An island micro-grid resilience evaluation method and system fusing multi-scene feature coding and physical constraints

Abstract

The application discloses a kind of fusion multi-scene feature coding and island micro-grid resilience evaluation method and system of physical constraint, the island micro-grid resilience evaluation method first constructs the five-dimensional mechanical resilience evaluation index system covering structural strength, anticorrosion performance, waterproof capacity, equipment health degree and load balance;Second, design multi-scene feature encoder, the heterogeneous monitoring data of different typical scenes are extracted and fused, and a unified scene representation is generated;Then, introduce the physical constraint integration layer, based on the environment parameters, apply the differentiable physical rule penalty to the prediction weight, to ensure that the weight meets the domain knowledge;Finally, the index weight vector that satisfies the normalization constraint is output through the weight prediction head, and the scores of the three comprehensive dimensions of robustness, economy and extreme weather survivability are aggregated and calculated.The application significantly enhances the resilience evaluation accuracy and scene adaptive ability of island micro-grid in complex marine environment.

Description

Technical Field

[0001] This invention belongs to the field of island microgrid resilience assessment technology, specifically involving an island microgrid resilience evaluation method and system that integrates multi-scenario feature coding and physical constraints. Background Technology

[0002] Unlike onshore power grids, island microgrids are exposed to extreme marine climates such as high humidity, high salt spray, and strong storms. Their physical infrastructure (including wind turbine towers, photovoltaic supports, and energy storage system casings) faces continuous risks of corrosion, fatigue, and mechanical damage. This mechanical degradation not only increases equipment failure rates but can also trigger system power outages, severely weakening the microgrid's survivability and resilience under extreme events. Therefore, establishing a systematic evaluation system to quantify the mechanical resilience of island microgrid physical infrastructure is of significant theoretical and engineering value for guiding the planning, design, and operation and maintenance of island microgrids and improving their ability to withstand disturbances in extreme environments. Existing microgrid resilience research mainly focuses on electrical operation strategies, such as islanding, load shedding, and network reconfiguration, while lacking a systematic integration of mechanical performance indicators for physical infrastructure (such as structural strength, corrosion resistance, and waterproofing). Regarding resilience assessment methods, scholars have proposed diverse indicators from the perspectives of complex networks and probabilistic simulations, but the core remains focused on electrical indicators such as load recovery and power supply continuity, failing to fully reflect the mechanical condition of the equipment itself. Therefore, there is an urgent need for a mechanical resilience evaluation method for island microgrids that can integrate monitoring data from multiple scenarios with knowledge from the physical domain and achieve adaptive generation of evaluation index weights. Summary of the Invention

[0003] The purpose of this invention is to provide a resilience evaluation method for island microgrids that integrates multi-scenario feature encoding and physical constraints, addressing the problems of missing mechanical performance indicators, static weight allocation, and lack of scenario adaptability in existing technologies. The core innovation of this invention lies in proposing a learnable physical constraint integration mechanism. Unlike existing technologies that apply penalty terms using fixed empirical thresholds, this invention, for the first time, defines the threshold parameters in the physical constraint formula as learnable network parameters. These parameters no longer rely on subjective presets but are optimized collaboratively during model training through backpropagation and a composite loss function, enabling the physical constraint boundaries to adaptively and dynamically adjust according to data distribution and scenario characteristics. This design significantly improves the physical rationality and generalization ability of the model under complex boundary conditions such as typhoons and high salt spray.

[0004] In a first aspect, the present invention provides a method for evaluating the resilience of island microgrids that integrates multi-scenario feature coding and physical constraints, the method comprising:

[0005] The raw monitoring data of the microgrid is collected, and basic evaluation indicators are calculated based on the raw monitoring data. The basic evaluation indicators include structural strength indicators, corrosion resistance indicators, waterproofing indicators, equipment health indicators, and load balance indicators.

[0006] A weight generation model is constructed, comprising a multi-scene feature encoder, a feature fusion module, a physical constraint integration layer, and a weight prediction head. The multi-scene feature encoder includes multiple parallel scene-specific feature extraction branches and a fusion module. The scene-specific feature extraction branches process different types of raw monitoring data, and the multi-scene information extracted by the multi-scene feature extraction branches is fused by the feature fusion module to generate a scene representation. The physical constraint integration layer is used to add physical constraints during model training. The weight prediction head maps the scene representation to a weight vector.

[0007] The basic evaluation indicators are weighted and fused based on the weight vector to generate comprehensive dimensional indicators, which complete the evaluation of the resilience level of microgrids at different levels. The comprehensive dimensional indicators include robustness, economy and extreme weather survival.

[0008] Preferably, the feature fusion module includes a feature interaction module and a multi-head attention module; the feature interaction module is used to process the combined feature matrix of multi-scene information fusion to obtain a global representation; the input of the multi-head attention module is the combined feature matrix and scene type label, which is used to dynamically focus on the features most relevant to the current scene to obtain a multi-head attention output; the global representation and the multi-head attention output are fused to obtain a scene representation.

[0009] Preferably, in the feature interaction module, the combined feature matrix is ​​processed sequentially through a first linear layer, layer normalization, activation function, and a second linear layer to obtain a global representation;

[0010] In the multi-head attention module, the scene type label is projected into a query vector, and the combined feature matrix is ​​projected into a key vector and a value vector; the multi-head attention output is obtained based on the query vector, key vector, and value vector.

[0011] Preferably, the structural strength index consists of stress-strain parameters, vibration frequency parameters, displacement parameters, and fatigue damage parameters; the corrosion resistance index consists of salinity parameters, corrosion rate parameters, coating condition parameters, and rust percentage parameters; the waterproofing index consists of rainfall parameters, humidity parameters, and wave height parameters; the equipment health index consists of health index parameters, equipment temperature parameters, and remaining life parameters; and the load balance index consists of power output parameters, stress-strain parameters, and vibration frequency parameters.

[0012] Preferably, the robustness dimension consists of all the basic evaluation indicators; the economic dimension consists of the equipment health index and the load balance index; and the extreme weather survivability dimension consists of the structural strength index, the corrosion resistance index, and the waterproof capability index.

[0013] Preferably, the weight generation model incorporates a smoothness loss for optimization during training; the smoothness loss is constructed based on the weight vectors obtained by the model at adjacent time points and is used to smooth out the changes in weights over time.

[0014] Preferably, the physical constraints include structural strength constraints, corrosion resistance constraints, waterproofing constraints, load balancing constraints, and scenario-specific constraints.

[0015] Preferably, the scene type label is obtained based on the current environmental parameter data and the corresponding physical threshold.

[0016] Preferably, the scene-specific feature extraction branch includes a three-layer fully connected network; the first fully connected network includes a linear layer, a layer normalization layer, an activation function, and a Dropout layer connected in sequence; the second fully connected network includes a linear layer, an activation function, and a Dropout layer connected in sequence; and the third fully connected network uses a linear layer.

