Method for predicting service life of steam turbine rotor under deep peak regulation working condition of coal power unit

Abstract

The invention provides a steam turbine rotor life prediction method under a deep peak regulation working condition of a coal power unit. The method comprises the following steps: collecting operation parameters in real time and preprocessing the operation parameters to form input feature vectors; simulating a transient temperature field and a stress field through finite element analysis, and constructing a training data set; training the neural network model to input the feature vector to predict the transient temperature and equivalent stress of the key part; calculating low-cycle fatigue damage based on a Manson-Coffin model, calculating creep damage based on a Norton creep law, and calculating total damage by adopting a Miner linear cumulative damage theory; and calculating residual life according to the total damage and performing early warning. The method is suitable for the coal power unit with frequent deep peak regulation operation, the life risk can be found in time, and non-planned shutdown is reduced.

Description

Technical Field

[0001] This application belongs to the field of power equipment condition monitoring and life management technology, and in particular relates to a method for predicting the life of a steam turbine rotor under deep peak shaving conditions of a coal-fired power unit. Background Technology

[0002] With the advancement of my country's "dual carbon" goals and the construction of new power systems, the installed capacity of new energy sources such as wind power and photovoltaics has grown rapidly, significantly increasing the volatility of power grid load. As the main regulating power source of the power system, coal-fired power units are gradually shifting from traditional base load operation to deep peak-shaving operation (load reduced to 20%-50% of rated load), and frequent start-ups and shutdowns and large-scale load changes have become the norm.

[0003] Under deep peak shaving conditions, turbine rotors (especially high- and intermediate-pressure rotors) are subjected to drastic temperature and pressure changes, resulting in significant alternating thermal stress in key parts (such as the root of the regulating stage impeller, the fillet of the impeller disk, and the shaft seal section of the intermediate-pressure cylinder), which accelerates low-cycle fatigue damage. At the same time, creep damage is aggravated under high-temperature conditions, and the interaction between low-cycle fatigue and creep significantly shortens the service life of the rotor.

[0004] In existing technologies, turbine rotor life assessment mainly relies on offline finite element analysis and traditional empirical models (such as life consumption estimation based on operating hours), which makes it difficult to achieve real-time online prediction and accurately reflect the complex damage mechanisms under deep peak shaving transient conditions, resulting in delayed operation and maintenance decisions and increased risk of unplanned downtime.

[0005] Therefore, there is an urgent need for a method that can predict the lifespan of turbine rotors under deep peak shaving conditions in real time and with relatively high accuracy, so as to ensure the safe and stable operation of the unit and support the construction of new power systems. Summary of the Invention

[0006] The purpose of this application is to overcome the problem of poor real-time performance in the prediction of turbine rotor life in the prior art, and to provide a method for predicting the life of turbine rotor under deep peak shaving conditions of coal-fired power units. This method uses neural networks to quickly predict the transient temperature and stress field of key parts, and combines the classical fatigue-creep damage model to achieve online prediction of total damage and remaining life.

[0007] This application provides a method for predicting the lifespan of a steam turbine rotor under deep peak-shaving conditions in coal-fired power units, including the following steps: S1. Real-time acquisition of operating parameters of coal-fired power units under deep peak shaving conditions, and preprocessing to form input feature vectors; S2. Use finite element analysis to simulate the transient temperature field and stress field of the turbine rotor under deep peak shaving conditions, extract the temperature and equivalent stress of key parts as label data, and construct a training dataset. S3. Design and train a neural network model based on the training dataset. The neural network model uses the input feature vector as input to predict the transient temperature and equivalent stress of key parts of the turbine rotor. S4. Input the real-time collected input feature vector into the trained neural network model to obtain the transient temperature and equivalent stress of the key parts; S5, based on Manson-Coffin The model calculates low-cycle fatigue damage based on Norton The creep law is used to calculate creep damage, and employs... Miner The linear cumulative damage theory calculates the total damage; S6. Calculate the remaining lifespan based on the total damage and issue an early warning.

