High-proportion new energy oriented power system toughness quantitative evaluation method and system

Abstract

The invention relates to the technical field of power system toughness evaluation, and discloses a high-proportion new energy oriented power system toughness quantitative evaluation method and system. The method comprises the following steps: constructing a multi-dimensional association database based on historical data; establishing a weather-output mapping relation through feature selection and model training; an extreme weather scene set is generated, hierarchical index calculation is performed, and a toughness evaluation matrix is obtained in combination with mixed weighting; dynamic benefit analysis and investment priority ranking are carried out in combination with economic factors; and finally, closed-loop updating of the model parameters is realized. According to the method, the toughness evaluation precision of the high-proportion new energy power system in extreme weather can be effectively improved, economy and adaptability are considered, and a reliable basis is provided for system planning and risk prevention and control.

Description

Technical Field

[0001] This invention relates to the field of power system resilience assessment technology, and in particular to a method and system for quantitative assessment of power system resilience with a high proportion of renewable energy. Background Technology

[0002] In the field of power system resilience assessment technology, existing solutions related to a quantitative assessment method for power system resilience with a high proportion of renewable energy typically construct renewable energy output models based on historical data or single meteorological conditions, and use static scenario sets for system anti-interference capability analysis. These solutions suffer from limitations such as insufficient modeling of extreme weather correlations, limited multi-dimensional data feature fusion capabilities, and low dynamic scenario coverage. Existing methods largely rely on independent model training and manual scenario selection, which easily leads to insufficient model generalization and high scenario redundancy under the coupled effects of renewable energy output volatility and extreme weather events, making it difficult to meet the accuracy requirements for quantitative assessment of power system resilience under high-proportion renewable energy integration. Regarding the joint processing of weather-output mapping relationships, load fluctuation characteristics, and extreme event generation conditions, existing technologies generally lack cross-domain data correlation modeling and adaptive scenario optimization mechanisms. This makes it difficult to establish a consistent process of model training—scenario generation—redundancy selection—resilience assessment under the correlation of renewable energy output volatility and extreme weather, resulting in insufficient adaptability of assessment results to complex meteorological conditions. Summary of the Invention

[0003] This invention provides a method and system for quantitative assessment of power system resilience in the context of high proportion of renewable energy sources. It addresses the problem of how to achieve quantitative assessment of power system resilience under the correlation between renewable energy output fluctuations and extreme weather by using multidimensional correlation models and extreme scenario generation algorithms, through feature selection and model fusion.

[0004] To address the aforementioned technical problems, this invention provides a method for quantitatively assessing the resilience of power systems with a high proportion of renewable energy sources, comprising:

[0005] Historical operation data of new energy power stations are obtained, and missing values ​​are filled by time series interpolation and similarity matching of data in adjacent time periods. Weather-related features are extracted and normalized by combining min-max normalization and Z-score standardization. Spatial mapping is performed by combining Kriging interpolation or inverse distance weighting to generate a multidimensional relational database.

[0006] The weather-output mapping relationship is retrieved from the multidimensional association database. Feature selection is performed using the improved Shannon entropy algorithm. Gradient boosting decision tree training is used to obtain the initial model parameters. The improved PageRank algorithm is used to cross-validate the initial model parameters to generate optimized model weights and construct the multidimensional association model.

[0007] The wind speed fluctuation and radiation intensity gradient of the multidimensional correlation model are called to perform Monte Carlo algorithm initialization, generate random seed set and extreme weather scenarios, and use dynamic time warping algorithm combined with time decay coefficient to perform redundancy screening to generate extreme weather scenario set.

[0008] Load fluctuation data are extracted from extreme weather scenarios, and hierarchical indicators are calculated at the system level, regional level and equipment level. The weights are assigned by a combination of the analytic hierarchy process and the entropy weight method to generate a resilience assessment matrix.

[0009] By combining investment budgets with policy incentives, economic parameters are matched, dynamic benefit analysis is performed, and investment priority plans are generated.

[0010] Based on the update requirement list, adjust the key parameters of the multidimensional association model, including weights, thresholds, and regularization coefficients, to generate closed-loop update instructions.

[0011] Furthermore, the weather-output mapping relationship is retrieved from the multidimensional relational database, specifically including:

[0012] Based on the spatiotemporally aligned renewable energy output and key weather parameter data stored in the multidimensional association database, weather variables closely related to renewable energy power output are automatically retrieved and filtered.

[0013] The extraction of the weather-output mapping relationship adopts a feature selection algorithm, including methods based on mutual information and recursive feature elimination, combined with domain knowledge rules, to remove redundant and weakly correlated features and form a set of highly correlated weather variables;

[0014] For highly correlated variables, a correspondence between the input feature matrix and the output power sequence is constructed, which serves as the training sample for the machine learning model.

[0015] Furthermore, the process of using training samples for machine learning models also includes:

[0016] The training samples were augmented before training using data augmentation techniques, including time series slices, random noise injection, and weather condition simulation.

[0017] The machine learning model training adopts supervised machine learning algorithms, including gradient boosting decision trees, support vector machines and deep neural networks, which are trained in parallel. The model performance is evaluated by combining cross-validation strategies.

[0018] The training process adjusts hyperparameters, including learning rate, tree depth, and regularization coefficient, through an automatic parameter tuning module, using grid search and Bayesian optimization methods.

[0019] Furthermore, cross-validation is performed on the initial model parameters, specifically including:

[0020] The K-fold cross-validation method is adopted, which divides the training data into K subsets, uses K-1 subsets to train the model in turn, and uses the remaining 1 subset for validation. This process is repeated K times, and the performance indicators of each round of validation are calculated, including mean squared error, coefficient of determination, and mean absolute error.

[0021] For model parameters that perform poorly in cross-validation, perform model pruning and regularization adjustments;

[0022] Based on the validation results, an automatic model weight adjustment scheme is generated.

[0023] Furthermore, based on the validation results, a model weight adjustment scheme is automatically generated, specifically including:

[0024] We utilize model ensemble techniques to weight and fuse the weights of multiple trained models, employing weighted averaging and stacking methods.

[0025] The weight fusion process uses optimization algorithms, including genetic algorithms or particle swarm optimization, to determine the optimal weight combination under the condition of meeting a preset performance threshold.

[0026] Furthermore, the wind speed fluctuations and radiation intensity gradients from the multidimensional correlation model are invoked, specifically including:

[0027] Call the weather fluctuation parameters of the multidimensional correlation model, and extract the set of key parameters related to weather changes from the multidimensional correlation model, including wind speed fluctuation amplitude, solar radiation intensity change rate, temperature gradient and humidity fluctuation range;

[0028] Weather fluctuation parameters are automatically calculated through the time series analysis module and spatial distribution module inside the model. Combined with the statistical characteristics of historical data and the output of the prediction model, a multidimensional parameter matrix is ​​formed.

[0029] Furthermore, the Monte Carlo algorithm is initialized to generate a random seed set and extreme weather scenarios, specifically including:

[0030] The Monte Carlo algorithm module uses a random seed set and sampling parameters to generate extreme weather scenarios covering different time periods and spatial regions through a large number of random simulation iterations;

[0031] The generation process includes multi-stage simulation:

[0032] A preliminary weather variable sequence is generated, and the weather variables of each geographic unit are spatially coupled and adjusted by combining the spatial correlation model.

