A steam turbine, a regenerative system and a control method thereof

Abstract

The present invention proposes a steam turbine, a regenerative system and a control method thereof, which includes collecting data by using multi-modal sensors such as steam pressure sensors, fusing through a spatio-temporal alignment algorithm, and constructing a digital twin model containing thermodynamic equations and equipment aging parameters. The LSTM network is used to predict model deviations and dynamically calibrate. The proximal policy optimization algorithm is adopted to solve the Pareto optimal solution of maximizing heat recovery efficiency and minimizing corrosion rate in the digital twin environment, and generate control instructions. When the corrosion prediction risk exceeds the threshold, the control strategy is dynamically adjusted to achieve health-efficiency collaborative control. At the same time, it also involves corresponding collaborative optimization devices and corrosion monitoring equipment, among which the multi-modal data fusion module is deployed on the edge computing node to achieve efficient data processing and analysis. The present invention effectively improves the thermal cycle efficiency of the steam turbine regenerative system, reduces the corrosion risk, prolongs the service life of the equipment, and has significant economic and technical benefits.

Description

Technical Field

[0001] The present invention relates to the field of steam turbines, and particularly to a steam turbine, a regenerative system and a control method thereof. Background Art

[0002] As a core part of thermal power generation, the efficiency and reliability of the steam turbine regenerative system directly affect the energy utilization rate and the equipment life. The prior art (CN215860369U) discloses a steam turbine regenerative system (such as Figure 2 and Figure 3 ), which drives a steam turbine to generate electricity by heating steam in a boiler, and the steam is transported to a condensation device to recover heat through a gas pushing fan and a one-way valve in a gas transmission pipeline. This system has the following defects:

[0003] 1. Insufficient efficiency optimization: relying on a fixed mechanical structure (such as the fan speed and the one-way valve opening degree), it cannot adapt to dynamic working conditions, resulting in low heat recovery efficiency;

[0004] 2. Lack of health management: it does not have corrosion monitoring, and only relies on a single sensor or manual inspection through conventional technical means, and cannot predict the pipeline corrosion risk in real time.

[0005] 3. Lack of efficiency optimization: it does not have efficiency optimization, and only relies on separate digital twin modeling through conventional technical means to achieve efficiency optimization, and cannot cooperate with corrosion monitoring.

[0006] The prior art does not involve technical means such as digital twin modeling, multimodal sensor fusion or reinforcement learning control, and cannot meet the intelligent requirements under complex working conditions. The present invention solves the above problems by introducing digital twin dynamic optimization and multimodal perception technology, and realizes the dual goals of improving the thermal cycle efficiency and extending the equipment life. Summary of the Invention

[0007] Embodiments of the present invention provide a steam turbine, a regenerative system and a control method thereof, aiming at the problems existing in the current technology, such as the technical means involving digital twin modeling, multimodal sensor fusion or reinforcement learning control cannot meet the intelligent requirements under complex working conditions.

[0008] The core technology of the present invention mainly realizes the coordinated control of improving the thermal cycle efficiency of the steam turbine regenerative system and predicting equipment corrosion through digital twin dynamic optimization and multimodal sensor data fusion.

[0009] In a first aspect, the present invention provides a control method for a steam turbine regenerative system, and the method includes the following steps:

[0010] S00. Construct a digital twin model: Integrate thermodynamic equations with equipment aging parameters to establish a virtual model of steam flow, condensation heat transfer, and corrosion evolution. The equipment aging parameters include the porosity decay coefficient of the sponge pad and the pipeline roughness growth model;

[0011] S10. Multimodal data acquisition: Real-time obtain steam pressure, condensate flow rate, pipeline inner wall images, acoustic signals, and water quality data through a synchronously deployed sensor array;

[0012] S20. Data fusion and dynamic calibration: Use a spatio-temporal alignment algorithm to align the timestamps of multi-source data, and predict model deviations and update the parameters of the digital twin model through an LSTM network prediction model;

[0013] S30. Reinforcement learning optimization: Run a multi-objective reinforcement learning agent in the digital twin environment to solve the Pareto optimal solution of maximizing heat recovery efficiency and minimizing corrosion rate, and generate control instructions;

[0014] S40. Health-efficiency collaborative control: When the corrosion prediction risk exceeds the threshold, dynamically adjust the control strategy, and optimize the digital twin model through a virtual-real interactive closed-loop feedback.

[0015] Furthermore, in step S10, the sensors used for multimodal data acquisition include:

[0016] A steam pressure sensor for real-time monitoring of the steam pressure in the gas transmission pipeline;

[0017] A condensate flowmeter for measuring the condensate flow rate of the liquid in the condensate output pipeline;

[0018] A fan speed encoder for detecting the rotation speed of the gas pushing fan;

[0019] A high-temperature resistant endoscope camera for obtaining images of the inner wall of the curved gas flow pipeline;

[0020] A broadband acoustic sensor for collecting acoustic signals in the 10 Hz - 200 kHz frequency band in the gas activity room;

[0021] A water quality sensor for detecting the Cl - concentration and pH value of the liquid in the condensate output pipeline.

