Biomass engine data coordination method and system

Abstract

This invention relates to the field of biomass gas turbine system technology, and discloses a biomass gas turbine data coordination method and system. The method includes: constructing a dynamic mathematical model to obtain the prior predicted values ​​of each measuring point at the current time; abstracting components, measuring points or state points and their connections into graph nodes and graph edges to construct a graph structure; using the prior predicted values, the real-time measured values ​​collected by sensors, and the residuals between the measured values ​​and the prior predicted values ​​as the initial features of each node to form a node feature matrix; training a graph convolutional network through the node feature matrix and supervised labels to enable it to learn the mapping law from prior predictions and real-time measurements to the coordinated state; inputting the real-time measured values ​​collected by sensors and the prior predicted values ​​provided by the dynamic mathematical model into the graph convolutional network, and directly outputting the corrected coordinated state parameters through one forward propagation calculation of the graph convolutional network, which can be used for the control or monitoring of biomass gas turbines and is suitable for complex nonlinear and time-varying systems.

Description

Technical Field

[0001] This invention relates to the field of biomass gas turbine system technology, and in particular to a biomass gas turbine data coordination method and system. Background Technology

[0002] Biomass gas turbines are a type of power and power generation equipment that uses biomass such as agricultural and forestry waste, straw, and sawdust as fuel, converting its chemical energy into thermal energy and further into mechanical and electrical energy. Common forms include biomass gas internal combustion engines, gas turbines, and combined systems coupled with waste heat recovery. Compared to fossil fuel units, biomass gas turbines are characterized by dispersed fuel sources, large fluctuations in composition, and significant variations in calorific value and moisture content, resulting in stronger nonlinearity and time-varying characteristics in various processes. To achieve high efficiency, low emissions, and stable grid-connected operation, the system typically deploys numerous sensors and actuators to collect and regulate multi-source data such as speed, torque, pressure, exhaust temperature, and vibration in real time, thus placing higher demands on dynamic state estimation and data consistency.

[0003] During the dynamic operation of biomass gas turbines (start-up, shutdown, load surges, grid disturbances, etc.), if key measurement data drifts, is delayed, missing, or inconsistent between different measurement points, it can easily cause and amplify deviations in control quantity calculations. For example, deviations in exhaust temperature and boost pressure estimation can lead to overheating, surge risks, and reduced efficiency; torque and speed measurement errors can cause fluctuations in speed regulation and grid-connected power, affecting frequency and power quality; inaccurate health monitoring data such as vibration and bearing temperature can lead to false alarms or missed alarms, delaying fault handling. In more serious cases, errors in multivariate coupling links can propagate between nodes, causing model predictions to deviate from reality, resulting in energy distribution imbalances, increased emissions, increased component heat loads, and even triggering protection actions and unplanned shutdowns, leading to economic losses and safety hazards.

[0004] To address the aforementioned issues, current research primarily employs two types of methods: dynamic data coordination and state estimation algorithms, and purely data-driven machine learning methods. The former relies on a system mechanism model combined with noise assumptions to recursively estimate the system state; the latter directly utilizes operational data to establish a mapping relationship between input and state. However, in scenarios like biomass gas turbines—characterized by strong nonlinearity, strong coupling, non-Gaussian noise, and time-varying parameters—traditional dynamic data coordination and state estimation methods often face challenges such as heavy modeling and parameter tuning burdens, sensitivity to noise statistical assumptions and linearization accuracy, and susceptibility to outliers and sensor malfunctions. Furthermore, as the number of measurement points increases, matrix operations and iterative updates in high-dimensional states significantly increase real-time pressure, manifesting as high computational overhead, slow convergence, or tracking lag under rapidly changing operating conditions. In contrast, purely machine learning methods, lacking physical mechanism constraints, suffer from insufficient generalization ability. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a biomass gas turbine data coordination method and system, which not only reduces the reliance on linearization and high-dimensional matrix operations, but also significantly improves real-time performance and computational efficiency, making it suitable for complex nonlinear and time-varying systems.

[0006] The first aspect of this invention provides a biomass gas turbine data coordination method, comprising the following steps: Graph Construction: A dynamic mathematical model is constructed based on the thermodynamic mechanism of the biomass gas turbine. The prior predicted values ​​of each measuring point at the current moment are obtained based on the model. The components, measuring points, or state points in the biomass gas turbine are abstracted as graph nodes, and the connections and coupling relationships between graph nodes or the internal thermodynamic processes of the equipment are abstracted as graph edges to construct a graph structure that reflects the system topology. The prior predicted values, the real-time measured values ​​collected by the sensors, and the residuals between the measured values ​​and the prior predicted values ​​are used together as the initial features of each node to form a node feature matrix. Offline training: The historical running data is processed using a pre-defined recursive state estimation algorithm to generate the coordinated state estimates at each time point as supervision labels; the graph convolutional network is trained with the node feature matrix as input and the supervision labels as expected output, so that the graph convolutional network learns the mapping rules from prior prediction and real-time measurement to the coordinated state. Online application: The measured values ​​collected in real time by the sensor and the prior predicted values ​​provided by the dynamic mathematical model are input into the trained graph convolutional network. Through one forward propagation calculation of the graph convolutional network, the corrected coordinated state parameters are directly output for the control or monitoring of biomass gas turbines.

[0007] Optionally, the dynamic mathematical model can be represented as a discrete-time dynamic state-space model: , , in, For the first Discrete-time index of each sampling time; These are the system state vector and the dynamic parameter vector to be coordinated; To control the input vector; The sensor measurement vector; These are state transition functions, constructed based on thermodynamic or fluid dynamic mechanisms. For measurement functions; This is process noise; For measuring noise; The prior predictions before the current time step, without fusion, are obtained from the discrete-time dynamic state-space model: , in, For a moment Prediction of prior states, For a moment -1 is the posterior estimate.

[0008] Optionally, the graph structure can be represented by an adjacency matrix, which is constructed as follows: Construction Graph: , in, Representation diagram; For a set of nodes, Represents the set of edges; Construct the corresponding adjacency matrix: , , in, For the first 1 node For the number of nodes, Represents a node With nodes There is a connection or coupling; To enhance the retention of node information, a self-loop is introduced: , in, for An identity matrix of order 1.

[0009] Optionally, the node feature matrix can be constructed as follows: At every moment For each node Construct node feature vectors: , Stack all node features row-wise to obtain the node feature matrix: , in, For nodes At any moment eigenvectors; The feature dimension for each node is independent of the degree matrix; For a moment The node feature matrix includes: node measured values, prior predicted values, residuals between measured values ​​and prior predicted values, global operating condition features of operating conditions and control variables, and coordination results from the previous time step. Normalize or standardize various features to obtain the normalized node feature matrix. or As input to a graph convolutional network.

[0010] Optionally, the recursive state estimation algorithm includes one of the following: Kalman filter algorithm, information filter algorithm, or particle filter algorithm; When the recursive state estimation algorithm is the unscented Kalman filter algorithm in the Kalman filter class, Sigma points are constructed through unscented transformation, and prior statistics are calculated after propagation through the dynamic mathematical model. Then, the measurement is updated by combining real-time measurements to obtain the coordinated posterior estimate as the supervision label.