[0008] In one optional implementation, in step S1, the operating parameters include the main steam temperature. T main Main steam pressure P main Load factor L Rotor shaft vibration value V Steam inlet flow rate Q Load change rate dL / dt and the rate of change of main steam temperature dT main / dt ; The preprocessing includes normalization, outlier removal, and missing value imputation to form the input feature vector. X =[ T main , P main , L , V , Q , dL / dt , dT main / dt ].

[0009] In an optional implementation, in step S2, the finite element analysis uses a two-dimensional axisymmetric model or a three-dimensional model to simulate deep peak shaving conditions, including load reduction to 20%-50% of rated load and a load change rate of 4-20 MW / min; the key components include the root of the regulating stage impeller, the fillet of the impeller disk, and the shaft seal section of the intermediate pressure cylinder; the training dataset includes samples under steady-state, transient, and deep peak shaving conditions, with a sample size of not less than 10,000.

[0010] In one optional implementation, in step S3, the neural network model is a Long Short-Term Memory network, including an input layer, a hidden layer, and an output layer; the hidden layer employs... ReLUActivation function; loss function is mean squared error; optimizer is Adam The algorithm has a learning rate of 0.001, 200-500 training epochs, and a validation set error of no more than 5%.

[0011] In an optional implementation, in step S5, the... Manson-Coffin The calculation formula for the model is: ; in, For plastic strain amplitude, For the first i Fatigue life cycle count at stress level 1 The fatigue ductility coefficient, The fatigue strength coefficient, b and c The fatigue index of the material. E It is the elastic modulus.

[0012] In an optional implementation, in step S5, the... Norton The formula for calculating the law of creep is: ; in, ε c For creep strain, A , n , Q For material constants, σ For equivalent stress, R The gas constant is T This is the absolute temperature obtained by converting the transient temperature of the key component. t This refers to the creep time.

[0013] In an optional implementation, in step S5, the... Miner The calculation formula for linear cumulative damage theory is as follows: ; in, D Total damage, For the first i Number of fatigue cycles For the first i Fatigue life cycle count at stress level 1 t j For the first j creep time T j This corresponds to the creep life.

[0014] In one optional implementation, in step S6, the formula for calculating the remaining lifetime is: ; in, RL For remaining lifespan, D Total damage, L 0 Design life; when D An early warning is issued when the value is ≥0.8. RL An alarm is triggered when the power plant's operating time is less than 5000 hours, and the alarm is pushed to the power plant's smart platform.

[0015] In one alternative implementation, the method is applicable to the high- and medium-pressure rotors of steam turbines in 300MW-700MW coal-fired power units.

[0016] In one alternative implementation, the method further includes dynamically adjusting the load change rate based on the remaining lifetime to reduce lifetime loss.

[0017] Compared with the prior art, this application has the following beneficial effects: This application provides a method for predicting the life of a steam turbine rotor under deep peak-shaving conditions in coal-fired power units. It replaces the time-consuming offline calculations of traditional finite element methods by employing neural networks to rapidly predict transient temperatures and stress fields in key components, achieving real-time online prediction. Simultaneously, it combines... Manson-Coffin , Norton and Miner These classic mechanistic models ensure the physical basis and engineering accuracy of damage assessment. This method is applicable to coal-fired power units that frequently operate under deep peak-shaving conditions, enabling timely detection of potential life risks, reducing unplanned downtime, and improving the unit's peak-shaving capacity and overall operational economy. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating a method for predicting the lifespan of a steam turbine rotor under deep peak-shaving conditions in a coal-fired power unit, provided as an embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions in the embodiments of this application are described clearly and completely below. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are also within the scope of protection of this application.

[0021] First, let me explain the terms used in this application: Deep peak shaving: refers to the operating condition where the load of a coal-fired power unit drops to 20%-50% of its rated load and is frequently started and stopped or undergoes significant load changes; Low-cycle fatigue damage refers to the damage that occurs to a component under cyclic loading. Creep damage refers to the slow deformation damage that occurs in materials under continuous stress at high temperatures. Miner Linear cumulative damage theory: refers to the principle that fatigue damage and creep damage can be linearly superimposed, and the component fails when the cumulative damage reaches a critical value; Long Short-Term Memory Network (LSTM) LSTM : is a recurrent neural network that can process time-series data.

