Automatic scoring method and system of thermal generator set simulation operation platform

Abstract

The invention relates to the technical field of electric power system simulation and automatic control, and discloses an automatic scoring method and system for a thermal generator set simulation operation desk, and the method comprises the steps: collecting an operation instruction stream and physical parameter monitoring data through a multi-source sensor, and carrying out the time-space alignment and abnormality cleaning of the collected multi-source data, then constructing a space-time four-dimensional tensor model representing the running state of the unit, wherein the dimensions of the space-time four-dimensional tensor model comprise a time axis, a space subsystem, running parameters and equipment health degree; constructing a meta reinforcement learning framework, generating a dynamic scoring weight vector by using the space-time four-dimensional tensor, and performing real-time scoring in combination with an attenuation factor in the equipment health degree; through TSN protocol dynamic priority distribution and a multi-source data synchronous compensation technology, microsecond-level space-time alignment of operation instruction streams, physical sensing parameters and virtual data is realized, the bottleneck of time sequence mismatch caused by a traditional heterogeneous protocol is broken through, and the problem of feature modeling distortion caused by a multi-source data island is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of power system simulation and automatic control, in particular to an automatic scoring method and system for a simulation operating console of a thermal power generating set. Background Art

[0002] As the intelligent transformation of thermal power generation units continues, simulation-based operation scoring systems, as core tools for personnel training and operational optimization, face higher demands for accuracy and real-time performance. Traditional scoring methods rely on manual experience to set fixed-weight rule bases, employing a single sensor data source and linear modeling techniques (such as PCA and SVD). These methods struggle to adapt to the complex demands of multi-source heterogeneous data fusion, dynamic operating condition response, and real-time decision-making at the edge.

[0003] Existing technologies use discrete sensor networking, resulting in timing deviations between physical parameters (pressure, temperature), virtual sensor data (stress field, temperature field), and operational instruction streams due to protocol heterogeneity (Modbus / OPC UA). Furthermore, when processing 32 subsystems and 128-dimensional parameters, traditional linear modeling methods ignore cross-dimensional correlations, leading to loss of effective information (such as the nonlinear coupling effect between vibration spectrum and thermal stress field). Furthermore, fixed-weight rule bases can generate scoring delays when operating conditions change suddenly (such as peak load start-stop operations), failing to meet the needs of real-time operational feedback.

[0004] What is particularly noteworthy is that existing methods lack quantitative modeling of the dynamic attenuation of equipment health and compliance with operating procedures, and are unable to reflect the real-time constraints of equipment status on operational feasibility; resulting in a high rate of missed compliance risk detection.

[0005] The above technical bottlenecks seriously restrict the training efficiency and practical application value of the thermal power generation simulation scoring system, and there is an urgent need to achieve technological breakthroughs through multi-source data fusion, high-dimensional feature decoupling, dynamic weight optimization and edge computing architecture. Summary of the Invention

[0006] In response to the shortcomings of the existing technology, the present invention provides an automated scoring method and system for a thermal power generation unit simulation console, which solves the problems of feature inaccuracy caused by multi-source data islands, loss of effective information caused by dimensional collapse of high-dimensional parameter modeling, and dynamic adaptation failure caused by response hysteresis of the fixed weight rule base in the existing thermal power generation simulation scoring.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions: an automated scoring method for a thermal power generator simulation console comprises the following steps:

[0008] S1: collects operation instruction streams and physical parameter monitoring data through multi-source sensors;

[0009] S2: Perform spatiotemporal alignment and anomaly cleaning on the collected multi-source data, and then construct a spatiotemporal four-dimensional tensor model to represent the unit's operating status. The dimensions of the spatiotemporal four-dimensional tensor model include the time axis, spatial subsystem, operating parameters, and equipment health;

[0010] S3: Construct a meta-reinforcement learning framework, use the spatiotemporal four-dimensional tensor to generate a dynamic scoring weight vector, and combine it with the attenuation factor in the device health to perform real-time scoring;

[0011] S4: Construct a multi-objective optimization algorithm with sparse constraints to optimize the scoring weight distribution strategy and output the scoring results. The multi-objective optimization algorithm includes a loss function and an iteration rule.

[0012] Preferably, in step S1:

[0013] The operation instruction stream is collected through the OPCUA protocol;

[0014] The physical parameter monitoring data is obtained through a sensor array deployed at key nodes of the unit.