[0017] Preferably, the weight prediction head includes a three-layer fully connected network; the first fully connected network includes a fully connected layer, a layer normalization, and an activation function connected in sequence; the second fully connected network includes a fully connected layer, a Dropout layer, and an activation function connected in sequence; and the third fully connected network includes a fully connected layer and a Softmax function connected in sequence.

[0018] Secondly, this invention provides a resilience evaluation system for island microgrids that integrates multi-scenario feature encoding and physical constraints, used to execute the aforementioned island microgrid resilience evaluation method. The island microgrid resilience evaluation system includes a data acquisition module, an index calculation module, a weight generation module, and a resilience evaluation module. The data acquisition module is used to collect monitoring data of the island microgrid. The index calculation module is used to obtain basic evaluation indicators. The weight generation module has an embedded weight generation model, used to adaptively generate weight vectors for the basic evaluation indicators based on the monitoring data of the current scenario. The resilience evaluation module is used to output comprehensive dimensional indicators based on the basic evaluation indicators and the weight vectors.

[0019] The beneficial effects of this invention are:

[0020] 1. This invention constructs a five-dimensional mechanical toughness evaluation index system covering structural strength, corrosion resistance, waterproofing, equipment health, and load balance. It comprehensively characterizes the toughness features of island microgrids from two levels: equipment status and system operation. It overcomes the limitations of existing technologies that only focus on electrical operation indicators and provides a scientific basis for assessing the physical disaster resistance capabilities of island microgrids in harsh environments.

[0021] 2. This invention employs a weight generation model for multi-scenario adaptive weight generation. It fuses heterogeneous input data through a multi-scenario feature encoder, introduces a physical constraint layer to ensure weight rationality, and designs a composite loss function to balance data fitting and domain knowledge. This achieves end-to-end dynamic allocation from raw monitoring data to evaluation index weights. Compared to traditional static weighting methods, this method can automatically adjust index weights according to different operational scenarios such as typhoons, high salt spray, and heavy rain, significantly improving the scenario adaptability of the weights.

[0022] 3. This invention integrates domain knowledge into model training as a differentiable loss term through a physical constraint integration layer, ensuring that the model output conforms to basic physical laws and operational experience. Experimental results show that the method of this invention outperforms traditional methods in terms of robustness and survival in extreme weather, verifying its adaptive capability and physical rationality in complex environments. Attached Figure Description

[0023] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used according to the embodiments will be briefly described below. Obviously, the descriptions in the drawings are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is the overall flowchart of the present invention.

[0025] Figure 2 This is a schematic diagram illustrating the basic evaluation indicators of the mechanical resilience of the island microgrid in this invention.

[0026] Figure 3 This is a diagram showing the overall architecture of the weight generation model in this invention.

[0027] Figure 4 The following is a distribution diagram of the scene composition and environmental parameters of the dataset in this invention; wherein, (a) is a pie chart of scene composition ratio, (b) is a histogram of wind speed distribution, (c) is a two-dimensional scatter plot of salinity and humidity, (d) is a box plot of rainfall intensity for each scene, (e) is a distribution diagram of expert weights divided by scenario, and (f) is a distribution diagram of equipment health and remaining life.

[0028] Figure 5The diagram shows the weight comparison of the three weight generation methods in this invention under various scenarios; where (a) is the structural strength weight comparison diagram, (b) is the corrosion resistance performance weight comparison diagram, (c) is the waterproof capability weight comparison diagram, (d) is the equipment health weight comparison diagram, and (e) is the load balance weight comparison diagram.

[0029] Figure 6 A three-dimensional bar chart comparing the simulation results of the present invention and the comparative method in three dimensions: robustness, economy, and extreme weather survival. Detailed Implementation

[0030] The present invention will be further described below with reference to the accompanying drawings.

[0031] A resilience evaluation method for island microgrids integrating multi-scenario feature encoding and physical constraints is proposed. The island microgrid resilience evaluation system includes a data acquisition module, an index calculation module, a weight generation module, and a resilience evaluation module. The data acquisition module collects structural response data, environmental parameter data, and equipment status data for key facilities in the island microgrid. The index calculation module calculates scores for structural strength, corrosion resistance, waterproofing, equipment health, and load balance based on the collected data. The weight generation module embeds a weight generation model to adaptively generate weight vectors for the five basic evaluation indices based on monitoring data from the current scenario. The resilience evaluation module calculates scores for robustness, economic efficiency, and extreme weather survivability dimensions based on the basic evaluation index scores and weight vectors, outputting the resilience evaluation results.

[0032] like Figure 1 As shown, the resilience evaluation method for island microgrids includes the following steps:

[0033] Step 1: Constructing an evaluation index system for the mechanical resilience of island microgrids

[0034] like Figure 2 As shown, the evaluation index system defines basic evaluation indicators from five aspects: structural status, corrosion resistance, waterproofing, equipment health, and load balance of key microgrid facilities. These indicators reflect the technical status of the microgrid's mechanical system from different perspectives. The calculation of each basic evaluation indicator relies on several directly or indirectly measurable raw monitoring data. Based on this, the data is further aggregated into three comprehensive dimensions: robustness, economy, and extreme weather survivability. All indicator scores are normalized to... Within a given range, higher scores indicate better performance. The specific methods for constructing the basic evaluation indicators and comprehensive dimensional indicators are as follows:

[0035] 1-1. Basic Evaluation Indicators

[0036] 1) Structural strength index

[0037] Structural strength indices are used to measure the ability of key load-bearing components of a microgrid (such as wind turbine towers, photovoltaic supports, and transmission towers) to resist mechanical loads and fatigue damage. The calculation formula is:

[0038]

[0039] in, The stress-strain parameters are obtained by normalizing the real-time strain values ​​collected by the strain sensor. The vibration frequency parameter is obtained by normalizing the dominant frequency of structural vibration acquired from the accelerometer. The displacement parameter is obtained by normalizing the relative displacement of key points monitored by a laser displacement meter. The fatigue damage parameters are obtained by normalizing the cumulative damage factor extracted from the stress-time history using the rainflow counting method. Weight allocation. This demonstrates the dominant role of stress and strain in structural safety assessment.

[0040] 2) Corrosion resistance performance indicators

[0041] Corrosion resistance performance indicators are used to evaluate the ability of equipment to resist salt spray corrosion. The calculation formula is:

[0042]

[0043] in, The salinity parameter is normalized from the salinity meter measurement. The corrosion rate parameter was determined and normalized using corrosion test pieces or electrochemical sensors. The coating condition parameters are normalized by scoring through a coating thickness gauge or visual inspection. The percentage of corrosion is a parameter normalized to the proportion of corroded area on the component surface. Since higher salinity, corrosion rate, and percentage of corrosion correlate with poorer corrosion resistance, the normalized value is subtracted from the value by 1 for conversion. Weighting This highlights the dominant role of salinity and corrosion rate.