[0011] Optionally, the graph convolutional network adopts a multi-layer structure, and its propagation method is as follows: Define the normalized graph convolution operator: , No. layer, The spread is as follows: , The output layer obtains the coordination output matrix: , in, For the first Layer hidden representation matrix, This layer hides the dimension; For the first Layer trainable weight matrix; For activation functions; for The corresponding degree matrix, ; To coordinate the output matrix; The number of physical quantities / parameters output for each node; The coordination output matrix is ​​restored to the system's global coordination state vector according to a preset mapping relationship: , in, This is a function for assembling / mapping node outputs to the global state vector.

[0012] Optionally, in offline training, mean squared error can be used as the loss function: , in, Number the nodes. Output component numbers for the nodes. For a moment node middle The predicted value of the component. For a moment node middle The true value of the component; When there are missing or invalid labels, the output mask matrix is ​​introduced into the loss calculation: , in, Indicates time No. In each node The output mask of the component. Indicates time The number of valid labels used in training.

[0013] Optionally, during offline training, a physical consistency constraint regularization term is also introduced to ensure that the coordination results meet preset physical constraints. The comprehensive loss function is: , in, These are weighting coefficients used to balance data supervision terms and physical constraint terms; This is a penalty term for physical consistency loss or constraint. This is the weighted mean squared error loss function.

[0014] Optionally, a single forward propagation computation can replace the iterative update process of the traditional recursive state estimation algorithm, enabling real-time data coordination.

[0015] A second aspect of the present invention provides a biomass gas turbine data coordination system based on graph neural networks for implementing the above-described method. The system includes: The graph construction module is used to construct a dynamic mathematical model based on the thermodynamic mechanism of the biomass gas turbine, and obtain the prior predicted values ​​of each measuring point at the current moment based on the model. It abstracts the components, measuring points or state points in the biomass gas turbine into graph nodes, and the connections and coupling relationships between graph nodes or the thermodynamic processes inside the equipment into graph edges to construct a graph structure that reflects the topology of the system. The prior predicted values, the real-time measured values ​​collected by the sensors, and the residuals between the measured values ​​and the prior predicted values ​​are used together as the initial features of each node to form a node feature matrix. The offline training module is used to process historical running data using a preset recursive state estimation algorithm to generate coordinated state estimates at each time point as supervision labels. The graph convolutional network is trained with the node feature matrix as input and the supervision labels as expected output, so that the graph convolutional network learns the mapping law from prior prediction and real-time measurement to the coordinated state. The online application module is used to input the measured values ​​collected by the sensor in real time and the prior predicted values ​​provided by the dynamic mathematical model into the trained graph convolutional network. Through one forward propagation calculation of the graph convolutional network, the corrected coordinated state parameters are directly output for the control or monitoring of biomass gas turbines. Attached Figure Description

[0016] Figure 1 A flowchart of a biomass gas turbine data coordination method and system provided in an embodiment of the present invention; Figure 2 A diagram of a biomass gas turbine system with a regenerative cycle provided in an embodiment of the present invention; Figure 3 The adjacency matrix and degree matrix of the biomass gas turbine system with regenerative cycle provided in the embodiments of the present invention are constructed.

[0017] Explanation of reference numerals in the attached figures: 1. First state point; 2. Second state point; 3. Third state point; 4. Fourth state point; 5. Fifth state point; 6. Sixth state point; 7. Seventh state point; 8. Eighth state point; 9. Compressor; 10. Regenerator; 11. Biomass boiler; 12. Turbine; 13. Generator; 14. Evaporator; 15. Booster pump; 16. Steam; 17. Water; 18. Heating heat exchanger; 19. Drain; 20. Return. Detailed Implementation

[0018] The following detailed description of a specific embodiment of the present invention is provided in conjunction with the accompanying drawings. However, it should be understood that the scope of protection of the present invention is not limited to the specific embodiment.

[0019] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the technical solution of this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0020] The present invention will be described below through several specific embodiments.

[0021] This invention provides a method and system for coordinating data in biomass gas turbines. This method is applicable to biomass gas turbine systems with different structural forms (e.g., whether they are regenerative, whether they recover waste heat for heating / steam generation, whether they are multi-stage compression / expansion, and different numbers and layouts of measuring points). Its core is: establishing a dynamic system model to obtain prior predictions, with each state variable as a node, and each state variable corresponding to several sensors. State variables are connected through dynamic relationships between devices, corresponding to "edges," abstracting the system into a graph structure to generate an adjacency matrix; constructing node feature matrices and inputting them into a graph convolutional network, outputting the coordinated dynamic parameters of each node; during the training phase, Kalman filtering and other methods can be used to generate supervisory labels, enabling the graph convolutional network to learn and replace the traditional recursive update process (a complete filtering process needs to be run at each time step, suitable for large-scale training data. Sampling is performed several times using the filtering process, and after collecting large-scale training data, it is fed into the neural network for training). It mainly consists of a data preparation and graph construction module, a network training module, and a rapid coordination module.

[0022] like Figure 1 As shown, the first part of the embodiments of the present invention provides a biomass gas turbine data coordination method based on graph neural networks, including the following steps: Graph Construction: A dynamic mathematical model is constructed based on the thermodynamic mechanism of the biomass gas turbine, and the prior predicted values ​​of each measuring point at the current moment are obtained based on the model. In this embodiment of the invention, the measured values, the prior predicted values ​​of the dynamic mathematical model, and their residuals are used together as initial features of the nodes, enabling the graph convolutional network to intelligently identify the source of data deviation: for faulty sensors, the network corrects based on the prior information of the physical model and adjacent nodes; for model mismatch, dynamic calibration is performed based on real-time measurements. This dual-source input mechanism ensures that the coordinated state parameters remain highly accurate and consistent even under strong noise and non-Gaussian interference, avoiding missed detections or unplanned shutdowns caused by false alarms due to vibration or bearing temperature. This invention abstracts components, measuring points, or state points in a biomass gas turbine as graph nodes, and the connections and coupling relationships between these nodes, or the internal thermodynamic processes, as graph edges to construct a graph structure reflecting the system topology. Prior predictions, real-time sensor measurements, and the residuals between these measurements and prior predictions are used as initial features for each node, forming a node feature matrix. This embodiment of the invention introduces a graph convolutional network to encode the complex nonlinear thermodynamic processes and time-varying characteristics of the biomass gas turbine into the learning process of the graph structure. Compared to traditional dynamic data coordination methods that rely on linear approximations (such as extended Kalman filtering) leading to model mismatch, the graph convolutional network, through end-to-end supervised learning, directly fits the highly nonlinear mapping relationship from "prior prediction + real-time measurement" to "coordinated state." It can adapt to parameter drift caused by fuel composition fluctuations and calorific value changes, solving the risks of overheating, surge, and control deviations caused by model inaccuracies. Traditional methods often suffer from amplified errors propagating between nodes when dealing with multivariate coupling (such as the mutual influence between exhaust temperature, pressure, and speed). This invention abstracts the system into a graph structure reflecting the physical topology, defining components and measurement points as nodes and connections / coupling relationships as edges. The graph convolutional network explicitly learns the higher-order interactive effects across measurement points through information propagation and feature aggregation between nodes, achieving global optimal coordination rather than local correction. This effectively suppresses speed fluctuations and grid-connected power oscillations caused by torque-speed measurement errors, ensuring frequency and power quality.

