Process whole process voltage sag sensitivity assessment method based on physical guidance graph neural network

Abstract

The invention provides an overall process voltage sag sensitivity evaluation method based on a physical guidance graph neural network, which comprises the following steps of: firstly, analyzing and acquiring key characteristics of equipment voltage sag influence, and types and connection relationships of sensitive equipment based on characteristics when voltage sag occurs; the technological process structure is converted into graph data suitable for the prediction model; secondly, constructing a physical basis function based on a transmission mechanism; and finally, a physically guided message transfer graph neural network (PGNN) is established, voltage sag features are used as global variables of the graph neural network, sample data and physical knowledge are combined, enough information is mined from two dimensions of data and a mechanism so as to improve the model prediction precision, and the prediction accuracy of the model is improved. Therefore, accurate evaluation of the voltage sag sensitivity in the whole process is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution networks, and particularly to a method for evaluating the voltage sag sensitivity of the entire process of a process based on a physics-guided graph neural network. Background Art

[0002] With the development of high-tech science and technology, more and more power electronics sensitive devices are applied to various industries. However, the resulting problem of power quality is becoming increasingly prominent. Among them, voltage sag, as one of the most serious power quality problems, has attracted great attention from the power grid and users. The impact of voltage sag mainly acts on the user side. When user equipment suffers from voltage sag, it may lead to problems such as equipment shutdown, product scrapping, and data loss. A single component failure may also cause the products of the entire production line to be scrapped, resulting in a large amount of economic property losses.

[0003] Currently, Message Passing Graph Neural Networks (MP-GNNs), as an emerging deep learning algorithm, its unique structure enables it to efficiently process network data, capture the complex relationships between nodes (sensitive devices), and thus more accurately reflect the dynamic characteristics of the power system.

[0004] Currently, although there have been certain research progresses on the voltage sag sensitivity of users, there are still some deficiencies. For example, traditional evaluation objects often only target single devices, ignoring the mutual influence between devices. For industrial users, the evaluation of the voltage sag sensitivity of the overall process flow is often more valuable. Moreover, general evaluation models only consider the impact of basic voltage sag characteristics on devices, without considering the requirements of devices for other physical conditions. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a method for evaluating the voltage sag sensitivity of the entire process of a process based on a physics-guided graph neural network. By constructing a physics-induced graph neural network (PGNN) that combines the characteristics of the message passing neural network, through its unique message passing mechanism, it can effectively aggregate and transmit node information, capture the complex relationships between devices in the process flow, and thus more accurately predict the impact degree of voltage sag. On this basis, physical guidance is added, that is, in the edge update process of the GN layer, the constructed physical information basis function is adopted. This method can not only improve the evaluation efficiency, but also realize the transmission and fusion of physical information in different-level updates, enhancing the model's understanding and expression ability of physical processes. It has broad application prospects and important theoretical significance.

[0006] To achieve the above object, the present invention adopts the following technical solutions: A method for evaluating the voltage sag sensitivity of the entire process of a process based on a physics-guided graph neural network, comprising the following steps:

[0007] Step 1: Construct process flow equipment feature map data;

[0008] Step 2: Construct a physical information basis function;

[0009] Step 3: Build a physics-guided message passing graph neural network PGNN;

[0010] Step 4: Evaluate the voltage sag sensitivity of the entire process of the process.

[0011] In a preferred embodiment, the key information of each element of the process flow equipment feature map is described as follows:

[0012] G = {N, E, g}

[0013] Wherein, N is the node feature, used to represent each device in the production process, N = {n i}, n i is the vector representation of the feature of device i, used for the expression of electrical physical parameters; E is the edge feature, E = {e ij}, where e ij is the vector representation of the connection relationship between device i and device j, used to characterize the coupling mechanism between devices; g is the global feature.