[0022] like Figure 1 As shown in the figure, this application provides a method for predicting the life of a steam turbine rotor under deep peak-shaving conditions in coal-fired power units, including the following steps: S1. Real-time acquisition of operating parameters of coal-fired power units under deep peak shaving conditions, and preprocessing to form input feature vectors; S2. Use finite element analysis to simulate the transient temperature field and stress field of the turbine rotor under deep peak shaving conditions, extract the temperature and equivalent stress of key parts as label data, and construct a training dataset. S3. Design and train a neural network model based on the training dataset. The neural network model uses the input feature vector as input to predict the transient temperature and equivalent stress of key parts of the turbine rotor. S4. Input the real-time collected input feature vectors into the trained neural network model to obtain the transient temperature and equivalent stress of key parts; S5, based on Manson-Coffin The model calculates low-cycle fatigue damage based on Norton The creep law is used to calculate creep damage, and employs... Miner The linear cumulative damage theory calculates the total damage; S6. Calculate the remaining lifespan based on the total damage and issue an early warning.

[0023] This embodiment addresses the problem of poor real-time performance in life prediction in existing technologies. Existing technologies rely on offline finite element analysis, making it difficult to acquire transient temperature and stress field data in real time under deep peak-shaving conditions. This embodiment first acquires operating parameters in real time in step S1 and preprocesses them to form input feature vectors. Then, in step S2, it uses finite element analysis to simulate various deep peak-shaving conditions to construct a training dataset. Next, in step S3, it trains a neural network model based on this dataset. Finally, in step S4, it directly substitutes the real-time input into the model to obtain the transient temperature and equivalent stress of key components. This data acquisition-to-model prediction processing method transforms the offline calculation of finite element simulation into rapid inference of a pre-trained model, reducing the repetitive simulation time for each change in operating conditions and helping to improve prediction response speed.

[0024] Furthermore, in step S5 of this embodiment, a solution is provided to address the problem that existing technologies cannot accurately reflect the complex damage mechanisms under deep peak-shaving transient conditions. Existing technologies are mostly based on empirical estimations of operating hours, neglecting the interaction between low-cycle fatigue and creep. In step S5 of this embodiment, based on... Manson-Coffin The model calculates low-cycle fatigue damage based on Norton The creep law is used to calculate creep damage, and employs... Miner The linear cumulative damage theory calculates the total damage. Specifically, the transient temperature and equivalent stress obtained from step S4 are used as inputs and directly substituted into these models for damage quantification. This model integration process considers the influence of alternating thermal stress and high-temperature bearing time during deep peak shaving, reducing the simplification error of empirical models and thus providing a more reliable basis for damage assessment.

[0025] Furthermore, this embodiment addresses the issue of delayed operation and maintenance decision-making caused by existing technologies through step S6. Existing technologies increase the risk of unplanned downtime due to untimely predictions. In step S6 of this embodiment, the remaining lifespan is directly calculated based on the total damage calculated in step S5, and an early warning is issued. Specifically, the alarm mechanism is triggered by comparing the total damage value with a threshold. This continuous process from damage calculation to early warning transforms the prediction results into an operation and maintenance reference, shortens the interval between analysis and action, reduces the probability of sudden rotor failure, and helps improve the safety of unit operation.

[0026] In practical applications, the operating parameters collected in step S1 of this embodiment originate from the power plant's DCS system. These parameters directly reflect the current operating characteristics of the unit. The input feature vector formed after preprocessing can better adapt to the input requirements of the neural network, facilitating subsequent prediction. The finite element simulation in step S2 can utilize ANSYS software to establish a two-dimensional axisymmetric model or a three-dimensional model.

[0027] In step S1, the operating parameters include the main steam temperature.T main Main steam pressure P main Load factor L Rotor shaft vibration value V Steam inlet flow rate Q Load change rate dL / dt and the rate of change of main steam temperature dT main / dt ; Preprocessing includes normalization, outlier removal, and missing value imputation to form the input feature vector. X =[ T main , P main , L , V , Q , dL / dt , dT main / dt ].