[0023] Offline Training: A pre-defined recursive state estimation algorithm is used to process historical operating data, generating coordinated state estimates at each time point as supervision labels. Using the node feature matrix as input and the supervision labels as the expected output, the graph convolutional network is trained, enabling it to learn the mapping from prior predictions and real-time measurements to the coordinated state. Traditional state estimation methods require precise noise statistics assumptions and tedious manual parameter tuning, placing high demands on on-site technicians. This invention employs an offline training mode, using historical operating data to generate supervision labels, allowing the graph convolutional network to autonomously learn error distribution patterns and coordination logic. Once training is complete, online applications do not require maintaining a noise model or repeatedly adjusting parameters, significantly reducing the workload of on-site debugging and maintenance. This solves the real-time pressure and convergence problems caused by increased measurement points and high-dimensional states, improving the system's engineering deployability.

[0024] Online Application: Real-time sensor measurements and prior predictions from a dynamic mathematical model are input into a trained graph convolutional network. Through a single forward propagation calculation, the corrected coordinated state parameters are directly output for the control or monitoring of biomass gas turbines. Traditional dynamic data coordination requires high-dimensional matrix operations and iterative updates, facing real-time bottlenecks in biomass gas turbine scenarios with multiple measurement points and high sampling rates. This invention utilizes the parallel computing capabilities of graph convolutional networks to simplify the complex filtering and update process into a single forward propagation calculation, significantly reducing online computational overhead. Even under rapid operating conditions such as start-up, shutdown, load surges, and grid disturbances, millisecond-level response and tracking can be achieved, eliminating the lag problem of traditional recursive algorithms and providing reliable data support for closed-loop control.

[0025] Pure machine learning methods lack physical constraints and are prone to failure when extrapolating operating conditions. This invention injects physical mechanisms in two ways: first, it uses a dynamic mathematical model to generate prior predictions, which serve as input features to guide the network's learning direction; second, it introduces physical constraints (such as energy conservation and mass conservation) during training, embedding thermodynamic laws into the loss function. Through this fusion strategy of physical mechanisms and data-driven approaches, the network can still output consistent results that conform to physical laws even when encountering unexperienced or extreme operating conditions, significantly enhancing the model's generalization ability and engineering applicability.

[0026] Specifically, step 1: Establish the system dynamic model and measurement model. For the coordinated biomass gas turbine system, a dynamic mathematical model is constructed, which is represented as a discrete-time dynamic state-space model: , , in, For the first Discrete-time index of each sampling time; Let be the system state vector and the dynamic parameter vector to be coordinated. For having state dimensions real numbers, It is a state dimension, including any dynamic variables that need to be coordinated, such as temperature, pressure, flow rate, power, efficiency, and heat transfer coefficient; To control the input vector, For input dimensions, this may include fuel commands, speed / load commands, guide vane or valve opening; For sensor measurement vectors, For having measurement dimensions real numbers, The number of dimensions to be measured; The state transition function is constructed based on thermodynamic or fluid dynamics mechanisms and outputs the state prediction for the next time step. For measurement functions (nonlinear mappings from state space to measurement space); Process noise is determined based on the specific system and can be obtained in several ways: noise assumptions based on physical models; process noise estimation through experimental data; estimation through system identification methods; and theoretical assumptions and hypothesis verification, which are noise estimates and assumptions based on human experience. To measure noise, the measurement noise is determined based on the sensor. Different sensors have different uncertainties, and sensors produced by different manufacturers also have different uncertainties, but there will be a manufacturer's mark on the sensor when it leaves the factory. The prior prediction value (unfused time) before the measurement at the current time is obtained from the discrete-time dynamic state-space model. Pre-measurement prediction): , in, For a moment Prediction of prior states, For a moment -1 is the posterior estimate.

[0027] Step 2: Abstract the system topology and coupling relationships into a graph structure and generate an adjacency matrix. By abstracting components, measuring points, or state points in the system as graph nodes, and system connections, coupling relationships, or internal thermal processes as graph edges, a graph is obtained: , And construct the corresponding adjacency matrix: , , in, Representation diagram; For a set of nodes, For the first 1 node The number of nodes; Denotes the set of edges. Represents a node With nodes There is a connection or coupling; Representation and diagram The corresponding adjacency matrix, if Unauthorized map ,like Weighted diagram The weights are used to represent coupling strength or confidence level, etc. The edge weights are used to characterize the coupling strength or information confidence level between nodes. Their values ​​can be determined based on at least one of the system model parameters, node state similarity, statistical correlation or noise characteristics, and can be normalized as needed.

[0028] Definition diagram Degree matrix: , , in, For nodes The degree.

[0029] To enhance the preservation of node information, self-loops can be introduced and used for subsequent graph convolution normalization: , in, for An identity matrix of order 1.

[0030] The above graph abstraction and adjacency matrix construction rules are independent of the specific system configuration: when new components, loops and heat exchangers are added to the system, it only manifests as an increase in nodes and edges or a change in weights, and the method can still be uniformly applied.

[0031] Step 3: Construct the node feature matrix and form the input of the graph neural network. At every moment For each node Construct node feature vectors: , Stack all node features row-wise to obtain the node feature matrix: , in, For nodes At any moment eigenvectors; The feature dimension for each node is independent of the degree matrix; For a moment The node feature matrix.

[0032] Node features include: node measurements, prior predictions, residuals (measurement-prediction), global operating condition features of operating conditions and control variables, and coordination results from the previous time step. These node features are organized after time alignment and concatenated to form a node feature vector, which serves as the input to the coordination model. To avoid the impact of features with different dimensions on model performance, various features can be normalized or standardized before input based on their statistical characteristics or physical range. The specific method can be set according to the actual application requirements. When measurements are lacking, a mask matrix can be introduced. Mark the locations of valid features. During model training, the mask is not only used to distinguish between valid and missing features, but also to mask missing locations during the loss function calculation stage, accumulating only the prediction error corresponding to valid measurements, thereby avoiding interference from missing data in model training.

[0033] Network input can be represented as: or .

[0034] Recursive state estimation algorithms include one of the following: Kalman filter algorithms, information filter algorithms, or particle filter algorithms; When the recursive state estimation algorithm is the unscented Kalman filter algorithm in the Kalman filter class, Sigma points are constructed through unscented transformation, and prior statistics are calculated after propagation through the dynamic mathematical model. Then, the measurement is updated by combining real-time measurements to obtain the coordinated posterior estimate as the supervision label.

[0035] Specifically, step 4 involves using a graph convolutional network to output the coordinated dynamic parameters. To fully utilize the topological connections and coupling information among multiple measurement points and achieve fast and robust coordination of dynamic parameters under conditions such as missing measurements, noise disturbances, and time-varying parameters, a graph convolutional neural network is used to predict the coordinated dynamic parameters. Specifically, each measurement point is treated as a node in a graph. Connections between nodes are constructed based on physical connectivity or statistical correlation. Using the observation and auxiliary information of each node as input, the network performs neighborhood information aggregation and feature extraction through graph convolution, outputting the coordinated dynamic parameter results. The implementation method is as follows:

[0036] use For a layered graph convolutional network, let: , No. layer( The spread is as follows: , The output layer yields: , in, For the first Layer hidden representation matrix, This layer hides the dimension; For the first Layer trainable weight matrix; For activation functions, GELU, ReLU, etc. can be used; for The corresponding degree matrix, ; To coordinate the output matrix; The number of physical quantities / parameters output for each node (e.g., output temperature, pressure, etc.) ).