[0044] 3) Waterproofing capability index

[0045] Waterproofing performance indicators assess the sealing and weather resistance of microgrid equipment under conditions of rainfall, high humidity, and wave impact. Waterproofing performance indicators The calculation formula is:

[0046]

[0047] in, The rainfall parameter is normalized from rain gauge records; The humidity parameter is monitored and normalized using a humidity sensor. The wave height parameter is obtained by normalizing the data using a wave buoy or radar wave meter.

[0048] 4) Equipment health indicators

[0049] Equipment health indicators are used to evaluate the current technical condition and remaining lifespan of critical power equipment (such as inverters, energy storage batteries, diesel generators, etc.). The calculation formula is:

[0050]

[0051] in, The health index parameters are output and normalized by the equipment management system. For equipment temperature parameters, the normalized values ​​of the operating temperatures of key components; The remaining service life parameter is the normalized value of the proportion of remaining service time to design life predicted based on the degradation model. The health index is assigned the highest weight of 0.50, reflecting its core position in the overall condition characterization.

[0052] 5) Load balance index

[0053] Load balancing metrics are used to measure the uniformity of power distribution among different power sources and loads in a microgrid. The calculation formula is:

[0054]

[0055] in, The power output parameters are normalized from the meter measurements. The standard deviation function indicates that the larger the standard deviation, the more uneven the load distribution. Therefore, the balance score is obtained by subtracting the standard deviation from 1.

[0056] 1-2. Calculation of Comprehensive Dimensional Indicators

[0057] 1) Robustness dimension

[0058] Robustness dimension reflects the ability of a microgrid to maintain stable operation under normal and disturbance conditions. The weighted sum of the scores for all basic indicators is expressed as follows:

[0059]

[0060] in, These are the weighting coefficients for different basic evaluation indicators.

[0061] 2) Economic dimension

[0062] Economic dimension The focus is on evaluating the operating costs and maintenance efficiency of microgrids, which is expressed as follows:

[0063]

[0064] in, The maintenance factor is obtained by normalizing data such as historical maintenance frequency and spare parts consumption.

[0065] 3) Survival dimension in extreme weather

[0066] Extreme weather survival dimension It is obtained by a weighted calculation of structural strength, corrosion resistance, and waterproofing capability, and is expressed as follows:

[0067]

[0068] Step 2: Construct the weight generation model (PC-AWNet)

[0069] A physically constrained weight generation model is constructed to generate weight vectors for five basic evaluation indicators from multi-source heterogeneous monitoring data in an end-to-end manner. The input to the weight generation model is the original monitoring data, and the output is the weight vectors of the five basic evaluation indicators. The raw monitoring data includes physical parameter data and environmental parameter data. and , .

[0070] like Figure 3 As shown, the weight generation model adopts an encoder-predictor structure, which includes a multi-scene feature encoder, a feature fusion module, a physical constraint integration layer, and a weight prediction head. The multi-scene feature encoder extracts raw monitoring data for four scenarios: typhoon, high salt spray, rainstorm, and normal conditions, and then fuses them through the feature fusion module to generate a unified scene representation. The physical constraint integration layer applies physical rules to the environmental parameters and scene representation to ensure the rationality of the predicted weights. The weight prediction head maps the scene representation to a weight vector and uses Softmax to ensure normalization.

[0071] 2-1 Multi-scene feature encoder

[0072] The dimensions and types of monitoring data available for island microgrids vary depending on the specific scenario. For example, typhoon scenarios focus on structural response data such as wind speed and vibration; high salt spray scenarios focus on environmental parameters such as salinity and corrosion rate. For each scenario... Define its input dimensions For example, a typhoon scene contains 10-dimensional features, while a high-salt spray scene has 8-dimensional features. A dedicated feature extraction branch is constructed for each scene. ,in The hidden layer dimension is set to 256 in the code.

[0073] The multi-scene feature encoder includes multiple parallel scene-specific feature extraction branches; these branches are divided into typhoon, high salt spray, rainstorm, and normal branches. The input data for each scene-specific feature extraction branch is shown in Table 1.

[0074] Table 1 Input data for scene-specific feature extraction branch

[0075]

[0076]

[0077] Regarding the input design for each branch in a multi-scene feature encoder, two points need to be clarified: First, the input sequences for different scene branches are different—including differences in both dimensionality and the physical meaning of the features. Therefore, each branch must independently process its unique input vector. Second, during validation or actual deployment, the model requires each sample to be pre-labeled with a scene tag (i.e., the `current_scenario` parameter). This is because forward propagation needs to feed the sample into the corresponding branch, and simultaneously generate an attention query vector based on the scene tag (mapped to the latent space through one-hot scene encoding).

[0078] In this embodiment, scene labels are obtained using physical thresholds of environmental parameters. Specifically, when the wind speed exceeds 15 m / s (level 8 wind), it is determined to be a typhoon scene; when the salinity is higher than 35 ppt and the humidity is greater than 80%, it is determined to be a high salt spray scene; when the rainfall exceeds 50 mm / h, it is determined to be a rainstorm scene; otherwise, it is a normal scene.

[0079] In some embodiments, end-to-end automatic scene recognition is achieved through joint training of the scene classification module and the weight prediction network.

[0080] The structure of multiple scene-specific feature extraction branches is identical, all consisting of a three-layer fully connected network. In the scene-specific feature extraction branch, the input first passes through a linear layer to map to the hidden dimension, then layer normalization is applied to stabilize training, followed by the introduction of non-linearity through the GELU activation function, and then a Dropout layer to prevent overfitting; finally, a linear layer reduces the input to the hidden dimension. Dimension, reactivated with Dropout; finally restored to its original state through a third linear layer. The output dimension provides specific scene information. This design allows each scene branch to learn scene-specific feature representations, while unifying the output dimension to the same hidden space, resulting in a set of feature tensors. This facilitates subsequent integration.

[0081] 2-2. Feature Fusion Module

[0082] The feature fusion module includes a feature interaction module and a multi-head attention module. The feature interaction module is used to fully interact with information from different scenarios to obtain a global representation that integrates information from multiple scenarios. The input to the feature interaction module is a combined feature matrix obtained by concatenating information from multiple scenes along the scene dimension. The multi-head attention module is used to dynamically focus on the features most relevant to the current scene, obtaining the multi-head attention output. Global representation and multi-head attention output After fusion, layer normalization is performed to obtain the scene representation. It integrates information from multiple scenarios and highlights the specific needs of the current scenario, laying the foundation for subsequent weight generation. Scenario Representation This can be expressed by the formula:

[0083]

[0084] In the feature interaction module, the feature matrix is ​​combined sequentially through a first linear layer, layer normalization, GELU activation function, and a second linear layer. Processing is performed to obtain a global representation. .