[0014] In a preferred embodiment, Step 2 includes defining the energy transfer efficiency between devices as a composite function of line impedance, device power demand, simulated distance, and trainable parameters:

[0015]

[0016] Wherein, α ij is the energy attenuation coefficient from device i to j, θ 1 , θ 2 , θ 3 are trainable parameters, respectively adjusting the weights of the power ratio term and the distance attenuation term; Z ij is the line impedance between sensitive devices i and j, d ij is the physical transmission distance between devices, P i out and P j req are the output power of device i and the rated power of device j, respectively.

[0017] In a preferred embodiment, Step 3 constructs a "electrical - physical" coupled PGN layer, and uses the physical information basis function constructed in Step 2 as the weight update function in the edge update link of the process flow diagram evaluation model.

[0018] In a preferred embodiment, the PGN layer utilizes the sensitive device feature update function φ n , the device connection feature update function φ e , the process flow global feature update function φ g , and the physical energy attenuation coefficient update function φ w to complete the information update during the message passing process. The sensitive device feature update function φ n , the device connection feature update function φ e , the process flow global feature update function φ g , and the physical energy attenuation coefficient update function φ w are multi-layer perceptrons (MPLs) in the process flow evaluation model. The sensitive device feature update function φ n , the device connection feature update function φ e , the process flow global feature update function φ g , and the physical energy attenuation coefficient update function φ w each contain their respective aggregation functions ρ n , ρ e , ρ g , and ρ w for aggregating the information received from other devices, where n, e, g, and w represent the sensitive device node feature, the device connection feature, the process global feature, and the physical energy attenuation coefficient, respectively. Define I as the index set of all sensitive device node indices; also define the index sets R i and S i as the sets of all receiver node indices and sender node indices of device i, respectively, and U as the index set containing all device connection relationships.

[0019] In a preferred embodiment, the device connection feature update function φ e updates all the edges of the graph, i.e., the device connection features, according to the following formula. Using the sending node device feature n j , the receiving node device feature n i , and the connection feature e ij between the two nodes to generate the updated device connection feature e i ′ j ; the PGN layer utilizes the physical energy attenuation coefficient update function φ w to dynamically adjust the weights according to the connection relationship between devices; selects the device location as the input of φ w , and φ w calculates the weight W ij for obtaining the updated connection feature e i ′ j , where θ 1 and θ 2 are trainable parameters;

[0020]

[0021] In a preferred embodiment, the aggregation function ρ e aggregates the edge features adjacent to the target device node in the SUM aggregation manner; then concatenates the original node features and the aggregated edge features; finally, puts the concatenated device node features into a feature update network with the same structure as the edge update to complete the update of the device node features.

[0022] In a preferred embodiment, the sensitive device node feature update function φ n uses the current input device node feature n i , the updated device connection feature e i ′ and the process flow global feature g to update the node features; the PGN layer uses the edge aggregation function ρ e to reduce the set of all connection features connected to the device to a single aggregated connection feature e i ′ j ; the sensitive device node feature update function φ n processes the aggregated device connection feature e i ′ j and the current device node feature n i to generate the updated device node feature n i ′, where θ 3 is a trainable parameter; the edge aggregation and node update processes are as follows:

[0023]

[0024] In a preferred embodiment, the process flow global update function φ g is used to update the global feature; the PGN layer uses the updated edge e′ and node feature n′ to update the global feature; the PGN layer uses the aggregation functions ρ e and ρ n , and aggregates all device connection features E′ and all sensitive device node features N′ into two vectors e′ and n′ in the way of sum-average, where θ 4 is a trainable parameter; the process flow global feature update process is:

[0025]

[0026] In a preferred embodiment, the physical energy attenuation coefficient W ij is updated, which represents the attenuation degree of the intermediate voltage sag feature and the physical quantity transfer process. The physical energy attenuation coefficient update function φ w is used for updating the edge information between nodes in the graph neural network. The expression is the constructed physical information basis function, which represents the attenuation degree of energy:

[0027]

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] 1) Innovation: When evaluating the voltage sag tolerance of sensitive users, the present invention considers the coupling relationship of "equipment - equipment" with the overall process flow as the evaluation object; at the same time, physical guidance is added to the model of the present invention, considering the relationship of "electrical - physical" coupling.