[0015] Preferably, in step S2, the construction of the spatiotemporal four-dimensional tensor satisfies:

[0016]

[0017] Where T is the time axis; S is the space subsystem; P is the operating parameter; D is the equipment health;

[0018] The time axis T is replaced by sliding window filtering;

[0019] The spatial subsystem S is divided into 32 independent areas, including the boiler combustion area, the steam turbine high-pressure cylinder and the deaerator;

[0020] The operating parameters P contain 128 dimensions, of which 64 dimensions are physical sensing parameters, including pressure, temperature, and vibration acceleration; 32 dimensions are virtual sensing parameters, including stress field and temperature field simulation values; and 32 dimensions are control instruction parameters, including valve opening and fuel flow.

[0021] The generation of equipment health D is based on the historical data of the current subsystem parameters. The input data covers 64-dimensional physical parameters such as pressure, temperature, and vibration acceleration, as well as stress field simulation values.

[0022] Preferably, in step S2, after constructing the spatiotemporal four-dimensional tensor, the core feature tensor is extracted by improved Tucker decomposition. The decomposition formula is:

[0023]

[0024] Among them: U(T) 、U (S) 、U (P) 、U (D) are the factor matrices of time, space, parameters and health respectively; is the core tensor used to capture potential cross-dimensional correlations; ε is the residual tensor.

[0025] Preferably, in step S2, the core feature tensor is extracted by the improved Tucker decomposition. When the spatial factor matrix U (S) Apply L1 regularization constraint:

[0026] min||U (S) ||1;

[0027] Among them, ||U (S) ||1 represents the sum of the absolute values of all elements in the matrix.

[0028] Preferably, in step S3, the meta-reinforcement learning framework uses the core feature tensor of the spatiotemporal four-dimensional tensor as input value and generates a dynamic weight vector through a model-independent meta-learning MAML algorithm, specifically including:

[0029] Pre-trained scenario association: unit start-up and shutdown, peak load regulation operation, and combined fault handling;

[0030] Dynamic weight calculation: Based on the dimensional data of the device health, an attenuation factor is introduced to adjust the scoring weight;

[0031] Dynamic weight calculation introduces device health attenuation factor:

[0032]

[0033] Where Q(s,a) is the comprehensive score of executing action a in state s, which is used to quantify the quality of the operation behavior; γ is the health attenuation factor; w i is the scoring weight of item i; f i (s) is the scoring sub-item corresponding to state s; HealthLoss i is the health loss value of the i-th device; exp(―γ·HealthLoss i ) indicates exponential decay of the device health loss value, which is used to dynamically adjust the scoring weight.

[0034] Preferably, in step S3, after introducing the device health attenuation factor, a reward function is constructed to further implement dynamic weight calculation, and the reward function is designed as follows:

[0035] r t=0.4·Stability+0.3·Efficiency―0.2·Risk―0.1·HealthCost

[0036] in:

[0037]

[0038] λ i is the i-th eigenvalue of the system Jacobian matrix J; N is the number of subsystems; |Re(λ i )|, the larger the absolute value of the real part of the eigenvalue, the worse the subsystem stability; Efficiency is the unit thermal efficiency; Risk is the operational risk score; HealthCost is the cumulative health loss rate.

[0039] Preferably, in step S4, the loss function implements a distribution strategy for optimizing the dynamic scoring weight vector:

[0040]

[0041] Among them, Y is the expert's pre-labeled rating label; X is the tensor feature matrix; w is the weight vector to be optimized; λ is the L1 regularization coefficient; min w is the objective function.

[0042] Preferably, in step S4, after constructing the loss function, the optimal weight is iteratively solved by the proximal gradient descent algorithm, and the iteration rule is:

[0043]

[0044] Among them, η = 0.01 is the learning rate; k max =1000 is the maximum number of iterations; is the gradient of the loss function; ⊙ is the Hadamard product; sign is the sign function; w k is the weight component; λη represents the intensity of weight sparsification

[0045] Preferably, including:

[0046] The data acquisition and processing module, including the OPCUA protocol parsing unit, physical sensor interface, and digital twin virtual sensor channel, transmits the real-time collected operation instruction stream, physical parameters, and equipment health data to the spatiotemporal modeling module through the data bus;

[0047] The spatiotemporal modeling module receives data from the data acquisition and processing module through high-speed memory sharing, performs spatiotemporal tensor construction, improved Tucker decomposition and coupling strength matrix calculation, and generates core feature tensors

[0048] Dynamic scoring module, receiving core feature tensors through PCIe interface Deploy the MAML meta-reinforcement learning algorithm and multi-objective optimizer to output dynamic scoring results;

[0049] Among them, the data flow and control signals between modules are synchronized through the time-sensitive network TSN protocol.