[0022] Furthermore, in step S00, the thermodynamic equations of the digital twin model include the continuity equation, the momentum equation, and the energy equation, and the equipment aging parameters are obtained by fitting historical operation data.

[0023] Furthermore, in step S20, the data fusion steps include:

[0024] Reconstruct the time series of high-frequency acoustic signals using cubic spline interpolation;

[0025] Using the attention mechanism to weightedly fuse the proportion of corrosion area, the peak value of acoustic energy, and the Cl - concentration feature.

[0026] Furthermore, in step S30, the reinforcement learning agent adopts the proximal policy optimization algorithm, and the reward function R is:

[0027] R = 0.7×efficiency improvement rate - 0.3×corrosion rate + 0.1×(1 - control action frequency).

[0028] Furthermore, in step S10, the high-temperature resistant endoscope probe integrates a micro camera and a MEMS microphone, and can rotate 360° to scan the inner wall of the curved gas flow pipeline.

[0029] Furthermore, in step S40, the health-efficiency control adjusts the control strategy when any of the following conditions is triggered:

[0030] The corrosion rate ≥ 0.1 mm / year;

[0031] The virtual model predicts that the thermal resistance of a certain section of the pipeline increases by more than 15% of the benchmark value.

[0032] Furthermore, in step S10, the multi-modal data acquisition uses the IEEE1588v2 protocol to achieve the time stamp alignment of multi-source data, and the synchronization accuracy ≤ 10 μs.

[0033] In the second aspect, the present invention provides a steam turbine regenerative system, including a control end, and the control end is equipped with the above-mentioned steam turbine regenerative system control method.

[0034] In the third aspect, the present invention provides a steam turbine, including the above-mentioned steam turbine regenerative system.

[0035] The main contributions and innovations of the present invention are as follows:

[0036] 1. Significant efficiency optimization

[0037] Dynamic adaptive regulation: The prior art relies on a fixed mechanical structure and is difficult to adapt to changes in working conditions. The present invention uses multi-modal sensors to collect data such as steam pressure and condensate flow in real time, and through the digital twin model and the reinforcement learning algorithm, dynamically adjusts the rotation speed of the gas pushing fan and the frequency of the condensate pump, and the heat recovery efficiency is significantly improved. For example, when the steam pressure suddenly changes, the system parameters can be quickly optimized to maintain efficient operation.

[0038] Precise thermodynamic cycle control: The digital twin model that fuses the thermodynamic equation and the equipment aging parameters accurately simulates the steam flow and condensation heat transfer processes, optimizes the gas residence time and heat exchange efficiency, and the energy utilization rate is greatly improved compared with the prior art.

[0039] 2. Real-time and precise health management

[0040] Multi-modal fusion prediction: The existing corrosion monitoring methods are single and lagging. The present invention uses a multi-modal sensor array composed of an endoscope camera, an acoustic sensor, and a water quality sensor to collect pipeline inner wall images, acoustic signals, and water quality data in real time. Through advanced image processing and machine learning algorithms, the corrosion risk is accurately predicted in advance, and the corrosion false alarm rate is significantly reduced.

[0041] Active maintenance decision-making: Once it is predicted that the corrosion rate exceeds the safety threshold, the system automatically triggers a health-efficiency collaborative optimizer to actively adjust the control strategy, such as reducing the flow rate of high-risk pipelines, avoiding further damage to equipment, and reducing maintenance costs.

[0042] 3. Strong system coordination

[0043] Cross-level coordination mechanism: In the existing technology, efficiency optimization and health management are independent of each other. The present invention realizes deep coordination between digital twin optimization and corrosion prediction modules, shares sensor data, and extends the equipment life while ensuring the heat recovery efficiency. For example, when the corrosion risk of a certain section of the curved pipeline is predicted to increase, the system reduces the steam flow rate in this area while increasing the load compensation in other areas to compensate for the efficiency loss, achieving a dynamic balance of the dual objectives of "efficiency-life".

[0044] 4. High prediction and control accuracy

[0045] Digital twin dynamic calibration: The LSTM network is used to predict the model deviation in real time and update the digital twin model parameters, so that the virtual model can more accurately reflect the physical system state. The prediction error of the digital twin model ≤ 2.5%, which is much lower than the traditional method, providing a reliable basis for optimal control.

[0046] Reinforcement learning multi-objective optimization: The proximal policy optimization algorithm is used to solve the Pareto optimal solution, comprehensively considering the heat recovery efficiency and corrosion rate, and the generated control instructions are more scientific and reasonable, improving the overall performance of the system.