[0033] During the simulation, the system dynamically adjusts the sampling strategy to increase the probability of extreme values ​​occurring.

[0034] Furthermore, the expression for constructing a multidimensional association model includes:

[0035] Retrieve the spatiotemporally aligned new energy output sequence P from a multidimensional relational database t With weather feature matrix W t We construct a feature selection model for the weather-output mapping relationship, and perform feature selection and calculate feature importance using improved Shannon entropy:

[0036]

[0037] Among them, H s (w i ) represents the weather variable w i The feature importance index; w i Let be the i-th weather variable; i is the index of the weather variable; k is the index of the time window; n is the total number of time windows; p k (w i ) represents the weather variable w i The probability distribution value at the k-th time window; μ i For weather variables w i Historical mean; σ i For weather variables w i Historical standard deviation;

[0038] Furthermore, an improved PageRank algorithm is used to cross-validate the initial model parameters to generate optimized model weights, based on the initial model parameters θ. 0 Perform K-fold cross-validation and define an improved PageRank algorithm to fuse weights from multiple models:

[0039]

[0040] Among them, PR v (θ m The output is the model parameters θ. m PageRank weights; θ m Here are the model parameters obtained from training the m-th fold cross-validation; α is the damping factor; K is the number of folds in the K-fold cross-validation; θ n This is the nth model parameter in the neighborhood model set; To be related to the model parameters θ m A set of neighborhood models with similar performance; L(θ) n ) represents the model parameters θ n The loss function value;

[0041] Furthermore, the optimized model weights θ are loaded into a deep neural network to construct a spatiotemporal attention mechanism:

[0042]

[0043] Among them, A tThe output is the spatiotemporal attention weight matrix; softmax is the standard normalization function; Q t For query matrix; W t θ is the weather feature matrix; K is the weight of the optimized model; t P is the key matrix; t Contributing to new energy sources; d k (P) is the scaling factor; t ) T This represents the new energy output sequence vector P. t Transpose.

[0044] Furthermore, the expressions for generating extreme weather scenario sets include:

[0045] Calling the weather fluctuation parameter Ψ in the multidimensional correlation model t Including wind speed fluctuation δ v With radiation intensity gradient Define improved Monte Carlo initialization:

[0046]

[0047] Among them, s j Use the random seed for the j-th weather parameter; μ is the fluctuation component of the j-th weather parameter; j ψ For fluctuation components Historical mean; σ j ψ For fluctuation components The historical standard deviation; k2 is the time step index; T is the total length of the time series; Let be the radiation intensity gradient at the k2th time step; Let be the wind speed fluctuation at the k2-th time step; exp is the standard exponential function;

[0048] Based on the j-th weather parameter, a random seed s j Construct a random seed s for the j2th scene. j2 :

[0049] s j2 =SpatialInterp(s j ,Loc j2 )

[0050] SpatialInterp() is a geolocation-based interpolation function, while Loc... j2 These are the spatial coordinates of the j2th scene, where j2 is the scene index;

[0051] Furthermore, based on the random seed s of the j2th scene j2Define extreme scenario coverage verification:

[0052]

[0053] Among them, C s Score for extreme scene coverage; j2 is the scene index; N is the total number of scenes generated in Monte Carlo; Φ(s) j2 ) represents the statistical feature vector of the j2th scene; Φ hist represents the statistical feature vector of historical extreme events; ||·||1 is the L1 norm; I(·) is the indicator function; τ represents the maximum wind speed fluctuation in the j2th scene; v The wind speed threshold;

[0054] Furthermore, a dynamic time warping algorithm combined with a time decay coefficient is used to perform redundancy filtering, and an improved similarity metric is defined:

[0055]

[0056] Among them, D s p represents the scene similarity distance; p and q represent time point indices. To optimize the variable π, minimize the objective function; This represents the value of the j2th scene at the p-th time point; The value of the k2th scene at time q; γ is the time decay coefficient; t p t is the timestamp of the p-th time point; q λ is the timestamp of the q-th time point; λ is the path penalty term; len(π) is the length of the alignment path π.

[0057] Furthermore, a quantitative assessment system for the resilience of power systems with a high proportion of renewable energy, applied to any of the methods described above, includes:

[0058] The data acquisition module is used to obtain installation parameters and historical weather data of new energy generating units;

[0059] The preprocessing module is used to perform data cleaning and outlier correction operations;

[0060] The timing alignment module is used to generate synchronously associated datasets;

[0061] The strategy decision-making module is used to call the resilience assessment strategy table to output control instructions;

[0062] The control execution module is used to transmit output threshold adjustment commands to the new energy power station controller;

[0063] The database management module is used to store and update parameters of multidimensional relational databases.

[0064] The key innovations of this invention include:

[0065] (1) The initial model parameters are obtained by calling the weather-output mapping relationship from the multidimensional association database to train the machine learning model, and the optimized model weights are generated by cross-validation.

[0066] (2) Based on the optimized model weights, a new energy-weather-load correlation model is constructed, and the Monte Carlo algorithm is used to initialize it to generate a random seed set, thereby generating extreme weather scenarios.

[0067] (3) Redundancy screening is performed on the generated candidate scene set to form an extreme weather scene set, so as to improve the dynamic matching and coverage of the scene.

[0068] The following are its main beneficial effects:

[0069] (1) By cross-validating the initial model parameters, the optimized model weights can more accurately reflect the correlation between new energy output and weather changes, and improve the applicability of the model under the fluctuation of new energy output and extreme weather conditions.

[0070] (2) The extreme weather scenarios generated by the Monte Carlo algorithm are effectively increased in terms of diversity and coverage by initializing the random seed set, making them suitable for quantitative assessment of resilience under complex meteorological conditions.

[0071] (3) After redundancy screening of the candidate scenario set, the extreme weather scenario set formed reduces scenario redundancy, improves the accuracy and reliability of the evaluation results, and is suitable for power systems with a high proportion of new energy access. Attached Figure Description

[0072] Figure 1 A flowchart illustrating a method for quantitatively assessing the resilience of power systems with a high proportion of renewable energy, provided in an embodiment of this application;

[0073] Figure 2 This is a structural block diagram of a power system resilience quantitative assessment system for high-proportion renewable energy sources, provided as an embodiment of this application. Detailed Implementation

[0074] Example 1: Refer to Figure 1 This is a flowchart illustrating a method for quantitatively assessing the resilience of power systems with a high proportion of renewable energy, provided by an embodiment of the present invention. The process may include at least steps S100-S600:

[0075] S100. Obtain historical operation data of new energy power stations, use time series interpolation and similarity matching of adjacent time periods to fill missing values, extract weather-related features and perform normalization processing by combining minimum-maximum normalization and Z-score standardization, combine Kriging interpolation or inverse distance weighting method for spatial mapping, and generate a multidimensional association database.

[0076] S200: Retrieve the weather-output mapping relationship from the multidimensional association database, perform feature selection through improved Shannon entropy, train the gradient boosting decision tree to obtain the initial model parameters, and use the improved PageRank algorithm to cross-validate the initial model parameters to generate optimized model weights and construct the multidimensional association model.

[0077] S300: Call the wind speed fluctuation and radiation intensity gradient of the multidimensional correlation model, perform Monte Carlo algorithm initialization, generate random seed set and extreme weather scenarios, and use dynamic time warping algorithm combined with time decay coefficient to perform redundancy screening to generate extreme weather scenario set.