[0028] These operating parameters are derived from real-time data collected by the power plant's DCS system, including the main steam temperature. T main and main steam pressure P main Directly reflects the boiler outlet operating conditions and load rate L and load change rate dL / dt The rotor shaft vibration value reflects the peak-shaving depth and speed of the unit. V and steam flow Q This is used to monitor rotor vibration and abnormal steam flow. The normalization process can employ a min-max normalization method, mapping each parameter to the [0, 1] interval; outlier removal can be achieved through 3... σ Criteria or box plots identify and replace values ​​with nearest neighbor values ​​or the median; missing values ​​are filled using linear interpolation or by using the average of historical data. This preprocessing method has been widely validated in the field and can effectively eliminate dimensional differences and noise interference, making the input feature vector more suitable for subsequent neural network model processing.

[0029] The min-max normalization method is used for normalization, mapping each parameter to the interval [0, 1], as shown in the formula. X' = ( X - min ) / ( max - min This ensures that parameters with different dimensions are compatible with neural network inputs. Through 3 σ The criterion for outlier removal is the "three sigma rule" in statistics, based on a normal distribution. Outliers exceeding the mean by ±3 standard deviations are excluded.σ Values ​​exceeding a certain threshold are considered outliers and are used for outlier removal. Box plots are commonly used to visualize data distribution and detect outliers. Those skilled in the art can choose the appropriate outlier removal method based on the specific circumstances.

[0030] Through the acquisition and preprocessing in step S1 above, this embodiment obtains high-quality input data, providing a reliable foundation for subsequent neural network prediction. In practical applications, the real-time performance of these parameters directly determines the model's response speed and prediction accuracy.

[0031] Furthermore, in this embodiment, the input feature vector X obtained through the acquisition and preprocessing in step S1 is a time-series feature vector suitable for the LSTM model. This vector can be updated once per second or per minute in actual operation, synchronized with the real-time operating conditions of the unit, and is more suitable for rotor life prediction scenarios of 300MW to 700MW coal-fired power units.

[0032] In step S2, the finite element analysis uses a two-dimensional axisymmetric model or a three-dimensional model to simulate deep peak shaving conditions, including load reduction to 20%-50% of the rated load and a load change rate of 4-20 MW / min. Key components include the root of the regulating stage impeller, the fillet of the impeller disk, and the shaft seal section of the intermediate pressure cylinder. The training dataset includes samples under steady-state, transient, and deep peak shaving conditions, with a sample size of not less than 10,000.

[0033] Finite element simulations were performed using ANSYS software. Taking a 660MW coal-fired power unit as an example, the rotor model was established based on the actual geometric dimensions and material properties of the 660MW unit. The material was 30Cr1Mo1V high-temperature alloy. Boundary conditions included loading curves showing steam temperature and pressure variations with load, as well as the convective heat transfer coefficient. The simulation process first calculated the temperature and stress fields under steady-state conditions as initial conditions, then applied transient loads with varying loads. The time step was adjusted according to the load variation rate, typically 1-10 seconds, to capture temperature gradients and thermal stress changes. During the simulation, transient temperatures and von Mises equivalent stresses at key locations were extracted as output indicators for subsequent neural network training and damage calculation. During mesh generation, local refinement was performed in stress concentration areas such as the impeller root and disk fillets, with the mesh size controlled between 100,000 and 200,000 to balance computational accuracy and efficiency.

[0034] In step S3, the neural network model is a Long Short-Term Memory network, including an input layer, hidden layers, and an output layer; the hidden layer uses... ReLU Activation function; loss function is mean squared error; optimizer is Adam The algorithm has a learning rate of 0.001, 200-500 training epochs, and a validation set error of no more than 5%.

[0035] In the above embodiments, the Long Short-Term Memory network trained in step S3 ( LSTM Its hidden layer uses ReLU The activation function is the mean squared error loss function, and the optimizer is... Adam The algorithm has a learning rate of 0.001 and 200-500 training epochs until the validation set error is no greater than 5%. This configuration is an effective way to process time series data and can capture dynamic features such as load change rate and temperature change rate.