[0030] 2) Precise evaluation: Using PGNN to study the problem of voltage sag sensitivity evaluation can enhance the identification of the fault risk of sag - sensitive equipment, improve the evaluation accuracy, and provide important support for sensitive users.

[0031] 3) Comprehensive analysis: The connection and attribute information between nodes help to establish a more accurate voltage sag sensitivity evaluation model, enabling the model to consider the influence of the entire network and enhancing the robustness of the voltage sag sensitivity evaluation. Brief Description of the Drawings

[0032] Figure 1 It is the model working flow chart of the preferred embodiment of the present invention. Detailed Embodiments

[0033] The present invention will be further described below in conjunction with the drawings and embodiments.

[0034] It should be noted that the following detailed description is illustrative and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present application belongs.

[0035] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application; as used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0036] A method for evaluating the voltage sag sensitivity of the whole process of a process based on a physically - guided graph neural network, referring to Figure 1 , includes the following steps:

[0037] Step 1: Construct process flow equipment feature map data

[0038] Based on the power grid topology, measured data of voltage sags (amplitude, duration), operating parameters of equipment (rated voltage, power, starting current), and the connection relationships between equipment, node features (electrical parameters, physical sensitivity types, operating conditions), edge features (line impedance, power transmission direction, physical coupling coefficient), and global features (sag amplitude, duration) are constructed. Through standardization processing, directed graph data is formed to represent the energy flow and coupling relationships between equipment in the process flow.

[0039] The equipment feature graph data of the process flow is not only the training data for the physical-guided graph neural network model but also represents the flow direction and correlation degree of energy and matter in the production process. Therefore, it is particularly important to accurately construct the equipment feature graph of the process flow. To construct the equipment feature graph of the process flow, it is first necessary to clarify the key information of each element in the graph, and its mathematical description is as follows:

[0040] G = {N, E, g}

[0041] Where N is the node feature, used to represent each piece of equipment in the production process, N = {n i}, n i is the vector representation of the feature of equipment i, used for expressing electrical and physical parameters; E is the edge feature, E = {e ij}, where e ij is the vector representation of the connection relationship between equipment i and equipment j, used to characterize the coupling mechanism between equipment; g is the global feature.

[0042] 1. Sensitive equipment feature vector N

[0043] (1) Types of sensitive loads: According to the response differences of equipment to voltage sags and physical factors, sensitive loads can be divided into electrical-sensitive type, physical-parameter-sensitive type, and mixed-sensitive type. Electrically sensitive equipment is directly affected by the characteristic values of voltage sags, such as voltage sag amplitude, duration, etc. For example, some precision electronic instruments may have data errors or equipment failures even with a slight voltage fluctuation. Physically parameter-sensitive loads are sensitive to changes in physical parameters, and these physical parameters are subdivided into temperature, flow rate, rotational speed, pressure, etc. For example, in chemical production, some reaction equipment is extremely sensitive to temperature. Voltage sags may cause abnormalities in the cooling system or heating system, resulting in temperature fluctuations in the equipment, affecting the production process and even causing product quality problems. Mixed-sensitive equipment is affected by both electrical and physical factors.

[0044] (2) Electrical parameters: Electrical parameters such as the rated voltage, rated power, and starting current of equipment are key indicators reflecting the electrical energy demand and usage characteristics of the equipment. Equipment with different rated parameters have different sensitivities to voltage sags. For example, when the voltage sag amplitude is the same, high-rated-voltage equipment may be less affected than low-rated-voltage equipment.