[0037] Will Restore the system to its coordinated state according to the preset mapping relationship: , in, This is a function for assembling / mapping node outputs to the global state vector.

[0038] 5. Training data acquisition and supervised learning: Generate supervision signals using filtering or coordination algorithms. During the training and online phases, a recursive state estimator can be used to perform fusion estimation of the dynamic system to obtain a coordinated reference value. Recursive state estimators can be Kalman filter type, information filter type, particle filter type, etc. Their common characteristic is that they construct residuals using prior predictions from the dynamic model and real-time measurements, then recursively update the residuals to output a posterior estimate. And its uncertainty characterization.

[0039] In this type of implementation, the recursive estimator satisfies the following general update structure: Prior prediction: , Residual: , , renew: , in, The specific calculation method for updating the gain or equivalent weight matrix / operator varies depending on the selected filtering and estimation algorithm (e.g., obtained through covariance recursion, information matrix recursion, sampling weight update, or optimization solution). The tag matrix is ​​obtained by mapping nodes. .

[0040] Mean squared error (MSE) can be used as the loss function: , in, Number the nodes. Output component numbers for the nodes. Refers to time node middle The predicted value of the component. Refers to time node middle The actual value of the component.

[0041] For a training sequence of length The data can be averaged over time: , in, For a moment The sample loss; The total loss for the sequence (or batch); The length of the training sequence.

[0042] When certain nodes or certain physical quantities are at time... When no valid measurement exists or high-confidence labels cannot be generated, an output mask matrix is ​​introduced. This mask matrix is ​​defined as an N×P matrix containing only 0s and 1s. A 1 indicates that a valid label exists at that position and participates in training, while a 0 indicates that the position is ignored. This allows the loss function to be defined using the mask matrix, such as... : , in This indicates that a valid label exists at this location and is used in training. This indicates that the position should be ignored. Therefore, the mask MSE is:

[0043] , Sequence loss can also be averaged over time: , in, This is the output mask matrix used to handle missing measurements, bad pixels, or unlabeled locations. Refers to time No. In each node Output mask for the component; For a moment The number of valid labels used in training.

[0044] To enhance robustness to outliers, the MSE loss function can be replaced with any robust loss function. Taking the Huber loss function as an example, its definition is as follows:

[0045] , , The robust mask loss can then be written as: , in, The error between the prediction and the label; Huber loss function; , where is the Huber threshold parameter. The standard deviation of the estimation error is given for each output channel. ,Pick .in, Dimensionless coefficients are commonly used. .

[0046] To ensure that the coordination results satisfy the basic physical constraints, a regularization term can be added to the supervision loss: , in, These are weighting coefficients used to balance the supervision and physical constraint terms. This applies when both losses are averaged and the physical residuals are normalized. Usually taken If the physics term is used as a weak regularization, take 0.01~0.1; if data fitting and physical constraints are equally important, take 0.1~1; if stronger physical consistency is needed, take 1~10. If the two terms differ significantly in magnitude, the values ​​can be adjusted according to the initial training phase. set up Make the two contributions of the same magnitude; This can be a physical consistency loss or constraint penalty term, used to constrain things like temperature / pressure range, monotonicity, and energy conservation residuals. This is the weighted mean squared error loss function.

[0047] At the same time, set the maximum number of steps or the maximum number of sampling points. The value is set based on experience. When the condition is met... When the current iteration / sequence processing has reached the preset termination condition, the current training loop ends or the prediction / coordination process for the current data segment ends.

[0048] Optionally, a single forward propagation computation can replace the iterative update process of the traditional recursive state estimation algorithm, enabling real-time data coordination.

[0049] Step 6: Online Coordination and Generalization Instructions When running online, for any biomass gas turbine system, only: (1) Based on the system topology construction (or ); (2) From the model Obtain prior predictions and construct them with real-time measurements ; (3) Input network to obtain and As a result of coordination.

[0050] Since this process relies on a unified expression of "graph topology + node features", changes in the system structure will only change... , and The construction rules do not change the essence of the method, so they are applicable to different types of biomass gas turbine systems.

[0051] This invention provides a data coordination method for biomass gas turbines. By introducing a graph convolutional network (GCNN), the system is abstracted into a graph structure, replacing the traditional filtering and update process with information propagation and feature aggregation between nodes. The GCNN can directly capture the nonlinear and coupling relationships between nodes in the graph structure, improving the estimation accuracy and real-time performance of the system state. Based on the actual physical structure of the system, the state variable is defined as a node in the graph, and the edges between nodes are constructed according to physical connections. This mapping tightly integrates the physical properties of the system with the network structure, effectively preserving the physical laws and dynamic characteristics of the system. Compared to the high-dimensional matrix operations and repeated iterations required by traditional filtering methods, the GCNN significantly reduces computational overhead through parallel computation of graph convolution. By optimizing the information transmission process, the GCNN can respond rapidly under changing operating conditions and update the system state in real time. By introducing physical constraints into the training process of the GCNN, combined with dynamic mathematical model constraints and data-driven learning capabilities, the robustness of the model is further improved, especially under the influence of noise, missing data, and outliers, maintaining high accuracy and stability.

[0052] The second part of this invention provides a biomass gas turbine data coordination system based on graph neural networks to implement the above-described method. The system includes: The graph construction module is used to construct a dynamic mathematical model based on the thermodynamic mechanism of the biomass gas turbine, and obtain the prior predicted values ​​of each measuring point at the current moment based on the model. It abstracts the components, measuring points or state points in the biomass gas turbine into graph nodes, and the connections and coupling relationships between graph nodes or the thermodynamic processes inside the equipment into graph edges to construct a graph structure that reflects the topology of the system. The prior predicted values, the real-time measured values ​​collected by the sensors, and the residuals between the measured values ​​and the prior predicted values ​​are used together as the initial features of each node to form a node feature matrix. The offline training module is used to process historical running data using a preset recursive state estimation algorithm to generate coordinated state estimates at each time point as supervision labels. The graph convolutional network is trained with the node feature matrix as input and the supervision labels as expected output, so that the graph convolutional network learns the mapping law from prior prediction and real-time measurement to the coordinated state. The online application module is used to input the measured values ​​collected by the sensor in real time and the prior predicted values ​​provided by the dynamic mathematical model into the trained graph convolutional network. Through one forward propagation calculation of the graph convolutional network, the corrected coordinated state parameters are directly output for the control or monitoring of biomass gas turbines.

