predicted vibration vector, and then correct the internal parameters of the vibration prediction model online based on the prediction deviation.

2. The power plant turbine valve sequence detection and adjustment system according to claim 1, characterized in that, The original vibration vector signal includes vibration amplitude and vibration phase.

3. The power plant turbine valve sequence detection and adjustment system according to claim 1, characterized in that, The dynamic shaping window consists of a multi-dimensional parameter space composed of the opening rate of the speed regulating steam valve, the overlap time of the opening of adjacent steam valves, and the phase difference of the combined opening of multiple valves.

4. The power plant turbine valve sequence detection and adjustment system according to claim 3, characterized in that, The boundary of the dynamic shaping window is set based on the physical response limit of the valve and the preset turbine operation safety regulations.

5. A power plant turbine valve sequence detection and adjustment system according to claim 3, characterized in that, The optimal timing shaping parameters are the combination of opening rate, overlap time, and phase difference that minimize the amplitude of the predicted vibration vector within the dynamic shaping window.

6. The power plant turbine valve sequence detection and adjustment system according to claim 1, characterized in that, The prediction deviation is a prediction deviation vector, which is the value obtained by subtracting the predicted vibration vector output by the vibration prediction model under the same input parameters from the actual vibration vector collected by the sensor after the control command is executed.

7. The power plant turbine valve sequence detection and adjustment system according to claim 1, characterized in that, The model adaptive correction unit is used to construct a loss function based on the prediction bias, and to iteratively update the internal parameters of the vibration prediction model using a gradient descent algorithm based on the gradient of the loss function with respect to the current model parameters.

8. A power plant turbine valve sequence detection and adjustment system according to claim 7, characterized in that, The gradient descent algorithm uses a preset learning rate to control the magnitude of the internal parameter updates in each iteration.

Images