[0036] Optionally, the training process uses PyTorch or TensorFlow The framework implementation matches the number of nodes in the input layer with the dimension of the feature vector, sets 2-4 hidden layers with 128-256 neurons per layer, and the number of nodes in the output layer corresponds to the temperature and equivalent stress of key parts. During training, the dataset constructed by S2 is divided into training, validation, and test sets in an 8:1:1 ratio. A mini-batch gradient descent method with a batch size of 32-128 is used. After each training round, the mean squared error is calculated on the validation set. Training stops when the error no longer decreases or falls below 5% for 10 consecutive rounds.

[0037] Actual training can be performed on servers equipped with GPUs, and data loading is done using... PyTorch In DataLoader This approach aims to improve efficiency. The number of neurons in the hidden layer and the batch size can be adjusted according to the dataset size; for example, a larger batch size is used when the sample size is large to accelerate convergence. The early stopping mechanism is implemented by monitoring the validation set error to avoid overfitting the model on the training set. This training process enables the model to learn dynamic features such as the rate of load change, providing stable prediction results for transient temperature and equivalent stress in key components under deep peak shaving conditions.

[0038] In step S4, the real-time feature vector is input into the trained... LSTM The model obtains the transient temperature and equivalent stress of key components. This process has a short computation time and is suitable for online applications.

[0039] In step S5, Manson-Coffin The calculation formula for the model is: ; in, For plastic strain amplitude, For the first i Fatigue life cycle count at stress level 1 The fatigue ductility coefficient, The fatigue strength coefficient, b and c The fatigue index of the material. E It is the elastic modulus.

[0040] Based on Manson-Coffin When calculating low-cycle fatigue damage using the model, the plastic strain amplitude is calculated based on the equivalent stress at key locations. Substitute Manson-Coffin The calculation formula of the model, solving the first... i Level fatigue life cycle count Thus, with the first i Level fatigue cycle count With the i Level fatigue life cycle count ratio To calculate low-cycle fatigue damage.

[0041] Specifically, in low-cycle fatigue damage calculation, the equivalent stress obtained from S4 is first decomposed into two parts: elastic strain and plastic strain. The elastic strain is calculated using Hooke's law, and the plastic strain amplitude... Take the peak-to-valley difference under cyclic loading. Material constants. , , b , c Typical values ​​can be obtained from handbooks or experimental curves for rotor steel (such as 30Cr1Mo1V). ≈0.35 ≈1200Mpa b ≈-0.09、 c ≈-0.6、 E ≈200GPa. Among them, b The fatigue strength index reflects the slope of the stress-life relationship in the high-cycle zone, and is usually a negative value. c The fatigue ductility index reflects the slope of the plastic strain-life relationship in the low-cycle region and is also negative. Substituting into the formula, the corresponding solution can be obtained through numerical solutions (such as Newton's iteration method). The process is in MATLAB or Python It is implemented in a short time and is suitable for online applications.

[0042] In step S5, Norton The formula for calculating the law of creep is: ; in, ε c For creep strain, A , n , Q For material constants, σ For equivalent stress, R This is the gas constant (usually 8.314 J / mol·K); TThe absolute temperature (in Kelvin, K) is obtained by converting the transient temperature of the key component, and is used for calculation. Arrhenius Activation energy item ; t This refers to the creep time.

[0043] In Norton's law of creep, A , n , Q These three parameters are collectively referred to as "material constants" or "creep constants," among which, A This is the creep coefficient (or structural factor). n The stress index reflects the creep mechanism (diffusion creep). n ≈1, dislocation creep n =3~8); Q It is the creep activation energy.

[0044] Since the transient temperature of critical components (predicted from step S4) is usually expressed in degrees Celsius (°C), it must be converted to an absolute temperature before being substituted into the formula. This conversion (absolute temperature = transient temperature + 273.15) is a standard procedure in creep calculations in this field to ensure that the thermal activation process in the formula accurately reflects the material behavior.

[0045] Based on Norton When calculating creep damage using creep laws, the transient temperature of critical components is converted to absolute temperature and compared with equivalent stress. σ Let's put it into perspective Norton The creep law is used to calculate creep strain. ε c Then, the creep life is calculated based on the material creep failure criterion. T j Thus, with the j-th stage creep time t j and T j ratio To calculate creep damage.