[0047] Details of one or more embodiments of the present invention are set forth in the following drawings and description to make other features, objects, and advantages of the present invention more concise and understandable. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The illustrative embodiments and descriptions thereof are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0049] Figure 1 is a flowchart of a steam turbine regenerative system control method according to an embodiment of the present invention;

[0050] Figure 2 is the structural diagram of the prior art;

[0051] Figure 3 is the structural diagram of the condensation device of the prior art;

[0052] In the figure, 1 is a boiler; 2 is a steam turbine body; 3 is a generator; 4 is a gas transmission pipeline; 5 is a gas pushing fan; 6 is a check valve; 7 is a condensation device body; 71 is a condensation device end cover; 72 is a condensation protection outer cylinder; 73 is a gas activity room; 74 is a middle gas condensation pipeline; 75 is a bent gas circulation pipeline; 76 is a sponge pad; 8 is a condensate liquid transmission pipeline; 9 is a condensate liquid output pipeline; 10 is a gas circulation pipeline; 11 is a condensate device; 12 is a deaeration device; 13 is a clean water water tank; 14 is a water pump; 15 is a condensate water tank. Specific Embodiments

[0053] Here, the exemplary embodiments will be described in detail, and the examples are shown in the drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with one or more embodiments of this specification. On the contrary, they are merely examples of devices and methods consistent with some aspects of one or more embodiments of this specification as detailed in the appended claims.

[0054] It should be noted that: In other embodiments, the steps of the corresponding methods are not necessarily executed in the order shown and described in this specification. In some other embodiments, the steps included in the method may be more or less than those described in this specification. In addition, a single step described in this specification may be decomposed into multiple steps for description in other embodiments; and multiple steps described in this specification may also be combined into a single step for description in other embodiments.

[0055] As Figure 2-3 shown, the prior art (CN215860369U) discloses a steam turbine regenerative system, and its core structure includes: The boiler 1 heats to generate steam to drive the steam turbine body 2 and the generator 3, and the steam is transported to the condensation device body 7 through the gas pushing fan 5 and the check valve 6 in the gas transmission pipeline 4. The condensed liquid is recovered to the condensate water tank 15 through the condensate liquid transmission pipeline 8 and the condensate liquid output pipeline 9, and the remaining gas enters the condensate device 11, the deaeration device 12 and the water pump 14 through the gas circulation pipeline 10 for recycling. This system realizes the basic heat recovery function through mechanical structures such as check valve anti-backflow and sponge pad 76 assisting condensation. For the specific structure, see the content disclosed in CN215860369U, which will not be elaborated here.

[0056] The prior art relies on fixed mechanical structures such as a gas-pushing fan with a fixed rotational speed and a one-way valve with a preset opening degree, and cannot adapt to dynamic working conditions such as sudden changes in steam pressure and fluctuations in condensate flow rate. For example, when the load of the steam turbine changes, resulting in fluctuations in steam flow rate, the gas-pushing fan 5 in the gas transmission pipeline 4 cannot dynamically adjust its rotational speed, leading to a decrease in condensation efficiency. The measured heat recovery efficiency is only 65% - 70%. In addition, the curved gas flow pipeline 75 and the middle gas condensation pipeline 74 of the condensate device body 7 do not integrate a flow regulation mechanism, making it difficult to optimize the gas residence time and heat exchange efficiency. And there is also a lack of means such as water quality monitoring.

[0057] Based on this, the present invention is an improvement on the prior art, introducing digital twin dynamic optimization and multi-modal perception technologies to solve the problems existing in the prior art.

[0058] Embodiment 1

[0059] The present invention aims to propose a control method for the regenerative system of a steam turbine, introducing digital twin dynamic optimization and multi-modal perception technologies to solve the core problems of the prior art:

[0060] The digital twin model maps the physical system state in real time and predicts the deviation of thermodynamic parameters (such as the attenuation of the heat transfer coefficient);

[0061] The multi-modal sensor array (including an endoscope camera, an acoustic sensor, and a water quality sensor) synchronously collects efficiency and health data;

[0062] The reinforcement learning agent solves the Pareto optimal solution of efficiency - corrosion rate in the digital twin environment, generates collaborative control instructions, and realizes the dual goals of "efficiency improvement" and "life extension".