[0078] S400: Extract load fluctuation data from extreme weather scenarios, calculate hierarchical indicators at the system, regional and equipment levels, and use a combination of analytic hierarchy process and entropy weighting method to assign weights and generate a resilience assessment matrix.

[0079] S500 combines investment budgets with policy incentives to match economic parameters, performs dynamic benefit analysis, and generates investment priority plans.

[0080] S600: Based on the update requirement list, adjust the key parameters of the multidimensional association model, such as weights, thresholds, and regularization coefficients, to generate a closed-loop update instruction.

[0081] Step S100 includes at least steps S110-S130:

[0082] S110. Obtain historical operation data of new energy power stations, fill in missing values, and obtain a complete historical dataset.

[0083] To acquire historical operational data from new energy power plants, specifically, based on a pre-set collection cycle and data interface protocol, historical operational data including power output, wind speed, solar radiation intensity, and temperature are automatically collected from the monitoring system and data storage devices of the new energy power plants. The collection process includes both real-time data capture and offline batch download modes. For data loss caused by network transmission anomalies or equipment failures, a missing value imputation algorithm based on time series interpolation and similarity matching of data from adjacent time periods is used to ensure data continuity and integrity. Furthermore, outlier detection is performed on the collected raw data. A combination of statistical thresholding and a machine learning-based anomaly detection model is used to filter and remove abnormal data points that significantly deviate from the normal operating range. Abnormal data is recorded in a log file for subsequent auditing and review. After the above cleaning and imputation processing, the historical operational data is structured and stored according to a unified data format specification, forming a complete historical operational dataset. This dataset contains multi-dimensional variable values ​​with timestamps, meeting the input requirements for subsequent multi-dimensional correlation analysis. The historical running dataset is used as an output field name for the "historical running dataset" in the next step S120 to ensure the continuity and integrity of the data link.

[0084] S120. Extract weather-related features from historical operational datasets, perform normalization processing, and generate standardized data sequences;

[0085] Weather-related features are extracted from the historical operational dataset. Specifically, based on predefined feature extraction rules, key weather parameters closely related to new energy output are selected from the original variables such as temperature, humidity, wind speed, and solar radiation in the historical operational dataset. The feature extraction process uses a sliding window technique to calculate local statistical features of the time-series data, including the extraction of indicators such as mean, variance, and kurtosis. Furthermore, domain knowledge is combined to derive composite features from weather factors, enhancing the model's expressive power. For the extracted weather-related features, normalization is performed using a combination of min-max normalization and Z-score standardization to eliminate the influence of different dimensions and magnitudes, achieving data standardization. Specifically, the normalization process first calculates the historical maximum, minimum, and mean standard deviation of each feature, then maps the original features to a unified numerical range, ensuring the stability and convergence speed of subsequent machine learning model training. During the generation of the standardized feature sequence, normalization parameters are automatically recorded to support the model's inverse transformation and result interpretation. The standardized feature sequence is passed as the output field name to the "standardized feature sequence" in the next step S130, laying the data foundation for the construction of the multidimensional relational database.

[0086] S130. Perform spatiotemporal alignment on the standardized feature sequences to generate a multidimensional relational database;

[0087] The standardized feature sequences are spatiotemporally aligned. Specifically, through a timestamp synchronization mechanism and spatial coordinate matching method, various weather features are precisely mapped to new energy power output data. The spatiotemporal alignment process first resamples the data based on a unified time grid to address the issue of inconsistent sampling frequencies from different data sources. Linear interpolation and timestamp alignment algorithms are used to ensure that all features have corresponding values ​​at the same time point. Further, regarding the spatial dimension, combining the geographical location of new energy power plants and the spatial distribution of meteorological stations, spatial interpolation algorithms such as Kriging Interpolation (Kriging) or Inverse Distance Weighting (IDW) are used to spatially map the weather data, generating weather data points that match the locations of new energy power plants. During the generation of the multidimensional relational database structure, a multi-level data storage architecture is designed, including a time series layer, a spatial feature layer, and a relational layer, supporting fast querying and efficient updates. The aligned data structure supports multidimensional indexing and parallel access, meeting the performance requirements of subsequent multidimensional relational model construction. The multidimensional relational database is used as an output field name for subsequent steps in S210 to call the "multidimensional relational database", thereby realizing the closed loop and connection of the data processing link.

[0088] Step S200 includes at least steps S210-S230:

[0089] S210. Retrieve the weather-output mapping relationship from the multidimensional association database, train the machine learning model, and obtain the initial model parameters;

[0090] The weather-output mapping relationship is retrieved from the multidimensional association database. Specifically, based on the spatiotemporally aligned renewable energy output and key weather parameter data stored in the multidimensional association database, weather variables closely related to renewable energy power output, such as wind speed, solar radiation intensity, temperature, and humidity, are automatically retrieved and filtered. The extraction of the weather-output mapping relationship employs a feature selection algorithm, including methods based on mutual information and recursive feature elimination (RFE), combined with domain knowledge rules, to remove redundant and weakly correlated features, forming a set of highly correlated weather variables. Further, for these highly correlated variables, a correspondence between the input feature matrix and the output power sequence is constructed as training samples for the machine learning model. The training samples are augmented before training using data augmentation techniques, including time series slicing, random noise injection, and weather condition simulation, to improve the model's generalization ability and robustness.

[0091] During model training, supervised machine learning algorithms such as Gradient Boosting Decision Tree (GBDT), Support Vector Machine (SVM), and Deep Neural Network (DNN) are used for parallel training, combined with cross-validation to evaluate model performance. The training process utilizes an automatic hyperparameter tuning module to adjust hyperparameters, including learning rate, tree depth, and regularization coefficients, employing grid search and Bayesian optimization methods to ensure convergence and model optimality. During training, sample reweighting and undersampling are performed to address outliers and data skewness, preventing overfitting. Training logs and intermediate model parameters are periodically saved to the model management system, supporting subsequent model version control and backtracking.

[0092] After the machine learning model is trained, initial model parameters, including model structure, weight parameters, and training configuration, are automatically exported as the basis for subsequent model optimization and validation. These initial model parameters are defined as output field names for use in the next step S220's "Initial Model Parameters" function, forming a closed loop in the multi-dimensional correlation model construction chain. This supports subsequent optimization of model weight generation and the construction of a new energy-weather-load correlation model.

[0093] S220. Perform cross-validation on the initial model parameters to generate optimized model weights;

[0094] The initial model parameters are cross-validated. Specifically, K-Fold CrossValidation is used, where the training data is divided into K subsets. K-1 subsets are used sequentially for model training, and the remaining subset is used for validation. This process is repeated K times, and the performance metrics for each round of validation are calculated, including Mean Squared Error (MSE) and Coefficient of Determination (R²). 2 The model's stability and generalization ability are evaluated by statistically analyzing the cross-validation results and the mean absolute error (MAE). Furthermore, for model parameters that perform poorly in cross-validation, model pruning and regularization are performed to reduce model complexity and prevent overfitting.

[0095] Based on the validation results, an automatic model weight adjustment scheme is generated. Specifically, ensemble learning is used to weight and fuse the weights of multiple trained models, employing weighted averaging and stacking methods to improve model prediction accuracy. The weight fusion process uses optimization algorithms, such as genetic algorithms or particle swarm optimization (PSO), to determine the optimal weight combination while meeting preset performance thresholds. The weight distribution is dynamically adjusted for different new energy types and regional characteristics to achieve personalized model adaptation.