[0046] Specifically, in creep damage calculations, the transient temperature is converted to an absolute temperature and then substituted along with the von Mises equivalent stress. Into Norton Formulas. Material constants. A , n , Q Typical values ​​for 30Cr1Mo1V steel in the 550-600℃ range, as shown in material handbooks. A ≈10 -25 , n ≈8、 Q ≈500 kj / mol The creep failure criterion adopts the strain limit method, that is, when... εc The material's lifespan ends when it reaches the fracture strain (typically 0.01-0.05). The corresponding solution is then used to determine the lifespan. T j The calculation takes into account the changes in temperature and stress over time, and can perform piecewise integration to handle multiple levels of operating conditions, ensuring that the damage assessment is consistent with actual operation.

[0047] In step S5, Miner The calculation formula for linear cumulative damage theory is as follows: ; in, D Total damage, For the first i Number of fatigue cycles For the first i Fatigue life cycle count at stress level 1 For the first j creep time This corresponds to the creep life.

[0048] use Miner When calculating total damage using the linear cumulative damage theory, low-cycle fatigue damage is accumulated. and creep damage The total damage was obtained. D .

[0049] Through the damage calculation in step S5 above, this embodiment obtains the total damage. D Used for S6 remaining life assessment.

[0050] In step S6, the formula for calculating the remaining lifetime is: ; in, RL For remaining lifespan, D Total damage, L 0 Design life; when D An early warning is issued when the value is ≥0.8. RL An alarm is triggered when the power plant's operating time is less than 5000 hours, and the alarm is pushed to the power plant's smart platform.

[0051] Remaining lifespan RL Units and design life L 0 The units are consistent, usually expressed in operating hours (h) or start-stop cycles. In turbine rotor life assessment, the design life is... L 0The design life of a 30Cr1Mo1V steel rotor is determined based on material properties, manufacturing standards, and operating conditions. For example, the design life of a 30Cr1Mo1V steel rotor under normal operating conditions is typically 200,000-300,000 operating hours, or equivalent to 10... 5 The start-stop cycle can be obtained from manufacturer data or the DL / T 785-2001 standard. In actual calculations, if... L 0 In hours, RL It is also in hours; if measured in terms of the number of cycles, then... RL This corresponds to the remaining number of iterations. This formula is based on... Miner The linear extrapolation of the criterion is simple to calculate.

[0052] The selection of early warning and alarm thresholds takes into account the engineering safety margin. D A value ≥0.8 indicates that the cumulative damage has reached 80%, at which point an early warning should be issued so that maintenance personnel can intervene in advance. RL A threshold of <5000 hours is used for alarms, typically corresponding to a major overhaul cycle, facilitating scheduled downtime for maintenance. Alarm information is determined by software programs. D and RL The value is generated and pushed to the power plant's smart platform or operator station. The content includes the current damage value, remaining life estimate, and recommended measures.

[0053] Furthermore, this method is applicable to the high- and medium-pressure rotors of steam turbines in 300MW-700MW coal-fired power units.

[0054] Most coal-fired power units within this capacity range operate with subcritical or supercritical parameters, and their rotor materials primarily utilize high-temperature alloys such as 30Cr1Mo1V. They exhibit similar characteristics in temperature and pressure distribution under operating conditions, as well as transient thermal stress changes caused by deep peak shaving. This method, through finite element simulation and neural network training, can match the actual operating characteristics of units within this range using the boundary conditions and material properties employed. Taking the applicant as an example, the Phase I 2×350MW units and the Phase II 2×660MW units accumulated a large amount of real-time parameters and damage data during deep peak shaving operation, which were used for model training and verification. The prediction results were largely consistent with the findings discovered during unit maintenance.

[0055] In some embodiments, the method further includes dynamically adjusting the load change rate based on the remaining lifetime to reduce lifetime loss.

[0056] Optionally, the load change rate can be dynamically adjusted based on the remaining lifespan using preset threshold rules. For example, when the remaining lifespan is less than 20,000 hours, the load change rate is limited to 10 MW / min; when the remaining lifespan is less than 10,000 hours, it is further limited to 6 MW / min. These thresholds and rate ranges are determined based on the unit's historical operating records and the safety margins in the DL / T 785-2001 standard, and are written in the software program as conditional statements. Adjustment commands are sent through the interface with the DCS system and are automatically executed by the control logic or take effect after operator confirmation. This method can reduce thermal stress shocks caused by rapid load changes in actual operation, thereby reducing additional lifespan losses.