[0063] Specifically, the embodiment of the present invention provides a control method for the regenerative system of a steam turbine. Specifically, referring to Figure 1 , the method includes:

[0064] S00. Construct a digital twin model: Integrate thermodynamic equations and equipment aging parameters to establish a virtual model of steam flow, condensation heat transfer, and corrosion evolution. The equipment aging parameters include the attenuation coefficient of the porosity of the sponge pad and the pipeline roughness growth model;

[0065] In this embodiment, the digital twin model includes:

[0066] 1. Digital twin layer:

[0067] 1.1 Virtual model construction:

[0068] 1.1.1 Thermodynamic model:

[0069] Based on ANSYS Fluent, establish a three-dimensional unsteady flow and heat transfer model, including the following equations:

[0070] Continuity equation: ; The continuity equation is the manifestation of the law of conservation of mass in fluid mechanics. Here, ρ represents the density of the fluid, and u is the velocity vector of the fluid. is the divergence operator, represents the divergence of the product of density and velocity. This equation indicates that in a steady-flowing fluid system, the mass of the fluid flowing into a certain control volume per unit time is equal to the mass of the fluid flowing out of this control volume, that is, the mass of the fluid will not be generated or disappear out of thin air within the system.

[0071] Momentum equation: ; The momentum equation is the application of Newton's second law in fluid mechanics. The left side of the equation represents the convective change rate of the fluid momentum, where ρuu is a second-order tensor representing the momentum flux. The meanings of the terms on the right side of the equation are as follows:

[0072] : Pressure gradient force, p is the pressure of the fluid, and the negative sign indicates that the direction of the pressure gradient force is opposite to the direction of increasing pressure, that is, the fluid flows from the high-pressure area to the low-pressure area.

[0073] : Viscous force, μ is the dynamic viscosity of the fluid, and this term describes the stress generated inside the fluid due to viscous action.

[0074] ρg: Gravity, g is the gravitational acceleration vector.

[0075] This equation describes that the change in fluid momentum is the result of the combined action of the pressure gradient force, viscous force, and gravity.

[0076] Energy equation: ; The energy equation is the expression of the law of conservation of energy in fluid mechanics. The left side of the equation represents the convective transport rate of the fluid enthalpy, where h is the specific enthalpy of the fluid. The meanings of the terms on the right side of the equation are as follows:

[0077] : Heat conduction term, k is the thermal conductivity of the fluid, T is the temperature of the fluid, and this term describes the heat conduction caused by the temperature gradient.

[0078] S_h: Heat source term, representing other heat sources or heat sinks within the system, such as the heat generated by chemical reactions, etc.

[0079] This equation describes that the change in fluid energy is the result of the combined action of the convective transport of enthalpy, heat conduction, and other heat sources.

[0080] Embedded device aging parameter:

[0081] Porosity decay model of the sponge pad 76:

[0082] ε(t)=ε 0 ×exp(-0.01t) (t is the number of operating hours); this formula describes the decay of the porosity of the sponge pad over time. Here, ε(t) is the porosity of the sponge pad at time t, and ε 0 is the initial porosity, and t is the number of operating hours. The exponential function exp(-0.01t) indicates that the porosity decays exponentially over time, with a decay coefficient of 0.01.

[0083] Inner wall roughness growth model of the pipeline: k s (t)=k s0 +0.002t (mm). This formula describes the growth of the inner wall roughness of the pipeline over time. Here, k s (t) is the roughness of the inner wall of the pipeline at time t, and k s0 is the initial roughness, and t is the number of operating hours. The linear function 0.002t indicates that the roughness grows linearly over time, with a growth rate of 0.002 mm / hour.

[0084] 1.1.2 Corrosion prediction model:

[0085] Use an improved MaskR-CNN model (ResNet50 backbone network, Focal Loss for handling sample imbalance) to segment the corrosion area in the endoscope image; for example:

[0086] Improvements to MaskR-CNN:

[0087] Backbone network: ResNet50 (pre-trained on the COCO dataset);

[0088] Loss function: Focal Loss (α = 0.25, γ = 2) + Dice Loss;

[0089] Input image size: 640×480, corrosion area annotation accuracy ±0.1mm;

[0090] Extract the energy features in the 150 - 200 kHz frequency band from the acoustic signal through wavelet packet transform, and input them into the SVM classifier (RBF kernel function) to determine the corrosion type (pitting corrosion, uniform corrosion). For example:

[0091] SVM kernel function: RBF, C = 10, γ = 0.1;

[0092] Feature extraction: Energy ratio in the 150 - 200 kHz frequency band, through 5-layer wavelet packet decomposition.

[0093] 1.2 Dynamic calibration module:

[0094] LSTM network structure: Input layer (10 sensor features) → Bidirectional LSTM layer (128 neurons) → Fully connected layer (output model deviation correction coefficient);

[0095] Training data: Historical operation data of a power plant (3 years, a total of 100,000 samples).