[0096] The optimized model weights are passed as output field names to the "Optimized Model Weights" in the next step S230, providing parameter support for the construction of the new energy-weather-load correlation model, ensuring the accuracy and reliability of the multi-dimensional correlation model, and promoting the continuity and closed-loop nature of the overall resilience quantitative assessment process.

[0097] S230. Construct a new energy-weather-load correlation model based on optimized model weights to generate a multi-dimensional correlation model;

[0098] A new energy-weather-load correlation model is constructed based on the optimized model weights. Specifically, firstly, based on the optimized model weights and combined with historical load data and weather parameters, multivariate regression analysis and deep learning methods are used to construct a multidimensional correlation model between new energy output, weather conditions, and electricity load. This model, by fusing time-series features and spatial distribution features, achieves an accurate description of the impact of new energy output on load under different weather scenarios. During the construction process, feature cross-referencing techniques and an attention mechanism are employed to enhance the model's ability to express complex nonlinear relationships.

[0099] Furthermore, model validation is performed using an independent test set to evaluate the performance of the multidimensional correlation model. Prediction error, goodness of fit, and robustness indices are calculated to ensure the model's applicability under different time periods and extreme weather conditions. The validation process automatically generates a detailed model validation report, including a description of the model structure, a comparison of training and validation results, error distribution analysis, and a description of the model's applicable scope. The validation report is output using a standardized template for easy review and technical evaluation.

[0100] The multidimensional correlation model, as the output field name, is called by the "multidimensional correlation model" in the next step S310. It supports the input of weather fluctuation parameters generated by the Monte Carlo extreme scenario and is a key technical link to realize the dynamic simulation and resilience assessment of the multidimensional correlation between new energy, weather and load.

[0101] In another embodiment, in S210, the spatiotemporally aligned new energy output sequence P is retrieved from the multidimensional association database. t With weather feature matrix W t A feature selection model for the weather-output mapping relationship is constructed. Formula ① performs feature selection using an improved Shannon entropy, specifically by calculating feature importance using the improved Shannon entropy:

[0102]

[0103] in:

[0104] w i : The i-th weather variable, where i is the index of the weather variable, for example, i=1 represents wind speed, i=2 represents solar radiation intensity;

[0105] p k (w i Weather variable w i The probability distribution value at the k-th time window represents the weather variable w within the time window k. i The probability of occurrence;

[0106] μ i Weather variable w i The historical average;

[0107] σ i Weather variable w i Historical standard deviation;

[0108] k: Time window index;

[0109] n: Total number of time windows;

[0110] H s w i ) represents the weather variable w i The characteristic importance index is used to assess the impact of weather variables on the output of new energy sources.

[0111] Wind speed and solar radiation intensity indices are extracted from a multidimensional relational database and mapped to w. 1 w 2 The output sequence is mapped to a timestamp-aligned power value P. t The feature importance H output by formula ① s (w i The initial model parameters θ are obtained by training a gradient boosting decision tree. Specifically, they are used to construct the feature input weights of the GBDT model, which are then trained in parallel to generate the initial model parameters θ. 0 This parameter contains the weights of the tree structure split points and leaf nodes, and is consumed by S220 as the output field name "Initial Model Parameters".

[0112] Furthermore, in S220, an improved PageRank algorithm is used to cross-validate the initial model parameters to generate optimized model weights. Specifically, based on the initial model parameters θ... 0 Perform K-fold cross-validation, and Equation ② defines the improved PageRank algorithm for fusing weights from multiple models:

[0113]

[0114] in:

[0115] PR v (θ m ): The output is the model parameters θ m PageRank weights;

[0116] θ m These are the model parameters obtained from the m-th fold cross-validation training.

[0117] α = 0.85: Damping factor, a fixed value;

[0118] L(θ n ) represents the model parameters θ n loss function value

[0119] To be related to the model parameters θ m A set of neighborhood models with similar performance

[0120] K: The number of folds in K-fold cross-validation, representing the number of groups for cross-validation. For example, in 5-fold cross-validation, K = 5;

[0121] θ n This is the nth model parameter in the neighborhood model set;

[0122] PR v (θ n ): Neighborhood model θ n PageRank value.

[0123] Specifically, H obtained from formula ① s (w i ) are used as feature weights in the calculation of L(θ) n The weight distribution PR output by formula ② v (θ m After Bayesian optimization, the optimized model weights θ are generated, and this field is used by S230 for model construction.

[0124] Furthermore, the optimized model weights θ are loaded into the deep neural network, and formula ③ constructs the spatiotemporal attention mechanism:

[0125]

[0126] in:

[0127] A t The output is a spatiotemporal attention weight matrix;

[0128] softmax is the standard normalization function;

[0129] Q t For query matrix;

[0130] W t This is a weather feature matrix;

[0131] θ represents the weights of the optimized model;

[0132] K t The key matrix;

[0133] P t Contributing to the development of new energy sources;

[0134] d k This is the scaling factor;

[0135] (P t ) T This represents the new energy output sequence vector P. t Transpose.

[0136] Specifically, θ obtained from formula ② is used as a weight parameter in matrix operations, and the output attention weights A t The data is fused with load data to generate a new energy-weather-load correlation model (multidimensional correlation model), and the output field name "multidimensional correlation model" is passed to S310.

[0137] Formula link closed-loop explanation:

[0138] Formula ① calculates the feature importance H s (w i ), which are used as weights in the model loss function L(θ) in formula ②. n The calculation enhances the relevance of model training.

[0139] Formula ② integrates the weights of multiple models (PR) v (θ m The final optimized model weights θ are generated through Bayesian optimization and used as input parameters for formula ③.

[0140] Formula ③ constructs the spatiotemporal attention weights A using optimized weights θ. t This weight is used to establish a multidimensional correlation model and also provides time correlation information for the calculation of the radiation intensity gradient in formula ④.

[0141] Step S300 includes at least steps S310-S330:

[0142] S310. Call the weather fluctuation parameters of the multidimensional correlation model, perform Monte Carlo algorithm initialization, and generate a random seed set;

[0143] The weather fluctuation parameters of the multidimensional correlation model are invoked. Specifically, a set of key parameters related to weather changes is first extracted from the multidimensional correlation model. This set of parameters includes, but is not limited to, wind speed fluctuation amplitude, solar radiation intensity change rate, temperature gradient, and humidity fluctuation range, used to describe the dynamic characteristics of new energy output affected by weather. The weather fluctuation parameters are automatically calculated through the time series analysis module and spatial distribution module within the model, and combined with the statistical characteristics of historical data and the output results of the prediction model to form a multidimensional parameter matrix. Specifically, the multidimensional correlation model, based on the input historical weather data and new energy output data, uses the weight parameters obtained through training to decompose the short-term and long-term fluctuation trends of weather variables, extracting seasonal changes, intraday fluctuations, and anomalous fluctuation components respectively, ensuring that the weather fluctuation parameters can comprehensively reflect the weather change characteristics at different time scales.