[0085] In the multi-head attention module, the current scene label is mapped to the hidden dimension through a linear projection layer to obtain the query vector. Combine the feature matrix Projection is performed using key and value vectors; multi-head attention output is computed based on the query vector, key vector, and value vector. This achieves a weighted sum of the importance of other scene features relative to the current scene.

[0086] 2-3. Physical Constraint Integration Layer

[0087] The physical constraint integration layer is used to overcome the poor adaptability of traditional fixed-threshold physical constraints. The physical constraint integration layer not only applies penalties to the prediction weights based on environmental monitoring data (such as wind speed, salinity, humidity, and rainfall intensity), but more importantly, it parameterizes the constraint boundary to ensure that it can maintain physical rationality under extreme conditions, thus achieving a deep integration of physical knowledge and data-driven approaches.

[0088] The environmental parameter vector is denoted as Including wind speed ,salinity ,humidity Rainfall intensity Etc. Define a learnable threshold parameter. The initial values ​​are set empirically (0.3, 0.15, 0.25, 0.2 respectively) and automatically optimized during training using gradient descent. A penalty coefficient is also set. Used to balance the importance of different constraints.

[0089] The degree of violation of each constraint is expressed by the ReLU function. Quantification, specifically defined as follows:

[0090] 1) Structural strength constraints

[0091] When wind speed Exceeding the typhoon threshold ( When this is the case, the structural strength weight is required to be no less than [amount missing]. Otherwise, a penalty will be imposed:

[0092]

[0093] in, For structural strength constraints; This is an indicator function that takes the value 1 if the condition is true, and 0 otherwise.

[0094] 2) Corrosion resistance constraints

[0095] When salinity More than 35 ppt and humidity When the percentage exceeds 80%, the weight of corrosion resistance should be at least one learning margin greater than the weight of waterproofing capability. , The violation is as follows:

[0096]

[0097] in, These are the performance constraints for corrosion resistance.

[0098] 3) Waterproofing constraints

[0099] When rainfall intensity Exceeding the rainstorm threshold ( When applying this requirement, the weight of waterproof capability should not be less than [a certain percentage]. :

[0100]

[0101] 4) Load balance constraint

[0102] In any scenario, the load balancing weight should be maintained within a reasonable range. Internally, the two-way penalty is:

[0103]

[0104] 5) Scenario-specific constraints

[0105] Additional constraints are imposed for different typical scenarios to strengthen the relative importance of specific indicators:

[0106]

[0107] In high-salt-spray environments, the impact on microgrid equipment is primarily through electrochemical corrosion. Corrosion directly affects the equipment itself, leading to material degradation, reduced insulation, and shortened lifespan. Therefore, there is a direct causal relationship and trade-off between corrosion resistance and equipment health, requiring corrosion resistance to be weighted more heavily than equipment health.

[0108]

[0109] In heavy rain scenarios, the weight of waterproofing capability should be higher than the average weight of the other indicators:

[0110]

[0111] By summing the above violations with weights, we obtain the total physical constraint loss:

[0112]

[0113] In this embodiment, the penalty coefficients for physical constraints are set as follows: structural strength constraint penalty coefficient 0.5, corrosion resistance constraint penalty coefficient 0.3, waterproof capability constraint penalty coefficient 0.4, load balance constraint penalty coefficient 0.2, and scenario-specific constraint penalty coefficient 0.3.

[0114] 2-4. Weighted Prediction Head

[0115] The weight prediction head consists of three fully connected layers. The first fully connected layer connects the weights to the weights of the predicted weights. The first layer maps the feature values ​​to 128 dimensions, followed by layer normalization and GELU activation. The second fully connected network reduces the dimensions to 64, also passing through GELU and Dropout. The final fully connected network outputs a 5-dimensional log-odds probability, which is then normalized using the Softmax function to obtain the weight vector. :

[0116]

[0117]

[0118]

[0119] In this embodiment, the core parameters of the weight generation model are: hidden layer dimension HIDDEN_DIM = 256, used to unify the feature representation of each scene; attention heads ATTENTION_HEADS = 4, used for multi-scene feature fusion; dropout rate DROPOUT_RATE = 0.2, to prevent overfitting; and output dimension NUM_CRITERIA = 5, corresponding to the weights of the five evaluation metrics.

[0120] Step 3: Train the weight generation model

[0121] Step 3-1. Construct the dataset

[0122] Since real long-term operational data of island microgrids is difficult to obtain and authoritative expert weight labels are lacking, this embodiment uses a simulation-based generation method to construct the dataset. Based on the physical characteristic ranges of four typical scenarios (typhoon, high salt spray, rainstorm, and normal), the dataset generates raw monitoring data containing physical and environmental parameters for each sample, and assigns a set of expert weights as supervisory labels according to the scenario type.

[0123] In some embodiments, the methods for generating supervision labels in different typical scenarios are as follows:

[0124] 1) Typhoon scenario: Structural strength has the highest weight (approximately 0.35), reflecting the primary importance of structural safety under strong wind loads; waterproofing capability and equipment health are secondary; corrosion resistance has the lowest weight (approximately 0.10).

[0125] 2) High salt spray scenario: Corrosion resistance has the highest weight (about 0.35), reflecting the dominant impact of salt spray corrosion on equipment life; equipment health is the second highest; structural strength has a moderate weight.

[0126] 3) Rainstorm scenario: Waterproofing capability has the highest weight (approximately 0.30), followed by equipment health (approximately 0.25), and structural strength has a moderate weight.

[0127] 4) Normal scenario: The weight distribution is relatively balanced, with structural strength and equipment health each around 0.25, and the remaining indicators between 0.15 and 0.20.

[0128] In some embodiments, to increase the authenticity of the data, 5% random noise is added during the generation process, and the weights are normalized to ensure that the sum of the weights in each scenario is 1.

[0129] The dataset contains 5000 samples, distributed across the following scenarios: 1000 typhoon scenarios, 1000 high-salt-fog scenarios, 1000 rainstorm scenarios, and 2000 normal scenarios. The data is randomly divided into training, validation, and test sets at a ratio of 70%, 15%, and 15%, respectively. Each sample contains a complete original feature vector and a corresponding expert weight vector for subsequent model training and comparison.

[0130] Step 3-2. Construct the loss function

[0131] To drive the model to learn accurate weights and obey physical rules, a composite loss function was designed. It comprises multiple components, each focusing on different optimization objectives. By minimizing the composite loss function, the model aims to fit the data while satisfying physical constraints as much as possible and maintaining temporal smoothness. Composite Loss Function Represented as:

[0132]

[0133] in, The main loss; Loss due to physical constraints; For smoothness loss; For L2 regularization terms; These are hyperparameters, and are set to 0.5, 0.1, and 0.05 respectively.

[0134] Main loss Used to measure generation weight With target weight The differences between them. The main loss is a weighted combination of three parts: KL divergence measures the distance between the two probability distributions; mean squared error directly measures the Euclidean distance between the weight vectors; and cosine similarity loss encourages generation in the same direction as the target. The specific expression is:

[0135]

[0136] Smoothing loss To encourage weights to change smoothly over time (avoiding drastic jumps), it is represented as:

[0137]

[0138] in, The weights generated at the current moment; The weights generated in the previous time step.