[0050] Preferably, the system further comprises a computer-readable storage medium storing a computer program, which, when executed by the processor of the system, implements the steps of the method provided by the present invention.

[0051] The present invention provides an automated scoring method and system for a thermal power generator simulation console. It has the following beneficial effects:

[0052] 1. This invention uses TSN protocol dynamic priority allocation and multi-source data synchronization compensation technology to achieve microsecond-level spatiotemporal alignment of operation instruction streams, physical sensor parameters, and virtual data, breaking through the timing mismatch bottleneck caused by traditional heterogeneous protocols, solving the feature modeling distortion problem caused by multi-source data islands, and improving the equipment health monitoring accuracy to millimeter-level spatial resolution.

[0053] 2. Based on the four-dimensional space-time tensor and improved Tucker decomposition technology, this invention captures the nonlinear coupling correlation of 32 subsystems and 128-dimensional operating parameters, overcomes the high-dimensional information loss defect of traditional linear methods such as PCA / SVD, and effectively retains the core tensor dimension of key features.

[0054] 3. This invention adopts a dynamic weight optimization mechanism that couples the meta-reinforcement learning (MAML) framework with the equipment health attenuation factor. Under complex working conditions such as peak load regulation and start-stop, the scoring weight is adjusted in real time through an exponential decay model, thereby reducing the response delay of the scoring system and improving its adaptability to dynamic scenarios.

[0055] 4. The present invention constructs a modular hardware architecture (data acquisition / spatiotemporal modeling / dynamic scoring module) and a PCIe bus-TSN protocol collaborative transmission mechanism to support real-time processing of high-frequency data streams and millisecond-level scoring feedback, breaking through the bandwidth bottleneck of traditional edge computing architecture and providing industrial-grade real-time decision support for thermal power generation simulation operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 is a flow chart of the method of the present invention;

[0057] Figure 2 It is a system module diagram of the present invention. DETAILED DESCRIPTION

[0058] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0059] The multi-source data collection in step S1 is achieved through a hardware sensor network.

[0060] Exemplarily, the operational command stream is collected in real time via the OPCUA protocol, specifically including valve opening setpoints and fuel quantity adjustment commands output by the DCS system. The sampling rate is set to 1kHz, and the timestamp synchronization accuracy is ±1ms. Preferably, the physical parameter monitoring data is acquired via a sensor array deployed at key nodes of the unit, including a high-temperature pressure sensor installed on the main steam pipe, an infrared thermal imager in the boiler furnace, and a turbine bearing vibration sensor.

[0061] The multi-source data is clock-synchronized via a Time Sensitive Network (TSN) protocol.

[0062] Preferably, the synchronization process adopts a dynamic priority allocation algorithm, and the operation instruction stream and physical parameters are used to jointly allocate time slot resources.

[0063] An exemplary synchronous compensation formula is:

[0064]

[0065] Among them, w i is the data flow priority weight; v i is the operation instruction stream; d max is the maximum allowed delay; d current is the front link delay; t sensor is the sensor data timestamp; t sync A timeline for synchronizing multi-source data.

[0066] The data acquisition module is connected to the spatiotemporal modeling module via a PCIe bus, and the transmission bandwidth meets the real-time requirements.

[0067] During the peak-shaving operation of the unit, the feedwater valve opening instruction is first collected through the OPCUA protocol, and the main steam pressure fluctuation data is obtained by the pressure sensor. After synchronous compensation, the multi-source data can achieve accurate input for subsequent space-time tensor modeling.

[0068] This implementation solves the feature extraction distortion problem caused by multi-source data timing mismatch in traditional methods by combining hard sensing with the high-precision synchronization of the TSN protocol, and achieves millimeter-level spatial resolution monitoring of device health status.

[0069] The construction of the spatiotemporal four-dimensional tensor in step S2 includes data preprocessing, tensor dimension definition and improved Tucker decomposition.