[0045] (3) Operating condition parameters: The operating conditions of the equipment, such as full load, half load or light load states, have a significant impact on its voltage sag tolerance. Taking a motor as an example, when operating at full load, the motor is more sensitive to voltage changes. Voltage sags are likely to cause a decrease in motor speed and torque, and in severe cases, stalling, which in turn affects the entire process flow.

[0046] 2. Eigenvector E of the connection (edge) of sensitive equipment

[0047] (1) Electrical connection characteristics: The line impedance of the connected equipment is an important electrical connection characteristic, which directly affects the voltage drop when voltage is transmitted between equipment. The greater the line impedance, the more complex the attenuation and change of the voltage sag during transmission, and the impact on the equipment will also change accordingly. In addition, the transmission delay reflects the time required for the voltage sag signal to be transmitted from one equipment to another. In a time-sensitive process flow, the transmission delay may cause differences in the response of different equipment to voltage sags, thus affecting the coordination of the entire process. The power transmission capacity between connected equipment is also crucial, which determines the degree of mutual support and influence between equipment under voltage sag conditions.

[0048] (2) Physical coupling relationship: In addition to electrical connections, there is also a physical coupling relationship between equipment. For example, when some equipment has temperature control functions and others are sensitive to temperature, even if there is no direct electrical and process flow connection between them, there is an indirect coupling relationship. This relationship can be reflected by temperature-related characteristics, such as the temperature transfer coefficient, which represents the degree of influence of the temperature change of one equipment on the temperature of another equipment; and the temperature response time, which reflects how quickly the affected equipment responds to temperature changes. For couplings involving physical parameters such as flow rate and pressure, there are also corresponding parameters to describe, such as the flow rate influence coefficient and the pressure transfer ratio. These parameters help to more accurately describe the mutual influence of physical parameters between equipment, thus comprehensively reflecting the coupling relationship between equipment and providing a more accurate basis for analyzing the propagation and influence of voltage sags in the process flow.

[0049] 3. Directed edge index set

[0050] The directed edge index set determines the direction of energy and material flow between equipment in the process flow. In actual production, the flow of energy and materials often has a clear direction. For example, in an assembly line, raw materials pass through multiple processing equipment in sequence from the starting equipment and finally become finished products. In this process, energy and materials flow unidirectionally. The directed edge index set accurately describes this flow characteristic by defining and indexing the direction of the edges. It can be represented by index pairs or index lists in mathematics. Each index corresponds to a directed edge, clearly indicating the connection relationship from the source equipment to the target equipment, providing a basis for analyzing the propagation path of voltage sags in the process flow.

[0051] 4. Global Feature Vector \(g\) of the Process Flow

[0052] The global features are selected as the voltage sag amplitude and the voltage sag duration. The voltage sag amplitude reflects the degree of voltage drop during a voltage sag event, and a single voltage sag event can cause disturbances to all electrical equipment in the factory, making it suitable as a global feature. At the same time, different voltage sag amplitudes have different impacts on equipment. A larger voltage sag amplitude may directly cause equipment to malfunction due to low voltage. The voltage sag duration reflects the length of time the voltage remains at a low level. Even if the voltage sag amplitudes are the same, different durations will result in different degrees of impact on the equipment. The longer the duration, the more severe the negative impact on the equipment may be. These two features are input into the trained model, and the model predicts the fault risks of each device based on this, thereby realizing the assessment of the voltage sag sensitivity throughout the entire process.

[0053] Step 2: Construct the Physics-Informed Basis Function

[0054] Combining the power grid transmission mechanism and the dynamic coupling characteristics of equipment, design the physics-informed basis function (physical quantity attenuation model, energy transfer equation), and embed it into the edge update process of the graph neural network as a weight update function to dynamically adjust the intensity of the influence between devices, ensuring that the model conforms to physical laws on the basis of data-driven and enhancing interpretability.