[0057] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for predicting the service life of a steam turbine rotor in a deep peak shaving operating condition of a coal-fired power unit, characterized in that, The method comprises the following steps: S1, collecting and preprocessing the operating parameters of a coal-fired unit in deep peak regulation conditions in real time to form an input feature vector; S2, simulating the transient temperature field and stress field of a steam turbine rotor in deep peak regulation conditions by finite element analysis, extracting the temperature and equivalent stress of key positions as label data, and constructing a training data set; S3, designing and training a neural network model based on the training data set, the neural network model taking the input feature vector as input to predict the transient temperature and equivalent stress of key positions of the steam turbine rotor; S4, inputting the input feature vector collected in real time into the trained neural network model to obtain the transient temperature and equivalent stress of the key positions; S5、based on Manson-Coffin model calculates low-cycle fatigue damage, based on Norton creep law calculates creep damage, and adopts Miner linear cumulative damage theory calculates total damage; S6, calculating the remaining life according to the total damage and giving a warning.

2. The method for predicting the service life of a steam turbine rotor in a deep load modulation condition of a coal-fired power unit according to claim 1, characterized in that, The operating parameter in the step S1 includes a main steam temperature T main , a main steam pressure P main , a load rate L , a rotor shaft vibration value V , an admission flow rate Q , a load change rate dL / dt and a main steam temperature change rate dT main / dt ; The preprocessing includes normalization processing, outlier rejection and missing value filling, forming an input feature vector X [ T main , P main , L , V , Q , dL / dt , dT main / dt ]。 3. The method of claim 1, wherein, In the step S2, the finite element analysis adopts a two-dimensional axisymmetric model or a three-dimensional model to simulate deep peak regulation conditions, including reducing the load to 20%-50% of the rated load, and the variable load rate is 4-20 MW / min; the key positions include the root of the regulating stage impeller, the fillet of the disc, and the shaft seal section of the intermediate pressure cylinder; the training data set includes samples under steady state, transient state, and deep peak regulation conditions, and the sample size is not less than 10,000.

4. The method of claim 1, wherein, In the step S3, the neural network model is a long short-term memory network, including an input layer, a hidden layer and an output layer; the hidden layer adopts ReLU an activation function; a loss function is a mean square error, and an optimizer is an algorithm, a learning rate is 0.001, a training round is 200-500, and a validation set error is not greater than 5%. Adam an algorithm, a learning rate is 0.001, a training round is 200-500, and a validation set error is not greater than 5%.

5. The method of claim 1, wherein, In the step S5, the Manson-Coffin The calculation formula of the model is: ； wherein, is the plastic strain amplitude, is the fatigue life cycle number at the i stress level, is the fatigue ductility coefficient, is the fatigue strength coefficient, b and c is the material fatigue exponent, E is the modulus of elasticity.

6. The method of claim 1, wherein, In the step S5, the Norton The calculation formula of the creep law is: ； wherein, ε c is the creep strain, A , n , Q is a material constant, σ is the equivalent stress, R is the gas constant, T is the absolute temperature converted from the transient temperature of the critical site, t is the creep time.

7. The method of claim 1, wherein, In the step S5, the Miner The calculation formula of linear cumulative damage theory is: ； wherein, D is the total damage, is the number of fatigue cycles at the i level, is the number of fatigue life cycles at the i stress level, t j is the time to creep at the j level, T j is the corresponding creep life.

8. The method of claim 1, wherein, In the step S6, the calculation formula of the remaining life is: ； wherein, RL is the remaining life, D is the total damage, L 0 is the design life; When D Warning is issued when ≥ 0.8, and RL Alarm is triggered when < 5000 hours, and pushed to the smart platform of the power plant.

9. The method of claim 1, wherein, The method is suitable for steam turbine high-pressure rotors of 300MW-700MW coal-fired units.

10. The method of claim 1-9, wherein, The method further comprises dynamically adjusting the load change rate according to the remaining life to reduce the life loss.

Images