[0144] Furthermore, the weather fluctuation parameters are used as initialization inputs for the Monte Carlo algorithm. Specifically, the Monte Carlo algorithm module generates a random seed set based on the parameter matrix. The generation process employs a high-quality pseudo-random number generator, combined with the statistical distribution characteristics of the weather fluctuation parameters, to ensure that the spatial and temporal distribution of the random seeds conforms to the actual weather fluctuation patterns. During the random seed generation process, the system automatically adjusts the seed length and distribution density of the random number generator according to preset sampling strategies and coverage requirements, supporting multi-threaded parallel computation to improve generation efficiency. For any abnormal seed distributions that may occur during the generation process, the system automatically executes a verification mechanism, including distribution uniformity checks and correlation detection. Abnormal results are recorded in a log file for subsequent analysis and adjustment.

[0145] The Monte Carlo algorithm initialization configuration, passed as an output field name, is passed to the "Algorithm Initialization Configuration" in the next step S320, providing basic input parameters for the generation of extreme weather scenarios. This initialization configuration includes not only a random seed set but also configuration information such as sampling parameters, time window settings, and spatial resolution, supporting multi-dimensional simulation of extreme weather scenarios in subsequent steps. The generation of this configuration effectively connects the multi-dimensional correlation model to the Monte Carlo algorithm, ensuring the scientific rigor and rationality of extreme scenario generation.

[0146] S320. Generate extreme weather scenarios based on the algorithm initialization configuration, perform coverage verification, and obtain a candidate scenario set;

[0147] Extreme weather scenarios are generated based on the algorithm's initial configuration. Specifically, the Monte Carlo algorithm module utilizes the random seed set and sampling parameters to generate extreme weather scenarios covering different time periods and spatial regions through numerous random simulation iterations. The generation process includes multi-stage simulations: first, a preliminary weather variable sequence is generated based on a statistical model of weather fluctuation parameters; then, combined with a spatial correlation model, the weather variables of each geographical unit are spatially coupled and adjusted to form weather scenarios that conform to actual geographical distribution. During the simulation, the system dynamically adjusts the sampling strategy to increase the probability of extreme values ​​occurring, thereby enhancing the coverage of extreme weather events.

[0148] Furthermore, a coverage verification process is executed for the generated extreme weather scenarios. This coverage verification employs a multi-indicator evaluation system, including statistical distribution consistency testing, extreme value coverage analysis, and historical extreme event matching detection, ensuring that the generated scenarios fully represent possible extreme weather conditions. During verification, the system automatically identifies weather types and time periods with insufficient coverage, adjusts sampling parameters accordingly, and re-triggers Monte Carlo simulation, forming an iterative optimization loop. Verification results and adjustment records are fully stored, supporting subsequent auditing and model improvement.

[0149] The candidate scenario set, as an output field name, is passed to the "candidate scenario set" in the next step S330, providing input for redundancy screening and the formation of the final extreme weather scenario set. The candidate scenario set contains diverse extreme weather simulation samples, covering various extreme events such as sudden increases in wind speed, radiation anomalies, and abrupt temperature changes, supporting multi-dimensional analysis for subsequent resilience index calculations.

[0150] S330. Redundancy screening is performed on the candidate scene set to generate an extreme weather scene set;

[0151] The system performs redundancy screening on the candidate scene set. Specifically, it compares candidate scenes pairwise based on a similarity measurement algorithm, employing a combination of Dynamic Time Warping (DTW) and multidimensional Euclidean distance to quantify the similarity between weather variable sequences. During the screening process, a similarity threshold is set to eliminate highly repetitive or similar scenes, reducing the impact of redundant data on subsequent calculations. The screening algorithm supports hierarchical processing, performing a coarse screening in the time dimension followed by a fine screening in the spatial dimension, thus improving screening efficiency.

[0152] Furthermore, by combining the importance weights of different scenarios, the system prioritizes retaining extreme weather scenarios that have a significant impact on renewable energy output. Specific weights are determined through historical data regression analysis and expert rules, ensuring that the selection results represent both diversity and highlight key risks. During the selection process, abnormal or marginal scenarios are individually marked for subsequent specialized analysis. The selection results automatically update the scenario library index, supporting rapid retrieval and retrieval.

[0153] The extreme weather scenario set, as the output field name, is passed to the "Extreme Weather Scenario Set" in the next step S410, providing high-quality input data for the calculation of hierarchical resilience indicators. This extreme weather scenario set is highly representative and has broad coverage, forming a key link in the multi-dimensional correlation model of new energy, weather, and load and the resilience assessment process, supporting subsequent modules in a comprehensive quantitative assessment of power system resilience.

[0154] In another embodiment, in S310, the weather fluctuation parameter Ψ in the multidimensional correlation model is called. t Including wind speed fluctuation δ v With radiation intensity gradient Based on this, the improved Monte Carlo initialization can be achieved using formula ④:

[0155]

[0156] in:

[0157] s j Use the random seed for the j-th weather parameter;

[0158] Let j be the fluctuation component of the j-th weather parameter;

[0159] μ j ψ For fluctuation components The historical average;

[0160] σ j ψ For fluctuation components Historical standard deviation;

[0161] k2 is the time step index;

[0162] T is the total length of the time series;

[0163] Let be the radiation intensity gradient at the k2th time step;

[0164] For the wind speed fluctuation at the k2th time step;

[0165] exp is the standard exponential function.

[0166] Specifically, radiation intensity gradient The temporal correlation is determined by the attention weight A output by formula ③. t Participate in the calculation. The random seed set s generated by formula ④ j After initialization and configuration by spatial interpolation algorithm, the output is sent to S320.

[0167] Based on the j-th weather parameter, a random seed s j Construct a random seed s for the j2th scene. j2 :

[0168] s j2 =SpatialInterp(s j ,Loc j2 )

[0169] SpatialInterp() is a geolocation-based interpolation function, while Loc... j2 These are the spatial coordinates of the j2th scene, where j2 is the scene index.

[0170] Furthermore, based on the random seed s of the j2th scene j2 Use formula ⑤ to define the coverage verification for extreme scenarios:

[0171]

[0172] in:

[0173] C s Score the coverage of extreme scenarios;

[0174] j2 is the scene index;

[0175] N is the total number of scenes generated in Monte Carlo;

[0176] Φ(s j2 ) represents the statistical feature vector of the j2th scene;

[0177] Φ hist This represents the statistical feature vector of historical extreme events;

[0178] ||·||1 is the L1 norm;

[0179] I(·) is an indicator function;

[0180] This represents the maximum wind speed fluctuation value in the j2th scenario;

[0181] τ v This is the wind speed threshold.

[0182] Specifically, the scene generation is driven by the random seed generated by formula ④, and the output coverage index C s Used to filter candidate scene sets and pass them to S330.

[0183] Furthermore, a dynamic time warping algorithm combined with a time decay coefficient is used to perform redundancy filtering. Specifically, the dynamic time warping algorithm is used to remove redundancy from the candidate scene set, and an improved similarity metric is defined using formula ⑥:

[0184]

[0185] in:

[0186] D s The distance is the scene similarity.

[0187] p and q are time point indices;

[0188] To optimize the variable π, minimize the objective function;

[0189] This represents the value of the j2th scene at the p-th time point;

[0190] This represents the value of the k2th scene at the qth time point;

[0191] γ is the time decay coefficient;

[0192] t p This is the timestamp of the p-th time point;

[0193] t q This is the timestamp of the q-th time point;

[0194] λ is the path penalty term;

[0195] len(π) is the length of the alignment path π.