[0096] 2. Decision-making layer:

[0097] 2.1 Reinforcement learning control agent:

[0098] The Proximal Policy Optimization (PPO) algorithm is adopted. The state space includes 15 parameters such as steam pressure, condensate flow rate, and the proportion of corroded area;

[0099] Action space: Rotation speed of gas push fan 5 (0 - 100%), frequency of condensate pump 14 (0 - 50 Hz);

[0100] Reward function: R = 0.7 × efficiency improvement rate - 0.3 × corrosion rate + 0.1 × (1 - control action frequency), where the efficiency improvement rate = (current efficiency - benchmark efficiency) / benchmark efficiency, and the unit of corrosion rate is mm / year. 0.7 × efficiency improvement rate represents the reward for the improvement of heat recovery efficiency, with a weight of 0.7. The higher the efficiency improvement rate, the greater the reward. -0.3 × corrosion rate represents the penalty for the corrosion rate, with a weight of -0.3. The higher the corrosion rate, the greater the penalty. 0.1 × (1 - control action frequency) represents the reward for the control action frequency, with a weight of 0.1. The lower the control action frequency, the greater the reward, encouraging the agent to minimize unnecessary control actions.

[0101] 2.2 Health - efficiency collaborative optimizer:

[0102] Trigger conditions:

[0103] Corrosion rate ≥ 0.1 mm / year;

[0104] The digital twin model predicts that the thermal resistance of a certain section of the pipeline increases by more than 15% of the benchmark value;

[0105] Control strategy: Reduce the steam flow rate of the corresponding pipeline by 10% - 20%, and at the same time increase the load in other areas to compensate for the efficiency loss.

[0106] S10. Multimodal data acquisition: Real-time acquisition of steam pressure, condensate flow rate, pipeline inner wall image, acoustic signal, and water quality data through a synchronously deployed sensor array;

[0107] In this embodiment, the multimodal sensor array:

[0108] Efficiency optimization sensor group:

[0109] Steam pressure sensor (such as model: Rosemount 3051S, accuracy ±0.075%FS, can be installed at the upstream flange of the gas transmission pipeline 4, and the pressure value is collected every 500 ms);

[0110] Condensate flowmeter (such as model: Krohne Promag53, range 0 - 50m 3 / h, can be installed at the elbow of the condensate liquid output pipeline 9, and the electromagnetic induction principle is used to measure the flow);

[0111] Rotary speed encoder of the gas push fan 5 (such as an incremental encoder (Omron E6B2-CWZ6C), resolution 1024 lines / revolution, integrated on the motor shaft of the gas push fan 5 through a coupling).

[0112] Health monitoring sensor group:

[0113] Pipeline endoscope camera (such as a customized high-temperature-resistant fiber optic probe, model: LumaScope360, 5 million pixels, installed at the inlet of the curved gas flow pipeline 75); or made of 316L stainless steel, temperature-resistant 300°C, with a built-in micro stepping motor (step angle 1.8°) to drive 360° rotation scanning; integrated with a MEMS microphone (model: Knowles SPM0408UD5H) and an LED lighting module (wavelength 450nm, light intensity 5000lux).

[0114] Broadband acoustic sensor (model: PCB377B10, frequency response 10Hz - 200kHz, can be pasted on the inner wall of the gas activity room 73 through high-temperature-resistant epoxy resin, sampling rate 200kHz);

[0115] Water quality sensor (model: Hach LDO101, detecting Cl - concentration (range 0 - 1000ppm, accuracy ±2%) and pH value (range 0 - 14, accuracy ±0.01), installed at the outlet of the condensate liquid output pipeline 9).

[0116] The above sensors are for reference only, and can be selected or customized according to the actual situation. Moreover, the installation method of the sensors can be installed according to the requirements or the method required in the manual, which belongs to conventional technical means. Therefore, the model and installation method of the sensors are not limited here.

[0117] Preferably, multi-modal data acquisition uses the IEEE 1588v2 (PTP) protocol to achieve time stamp alignment of multi-source data, and the synchronization accuracy ≤10μs; data acquisition frequency: pressure / flow sensor 1kHz, endoscope image 1Hz, acoustic signal 200kHz.

[0118] S20, Data Fusion and Dynamic Calibration: Use the spatio-temporal alignment algorithm to align the timestamps of multi-source data, predict the model deviation through the LSTM network, and update the parameters of the digital twin model;

[0119] In this embodiment, the specific method of data fusion and dynamic calibration is as follows:

[0120] 1. Multimodal Data Fusion Method:

[0121] 1.1 Spatio-Temporal Alignment Algorithm:

[0122] Interpolate the timestamps of the acoustic signal (200 kHz) to match the time scales of the endoscopic image (1 Hz) and the sensor data (1 kHz);

[0123] Reconstruct the time series of the high-frequency signal using cubic spline interpolation. For example, the acoustic signal (200 kHz) is downsampled to 1 kHz through cubic spline interpolation. The endoscopic image (1 Hz) and the sensor data (1 kHz) are matched by timestamps, and the missing values are filled using linear interpolation.