[0070] First, clock synchronization is performed. Through the Time Sensitive Stream (TSF) mechanism of the TSN protocol, the clock deviation of the pressure sensor (Modbus TCP), vibration sensor (OPCUA), and virtual temperature field (UDP) is controlled within ±1μs. For example, the vibration sensor is used as the reference clock source, and IEEE1588 synchronization messages are periodically sent. Then, spatial coordinate unification is performed. By establishing a three-dimensional coordinate system for the unit (with the origin located at the centerline of the turbine), the measurement values of each sensor are converted to a unified coordinate system:

[0071] p global =R i ·p local +t i ;

[0072] Among them, R i is the rotation matrix of the i-th sensor; t i Translation vector; p local is the measurement value of the local coordinate system of the i-th sensor; p global is the global coordinate in the unified coordinate system.

[0073] Then, data interpolation and alignment are performed. The pressure data, vibration data, and temperature field data are unified to the 10kHz time axis using the cubic spline interpolation algorithm:

[0074]

[0075] Among them, t k is the original data timestamp; a k is the cubic spline interpolation coefficient; x align (t) is represented as a continuous signal aligned to the 10 kHz time axis.

[0076] After completing the spatiotemporal alignment of the data, the multi-source data was cleaned for anomalies, specifically by using the isolation forest algorithm to detect outliers, with the subsampling size set to 256 and the anomaly score threshold set to 0.65.

[0077] Preferably, a 3σ filter replacement with a sliding window (window size N=10) is used for the continuous parameters:

[0078]

[0079] Where μ is the window mean; σ is the standard deviation; N=10 is the sliding window size; x t is the original data at the current moment; 3σ is the abnormality judgment threshold.

[0080] Then, a four-dimensional space-time tensor is constructed. For example, the time axis T is defined as a 300-second sliding window; the space axis S is divided into 32 subsystems and uses one-hot encoding; the parameter axis P contains 128 operating parameters, preferably including the derived indicator combustion efficiency:

[0081] η=Q output / Q input (Q output calculated by the heat balance equation);

[0082] Among them, Q output The effective heat output value of the unit; Q input Enter the calorific value for the fuel;

[0083] The health axis D contains 40-dimensional indicators, such as the thermal stress of the water-cooled wall tube σ thermal And rotor creep life loss rate:

[0084]

[0085] Where, σ(τ) is the real-time thermal stress of the rotor; ∈ c is the cumulative life loss rate;

[0086] Furthermore, the core feature tensor is extracted by the improved Tucker decomposition The decomposition formula is:

[0087]

[0088] Among them, U (T) 、U (S) 、U (P) 、U (D) is a factor matrix; is the core tensor; ε is the residual term. Preferably, the spatial factor matrix U (S) Apply L1 regularization constraint:

[0089] min||U (S) ||1;

[0090] To reduce the redundant associations between subsystems. The decomposition process is solved iteratively by the alternating least squares (ALS) method. The spatiotemporal modeling module and the dynamic scoring module are connected through a shared memory to transmit the core feature tensor.

[0091] Furthermore, the coupling strength matrix C of the operation instruction to the device health is calculated:

[0092]

[0093] Among them, the partial derivatives are calculated by the finite difference method; the time delay TimeDelayij Determined by the physical conduction model; HealthLoss i Operation is a comprehensive health loss indicator; j is the jth operation instruction; TimeDelay ij is the delay of operation instruction j to subsystem i.

[0094] This implementation utilizes the sparse constraints of the modified Tucker decomposition and coupling strength quantification to extract key correlation features from high-dimensional data, addressing the degradation of modeling accuracy caused by dimensionality explosion in traditional methods. The physical dimensional definition of the space-time tensor is strictly aligned with the device topology, achieving millimeter-level spatial resolution modeling.

[0095] The meta-reinforcement learning framework of step S3 is deployed on an edge computing unit.

[0096] First, a pre-trained scene library is loaded. Exemplarily, the framework uses the Model-Agnostic Meta-Learning (MAML) algorithm. Preferably, the dynamic weight generation is implemented via an Actor-Critic network, where the Actor network structure includes three convolutional layers and an LSTM layer, and the Critic network is a fully connected structure.

[0097] The scoring weight vector w t The calculation formula is:

[0098]

[0099] Among them, s t Core features extracted from spatiotemporal tensors The flattened vector dimensions (64×16×32×8); the temperature parameter 0.5 controls the sharpness of the weight distribution; ActorNet(s t ) is the original score weight output by the Actor network.

[0100] Furthermore, the Q value combined with the device health attenuation factor is calculated as follows:

[0101]

[0102] Among them, f i (s) is the scoring sub-item corresponding to state s (such as combustion efficiency deviation, steam pressure fluctuation rate); HealthLoss i is the health loss value of the i-th equipment (given by the formula calculation); w i is the scoring weight of item i.