[0196] Specifically, the coverage index C output by formula ⑤ s High-coverage scenes are prioritized and the final set of extreme weather scenes is passed to S410 after spatial indexing optimization.

[0197] Formula link closed-loop explanation:

[0198] Formula ④ is based on attention weight A t Generate a random seed s based on weather fluctuation parameters. j This drives the extreme scenario coverage verification in formula ⑤.

[0199] Formula ⑤ calculates the coverage index C s This is used to filter the candidate scene set and pass it to formula ⑥ for similarity measurement and redundancy removal;

[0200] Formula ⑥ is based on the coverage index C s Add time decay penalty, output the final set of extreme weather scenarios, and complete the closed loop.

[0201] Technical achievements in this section: By integrating improved entropy feature selection and spatiotemporal attention mechanisms, a high-precision multidimensional correlation model of new energy, weather, and load was constructed; combined with improved Monte Carlo initialization and dynamic time warping filtering, a high-quality scene set covering actual extreme events was generated, providing reliable and verifiable input data for subsequent resilience index calculations.

[0202] Step S400 includes at least steps S410-S430:

[0203] S410. Extract load fluctuation data from extreme weather scenarios, perform hierarchical index calculations, and generate a primary evaluation matrix;

[0204] Load fluctuation data is extracted from the extreme weather scenario set. Specifically, the "extreme weather scenario set" output from step S330 is first invoked. This scenario set contains diverse extreme weather simulation samples that have undergone redundancy filtering, covering multi-dimensional weather variables such as sudden increases in wind speed, abnormal solar radiation, and abrupt temperature changes. The extraction process automatically matches the load data corresponding to extreme weather events based on preset time and spatial windows, combined with the geographical distribution of new energy power plants and the location of load monitoring points. Specifically, the system uses a spatiotemporal mapping algorithm to align the timestamps in the extreme weather scenarios with the sampling time of the load data, and simultaneously uses spatial interpolation methods to map the weather-affected areas to load nodes, forming a load fluctuation dataset. The load fluctuation data includes indicators such as load peak-to-valley difference, load change rate, and load fluctuation amplitude. This data integrates historical load monitoring data and extreme weather response data through multi-source data fusion technology to ensure the completeness and accuracy of load fluctuation characteristics.

[0205] Further, based on the load fluctuation data, a hierarchical index calculation is performed. Specifically, the system divides the load fluctuation data into three levels—system level, regional level, and equipment level—according to a predefined resilience hierarchy, and calculates the corresponding resilience indexes for each level. The hierarchical index calculation employs a multi-index fusion method, combining statistical analysis, entropy method, and expert weighting mechanism to comprehensively evaluate the frequency, amplitude, and duration of load fluctuations. During the calculation process, the system automatically adjusts the index weights for abnormal fluctuations and extreme events, enhancing sensitivity to key risks. The results of the hierarchical index calculation undergo data consistency verification, eliminating outliers and noise, forming a structured primary evaluation matrix. This matrix contains numerical expressions for the resilience indicators at each level, supporting multi-dimensional resilience state descriptions. The primary evaluation matrix, as an output field name, is passed to the "Primary Evaluation Matrix" in the next step S420, providing basic data for subsequent weight allocation and standardization processing.

[0206] S420. Assign weights to the primary evaluation matrix and generate standardized resilience indicators.

[0207] The system extracts load fluctuation data and hierarchical indicators from the primary assessment matrix. Specifically, it assigns weights to each indicator in the primary assessment matrix according to a preset weight allocation strategy. The weight allocation employs a hybrid weighting model combining the Analytic Hierarchy Process (AHP) and entropy weighting, balancing expert experience with data-driven objectivity. Specifically, initial indicator weights are first constructed through expert questionnaires and historical data statistics. Then, the information entropy of the indicators is calculated using entropy weighting to adjust the weight distribution, reflecting the degree of variation and contribution of the indicators. During the weight allocation process, the system dynamically adjusts weight parameters based on different new energy types and regional characteristics, achieving personalized weight configuration. The weight assignment process incorporates correlation analysis between indicators to avoid redundant weight calculations and indicator redundancy, improving the scientific and rational nature of the weight allocation.

[0208] Furthermore, based on the weight allocation results, the primary evaluation matrix is ​​standardized. The standardized resilience index is generated using a combination of normalization and fuzzy comprehensive evaluation to eliminate the influence of different index dimensions and numerical ranges, achieving unified measurement of the indexes. Specifically, the system first performs range normalization on each index, mapping it to a unified numerical range. Then, it combines fuzzy membership functions to convert it into a fuzzy set for resilience evaluation, reflecting the uncertainty and fuzziness of the indexes. During the standardization process, the system automatically identifies outliers and marginal data, and uses robust statistical methods for correction to ensure the stability and representativeness of the standardized indexes. The standardized resilience index, as an output field name, is passed to the "Standardized Resilience Index" in the next step S430, providing input for constructing the hierarchical evaluation report and generating the resilience evaluation matrix.

[0209] S430. Construct a hierarchical assessment report based on standardized resilience indicators and generate a resilience assessment matrix;

[0210] Based on the standardized resilience indicators, a tiered assessment report is constructed. Specifically, the system summarizes and analyzes the standardized resilience indicators at each level according to a predefined report template. The construction process includes indicator aggregation, trend analysis, and anomaly identification, employing multi-dimensional data visualization technology to generate a multi-level expression of resilience status. The system combines historical assessment data to automatically calculate resilience change trends, identify key influencing factors for resilience improvement or decline, and form detailed tiered assessment content. Furthermore, the system performs consistency checks on the tiered assessment results, employing statistical consistency indicators and expert review mechanisms to ensure the accuracy and reliability of the report content. During the construction process, the system automatically generates text descriptions, charts, and risk warnings to facilitate decision-makers' understanding and application.

[0211] Based on the hierarchical assessment report, a resilience assessment matrix is ​​generated. This matrix is ​​a multi-dimensional matrix structure, containing comprehensive resilience indicators at the system, regional, and equipment levels, reflecting the resilience performance of the power system under extreme weather conditions. The matrix construction employs a multi-indicator fusion algorithm, combining weighted average and fuzzy comprehensive evaluation results to form a unified quantitative expression of resilience. The resilience assessment matrix is ​​output through a standard data interface, supporting seamless integration with the subsequent investment benefit analysis module. As an output field name, the resilience assessment matrix is ​​passed to the "Resilience Assessment Matrix" in the next step S510, providing key input data for dynamic investment benefit analysis and forming a closed-loop process from extreme weather scenarios to investment decisions.

[0212] Step 500 includes at least steps S510-S530:

[0213] S510. Obtain key indicators from the resilience assessment matrix, perform economic parameter matching, and generate investment simulation input;

[0214] Key indicators are obtained from the resilience assessment matrix. Specifically, the "resilience assessment matrix" output from step S430 is first invoked. This matrix contains multi-dimensional resilience indicators at the system, regional, and equipment levels. The acquisition process automatically filters out key indicators closely related to investment decisions, such as system reliability indicators, load response capacity, fault recovery time, and the stability of new energy output, using predefined economic parameter matching rules. Specifically, the system matches economic parameters stored in the investment strategy library, including investment budget, capital recovery cycle, risk tolerance, and policy incentives, and constructs investment simulation input by combining these with the indicator values ​​in the resilience assessment matrix. The matching process employs a multi-objective optimization algorithm to weightedly fuse key indicators and economic parameters, generating an input feature set that conforms to the current economic environment and investment objectives. Furthermore, for outliers and missing data in the input features, the system performs data repair and imputation processing, employing historical trend analysis and data inference methods from nearby time periods to ensure the completeness and accuracy of the investment simulation input. The investment simulation input is passed as the output field name to the "investment simulation input" in the next step S520, providing the necessary decision-making basis for dynamic investment benefit analysis, and forming a data feedback closed loop with the iterative update of model parameters in step S600.