[0124] 1.2 Feature Fusion Method:

[0125] The proportion of the corroded area (image segmentation result), the peak acoustic energy (wavelet packet transform), and the Cl⁻ concentration (water quality sensor) are weighted and fused through the attention mechanism:

[0126]

[0127] Among them, the weight w_i is obtained by learning from the training data. This formula is used for feature fusion in multimodal data fusion. Among them, f i is the i-th feature, such as the proportion of the corroded area, the peak acoustic energy, the Cl - concentration, etc. Here, i takes values from 1 to 3.

[0128] α i is the weight of the i-th feature, calculated through the attention mechanism. w i is a learnable weight parameter, exp(w i ) is the exponential function, and the denominator is the sum of the exponential weights of all features, used for normalization, so that the sum of all weights is 1. The final fused feature fusion_feature is the weighted sum of each feature, and the weight reflects the importance of each feature in the fusion.

[0129] 2. Virtual-Reality Interaction Closed-Loop Mechanism:

[0130] After the physical system executes the control instruction, the real-time feedback data (such as efficiency change, coordinates of new corrosion points) updates the digital twin model through the API interface;

[0131] Model iteration period: LSTM prediction and parameter correction are performed every 5 minutes.

[0132] Among them, the parameters of the LSTM model are as follows:

[0133] Input layer: 10 sensor features (pressure, flow rate, rotational speed, etc.);

[0134] Bidirectional LSTM layer: 128 neurons, dropout rate 0.2;

[0135] Output layer: Model deviation correction coefficient (such as heat transfer coefficient correction factor);

[0136] Training data: 3-year historical operation data of a certain power plant (100,000 samples), MSE loss function, Adam optimizer (lr = 0.001).

[0137] S30. Reinforcement learning optimization: Run a multi-objective reinforcement learning agent in the digital twin environment to solve the Pareto optimal solution of maximizing heat recovery efficiency and minimizing corrosion rate, and generate control instructions;

[0138] In this embodiment, for example, the parameters of the reinforcement learning optimization are as follows:

[0139] 1. PPO algorithm parameters:

[0140] Discount factor γ = 0.99, gae_lambda = 0.95;

[0141] Batch size 2048, update step 10;

[0142] Network structure: 2 fully connected layers (64 → 32 neurons), ReLU activation function;

[0143] 2. Reward function:

[0144]

[0145] Among them, η t is the current heat recovery efficiency, η 0 is the benchmark efficiency (65%), r t is the corrosion rate (mm / year), and Δa is the number of control actions within 10 minutes.

[0146] S40. Health-efficiency collaborative control: When the corrosion prediction risk exceeds the threshold, dynamically adjust the control strategy (such as reducing the flow rate of high-risk pipelines), and optimize the digital twin model through virtual-real interaction closed-loop feedback.

[0147] In this embodiment, the executing mechanism is as follows:

[0148] Gas pushing fan 5: Model ABB ACS880 - 01 - 072A - 3, variable frequency speed regulation range 0 - 1500 rpm;

[0149] Condensate pump 14: Grundfos CRN5 - 10 centrifugal pump, frequency adjustment range 0 - 50 Hz, controlled by a frequency converter (Siemens MM440).

[0150] In this embodiment, data fusion and dynamic calibration are achieved through an edge computing node, and the edge computing node plays the following core roles in the present invention:

[0151] 1. Multi - modal data pre - processing and fusion

[0152] Real - time processing of high - frequency data:

[0153] Integrate an image processing unit with GPU acceleration and an acoustic signal pre - processing unit to real - time process 5 - megapixel images (1 Hz) of an endoscope camera and 200 kHz signals of a broadband acoustic sensor. Reconstruct the time series through cubic spline interpolation to achieve spatio - temporal alignment of multi - source data.

[0154] Feature extraction and dimensionality reduction:

[0155] Deploy an improved MaskR - CNN model to segment the corrosion area and extract the area ratio feature; extract the energy feature in the 150 - 200 kHz frequency band from the acoustic signal through wavelet packet transform to reduce data redundancy.

[0156] 2. Model inference and decision support

[0157] Localized intelligent analysis:

[0158] Carry an LSTM network to predict the deviation of the digital twin model (such as heat transfer coefficient correction), and run an SVM classifier to judge the corrosion type (pitting corrosion / uniform corrosion), without relying on the cloud server, reducing latency.

[0159] Reinforcement learning policy generation:

[0160] Execute the proximal policy optimization (PPO) algorithm to solve the efficiency - corrosion rate Pareto optimal solution in the digital twin environment and generate control instructions (such as fan speed, pump frequency), with a response time ≤ 200 ms.