[0103] The reward function is designed as:

[0104] rt =0.4·Stability+0.3·Efficiency―0.2·Risk―0.1·HealthCost

[0105] in:

[0106]

[0107] λ i is the maximum real eigenvalue of the Jacobian matrix of subsystem i, which is updated by real-time sensor data; Stability is the stability index; |Re(λ i The larger the absolute value of the real part of the eigenvalue of )|, the worse the stability of the subsystem; the health cost HealthCost is defined as the ratio of the accumulated damage to the design life.

[0108] The dynamic scoring module receives the core tensor output by the spatiotemporal modeling module through the PCIe interface The scoring weights are calculated in real time and transmitted to the feedback execution module. For example, when a water pump failure is detected, the framework increases the cooling system health weight while decreasing the combustion efficiency weight.

[0109] This implementation achieves dynamic optimization of scoring weights under complex operating conditions by coupling the rapid adaptation mechanism of meta-reinforcement learning with the calculation of the equipment health attenuation factor. The multi-objective balance design of the reward function has a synergistic control effect of suppressing operational risk and health loss.

[0110] The multi-objective optimization algorithm of step S4 is deployed in the dynamic scoring module.

[0111] First, we receive the initial weight vector from the meta-reinforcement learning framework, and then construct a loss function with L1 regularization:

[0112]

[0113] Among them, Y is the expert-annotated score label (normalized from 0 to 100 points); X is the historical core feature tensor The expanded matrix of , w is the weight vector to be optimized.

[0114] Then, a proximal gradient descent algorithm is constructed in the optimization process, and the iterative update rule is:

[0115]

[0116] in, is the loss function gradient; learning rate η = 0.01; maximum number of iterations k max =1000; ⊙ is the Hadamard product; sign is the sign function; w kis the weight component. Preferably, the convergence condition is set to the adjacent iteration weight change || w k+1 ―w k ||2<10 ―4 .

[0117] The dynamic scoring module receives the real-time feature tensor output by the spatiotemporal modeling module through the shared memory interface And calculate the prediction score:

[0118]

[0119] in, is the value of the i-th feature at time t, w i is the dynamic weight of the i-th feature.

[0120] Furthermore, the scoring result is transmitted to the feedback execution module via the time-sensitive network (TSN) protocol.

[0121] This implementation achieves a balance between sparsification of scoring weights and error minimization through the coordinated optimization of L1 regularization constraints and proximal gradient descent. The real-time scoring calculation and transmission mechanism has millisecond-level response capabilities, addressing the overfitting problem caused by redundant weights in traditional methods.

[0122] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An automated scoring method for a thermal power generator simulation console, characterized in that: The following steps are involved: S1: collects operation instruction streams and physical parameter monitoring data through multi-source sensors; S2: Perform spatiotemporal alignment and anomaly cleaning on the collected multi-source data, and then construct a spatiotemporal four-dimensional tensor model to represent the unit's operating status. The dimensions of the spatiotemporal four-dimensional tensor model include the time axis, spatial subsystem, operating parameters, and equipment health; S3: Construct a meta-reinforcement learning framework, use the spatiotemporal four-dimensional tensor to generate a dynamic scoring weight vector, and combine it with the attenuation factor in the device health to perform real-time scoring; S4: Construct a multi-objective optimization algorithm with sparse constraints to optimize the scoring weight distribution strategy and output the scoring results. The multi-objective optimization algorithm includes a loss function and an iteration rule.

2. The automated scoring method for the thermal power generation unit simulation console according to claim 1, characterized in that: In the step S1: The operation instruction stream is collected through the OPCUA protocol; The physical parameter monitoring data is obtained through a sensor array deployed at key nodes of the unit.

3. The automated scoring method for a thermal power generator simulation console according to claim 1, characterized in that: In step S2, the construction of the spatiotemporal four-dimensional tensor satisfies: Where T is the time axis; S is the space subsystem; P is the operating parameter; D is the equipment health; The time axis T is replaced by sliding window filtering; The spatial subsystem S is divided into 32 independent areas, including the boiler combustion area, the steam turbine high-pressure cylinder and the deaerator; The operating parameters P contain 128 dimensions, of which 64 dimensions are physical sensing parameters, including pressure, temperature, and vibration acceleration; 32 dimensions are virtual sensing parameters, including stress field and temperature field simulation values; and 32 dimensions are control instruction parameters, including valve opening and fuel flow. The generation of equipment health D is based on the historical data of the current subsystem parameters. The input data covers 64-dimensional physical parameters such as pressure, temperature, and vibration acceleration, as well as stress field simulation values.