[0055] The physics-informed basis function of this method is constructed based on the law of conservation of energy. When conducting analysis, it is necessary to fully consider the influence of voltage sag propagation characteristics and physical coupling. Specifically, when the electrical distance between the sensitive device and the voltage sag source increases, or its equivalent impedance increases, the voltage sag amplitude at the location of the sensitive device will decrease accordingly, and the degree of influence by the voltage sag will also decrease. In addition, if a physical quantity control device fails, it will also affect other devices sensitive to this physical quantity, thereby breaking the original balance state of the system. Based on the above analysis, define the energy transfer efficiency between devices as a composite function of line impedance, device power demand, simulated distance, and trainable parameters:

[0056]

[0057] In the formula, \(\alpha\) ij is the energy attenuation coefficient from device \(i\) to \(j\), \(\theta\) 1 , \(\theta\) 2 , \(\theta\) 3 are trainable parameters that respectively adjust the weights of the power ratio term and the distance attenuation term; \(Z\) ij is the line impedance between the sensitive devices \(i\) and \(j\), \(d\) ij is the physical transmission distance between devices, \(P\) i out and \(P\) jreq are the output power of device i and the rated power of device j respectively.

[0058] Step 3: Build a Physics-Guided Message Passing Graph Neural Network (PGNN)

[0059] A PGNN architecture was designed, which included edge update (adjusting weights based on physical basis functions), node update (aggregating adjacent edge features and concatenating node information) and global update (integrating features of the entire network). The training set data was used to optimize the network hyperparameters (hidden layer dimension, learning rate), and regularization technology was used to prevent overfitting, thereby realizing nonlinear mapping modeling of device characteristics and the impact of voltage sag.

[0060] Specifically:

[0061] The physically guided Message Passing Graph Neural Network (PGNN) is based on the Message Passing Graph Neural Networks (MPNN), which is a deep neural network that extends convolution to graph structured data. It can extract feature information on nodes and edges based on the network topology and is widely used in graph state prediction and node state prediction problems. Therefore, based on this, this patent designs a PGN layer that reflects the mutual influence and "electrical-physical" coupling between devices. In the edge update link of the process flow chart evaluation model, the physical energy decay function of the device constructed above is used as the weight update function, so that the prediction model has a certain degree of physical interpretability. The following is a brief description of how to improve the process sensitivity assessment.

[0062] The PGN layer uses sensitive device characteristics to update the function φ n , device connection feature update function φ e , process flow global feature update function φ g , physical energy attenuation coefficient update function φ w The four functions complete the information update in the process of message passing. These four functions are multiple perceptrons (MPL) in the process flow evaluation model. Each update function also contains its own aggregation function ρ n , e , g , w Used to aggregate information received from other devices. Where n, e, g, and w represent sensitive device node characteristics, device connection characteristics, process global characteristics, and physical energy attenuation coefficients, respectively. Define I as the index set of all sensitive device node indexes; also define the index set R i and S i , they are the sets of all receiver node indexes and sender node indexes of node i, and U is the index set containing all device connection relationships.

[0063] 1). Update of device connection relationship (edge) features:

[0064] Device connection feature update function φ e Update all edges of the graph according to the following formula, that is, device connection features. It uses the sending node device feature n j , receiving node device feature n i and the connection feature e between the two nodes ij to generate the updated device connection feature e i ′ j . The PGN layer uses the physical energy attenuation coefficient update function φ w to dynamically adjust the weights according to the connection relationship between devices. Here, the device location is selected as the input of φ w , φ w calculates the weight W ij , which is used to obtain the updated connection feature e i ′ j , θ 1 and θ 2 are trainable parameters. The weights are obtained using the influence relationship in the attenuation model to adjust the intensity of the influence between devices and processes, making this calculation physically meaningful and highly interpretable.

[0065] e i ′ j ← φ e (e ij , n i , n j ; θ 1 )

[0066] W ij ← φ w (e ij , n i , n j ; θ 2 )

[0067] e i ′ j ← W ij × e ij ′

[0068] 2). Update of sensitive device node features:

[0069] In the update of sensitive device nodes, first, the aggregation of features between devices is performed, and the aggregation function ρ eAggregate the edge features adjacent to the target device node in the SUM aggregation manner. Then, splice the original node features and the aggregated edge features. Finally, put the spliced device node features into a feature update network with the same structure as the edge update to complete the update of the device node features. The change in the device node features reflects the combined influence of voltage sags and other devices.