[0053] The biomass gas turbine data coordination system based on graph neural networks provided by this invention offers enhanced real-time performance: Replacing high-dimensional matrix recursion and iterative calculations in traditional filtering with graph neural network inference significantly reduces computational load and latency, making it suitable for online coordination of rapid dynamic processes such as start-up, shutdown, and load surges, while maintaining high accuracy while maintaining real-time performance. It also exhibits better robustness and adaptability: By learning dynamic data distribution and inter-node correlation patterns, it demonstrates stronger adaptability to nonlinear, time-varying, and non-Gaussian noise; it possesses better fault tolerance and anti-interference capabilities against outliers, missing data, and sensor failures, preventing error propagation in the coupled link. Furthermore, it simplifies engineering deployment: Reducing reliance on precise noise statistical modeling, filter parameter tuning, and linearization approximations lowers model maintenance and parameter tuning costs, facilitating migration and application across different fuel qualities and turbine platforms. Improved operational safety and economy: Real-time dynamic parameter coordination results can provide timely feedback and rapid decision-making basis for grid-connected power regulation and health monitoring, shorten control and diagnostic response time, improve the ability to track rapid changes in operating conditions, reduce risks such as overheating and surge, reduce unplanned downtime, and maintain efficiency and emission control levels under frequent regulation and fluctuating operating conditions. Specific Implementation This paper uses a biomass gas turbine system with heat recovery as the research object to illustrate the implementation process of this method. The system diagram is as follows: Figure 2 As shown, the evaporator 14 and its associated equipment and piping (pressurization pump 15, steam 16 and water 17) and the heating heat exchanger 18 and its associated equipment and piping (outlet water 19, return water 20) are intended to demonstrate two applications of the biomass gasifier. In addition to these two devices, it can also be connected to other equipment components that can utilize waste heat.

[0055] (I) Data Preparation and Graph Construction Module The dynamic equilibrium equations of the following example systems correspond to those described in the above method introduction. .

[0056] 1. First, the state transition function must be constructed based on the system's dynamic equilibrium equations. The following four assumptions are made:

[0057] (1) Ignore the leakage of working fluid and heat loss in the connecting pipes between the components; (2) Assume that the external environment remains unchanged; (3) The operating conditions are all within the normal operating range and are not in a surge condition; (4) The established axis of rotation is a rigid body.

[0058] The above assumptions are used to establish dynamic models of each component.

[0059] Compressor 9 dynamic model: The dynamic characteristic relationship between the pressure ratio, efficiency, and operating conditions of compressor 9 is as follows: , , in, This indicates the reduced pressure ratio of compressor 9. , Given the current pressure ratio, For design pressure ratio; This indicates the equivalent flow rate of compressor 9. 0 indicates the design condition. This represents the temperature at state point 1. This represents the pressure at state point 1. This indicates the flow rate of compressor 9, which is the same as the flow rate of turbine 12, and is determined by the power rating of the biomass gas turbine. This indicates the equivalent rotational speed of compressor 9. , This indicates the compressor speed, calculated from the rotating shaft. This indicates the reduced efficiency of compressor 9. , The efficiency of compressor 9 under design conditions is a known boundary condition; , , , Here is the empirical coefficient, where , , The following calculation formula is given: , in, , It is a constant, and for centrifugal compressors it is usually taken as... empirical coefficient It is usually taken as 0.3.

[0060] It can be known , , , First, you need to: , Determine the isentropic exit state using the isentropic relation: , The formula for calculating the isentropic efficiency of compressor 9 is: , in, h The subscript s represents the enthalpy value, the subscript c represents the compressor, out represents the outlet, and in represents the inlet.

[0061] , The temperature is determined using the enthalpy-temperature relationship: , Compressor power consumption for: .

[0062] Dynamic model of biomass boiler 11: In this example system, the biomass boiler 11 is considered as an equivalent "constant power heater" that heats the working fluid in the main gas path, without explicitly describing the complex combustion chemical reaction process. The chemical energy input from the fuel side is converted into heating power for the main gas path with a given efficiency, thus obtaining the dynamic energy input of the biomass boiler 11.

[0063] (1) Equivalent heating power on the fuel side The fuel heats the working fluid in the main gas path at state point 7 (seventh state point 7) and is then discharged at state point 8 (eighth state point 8). The equivalent heating power provided by the fuel to the main gas path is defined as:

[0064] , in, Equivalent heating power provided to the main gas path of biomass boiler 11, in W; The equivalent thermal efficiency (EFE) of a biomass boiler (representing the proportion of fuel chemical energy converted into enthalpy rise in the main gas path, including incomplete release, heat dissipation, wall losses, etc., determined by the boiler manufacturer, and is dimensionless) is defined by boundary conditions. Fuel mass flow rate (control input), unit kg / s; The lower heating value of the fuel is expressed in J / kg.

[0065] (2) Energy conservation in the main gas path If the inlet state of the main gas path of biomass boiler 11 is denoted as the third state point 3, and the outlet state is denoted as the fourth state point 4, then the energy balance of the main gas path in biomass boiler 11 can be expressed as: , in, The mass flow rate through the main gas path of biomass boiler 11 (consistent with the gas mass flow rate of compressor 9 / turbine 12 circuit), in kg / s; The specific heat capacity at constant pressure of the working fluid in the main gas path of the biomass boiler 11 (which can be taken as a function of temperature or a constant value within the operating range), in J / (kg·K); Temperature of the main gas path at the inlet (third state point 3) of biomass boiler 11, in K; The temperature of the main gas path at the outlet of biomass boiler 11 (fourth state point 4) is in K.

[0066] (3) Pressure drop model of biomass boiler 11 The pressure drop of biomass boiler 11 is calculated using the pressure loss coefficient: , in, The pressure of the main gas path at the inlet of biomass boiler 11 (third state point 3) is in Pa; The pressure of the main gas path at the inlet of biomass boiler 11 (fourth state point 4) is in Pa; , is the relative pressure loss coefficient of biomass boiler 11 (the ratio of pressure loss from inlet to outlet to inlet pressure), dimensionless.

[0067] Turbo 12 dynamic model: Based on the derivation of the Frugel formula, the variable operating condition expression for turbine 12 is as follows: , in, This indicates the rated flow rate of turbine 12; This indicates the actual flow rate of turbine 12; This indicates the rated inlet temperature of turbine 12; This indicates the actual inlet temperature of turbine 12; This indicates the rated expansion ratio of turbine 12; This represents the actual expansion ratio of turbine 12; in the above formula... for: , in, This indicates the rated speed of turbine 12; This indicates a turbine speed of 12.

[0068] The formula for calculating the efficiency characteristics of turbine 12 is: , in, This indicates the reduced efficiency of turbine 12; This indicates the equivalent rotational speed of turbine 12; This represents the equivalent flow rate of turbine 12; These are undetermined coefficients, which are constants in the calculation, typically taken as 0.3~0.5. The calculation formulas for all converted parameters are the same as those for compressor 9.

[0069] In this example, the pressure loss on the hot side of heating heat exchanger 18 is... The hot-side pressure loss of evaporator 14 is The hot-side pressure loss of the regenerator 10 is The cold-side pressure loss of the regenerator 10 is The outlet of the heating heat exchanger 18 is at atmospheric pressure. The temperature is the ambient temperature. The pressure at the sixth state point 6 is obtained as follows:

[0070] , , Therefore, the expansion ratio of turbine 12 can be expressed as: .

[0071] Known , , , The isentropic exit state is: , , The formula for calculating the isentropic efficiency of turbine 12 is: , , Solving for the outlet temperature of turbine 12 using the equation of state: , Turbine 12 output power for: , Regenerator 10 Dynamic Model: During the calculation of regenerator 10, the outlet temperature of compressor 9 is known. hot end outlet Depending on the equipment conditions of the evaporator 14 and the cold end of the heating heat exchanger 18, it can be calculated according to the heat balance equation after it is determined, so it is set as a known quantity here.