[0139] The model employs an end-to-end training approach. During forward propagation, the encoder fuses multi-scene information, predicts the output weights of the head, and the physical constraint layer calculates weights based on the current scene environment parameters. Finally, the gradient is calculated using the total loss function, and the parameters are updated. During training, the model automatically adjusts the learnable thresholds in the constraint layers to adapt them to the data distribution.

[0140] In actual operation, the monitoring data of the island microgrid is collected continuously over time, and the weights at adjacent moments should generally not change drastically. To capture this temporal smoothness, the model maintains a scene memory dictionary. This dictionary uses scene names as keys and lists of values ​​to record the weight vectors generated for that scene in the last 10 iterations. Each time the model generates a new weight, it adds the currently output weight to the memory list for the corresponding scene and automatically removes the oldest record that is more than 10.

[0141] During the training phase, if the samples have a clear temporal order, such as monitoring data from multiple consecutive days of the same microgrid, a smoothing loss can be calculated using a memory mechanism. Specifically, for the current sample, the most recently confirmed historical weight of the same scenario is retrieved from the memory dictionary, and then the mean squared error between the currently generated weight and that historical weight is calculated as the smoothing loss term. This loss term is added to the total loss to encourage the model's output to remain stable across adjacent time steps. It should be noted that if there is no temporal correlation between samples, such as independent samples from different microgrids, the smoothing loss will not be used.

[0142] The memory mechanism also plays a role during the inference phase. The `get_scenario_history` method can retrieve the historical weight evolution trajectory of a specific scenario, allowing analysis of weight trends over time. For example, during typhoon season, it can be observed whether structural strength weights dynamically adjust with wind speed fluctuations; in long-term operation, it can assess whether the model's response to equipment health weights aligns with the equipment aging process. These analyses help verify the rationality of the model's behavior and provide a reference for operational and maintenance decisions.

[0143] In this embodiment, the model training hyperparameters are set as follows: batch size BATCH_SIZE = 32, balancing training speed and memory consumption; initial learning rate LEARNING_RATE = 1e-3, using the Adam optimizer; training epochs EPOCHS = 100, with an early stopping mechanism (patience=10) to prevent overfitting; and weight decay WEIGHT_DECAY = 1e-4 for L2 regularization.

[0144] Step 4: Toughness Evaluation

[0145] After completing the training and weight generation of the weight generation model, this step will comprehensively evaluate the mechanical resilience of the island microgrid based on the evaluation index system constructed in step one and the dynamic weight vector generated in step three. The resilience evaluation process includes three stages: basic index score calculation, weight vector acquisition, and comprehensive dimensional index aggregation. The final output is a quantitative result reflecting the resilience level of the system at different levels.

[0146] Specifically, for any monitoring sample to be evaluated, the structural strength index is first calculated from the original monitoring data one by one according to the calculation formula in step one. Corrosion resistance indicators Waterproofing performance indicators Equipment health indicators and load balance index The scores are calculated based on their respective underlying physical parameters, including stress and strain, vibration frequency, salinity, corrosion rate, rainfall, humidity, health index, remaining life, and power output. All parameters are normalized before calculation to ensure that the scores of each indicator are of uniform dimension. Range. A higher score for the basic indicator indicates better mechanical performance in that dimension.

[0147] After the basic indicator scores are calculated, the monitoring data of the current sample is input into the pre-trained weight generation model. The model adaptively outputs the weight vectors of the five basic indicators based on the scene type (typhoon, high salt spray, rainstorm, or normal) and its environmental parameter characteristics. This weight vector reflects the relative contribution of each indicator to the overall resilience of the system under the current operating environment. All weights are non-negative and their sum is 1, satisfying the basic constraints of comprehensive evaluation.

[0148] Based on this, the scores of the basic indicators are weighted and aggregated with the weight vector to calculate the scores of the three comprehensive dimensions. Robustness dimension It integrates all five basic indicators to measure the ability of a microgrid to maintain stable operation under normal and disturbance conditions; economic dimension The evaluation focuses on the system's operating costs and maintenance efficiency, with equipment health and load balancing playing a dominant role; extreme weather survivability is also a key dimension. Focusing on the system's survivability under extreme weather events such as typhoons, torrential rains, and storm surges, the evaluation is based on a weighted average of three indicators most directly related to weather resistance: structural strength, corrosion resistance, and waterproofing. The scores for these three dimensions are also normalized. The higher the score in a given range, the better the resilience level in that dimension.

[0149] Step 5: To verify the effectiveness of the PC-AWNet-based mechanical resilience evaluation method for island microgrids proposed in this invention, this embodiment designs two sets of interrelated experiments: a weight identification experiment and a simulation experiment. The weight identification experiment aims to compare the degree of agreement between the weight vectors output by the PC-AWNet method, the CRITIC method, and the entropy weight method in four typical scenarios and expert experience. The simulation experiment calculates the scores of five indicators for each sample based on actual monitoring data, and then aggregates them into three comprehensive dimensions. The performance differences of each dimension under the three weights are compared to verify the advantages of the PC-AWNet method in resilience assessment.

[0150] Step 5-1. Dataset Generation and Statistical Analysis

[0151] To support model training and experimental verification, this embodiment first simulates and generates a dataset of island microgrids encompassing four typical operating environments: typhoon, high salt spray, heavy rain, and normal conditions. In the dataset, the normal scenario accounts for the highest proportion (44.8%, approximately 224 samples), forming the baseline operating condition; the typhoon scenario is second, accounting for 24.4% (approximately 122 samples); the heavy rain scenario accounts for 16.0% (approximately 80 samples); and the high salt spray scenario accounts for the lowest proportion (14.8%, approximately 74 samples). Extreme scenarios (typhoon, heavy rain, and high salt spray) together account for 55.2%, roughly a 55:45 distribution with the normal scenario. This ensures both a performance benchmark under normal operating conditions and provides sufficient extreme stress samples for analyzing the impact of environmental factors on microgrid resilience. This distribution aligns with the objective law that extreme weather events have a low probability of occurrence but significant impact in the actual operation of island microgrids, and also provides sufficient data support for subsequent adaptive weight learning, such as... Figure 4 As shown in (a).