[0161] 3. Edge intelligence and cloud collaboration

[0162] Reduce cloud load:

[0163] Complete more than 90% of real-time data processing (such as corrosion prediction and control instruction generation) on the edge side, and only upload the key decision results (such as corrosion risk level and control parameters) to the cloud through the 5G module (Huawei ME909s-821) to reduce the network bandwidth pressure.

[0164] Offline emergency processing:

[0165] When the network is interrupted, rely on the 1TB SSD historical data and pre-trained models stored locally to continuously execute health monitoring and control strategies to ensure system reliability.

[0166] 4. Hardware integration and environment adaptation

[0167] Industrial-level deployment ability:

[0168] Adopt NVIDIA Jetson AGX Xavier edge computing nodes, support a wide temperature environment from -40°C to 70°C, and meet the requirements of harsh working conditions such as high temperature and vibration in the steam turbine regenerative system.

[0169] Modular expansion interface:

[0170] Expand the AI acceleration card (such as NVIDIA T4) through the PCIe interface to support future model iteration and upgrade without replacing the hardware platform.

[0171] 5. Virtual-real interaction closed-loop core

[0172] Real-time feedback correction:

[0173] Update the actual effect data (such as efficiency change and new corrosion point coordinates) after the physical system executes the control instruction to the digital twin model in real time through the API interface, and drive the model parameter iteration (once every 5 minutes).

[0174] 6. Health-efficiency collaborative trigger:

[0175] When the corrosion prediction risk exceeds the threshold, directly output a control signal to the actuator (such as a control valve) through the GPIO interface to achieve millisecond-level response without relying on the central controller.

[0176] Preferably, the following configurations can be selected for the edge computing node:

[0177] Hardware configuration: NVIDIA Jetson AGX Xavier, 32GB DDR4 memory, 1TB SSD storage, equipped with a 5G communication module (Huawei ME909s-821);

[0178] Software environment: Ubuntu 18.04 operating system, CUDA 11.4, cuDNN 8.2, TensorRT 8.0;

[0179] Integrated module:

[0180] GPU-accelerated image processing unit (Mask R-CNN inference engine optimized based on TensorRT);

[0181] Acoustic signal preprocessing unit (PyWavelets library for wavelet packet transform).

[0182] To verify the experimental effect of the present invention, the following test scenarios are provided:

[0183] The regenerative heating system of a 100MW steam turbine in a certain power plant, with a test period of 6 months;

[0184] Comparison schemes: Prior art (CN215860369U), single digital twin optimization (without corrosion prediction), single corrosion prediction (without efficiency optimization).

[0185] The experimental results are as follows in the table:

[0186]

[0187] Among them, statistical significance: verified by t-test, the result of efficiency improvement is p < 0.01, which is statistically significant.

[0188] Improvement in heat recovery efficiency (η): The difference between the heat recovery efficiency of the optimized system and the traditional system, formula:

[0189]

[0190] (Q is the heat recovery amount, unit: kJ / s)

[0191] False positive rate of corrosion (FPR): The proportion of normal samples misjudged as corroded by the model in the total normal samples, formula:

[0192]

[0193] Reduction in maintenance cost (C): The cost savings brought about by the reduction in the number of maintenance times under the optimized strategy, formula:

[0194]

[0195] Digital twin prediction error (ε): The root mean square error (RMSE) between the model prediction value and the true value, formula:

[0196]

[0197] Explanation of technical terms

[0198] SIMPLE algorithm: A semi - implicit method for solving the pressure - coupled equations (used for solving the thermodynamic equations in the digital twin model), which decouples the velocity and pressure through the pressure correction equation.

[0199] Dice Loss: A loss function for image segmentation, which measures the overlap between the predicted region and the ground - truth region.

[0200] Embodiment 2

[0201] Based on the same concept, the present invention also proposes a steam turbine regenerative system, including a control end, which is equipped with a control method for a steam turbine regenerative system in Embodiment 1.

[0202] In this embodiment, the structure of the existing steam turbine regenerative system is not completely improved. It only lies in adding sensors, edge computing nodes and their corresponding devices.

[0203] Embodiment 3

[0204] This embodiment also provides a steam turbine, including the steam turbine regenerative system in Embodiment 2 above.

[0205] Generally, various embodiments can be implemented in hardware or special - purpose circuits, software, logic, or any combination thereof. Some aspects of the present invention can be implemented in hardware, while other aspects can be implemented by firmware or software executed by a controller, microprocessor, or other computing device. However, the present invention is not limited thereto. Although various aspects of the present invention can be shown and described as block diagrams, flowcharts, or using some other graphical representation, it should be understood that, by way of non - limiting example, the blocks, devices, systems, techniques, or methods described herein can be implemented in hardware, software, firmware, special - purpose circuits or logic, general - purpose hardware or controllers, or other computing devices, or some combination thereof.