4. The automated scoring method for a thermal power generator simulation console according to claim 1, characterized in that: In step S2, after constructing the spatiotemporal four-dimensional tensor, the core feature tensor is extracted by improved Tucker decomposition. The decomposition formula is: Among them: U (T) 、U (S) 、U (P) 、U (D) are the factor matrices of time, space, parameters and health respectively; is the core tensor used to capture potential cross-dimensional correlations; ε is the residual tensor.

5. The automated scoring method for a thermal power generator simulation console according to claim 1, characterized in that: In step S2, the core feature tensor is extracted by the improved Tucker decomposition. When the spatial factor matrix U (S) Apply L1 regularization constraint min||U (S) ||1; Among them, ||U (S) ||1 represents the sum of the absolute values of all elements in the matrix.

6. The automated scoring method for a thermal power generator simulation console according to claim 1, characterized in that: In step S3, the meta-reinforcement learning framework uses the core feature tensor of the spatiotemporal four-dimensional tensor as input and generates a dynamic weight vector through the model-independent meta-learning MAML algorithm, specifically including: Pre-trained scenario association: unit start-up and shutdown, peak load regulation operation, and combined fault handling; Dynamic weight calculation: Based on the dimensional data of the device health, an attenuation factor is introduced to adjust the scoring weight; Dynamic weight calculation introduces device health attenuation factor: Where Q(s,a) is the comprehensive score of executing action a in state s, which is used to quantify the quality of the operation behavior; γ is the health attenuation factor; w i is the scoring weight of item i; f i (s) is the scoring sub-item corresponding to state s; HealthLoss i is the health loss value of the i-th device; exp(―γ·HealthLoss i ) indicates exponential decay of the device health loss value, which is used to dynamically adjust the scoring weight.

7. The automated scoring method for a thermal power generator simulation console according to claim 1, characterized in that: In step S3, after introducing the device health attenuation factor, a reward function is constructed to further implement dynamic weight calculation. The reward function is designed as follows: r t =0.4·Stability+0.3·Efficiency―0.2·Risk―0.1·HealthCost in: λ i is the i-th eigenvalue of the system Jacobian matrix J; N is the number of subsystems; |Re(λ i )|, the larger the absolute value of the real part of the eigenvalue, the worse the subsystem stability; Efficiency is the unit thermal efficiency; Risk is the operational risk score; HealthCost is the cumulative health loss rate.

8. The automated scoring method for a thermal power generator simulation console according to claim 1, characterized in that: In step S4, the loss function implements the allocation strategy for optimizing the dynamic scoring weight vector: Among them, Y is the expert's pre-labeled rating label; X is the tensor feature matrix; w is the weight vector to be optimized; λ is the L1 regularization coefficient; min w is the objective function.

9. The automated scoring method for a thermal power generator simulation console according to claim 1, characterized in that: In step S4, after constructing the loss function, the optimal weight is iteratively solved by the proximal gradient descent algorithm, and the iteration rule is: Among them, η = 0.01 is the learning rate; k max =1000 is the maximum number of iterations; is the gradient of the loss function; ⊙ is the Hadamard product; sign is the sign function; w k is the weight component; λη represents the intensity of weight sparsification.

10. An automated scoring system for a thermal power generation unit simulation console, applied to the automated scoring method for a thermal power generation unit simulation console according to any one of claims 1 to 8, characterized in that: include: The data acquisition and processing module, including the OPCUA protocol parsing unit, physical sensor interface, and digital twin virtual sensor channel, transmits the real-time collected operation instruction stream, physical parameters, and equipment health data to the spatiotemporal modeling module through the data bus; The spatiotemporal modeling module receives data from the data acquisition and processing module through high-speed memory sharing, performs spatiotemporal tensor construction, improved Tucker decomposition and coupling strength matrix calculation, and generates core feature tensors Dynamic scoring module, receiving core feature tensors through PCIe interface Deploy the MAML meta-reinforcement learning algorithm and multi-objective optimizer to output dynamic scoring results; Among them, the data flow and control signals between modules are synchronized through the time-sensitive network TSN protocol.

Images