[0152] From the distribution characteristics of key environmental parameters, such as Figure 4 As shown in (b), the main peak of wind speed data is located in the range of 5 to 15 m / s, with the highest frequency (approximately 40 samples) between 10 and 12.5 m / s, corresponding to common light to moderate wind conditions. As the wind speed increases, the number of samples gradually decreases, but approximately 25% of the samples are still retained above 15 m / s (the threshold for level 8 winds), with the highest wind speed extending to 35 m / s, ensuring coverage of extreme wind loads in typhoon scenarios. Figure 4Figure (c) shows a two-dimensional scatter plot of salinity and humidity. Samples from high-salt-fog scenarios clearly cluster in the range of salinity (35–42.5 ppt) and humidity (80%–105%), reflecting the typical characteristics of "high salinity + high humidity." Rainstorm scenarios are dominated by high humidity (80%–105%) and moderate salinity (30–37.5 ppt). Typhoon scenarios also have high humidity (85%–105%) but slightly lower salinity (29–35 ppt). Samples from normal scenarios are mostly located in the region of salinity (25–35 ppt) and humidity (50%–85%). This figure quantitatively reveals the separability of different scenarios in the salinity-humidity space, providing data support for the differentiated weighting design of corrosion-resistant and waterproof indicators. Figure 4 (d) Box plots of rainfall intensity were drawn according to different scenarios. In the heavy rain scenario, the median rainfall was approximately 80 mm / h, with the upper quartile and maximum values ​​all exceeding the heavy rain threshold of 50 mm / h, reaching a maximum of 120 mm / h. In the typhoon scenario, the median rainfall was approximately 40 mm / h, ranging from 20 to 90 mm / h. In the high salt fog scenario, the majority of rainfall was below 50 mm / h, with a median of approximately 15 mm / h. In the normal scenario, rainfall was extremely low, with a median close to 0 mm / h. The box plots clearly show that rainfall intensity is the key variable distinguishing between heavy rain / typhoon and normal / high salt fog scenarios, providing a quantitative basis for the scenario-dependent weighting of waterproofing capabilities.

[0153] Scenario-dependent analysis of expert weights shows that, for example Figure 4 As shown in Figure (e), the average weighting of the five evaluation indicators by experts in each scenario follows domain experience: structural strength has the highest weight in the typhoon scenario (approximately 0.35), reflecting the primary importance of structural safety under strong wind loads; corrosion resistance has a significantly prominent weight in the high salt spray scenario (approximately 0.35), reflecting the dominant impact of salt spray corrosion on equipment lifespan; waterproofing and equipment health have relatively high weights in the rainstorm scenario (approximately 0.30 and 0.25, respectively), reflecting the requirements of heavy rainfall and humid environments for the waterproofing and reliability of electrical equipment; the weight distribution is relatively balanced in the normal scenario, with structural strength and equipment health each around 0.25, and the remaining indicators ranging from 0.15 to 0.20. The sum of the weights for each scenario is 1, and there is an interpretable shift in weighting depending on the scenario type, providing physically meaningful target labels for supervised learning.

[0154] Correlation analysis between equipment health index and remaining lifespan, such as... Figure 4As shown in (f), the normal scenario samples are mainly concentrated in the region with a health index of 0.65 to 1.0 and a remaining lifespan of 15 to 25 years, showing a significant positive correlation and forming a baseline cluster of "high health - long lifespan". In contrast, the samples from rainstorm, high salt spray, and typhoon scenarios generally shift towards the region of low health (0.3 to 0.7) and short lifespan (5 to 17 years), while still maintaining a positive correlation trend. This indicates that extreme environments statistically correspond to worse equipment condition and shorter expected lifespan, consistent with the physical expectation that "harsh environments accelerate aging". This relationship not only verifies the rationality of the data generation but also provides a practical basis for the weighting of equipment health and load balance in resilience assessment.

[0155] Step 5-2. Weight Identification Result Analysis

[0156] In the weight recognition experiment, the CRITIC method, entropy weight method, and the PC-AWNet method proposed in this invention were used to calculate the weights of five evaluation indicators (structural strength, corrosion resistance, waterproofing, equipment health, and load balance) under four scenarios: typhoon, high salt spray, rainstorm, and normal conditions. The weight differences of the three methods under various scenarios are shown below. Figure 5 As shown in Tables 2 to 5, the weight allocation results for each scenario are listed respectively.

[0157] (1) In the typhoon scenario, the weight allocation results of the three methods are shown in Table 2.

[0158] Table 2. Weighting results of the three methods in typhoon scenarios.

[0159]

[0160] As shown in Table 2, the PC-AWNet method assigned the highest weight of 0.347 to structural strength, significantly higher than the CRITIC method's 0.183 and the entropy weight method's 0.230, indicating that the PC-AWNet model successfully captured the dominant role of structural strength under typhoon conditions. Meanwhile, the PC-AWNet method significantly reduced the weight of corrosion resistance to 0.102, consistent with the physical expectation that typhoons are primarily driven by wind rather than corrosion. The relatively balanced weight distributions of the CRITIC and entropy weight methods failed to highlight the extreme importance of structural strength, stemming from the fact that traditional methods rely solely on data variability and correlation without incorporating prior knowledge of the scenario. Notably, the entropy weight method assigned the highest weight of 0.234 to waterproofing capability, likely because the rainfall accompanying typhoons causes significant differences in this indicator among samples, which is amplified by the entropy weight method.

[0161] (2) The weight allocation results of the three methods in the high salt spray scenario are shown in Table 3.

[0162] Table 3. Weighting results of the three methods in high salt spray scenarios.

[0163]

[0164] Because high salt spray environments are typically accompanied by a combination of factors such as high humidity, salt spray deposition, and equipment insulation degradation, waterproofing, equipment health, and structural strength are also crucial factors affecting the resilience of island microgrids, in addition to corrosion resistance. Although the PC-AWNet method has a relatively low weighting for corrosion resistance in this scenario, it outperforms the CRITIC and entropy weight methods in waterproofing, equipment health, and structural strength. Furthermore, it appropriately reduces the weighting of load balancing, an indirect environmental tolerance indicator, making the weighting more focused on indicators related to equipment protection, condition maintenance, and structural reliability under high salt spray conditions. In contrast, the CRITIC method has a more balanced overall weighting distribution, making it difficult to highlight key influencing factors in complex marine environments; while the entropy weight method increases the weighting of waterproofing, its load balancing weighting remains too high. Overall, the PC-AWNet method better balances moisture intrusion, equipment health, and structural reliability in high salt spray scenarios, and its comprehensive weighting results are superior to the other two methods.

[0165] (3) In the rainstorm scenario, the weight allocation results of the three methods are shown in Table 4.

[0166] Table 4. Weighting results of the three methods in the rainstorm scenario.

[0167]

[0168] As shown in Table 4, the PC-AWNet method again prioritizes waterproofing (0.296), while also giving high weight to equipment health (0.248), which aligns with the physical characteristics of a rainstorm environment. However, the weight of corrosion resistance remains relatively low (0.109), contrasting with the CRITIC method (0.218) and the entropy weight method (0.163). This may be because the direct correlation between corrosion-related indicators in the data and the rainstorm scenario is weak, causing the model to fail to consider corrosion resistance as a key factor. The entropy weight method also emphasizes waterproofing (0.257) and equipment health (0.225), but its structural strength weight is relatively low. The CRITIC method has relatively even weights across multiple indicators, with slightly higher weights for equipment health (0.220) and corrosion resistance (0.218).