[0070] Sensitive device node feature update function φ n Use the current input device node feature n i , the updated device connection feature e i ′ and the process flow global feature g to update the node features. The PGN layer uses the edge aggregation function ρ e to reduce the set of all connection features connected to this device to a single aggregated connection feature e i ′ j . The sensitive device node feature update function φ n processes the aggregated device connection feature e i ′ j and the current device node feature n i to generate the updated device node feature n i ′, where θ 3 are trainable parameters. The edge aggregation and node update processes are as follows:

[0071]

[0072] n i ′ ← φ n (e i ′, n i , g; θ 3 )

[0073] 3). Process flow global feature update

[0074] The process flow global update function φ g is used to update the global feature. The PGN layer uses the updated edge e′ and node feature n′ to update the global feature. To this end, the PGN layer utilizes the aggregation functions ρ e and ρ n to aggregate all device connection features E′ and all sensitive device node features N′ into two vectors e′ and n′ in the way of summation and averaging, where θ 4 are trainable parameters. The process flow global feature update process is:

[0075]

[0076] g′ ← φ g (e′, n′, g; θ 4 )

[0077] 4). Update of Physical Energy Attenuation Coefficient

[0078] Physical energy attenuation coefficient W ij Update, which characterizes the attenuation degree of the intermediate voltage sag characteristics and the physical quantity transfer process. The physical energy attenuation coefficient update function φ w is used for the update of edge information between nodes in the graph neural network. The expression is the physical information basis function constructed above, which characterizes the attenuation degree of energy.

[0079]

[0080] Step 4: Evaluation of Voltage Sag Sensitivity in the Whole Process of the Process

[0081] Input the test set into the trained PGNN model, predict the failure risk probability and sensitivity of each device, analyze the sag propagation path in combination with global features, and generate a heat map of sensitivity distribution; verify the superiority of this method by comparing the indicators (accuracy rate, F1 value) of the traditional single-device evaluation method, and feedback the evaluation results to the power grid system to guide the optimization of device protection strategies and the iterative update of the model.

Claims

1. A method for evaluating voltage sag sensitivity during the entire process of a process based on a physical guided graph neural network, characterized in that: The following steps are involved: Step 1: Construct process equipment characteristic diagram data; Step 2: Construct physical information basis functions; Step 3: Build a physics-guided message passing graph neural network PGNN; Step 4: Evaluate the voltage sag sensitivity throughout the process.

2. The method for evaluating voltage sag sensitivity in the whole process of a process based on a physical guided graph neural network according to claim 1, characterized in that: The key information of each element of the process equipment characteristic diagram is described as follows: G={N,E,g} Where N is the node feature, which is used to characterize each device in the production process. i },n i is the vector representation of the device i feature, used to express the electrical physical parameters; E is the edge feature, E = {eij}, where e ij is the vector representation of the connection relationship between device i and device j, which is used to characterize the coupling mechanism between devices; g is the global feature.

3. The method for evaluating voltage sag sensitivity in the whole process of a process based on a physical guided graph neural network according to claim 1, characterized in that: Step 2 includes defining the efficiency of energy transfer between devices as a composite function of line impedance, device power requirements, simulation distance, and trainable parameters: In the formula, α ij is the energy attenuation coefficient from device i to j, θ1, θ2, and θ3 are training parameters, which adjust the weights of the power ratio term and the distance attenuation term respectively; Z ij is the line impedance between sensitive devices i and j, d ij is the physical transmission distance between devices, P i out With P j req are the output power of device i and the rated power of device j respectively.