[0072] (1) The heat expression for the regenerator 10 is: , in, The instantaneous heat exchange capacity (W) of the regenerator 10 is defined as the heat released at the hot end (a positive value indicates that the heat is transferred from the hot end to the cold end). The inlet temperature of the hot end of the regenerator 10 (fifth state point 5, unknown, calculated by the turbine model); The outlet temperature of the hot end of regenerator 10 (sixth state point 6, known); The common heat capacity factor (W / K) on both sides of the regenerator 10 is determined by the following formula: , in, Main gas path mass flow rate The specific heat of air at constant pressure can be found in the air property table by average temperature / pressure. , A suitable approximation of the inlet / outlet average (e.g.) or ).

[0073] (2) The cold end outlet temperature (third state point 3) is obtained from the heat exchange: , After sorting, we can obtain: , in, The outlet temperature of compressor 9 / cold end inlet temperature of regenerator 10 (second state point 2, known); The temperature at the cold end outlet of the regenerator 10 / the inlet temperature of the biomass boiler 11 (third state point 3, unknown and to be determined).

[0074] (3) The pressure relationship can be determined from the pressure loss of the regenerator 10 during both hot and cold cycles: Cold side: , Hot side: , (4) Enthalpy calculation Air properties can be calculated from tables: , , Subsequently, after coupling with biomass boiler 11 and turbine 12, iterative iterations can be performed to obtain... and .

[0075] Dynamic model of the rotating shaft: Output power of generator 13 for: , in, This indicates the rated generating power of generator 13; This indicates the rated speed of generator 13; This indicates the actual rotational speed of generator 13.

[0076] The energy balance of the rotating shaft is: , in, Indicates the moment of inertia; Indicates angular velocity; This indicates mechanical loss.

[0077] Power rating of biomass gas turbines: Assuming the power level of the biomass gas turbine is (W), the formula is: , Then the flow rate of the main circuit can be obtained.

[0078] In summary, all state points in this system are known. To unify and fuse real-time sensor measurements with the output of the dynamic model, and to provide a consistent measurement mapping for subsequent data coordination and graph neural network input feature construction, this embodiment further establishes a measurement model. It is used to describe the correspondence between system state and measured quantity.

[0079] (1) Discrete sampling and measurement sequence Let the sampling period be (Unit: seconds), the index of discrete sampling time is: At every moment The real-time measurement vector is obtained by the sensor:

[0080] , in, For a moment The measurement vector; The number of dimensions to be measured; For the first Observations from each measurement channel (e.g., temperature, pressure, rotational speed, power, etc.).

[0081] Simultaneously, the control input vector is acquired: , in, For a moment Control input; The number of input dimensions; control inputs may include fuel mass flow commands, load / speed commands, valve or guide vane openings, etc.

[0082] (2) General form of measurement model Define the system state vector as ,in For the state dimension, the measurement model adopts the following nonlinear form:

[0083] , in, This is a measurement function used to map the state space to the measurement space; To measure noise, the following conditions must be met: , in, The noise covariance matrix is ​​used to characterize the noise intensity of each measurement channel and the correlation between channels.

[0084] Similarly, process noise is introduced. satisfy: , in, The process noise covariance matrix ( (where is the number of state dimensions), used to characterize model simplification errors, unmodeled disturbances, and slow parameter drift.

[0085] (3) Measurement function Construction method In this example, the measurement function It consists of two parts: (a) Extraction of directly measurable quantities: If a measurement channel directly corresponds to a component in the state vector, then This corresponds to the extraction of that component; (b) Calculation of derived quantities: If a measurement channel corresponds to a derived quantity (e.g., power, efficiency, etc.) calculated from the state, then This corresponds to the deterministic functional expression of the derived quantity.

[0086] The system is equipped with the following measuring points, which can be added or removed according to the sensor configuration in actual engineering.

[0087] , in, The temperature measurement value is at point 2; The pressure measurement value is at point 2; The temperature measurement value is at point 4; The pressure measurement value is at point 4; The temperature measurement value is at point 6; This is the measured rotational speed value; This is a power measurement value (which can be obtained from the generator side or the electrical measurement system).

[0088] The corresponding measurement function can then be written as: , in, State vector The extracted value of the corresponding component; The power derivative obtained from the state vector (e.g., from the system power level / energy balance relationship) can be specifically expressed according to the aforementioned dynamic model: , , .

[0089] (4) Measure the noise covariance Sources and missing data handling Preferably, the noise covariance is measured. It can be obtained by converting the sensor's nominal accuracy or by estimating it statistically from historical data, and can be taken as a diagonal matrix: , in, This represents the noise variance for the corresponding measurement channel.

[0090] When certain measurement channels are at time When there are missing measurements or measurements are deemed invalid, a measurement validity mask vector can be introduced: , in, Indicates the first Each measurement channel at time efficient, This indicates that the channel is missing or invalid. For invalid channels, it is preferable to reduce weight and achieve robust fusion by increasing the corresponding noise variance or ignoring the channel in subsequent updates.

[0091] Preferably, the process noise covariance The parameters can be set empirically based on the importance and rate of change of the state components, or determined through offline parameter tuning, and can be taken as a diagonal matrix or a block diagonal matrix: , in, For the first Each state component at time... The process noise variance.

[0092] The initial filter value is set as follows: , in, For initial posterior state estimation; The initial state value can be directly assigned from the initial measurement, or a reasonable initial value can be given for the unmeasurable state or obtained by solving the mechanism model; This is the initial estimation error covariance matrix, used to characterize the initial value uncertainty.

[0093] Based on the above sampling settings, noise covariance, and initialization parameters, the dynamic model constructed in point 1 and the aforementioned measurement model can be recursively fused within a unified probabilistic framework. To generate supervisory labels for the training phase and provide an online reconciled benchmark, unscented Kalman filtering is used below. Perform prior predictions and combine them with real-time measurements. Complete the measurement update and obtain the coordinated posterior estimate. .

[0094] 2. Constructing the thermodynamic diagram structure Nodes are used to represent the inlet and outlet of each component and are assigned relevant parameter variables. Since pressure loss and heat loss in the piping within the system are ignored, the inlet and outlet of some components are considered to be the same node. The connection relationships of each node are characterized using the flow of the working fluid. Although the flow of the working fluid is directional, undirected graphs are used to represent these connections to simplify calculations. Furthermore, due to the special nature of heat exchange equipment, the inlet and outlet nodes of the hot and cold fluids are assumed to be connected.

[0095] Based on the adjacency matrix construction rules in the research content, the adjacency matrix of the biomass gas turbine system in this example is constructed as follows: The degree matrix is ​​constructed as follows: Abstract process such as Figure 3 As shown.

[0096] 3. Prior prediction and node feature matrix construction (graph network input) At every moment The prior prediction is obtained from the UKF time update process in point 2. The residuals are obtained from the measurement model: , in, The residuals between measurement and prediction are used to reflect the degree of inconsistency between the model and the measurement.

[0097] For each node Construct node feature vectors: , in, For node feature dimensions. Preferably, It should include at least a combination of the following available information: the measurement value corresponding to the node (if it exists), the prior prediction value corresponding to the node, the residual component corresponding to the node, and optional global load / control input features. For example, it can be written as:

[0098] , in, Represents nodes Associated measurement sub-vectors; Represents nodes The associated prior state subvector; For the corresponding residual subvectors; To control the input (which can be copied to each node or used as a global feature concatenation), the order of these variables can be set arbitrarily, as long as they remain consistent.