[0169] (4) Under normal circumstances, the weight allocation results of the three methods are shown in Table 5.

[0170] Table 5. Weight allocation results of the three methods in normal scenarios.

[0171]

[0172] As shown in Table 5, the PC-AWNet method ranks structural strength and equipment health highest (both 0.236), reflecting the fundamental role of structural integrity and equipment condition in system toughness under normal operating conditions. Corrosion resistance (0.202) and waterproofing capability (0.188) are next, while load balance is the lowest (0.138). The entropy weight method significantly emphasizes waterproofing capability (0.288), which may be due to the large variation in waterproofing capability of samples under normal scenarios. The CRITIC method has a relatively uniform weight distribution and slightly higher corrosion resistance (0.223). Compared with expert prior (relatively balanced weights under normal scenarios), the weight distribution of the PC-AWNet method is more in line with expectations, while the entropy weight method excessively amplifies waterproofing capability due to data fluctuations.

[0173] A comprehensive comparison of the weights across the four scenarios reveals that the PC-AWNet method exhibits stronger scenario adaptability: in the typhoon scenario, it successfully increased the structural strength weight to 0.347, significantly higher than traditional methods; in the rainstorm scenario, it prioritized waterproofing and equipment health; and in the normal scenario, it assigned equal weights to structural strength and equipment health. This indicates that by introducing physical constraints and multi-scenario feature encoding, the model can learn the non-linear relationship between scenario and indicator importance. Traditional methods tend to have an average weight distribution, making them susceptible to data fluctuations. The CRITIC method and entropy weight method show relatively small weight differences across scenarios and fail to highlight structural strength in the typhoon scenario or corrosion resistance in the high salt spray scenario. The entropy weight method is particularly susceptible to the degree of variation in a single indicator (e.g., waterproofing is overemphasized in normal scenarios), while the CRITIC method, although balancing contrast strength and conflict, still lacks physical guidance in extreme scenarios. Compared with the expert weight distribution, the PC-AWNet method's weight ranking in typhoon, rainstorm and normal scenarios is basically consistent with the expert prior, while the traditional method deviates to some extent. This verifies the effectiveness of the PC-AWNet model in integrating data-driven and domain knowledge.

[0174] Step 5-3. Simulation Comparison Analysis of Three Weighting Methods

[0175] To verify the effectiveness of the proposed PC-AWNet weighting model, this embodiment selects three weight generation mechanisms—CRITIC method, entropy weight method, and PC-AWNet method—and conducts simulation comparisons on the same sample data (N=363) of a microgrid on an island. The comparison dimensions include robustness, economy, and extreme weather survivability, and quantitative analysis is performed based on the statistical characteristics of the indicators. The statistical results of the five basic indicators are shown in Table 6.

[0176] Table 6 Statistical Results of Five Basic Indicators

[0177]

[0178] As shown in Table 6, from the perspective of dispersion, load balance has the largest standard deviation (0.2430), indicating the highest volatility; equipment health and waterproofing capability are the next most volatile; structural strength and corrosion resistance are relatively stable. This means that from a data-driven perspective, the impact of load balance on system performance has greater uncertainty and dynamism.

[0179] The weight vectors calculated by the three methods on the overall sample are shown in Table 7.

[0180] Table 7 shows the weight vectors calculated on the overall sample using the three methods.

[0181]

[0182] As shown in Table 7, the CRITIC method exhibits a relatively uniform weight distribution, with a maximum difference of only 0.026. This method considers both the comparative strength (standard deviation) and conflict (correlation) between indicators, thus exhibiting a "statistically balanced" characteristic and not highlighting a single core indicator. The entropy weight method shows a significantly higher weight concentration, with load balance having the highest weight (0.312), followed by waterproof capability (0.216), and a maximum-to-minimum difference of 0.166. Since load balance has the largest standard deviation (0.2430), the entropy weight method identifies it as the indicator with the most information based on the principle of information entropy, and therefore assigns it the highest weight. However, the entropy weight method only relies on the degree of data dispersion and cannot distinguish the true causal contribution of indicators to system survivability. The PC-AWNet method exhibits a "structure + state-dominated" weight distribution: structural strength (0.253) and equipment health (0.246) have the highest weights, while the load balance weight drops significantly to 0.150. This indicates that the PC-AWNet model does not simply allocate weights based on indicator volatility, but rather learns which indicators contribute the most to the overall objective through backpropagation of the loss function.

[0183] Three weights were applied to the index scores of the samples respectively to calculate the robustness, economy, and extreme weather survival scores of each sample. A three-dimensional bar chart was then generated. Figure 6 As shown in Table 8, the average values ​​are as follows.

[0184] Table 8. Average values ​​of simulation results for the three indicators

[0185]

[0186] As shown in Table 8, in the robustness dimension, the PC-AWNet method scores the highest (0.118), improving by approximately 11.3% compared to the CRITIC method and 15.7% compared to the entropy weight method. This indicates that PC-AWNet has a significant advantage in enhancing the system's resistance to disturbances, because it assigns higher weights to structural strength and equipment health, while reducing over-reliance on high-fluctuation load indicators. In the economic dimension, the PC-AWNet method performs similarly to the entropy weight method (both 0.121), outperforming the CRITIC method. This suggests a positive correlation between economic efficiency and load regulation, but PC-AWNet does not overemphasize load balancing like the entropy weight method, instead achieving "steady-state economic optimization" rather than "fluctuation-driven optimization." In the extreme weather survivability dimension, the PC-AWNet method scores significantly higher than the entropy weight method (0.082), with a relative improvement of 36.6%. This indicates that system failures under extreme weather conditions are usually triggered by structural instability and equipment degradation. The entropy weight method assigns high weights to load balancing, weakening the influence of structural indicators and resulting in a low survival score. In contrast, PC-AWNet correctly identifies "structure + health" as the dominant factors, thus obtaining a better survival assessment.

[0187] Observing the standard deviations across the three dimensions (robustness 0.0843, economy 0.1093, and extreme weather survivability 0.0866), the PC-AWNet method achieves the highest mean across all three dimensions, with stable variance control and no single-dimensional collapse. In contrast, the entropy weight method shows a significant decrease (0.082) in the extreme weather dimension, indicating that purely statistical discrete weighting carries a risk of distortion in extreme scenarios.

[0188] The essential difference between the three methods lies in the different weighting criteria, as shown in Table 9. Traditional methods are essentially static statistical mappings, deriving weights only from data distribution characteristics, and cannot capture the true causal contribution of indicators to system resilience. In contrast, the PC-AWNet method, through end-to-end training and optimization guided by a comprehensive objective, achieves the transformation from "data fluctuations" to "system causal contributions," thus performing better in key dimensions such as robustness and survival in extreme weather.