4. The method for evaluating voltage sag sensitivity in the whole process of a process based on a physical guided graph neural network according to claim 1, characterized in that: Step 3 constructs the "electrical-physical" coupled PGN layer, and uses the physical information basis function constructed in step 2 as the weight update function in the edge update link of the process flow chart evaluation model.

5. The method for evaluating voltage sag sensitivity in the whole process of a process based on a physical guided graph neural network according to claim 4, characterized in that: The PGN layer uses sensitive device characteristics to update the function φ n , device connection feature update function φ e , process flow global feature update function φ g , physical energy attenuation coefficient update function φ w The information is updated during the message transmission process, and the sensitive device feature update function φ n , device connection feature update function φ e , process flow global feature update function φ g , physical energy attenuation coefficient update function φ w In this process flow evaluation model, there are multiple perceptrons MPL; sensitive equipment feature update function φ n , device connection feature update function φ e , process flow global feature update function φ g , physical energy attenuation coefficient update function φ w Each contains its own aggregation function ρ n , e , g , w , used to aggregate the information received from other devices, where n, e, g, and w represent the sensitive device node characteristics, device connection characteristics, process global characteristics, and physical energy attenuation coefficients, respectively; define I as the index set of all sensitive device node indexes; also define the index set R i and S i , are the sets of all receiver node indexes and sender node indexes of device i, and U is the index set containing all device connection relationships.

6. The method for evaluating voltage sag sensitivity in the whole process of a process based on a physical guided graph neural network according to claim 5, characterized in that: Device connection feature update function φ e Update all edges of the graph according to the following formula, that is, the device connection characteristics; use the sending node device characteristics n j , receiving node device characteristics n i and the connection feature e between the two nodes ij Generate updated device connection characteristics i ' j ; The PGN layer uses the physical energy attenuation coefficient to update the function φ w Dynamically adjust the weight according to the connection relationship between devices; select the device location as φ w Input, φ w Calculate the weight W ij , used to obtain the updated connection feature e i ' j , θ1 and θ2 are trainable parameters; 7. The method for evaluating voltage sag sensitivity in the whole process of a process based on a physical guided graph neural network according to claim 5, characterized in that: Aggregation function ρ e Aggregate the edge features adjacent to the target device node using the SUM aggregation method; then concatenate the original node features and the aggregated edge features; finally, put the concatenated device node features into a feature update network with the same structure as the edge update to complete the device node feature update.

8. The method for evaluating voltage sag sensitivity in the whole process of a process based on a physical guided graph neural network according to claim 5, characterized in that: Sensitive device node feature update function φ n Use the current input device node feature n i , Updated device connection features i ′ and the global feature g of the process flow to update the node features; The PGN layer uses the edge aggregation function ρ e Reduce the set of all connection features connected to the device to a single aggregated connection feature e i ' j ; Sensitive device node feature update function φ n Processing aggregated device connection features i ' j and the current device node feature n i To generate updated device node characteristics n i ′, θ3 are trainable parameters; the edge aggregation and node update process is:

9. The method for evaluating voltage sag sensitivity in the whole process of a process based on a physical guided graph neural network according to claim 5, characterized in that: Process global update function φ g Used to update global features; The PGN layer uses the updated edge e′ and node features n′ to update the global features; the PGN layer uses the aggregation function ρ e and ρ n , all equipment connection features E′ and all sensitive equipment node features N′ are aggregated into two vectors e′ and n′ by summing and averaging, θ4 is a trainable parameter; the global feature update process of the process is:

10. The method for evaluating voltage sag sensitivity in the whole process of a process based on a physical guided graph neural network according to claim 5, characterized in that: Physical energy attenuation coefficient W ij Update, characterize the intermediate voltage sag characteristics and the attenuation degree of the physical quantity transfer process, and update the physical energy attenuation coefficient function φ w It is used to update the edge information between nodes in the graph neural network. The expression is the constructed physical information basis function, which represents the degree of energy attenuation:

Images