[0206] Embodiments of the present invention can be implemented by computer software, which is executable by a data processor of a mobile device, such as in a processor entity, or by hardware, or by a combination of software and hardware. A computer software or program (also referred to as a program product), including software routines, applets, and / or macros, can be stored in any device-readable data storage medium, and they include program instructions for performing specific tasks. The computer program product can include one or more computer-executable components configured to perform the embodiments when the program runs. One or more computer-executable components can be at least one software code or a part thereof. Additionally, in this regard, it should be noted that any box in the logical flow in the figure can represent a program step, or interconnected logical circuits, boxes, and functions, or a combination of program steps and logical circuits, boxes, and functions. The software can be stored on physical media such as memory chips or storage blocks implemented within the processor, magnetic media such as hard disks or floppy disks, and optical media such as, for example, DVDs and their data variants, CDs. The physical media are non-transitory media.

[0207] Those skilled in the art should understand that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as within the scope described in this specification.

[0208] The above embodiments merely represent several implementation manners of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be pointed out that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the appended claims.

Claims

1. A steam turbine heat recovery system control method, characterized in that: The following steps are involved: S00. Build a digital twin model: integrate thermodynamic equations with equipment aging parameters to establish a virtual model of steam flow, condensation heat transfer and corrosion evolution. The equipment aging parameters include the sponge pad porosity attenuation coefficient and the pipeline roughness growth model. S10, Multimodal data acquisition: Real-time acquisition of steam pressure, condensate flow, pipeline inner wall image, acoustic signal and water quality data through synchronously deployed sensor arrays; S20, data fusion and dynamic calibration: Use the spatiotemporal alignment algorithm to align the timestamps of multi-source data, predict the model deviation through the LSTM network and update the parameters of the digital twin model; S30, Reinforcement Learning Optimization: Run a multi-objective reinforcement learning agent in a digital twin environment to solve the Pareto optimal solution of maximizing heat recovery efficiency and minimizing corrosion rate, and generate control instructions; S40, Health-efficiency collaborative control: When the predicted corrosion risk exceeds the threshold, the control strategy is dynamically adjusted, and the digital twin model is optimized through virtual-reality interactive closed-loop feedback.

2. A steam turbine heat recovery system control method according to claim 1, characterized in that: In step S10, the sensors used for the multimodal data collection include: Steam pressure sensor, used to monitor the steam pressure in the gas pipeline in real time; A condensate flow meter, used to measure the condensate flow rate of the liquid in the condensate output pipeline; Fan speed encoder, used to detect the speed of the gas push fan; High temperature resistant endoscope probe, used to obtain images of the inner wall of curved gas flow pipes; Wideband acoustic sensor, used to collect acoustic signals in the frequency range of 10Hz-200kHz in the gas activity chamber; Water quality sensor, used to detect the Cl content of the liquid in the condensate output pipeline - Concentration and pH.

3. A steam turbine heat recovery system control method according to claim 1, characterized in that: In step S00, the thermodynamic equations of the digital twin model include the continuity equation, the momentum equation and the energy equation, and the equipment aging parameters are obtained by fitting the historical operation data.

4. A steam turbine heat recovery system control method according to claim 1, characterized in that: In step S20, the data fusion step includes: The time series of high-frequency acoustic signals are reconstructed using cubic spline interpolation; The attention mechanism is used to weight the fusion of corrosion area ratio, acoustic energy peak and Cl - Concentration characteristics.

5. A steam turbine heat recovery system control method according to claim 1, characterized in that: In step S30, the reinforcement learning agent adopts a proximal strategy optimization algorithm, and the reward function R is: R=0.7×efficiency improvement rate-0.3×corrosion rate+0.1×(1-control action frequency).

6. A steam turbine heat recovery system control method according to claim 2, characterized in that: In step S10, the high temperature resistant endoscope probe is integrated with a micro camera and a MEMS microphone, and can rotate 360 ​​degrees to scan the inner wall of the curved gas circulation pipeline.

7. A steam turbine heat recovery system control method according to claim 1, characterized in that: In step S40, the health-efficiency collaborative control adjusts the control strategy when any of the following conditions is triggered: Corrosion rate ≥ 0.1mm / year; The virtual model predicted that the thermal resistance of a certain section of pipe would increase by 15% over the baseline value.

8. A steam turbine heat recovery system control method according to any one of claims 1 to 7, characterized in that: In step S10, multimodal data acquisition uses the IEEE1588v2 protocol to achieve multi-source data timestamp alignment, with a synchronization accuracy of ≤10μs.

9. A steam turbine heat recovery system, comprising a control end, characterized in that: The control end is equipped with a steam turbine heat recovery system control method as described in any one of claims 1-8.

10. A steam turbine, characterized in that: It includes a steam turbine heat recovery system as described in claim 9.

Images