[0189] Table 9 Comparison at the Mechanism Level

[0190]

[0191] This invention utilizes a PC-AWNet model based on the fusion of a multi-scene feature encoder and an attention mechanism to learn the weight distribution of evaluation indicators from heterogeneous monitoring data end-to-end. By constructing feature extraction branches for four typical scenarios—typhoon, high salt spray, rainstorm, and normal conditions—the model can effectively capture the dynamic changes in the importance of indicators under different extreme environments. This allows for testing the system's resilience under the worst conditions, ensuring the reliability of the evaluation results, enhancing the model's scenario adaptability, and reducing the risk of evaluation distortion caused by weight solidification. A physical constraint integration layer is employed to incorporate domain knowledge into the model training in the form of differentiable loss terms. This achieves a balance between data-driven approaches and physical laws during weight generation. By introducing a composite loss function (including main loss, physical constraint loss, smoothness loss, and regularization terms), the penalty intensity of each constraint condition can be flexibly adjusted. In the first stage, the model fuses heterogeneous input data through a multi-scene feature encoder to form a unified scenario representation, laying the foundation for subsequent weight generation. In the second stage, the weight prediction head outputs weight vectors that satisfy normalization constraints, and the model parameters are adjusted based on the degree of violation feedback from the physical constraint layer, making the weight allocation results more consistent with actual physical laws. Meanwhile, in order to ensure that the generated weights maintain reasonable fluctuations over time and closely approximate the evolution characteristics of real-world operating scenarios, a smoothness loss and memory mechanism are introduced to limit drastic changes in weights, enabling the optimization model to consider a wider range of uncertainties and enhancing the model's adaptability to complex marine environments.

[0192] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for evaluating the resilience of island microgrids that integrates multi-scenario feature coding and physical constraints, characterized in that: The method includes: Raw monitoring data of the island microgrid is collected, and basic evaluation indicators are calculated based on the raw monitoring data. The basic evaluation indicators include structural strength indicators, corrosion resistance indicators, waterproofing indicators, equipment health indicators, and load balance indicators. A weight generation model is constructed, comprising a multi-scene feature encoder, a feature fusion module, a physical constraint integration layer, and a weight prediction head. The multi-scene feature encoder includes multiple parallel scene-specific feature extraction branches. These branches process different types of raw monitoring data and fuse the multi-scene information extracted by the multi-scene feature extraction branches through the feature fusion module to generate a scene representation. The physical constraint integration layer adds physical constraints during model training. The weight prediction head maps the scene representation to a weight vector. In the physical constraint integration layer, the threshold parameter in the physical constraint formula is defined as a learnable network parameter, and the physical constraint loss is calculated based on environmental monitoring data, weight vector, and threshold parameter. The physical constraint loss is incorporated into the composite loss function of model training, and the model parameters and threshold parameter are co-optimized through backpropagation. The basic evaluation indicators are weighted and fused based on the weight vector to generate comprehensive dimensional indicators, which complete the evaluation of the resilience level of island microgrids at different levels. The comprehensive dimensional indicators include robustness, economy and extreme weather survival.

2. The island microgrid resilience evaluation method integrating multi-scenario feature coding and physical constraints as described in claim 1, characterized in that: The feature fusion module includes a feature interaction module and a multi-head attention module; The feature interaction module is used to process the combined feature matrix of multi-scene information fusion to obtain a global representation; the multi-head attention module takes the combined feature matrix and scene type label as input, and is used to dynamically focus on the features most relevant to the current scene to obtain the multi-head attention output. By fusing global representation and multi-head attention output, a scene representation is obtained.

3. The island microgrid resilience evaluation method integrating multi-scenario feature coding and physical constraints according to claim 2, characterized in that: In the feature interaction module, the combined feature matrix is ​​processed sequentially through the first linear layer, layer normalization, activation function, and second linear layer to obtain the global representation; In the multi-head attention module, the scene type label is projected into a query vector, and the combined feature matrix is ​​projected into a key vector and a value vector; Multi-head attention output is obtained based on query vector, key vector, and value vector.

4. The island microgrid resilience evaluation method integrating multi-scenario feature coding and physical constraints as described in claim 1, characterized in that: The structural strength index is composed of stress-strain parameters, vibration frequency parameters, displacement parameters, and fatigue damage parameters; the corrosion resistance index is composed of salinity parameters, corrosion rate parameters, coating condition parameters, and rust percentage parameters; the waterproof capability index is composed of rainfall parameters, humidity parameters, and wave height parameters; the equipment health index is composed of health index parameters, equipment temperature parameters, and remaining life parameters; and the load balance index is composed of power output parameters, stress-strain parameters, and vibration frequency parameters.

5. The island microgrid resilience evaluation method integrating multi-scenario feature coding and physical constraints according to claim 1, characterized in that: The robustness dimension consists of all the basic evaluation indicators; the economic dimension consists of the equipment health index and the load balance index; the extreme weather survivability dimension consists of the structural strength index, corrosion resistance index, and waterproofing index.

6. The island microgrid resilience evaluation method integrating multi-scenario feature coding and physical constraints according to claim 1, characterized in that: The weight generation model incorporates a smoothness loss for optimization during training; the smoothness loss is constructed based on the weight vectors obtained by the model at adjacent time points and is used to smooth out the changes in weights over time.

7. The island microgrid resilience evaluation method integrating multi-scenario feature coding and physical constraints according to claim 1, characterized in that: The physical constraints include structural strength constraints, corrosion resistance constraints, waterproofing constraints, load balancing constraints, and scenario-specific constraints.

8. The island microgrid resilience evaluation method integrating multi-scenario feature coding and physical constraints according to claim 1, characterized in that: The scene type label is obtained based on the current environmental parameter data and the corresponding physical threshold.

9. The island microgrid resilience evaluation method integrating multi-scenario feature coding and physical constraints according to claim 1, characterized in that: The scene-specific feature extraction branch includes a three-layer fully connected network; the first fully connected network includes a linear layer, a layer normalization function, an activation function, and a Dropout layer connected in sequence; the second fully connected network includes a linear layer, an activation function, and a Dropout layer connected in sequence; and the third fully connected network uses a linear layer.

10. A resilience evaluation system for island microgrids integrating multi-scenario feature coding and physical constraints, characterized in that: This system is used to implement the island microgrid resilience evaluation method integrating multi-scenario feature encoding and physical constraints as described in claim 1. The island microgrid resilience evaluation system includes a data acquisition module, an index calculation module, a weight generation module, and a resilience evaluation module. The data acquisition module is used to collect monitoring data of the island microgrid. The index calculation module is used to obtain basic evaluation indicators. The weight generation module has an embedded weight generation model, used to adaptively generate weight vectors of the basic evaluation indicators based on the monitoring data of the current scenario. The resilience evaluation module is used to output comprehensive dimensional indicators based on the basic evaluation indicators and the weight vectors.

Images