[0099] Stack all node features row-wise to obtain the node feature matrix: , in, For a moment The graph neural network is input to the feature matrix.

[0100] (ii) Rapid Coordination Module 4. Design of graph convolutional networks In this embodiment, the graph convolutional network is used to learn the mapping from "node features constructed from prior predictions and real-time measurements" to "coordinated state / parameters", thereby replacing the traditional recursive update process with a single forward propagation in the online phase and achieving rapid data coordination.

[0101] (1) Input of graph convolutional network and definition of layer 0 At any moment The adjacency matrix with self-loops is obtained from the system topology. degree matrix And construct the node feature matrix: , in, The number of nodes; For node feature dimensions; The Behavior Nodes At any moment The feature vector (which may include measured values, prior predictions, residuals, control inputs / operating conditions, etc.).

[0102] Define the normalized graph convolution operator: , in, It is the symmetric normalized adjacency matrix.

[0103] To mitigate the adverse effects of differences in the dimensions and orders of magnitude of different physical quantities on the stability of network training, it is preferable to adjust the node feature matrix. Perform feature normalization or standardization to obtain the normalized feature matrix. When used as network input, that is: , in, This is the input to layer 0 of the graph convolutional network.

[0104] Preferably, z-score normalization based on feature dimensions can be used: , in, Indicates the first Each feature dimension is a column vector across all nodes; The first The mean and standard deviation of each feature dimension can be obtained statistically from the training data and used consistently during the online phase.

[0105] The above normalization process does not change the graph structure. By scaling only the input features, the network can learn the coupling relationships between various physical quantities more stably.

[0106] (2) No. Layer propagation and the first Layer output definition adopt common Layered graph convolutional network, the first layer( The propagation update is as follows: , in, For the first Layer hidden representation matrix, This layer hides the dimension; For the first Layer trainable weight matrix; For bias terms; This is the activation function used to introduce nonlinearity; options include ReLU, GELU, or Tanh.

[0107] Output layer (first) The layer obtains node-level coordination output: , in, This is the coordination result matrix output by the network; The physical quantity / parameter dimension output for each node (e.g., any combination of output temperature, pressure, enthalpy, flow rate, efficiency, etc.).

[0108] Furthermore, the node outputs are restored to the system's global coordination state vector according to a preset mapping relationship: , in, This is an assembly / mapping function for node outputs to the global state vector, used for processing output variables such as concatenation and sorting.

[0109] (3) Design of supervision signals and loss function during training phase During the training phase, a recursive estimator (such as UKF) is used to generate reconciliation reference values. And mapped to a label matrix: , in, For supervising the label matrix; This represents the inverse mapping / extraction mapping from the global coordination state to node labels.

[0110] Mean squared error is preferred as the monitoring loss: , in, This represents the Frobenius norm.

[0111] When there are missing tests or some nodes cannot generate valid labels, an output mask matrix is ​​introduced: , in, Represents a node The The output at time... Having valid labels and participating in training, This indicates that the term is ignored. Therefore, the mean square error of the mask is:

[0112] , in, This represents Hadamard element-wise multiplication.

[0113] (4) Physical consistency constraint regularization term To ensure that the coordination results satisfy basic physical laws (such as energy conservation residuals, temperature / pressure range, monotonicity, etc.), a physical consistency regularization term can be added to the supervision loss: , in, This is a physical constraint function (e.g., a residual function constructed from a dynamic model or component equilibrium equations). It is a 2-norm.

[0114] For example, for a regenerator, the heat transfer at the hot and cold ends can be written as: , , The physical consistency residual is then calculated as the difference in heat transfer between the hot and cold ends. , in, For the heat exchanger at all times Physical consistency residuals (constraint residuals).

[0115] That is, expanded as: , Corresponding to discrete time The physical consistency regularization term can be written as: , Thus, by arranging all the physical constraint functions into a vector, we obtain the physical constraint functions of the entire system.

[0116] The overall loss function is: , in, , which is the weighting coefficient used to balance the data supervision term and the physical constraint term.

[0117] For length of The training sequence (or batch) can be averaged over time to obtain the overall loss: , Update the network parameter set using gradient descent-type optimization methods (such as Adam). This yields the trained graph convolutional network model.

[0118] (III) Network Training Module 5. Recursive Coordination and Supervision Label Generation Based on Unscented Kalman Filter (UKF) UKF is used to generate labels in the training set of graph convolutional networks. UKF is run for each scenario to generate harmonized data. UKF is essentially a filtering and harmonization process. Graph convolutional networks aim to replace UKF, but since they are supervised learning systems, they still require labels.

[0119] The discrete dynamic model constructed in point 1 With measurement model Subsequently, this example employs unscented Kalman filtering (UKF) to recursively estimate the system, obtaining a consistent posterior estimate between the fusion mechanism model and real-time measurements. Furthermore, a node label matrix is ​​generated for training the graph convolutional network. UKF is an example of the "recursive state estimator" generally referred to in the network training module of the invention (iii).

[0120] (1) Construction of Unscented Transformation Parameters and Sigma Points Sigma points are a set of deterministic sampling points used in unscented transformations to characterize the uncertainty of random variables. The core idea is to approximate the probability distribution characteristics of a state using a small number of representative points selected in a specific way, thereby avoiding explicit linearization of the nonlinear model. To effectively characterize the state distribution after nonlinear propagation, several Sigma points need to be constructed around the current state estimate and its uncertainty, and then input into the system's nonlinear model for propagation. Subsequently, the propagation results are weighted and summarized to estimate the transformed mean, covariance, and other statistics, providing a basis for subsequent predictions and updates.

[0121] Let the state dimension be The parameters of the unscented transform are ,definition: , in, For scaling parameters; Controlling the dispersion of sigma points; Used to introduce prior distribution information (can be taken as an approximation of Gaussian distribution) ); This is the secondary scaling parameter.

[0122] At any moment posterior estimation With covariance When known, construct sigma points: , , , in, For the first 1 Sigma point; Represents the square root of a matrix; Indicates taking the first element of the matrix. List.

[0123] The corresponding mean weights and covariance weights are: , , , in, Weighted by mean, This represents the covariance weight.

[0124] (2) Prior prediction (time update) Propagate the sigma point to time according to the dynamic model. : , in, The prior sigma point after propagation; The system dynamic model function is constructed for the first point.

[0125] The prior state mean and prior covariance are respectively: , , in, For a moment Prior prediction; The prior error covariance; Let be the process noise covariance.

[0126] (3) Measurement prediction and measurement update (obtaining the reconciled posterior) Mapping the prior sigma point to the measurement space: , in, For the first One measurement sigma point; For measurement functions, describe the mapping relationship from state to measurement.

[0127] The prior measurement mean is: , The residual is defined as: , in, For a moment The residual.

[0128] The measurement covariance and the state-measurement cross-covariance are: , , in, For innovative covariance; For cross covariance; To measure the noise covariance.

[0129] Calculate the filter gain: , in, This is the UKF equivalent Kalman gain matrix.

[0130] We obtain the reconciled posterior estimate and posterior covariance: , , in, For a moment The coordinated state after fusion measurement; This is the covariance of the posterior estimation error.