[0215] S520. Perform dynamic benefit analysis on the investment simulation input and generate candidate investment schemes;

[0216] The system performs dynamic benefit analysis on the simulated investment inputs. Specifically, based on the input economic parameters and key resilience indicators, the system uses a dynamic simulation model to evaluate the benefits of the investment plan. This dynamic benefit analysis includes multiple stages such as investment cost prediction, revenue cash flow simulation, risk assessment, and sensitivity analysis. Specifically, firstly, based on the capital allocation plan and project cycle in the simulated investment inputs, a time-series cash flow model is constructed to simulate the dynamic changes in investment expenditures and expected returns. Subsequently, combined with indicators in the resilience assessment matrix, the impact of renewable energy output fluctuations on grid operation efficiency is simulated to evaluate the investment's effect on improving system resilience. The system uses the Monte Carlo simulation method to perform multiple random samplings of uncertain factors such as market price fluctuations, policy adjustments, and technological advancements to form a probability distribution of investment benefits. Further, the dynamic benefit analysis module performs risk quantification, using Value at Risk (VaR) and Conditional Value at Risk (CVaR) indicators to evaluate the economic performance of the investment plan under extreme scenarios. During the analysis, the system automatically identifies key risk points in the investment plan and adjusts model parameters based on historical data and expert experience to reflect the latest market and technological conditions. The candidate investment scheme is passed as the output field name to the "candidate investment scheme" in the next step S530, which supports subsequent error verification and priority sorting, forming the continuity and closed loop of the investment decision-making link.

[0217] S530. Based on the candidate investment schemes, perform error verification and generate investment priority schemes;

[0218] Based on the candidate investment schemes, error verification is performed. Specifically, the system first calls upon the "candidate investment schemes" from step S520 and conducts multi-dimensional error analysis on the schemes' economic benefits and risk indicators. The error verification uses historical investment project data and actual market feedback as benchmarks, identifying potential deviations and uncertainties in the investment schemes through residual analysis and error distribution statistics. Specifically, the system compares the scheme's predicted values ​​with the actual observed values, calculating indicators such as the root mean square error (RMSE) and mean absolute percentage error (MAPE) to quantify the error level. Further, for schemes with large errors, the system automatically triggers an adjustment mechanism to reassess investment parameters and risk assumptions, and conducts manual review and parameter correction in conjunction with an expert rule base. During the error verification process, the system records all adjustment and verification logs, forming a complete traceability chain. Based on the error analysis results, the system uses a multi-indicator ranking algorithm, combining investment return rate, risk level, resilience improvement, and policy priority to comprehensively rank the candidate investment schemes. The ranking process employs a hybrid approach combining the Analytic Hierarchy Process (AHP) and fuzzy comprehensive evaluation, taking into account both quantitative data and expert subjective judgment to generate an investment priority scheme. This investment priority scheme, as an output field name, is passed to the "Investment Priority Scheme" in the next step S610, providing a decision-making basis for iterative updates of model parameters and effectively linking investment benefit analysis with model optimization.

[0219] Step S600 includes at least steps S610-S630:

[0220] S610. Extract the verification results from the investment priority scheme, compare the model parameters, and generate an update requirement list.

[0221] The verification results are extracted from the investment priority scheme. Specifically, the "investment priority scheme" output from step S530 is first invoked. This scheme includes multi-dimensional investment ranking data and related economic benefit indicators after error verification. The extraction process is implemented through an automated data interface. The system filters out key verification results from the investment scheme based on preset verification rules and thresholds, including return on investment, risk indicators, rationality of capital allocation, and contribution to resilience improvement. Specifically, the system uses a data parsing module to parse the structured data in the investment priority scheme, combines it with a historical investment performance database, and automatically matches the corresponding actual feedback information to form a verification result set. The verification result set undergoes consistency verification, eliminating abnormal and incomplete data. A multi-dimensional comparative analysis method is used to identify deviations and potential problems in the investment scheme. For missing or abnormal data, the system triggers an exception handling process, invoking backup data sources or a manual review mechanism to ensure the completeness and accuracy of the verification results. The verification results serve as the basic input for subsequent model parameter comparison and are entered into the model parameter comparison module for in-depth analysis.

[0222] Furthermore, model parameters are compared, specifically, the verification results are compared item by item with the parameter configuration of the current multidimensional correlation model. This comparison process is based on a parameter mapping table, automatically matching the economic benefit indicators in the verification results with the key weights, thresholds, and adjustment factors in the model parameters. The system uses a difference detection algorithm to calculate the deviation and trend of each parameter, identifying areas for improvement and potential failure points in model performance. During the comparison, statistical analysis methods are used to assess the impact of parameter adjustments on model prediction accuracy and investment returns, generating parameter adjustment recommendations. For significant deviations detected, the system automatically generates an adjustment priority list, combining historical model versions and parameter evolution trajectories to determine the necessity and urgency of adjustments. The model parameter comparison results are recorded in the model management database, supporting version control and audit trails. After the comparison is completed, the system automatically compiles and generates an update requirement list, clearly defining the parameter items to be adjusted, the adjustment range, and the expected effects.

[0223] The update requirement list, as an output field name, is passed to the "Update Requirement List" in the next step S620, providing a clear basis for subsequent adjustments to the parameters of the multidimensional correlation model and achieving a closed-loop connection between the investment priority scheme and the iterative update of model parameters. The update requirement list contains detailed parameter adjustment suggestions, adjustment priority ranking, and corresponding verification basis, supporting automated execution and manual review of subsequent steps, and ensuring the systematic management of continuous model optimization.

[0224] S620: Adjust the parameters of the multidimensional association model based on the updated requirement list and generate a new version of the model configuration;

[0225] Based on the update requirement list, the system adjusts the parameters of the multidimensional association model. Specifically, the system first calls the "update requirement list" output from step S610, which details the model parameters to be adjusted and the corresponding adjustment strategies. The adjustment process is executed by the model parameter management module, which, in conjunction with the model version control system, automatically locates the parameter configuration file in the current multidimensional association model. Specifically, according to the parameter names and adjustment ranges in the update requirement list, the system calls the parameter modification interface to update key parameters such as model weights, thresholds, and regularization coefficients item by item. During parameter adjustment, the system employs a transaction management mechanism to ensure the atomicity and consistency of parameter modifications, preventing incomplete configuration due to abnormal interruptions. The adjustment operation supports batch processing and incremental updates, improving execution efficiency. For complex parameter dependencies, the system performs dependency checks to avoid model structure conflicts or performance degradation caused by parameter adjustments.