[0131] (4) Construction of the supervision label matrix Reconcile posterior states Mapping function by node Organized into a tag matrix: , in, The number of nodes in the graph; Output physical quantity dimensions for each node (e.g., output several quantities such as temperature, pressure, and enthalpy for each node). This means extracting the components belonging to each node from the global state vector and stacking them row by row.

[0132] 6. Online coordination The online phase does not require running the complete recursive estimator update process; rapid reconciliation can be achieved simply by following these steps: (a) Based on system topology construction ; (b) Obtaining prior predictions from the dynamic model and combined with real-time measurement Calculate residuals / residuals ; (c) Construct the node feature matrix And normalized to obtain ,make ; (d) Propagation according to the propagation formula in (2) Layer forward propagation obtained ; (e) by Restore the global coordination state as the coordination output at that moment, for use in control, monitoring, or diagnosis.

[0133] The above inventions are merely a few specific embodiments of the present invention. However, the embodiments of the present invention are not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A method of coordinating data for a biomass engine, the method comprising: Includes the following steps: Graph construction: A dynamic mathematical model is constructed based on the thermodynamic mechanism of biomass gas turbine, and the prior prediction values ​​of each measuring point at the current time are obtained based on the model; The components, measuring points, or state points in the biomass gas turbine are abstracted as graph nodes, and the connections and coupling relationships between graph nodes or the thermodynamic processes inside the equipment are abstracted as graph edges, so as to construct a graph structure that reflects the system topology. The prior predicted value, the real-time measured value collected by the sensor, and the residual between the measured value and the prior predicted value are used together as the initial features of each node to form a node feature matrix. Offline training: The historical running data is processed using a pre-defined recursive state estimation algorithm to generate the coordinated state estimate at each time point as a supervision label; the graph convolutional network is trained with the node feature matrix as input and the supervision label as the expected output, so that the graph convolutional network learns the mapping law from prior prediction and real-time measurement to the coordinated state. Online application: The measured values ​​collected in real time by the sensor and the prior predicted values ​​provided by the dynamic mathematical model are input into the trained graph convolutional network. Through one forward propagation calculation of the graph convolutional network, the corrected coordination state parameters are directly output for the control or monitoring of biomass gas turbines.

2. The biomass engine data reconciliation method of claim 1, wherein, The dynamic mathematical model is represented as a discrete-time dynamic state-space model: ， ， wherein, is a discrete time index for the th sampling instant; is a system state vector and a dynamic parameter vector to be coordinated; is a control input vector; is a sensor measurement vector; is a state transition function, constructed based on thermodynamic or fluid dynamic mechanisms; is a measurement function; is a process noise; is a measurement noise; The prior prediction values ​​before the current measurement without fusion are obtained from the discrete-time dynamic state-space model: ， in, For a moment Prediction of prior states, For a moment -1 is the posterior estimate.

3. The biomass gas turbine data coordination method as described in claim 1, characterized in that, The graph structure is represented by an adjacency matrix, and its construction method is as follows: Construction Graph: ， in, Representation diagram; For a set of nodes, Represents the set of edges; Construct the corresponding adjacency matrix: ， ， in, For the first 1 node For the number of nodes, Represents a node With nodes There is a connection or coupling; To enhance the retention of node information, a self-loop is introduced: ， in, for An identity matrix of order 1.

4. The biomass gas turbine data coordination method as described in claim 3, characterized in that, The node feature matrix is ​​constructed as follows: At every moment For each node Construct node feature vectors: ， Stack all node features row-wise to obtain the node feature matrix: ， in, For nodes At any moment eigenvectors; The feature dimension for each node is independent of the degree matrix; For a moment The node feature matrix includes: node measured values, prior predicted values, residuals between measured values ​​and prior predicted values, global operating condition features of operating conditions and control variables, and coordination results from the previous time step. Normalize or standardize various features to obtain the normalized node feature matrix. or As input to a graph convolutional network.

5. The biomass gas turbine data coordination method as described in claim 1, characterized in that, The recursive state estimation algorithm includes one of the following: Kalman filter algorithm, information filter algorithm, or particle filter algorithm; When the recursive state estimation algorithm is the unscented Kalman filter algorithm in the Kalman filter class, Sigma points are constructed through unscented transformation, and prior statistics are calculated after propagation through a dynamic mathematical model. Then, the measurement is updated in combination with real-time measurements to obtain a coordinated posterior estimate as a supervision label.

6. The biomass gas turbine data coordination method as described in claim 4, characterized in that, The graph convolutional network adopts a multi-layer structure, and its propagation method is as follows: Define the normalized graph convolution operator: ， No. layer, The spread is as follows: ， The output layer obtains the coordination output matrix: ， in, For the first Layer hidden representation matrix, This layer hides the dimension; For the first Layer trainable weight matrix; For activation functions; for The corresponding degree matrix, ; To coordinate the output matrix; The number of physical quantities / parameters output for each node; The coordination output matrix is ​​restored to the system's global coordination state vector according to a preset mapping relationship: ， in, This is a function for assembling / mapping node outputs to the global state vector.

7. The biomass gas turbine data coordination method as described in claim 1, characterized in that, In the offline training, mean squared error is used as the loss function: ， in, Number the nodes. Output component numbers for the nodes. For a moment node middle The predicted value of the component. For a moment node middle The true value of the component; When there are missing or invalid labels, the output mask matrix is ​​introduced into the loss calculation: ， in, Indicates time No. In each node The output mask of the component. Indicates time The number of valid labels used in training.

8. The biomass gas turbine data coordination method as described in claim 7, characterized in that, In the offline training, a physical consistency constraint regularization term is also introduced to ensure that the coordination results meet the preset physical constraints. The comprehensive loss function is: ， in, These are weighting coefficients used to balance data supervision terms and physical constraint terms; This is a penalty term for physical consistency loss or constraint. This is the weighted mean squared error loss function.

9. The biomass gas turbine data coordination method as described in claim 1, characterized in that, The single forward propagation calculation replaces the iterative update process of the traditional recursive state estimation algorithm, enabling real-time data coordination.

10. A biomass gas turbine data coordination system based on graph neural networks, characterized in that, The system for implementing the method of any one of claims 1 to 9 comprises: The graph construction module is used to construct a dynamic mathematical model based on the thermodynamic mechanism of the biomass gas turbine, and obtain the prior predicted values ​​of each measuring point at the current moment based on the model; it abstracts the components, measuring points or state points in the biomass gas turbine as graph nodes, and abstracts the connections and coupling relationships between graph nodes or the thermodynamic processes inside the equipment as graph edges to construct a graph structure that reflects the system topology; the prior predicted values, the real-time measured values ​​collected by the sensors, and the residuals between the measured values ​​and the prior predicted values ​​are used together as the initial features of each node to form a node feature matrix; The offline training module is used to process historical running data using a preset recursive state estimation algorithm to generate coordinated state estimates at each time point as supervision labels; the graph convolutional network is trained with the node feature matrix as input and the supervision labels as expected output, so that the graph convolutional network learns the mapping law from prior prediction and real-time measurement to the coordinated state. The online application module is used to input the measured values ​​collected by the sensor in real time and the prior predicted values ​​provided by the dynamic mathematical model into the trained graph convolutional network. Through one forward propagation calculation of the graph convolutional network, the corrected coordination state parameters are directly output for the control or monitoring of biomass gas turbines.

Images