[0226] Furthermore, during the adjustment process, an automated testing module is used to perform a preliminary performance evaluation of the updated model parameters. This performance evaluation includes rapid detection of metrics such as model prediction accuracy, training convergence speed, and overfitting risk. The system automatically compares the performance differences before and after the adjustment to determine whether the adjustment effect meets expectations. Adjustments that do not meet the expectations trigger a rollback mechanism, restoring the system to the previous stable version. For successfully adjusted parameter versions, the system generates a new version model configuration file, containing a complete parameter set, adjustment records, and version number information. The new version model configuration file is stored in the model management system, supporting subsequent access and version tracking.

[0227] The new version model configuration, as an output field name, is passed to the "New Version Model Configuration" in the next step S630. This provides basic data for the generation of model update instructions and the initialization parameters for the next evaluation cycle, enabling dynamic optimization and continuous iterative updates of the multidimensional correlation model. This configuration file has complete metadata descriptions, supports cross-module calls and system integration, and ensures the standardization and normalization of model parameter adjustments.

[0228] S630: Write the new version model configuration into the initialization parameters of the next evaluation cycle and generate a closed-loop update instruction;

[0229] Based on the new version model configuration, the initialization parameters for the next evaluation cycle are written. Specifically, the system calls the "new version model configuration" output from step S620 and uses it as the initialization input for the next evaluation cycle. This writing process is controlled by the model update management module, which synchronizes the new version model configuration file to the evaluation system's configuration storage area, replacing the old version parameter file. According to a predefined update strategy, the system automatically triggers the parameter loading process for the next evaluation cycle, ensuring that the evaluation module uses the latest model configuration for data processing and analysis. During the update process, the system performs integrity checks, verifying the configuration file format, parameter range, and dependencies to prevent configuration errors from causing evaluation failure.

[0230] Furthermore, the system generates a closed-loop update command, including an update task identifier, timestamp, and version information, serving as an operation log and task trigger signal. This update command, via a message queue or event-driven mechanism, notifies relevant modules to complete parameter switching and environment refresh. The system also records detailed logs of the update process, including parameter changes, execution time, and results, supporting subsequent auditing and problem tracking. The generation and distribution of update commands comply with the system security policy, employing access control and authentication mechanisms to prevent unauthorized configuration changes. Through this closed-loop update command, the system achieves dynamic iteration and automated management of model parameters, forming a complete closed loop from investment priority scheme feedback to model parameter updates.

[0231] The model update instruction, as an output field name, is passed to the "historical running dataset" in the next step S110, enabling the data acquisition and multidimensional correlation processing module to respond to model updates and ensuring the continuity and adaptability of the entire resilience quantitative assessment process. This instruction not only triggers the re-acquisition and preprocessing of historical data but also drives the retraining and optimization of the subsequent multidimensional correlation model, forming a dynamic assessment system based on investment feedback.

[0232] Example 2: Figure 2 This diagram illustrates a structural block diagram of a power system resilience quantitative assessment system for high-proportion renewable energy sources, according to an embodiment of the present invention. Figure 2 As shown, the structure may include:

[0233] Data acquisition module 01 is used to acquire installation parameters and historical weather data of new energy units. Specifically, it receives input objects from the new energy power station monitoring system and data storage devices, including multi-dimensional historical operating data such as power output, wind speed, solar radiation intensity, and temperature. The system automatically performs real-time data capture and offline batch downloading through a preset acquisition cycle and data interface protocol to form a raw dataset. To address data loss due to network transmission anomalies or equipment failures, the system employs a missing value filling algorithm based on time series interpolation and similarity matching of data from adjacent time periods to ensure data continuity and integrity. The raw dataset is recorded as a "historical operating dataset" and transmitted to the preprocessing module for use as a "historical operating dataset," while acquisition logs are retained for subsequent traceability.

[0234] Preprocessing module 02 is used to perform data cleaning and outlier correction operations. Specifically, it receives the "historical operating dataset" output from the data acquisition module, performs outlier detection and correction on the dataset, and filters and removes abnormal data points that significantly deviate from the normal operating range by combining statistical thresholding and a machine learning-based anomaly detection model. The abnormal data is recorded in a log file for auditing and review. After cleaning and outlier correction, the dataset is structured and stored according to a unified data format specification to form a complete "cleaned dataset". The "cleaned dataset" is then passed to the time-series alignment module for invocation, and the corresponding time information is registered in the data storage for the strategy decision module to read.

[0235] The temporal alignment module 03 is used to generate a synchronized and associated dataset. Specifically, it receives the "cleaned dataset" from the preprocessing module, performs timestamp synchronization and spatial coordinate matching in conjunction with real-time power output data, and accurately maps various weather features to new energy power output data. The system addresses the issue of inconsistent sampling frequencies from different data sources through time grid resampling and linear interpolation algorithms, ensuring that all features have corresponding values ​​at the same time point. Spatially, combining the geographical location of new energy power plants with the spatial distribution of meteorological stations, a spatial interpolation algorithm is used to generate weather data points that match the locations of new energy power plants, forming a "multidimensional associated database." This "multidimensional associated database" is then used by the strategy decision module as a "multidimensional associated database," and the database management module records the correspondence with parameter configurations.

[0236] The strategy decision module 04 is used to call the resilience assessment strategy table to output control instructions. Specifically, it completes strategy matching and instruction generation based on the "multi-dimensional relational database" from the time-series alignment module, and generates control instructions by combining the rules and thresholds of the resilience assessment strategy table. The control instructions include various control signals such as output adjustment, load allocation, and fault response; the control instructions are output to the control execution module, and the instruction effectiveness status is returned to the database management module for registration.

[0237] The control execution module 05 is used to transmit output threshold adjustment commands to the new energy power station controller. Specifically, it receives control commands from the strategy decision module, parses the command content, and generates specific control signals. These commands include output adjustment, load allocation, and fault response. The control signals are then output to the new energy power station controller interface to ensure accurate transmission and execution. After execution, the system sends the effective status or confirmation information back to the database management module to ensure the recording of execution results and subsequent analysis.

[0238] Database Management Module 06 is used to store and update parameters of the multidimensional relational database. Specifically, it receives the effective status from the control execution module and the control instruction logs from the strategy decision module, performs parameter updates and database maintenance, and ensures the timely update and integrity of the multidimensional relational database. The system performs version control and index management on the parameters in the database to ensure the accuracy and consistency of data retrieval. The updated database parameters are provided to the data acquisition module for parameter backfeeding, and the index relationship consistent with the time series is maintained to achieve a closed loop and dynamic update of the data link.

Claims

1. A method for quantitatively evaluating the resilience of a power system oriented to a high proportion of new energy, characterized in that, The method comprises the following steps: Obtain historical operation data of a new energy station, fill in missing values by time series interpolation and similarity matching of adjacent time period data, extract weather-related features, perform normalization processing by combining minimum-maximum normalization and Z-score standardization, perform spatial mapping by combining Kriging interpolation or inverse distance weighting method, and generate a multi-dimensional correlation database; Call the weather-power mapping relationship from the multi-dimensional correlation database, perform feature selection by improved Shannon entropy, train the initial model parameters by gradient boosting decision tree, generate the optimized model weight by cross-validation of the improved PageRank algorithm, and construct the multi-dimensional correlation model; Call the wind speed fluctuation and radiation intensity gradient of the multi-dimensional correlation model, initialize the Monte Carlo algorithm, generate a random seed set and an extreme weather scenario, perform redundant screening by combining the dynamic time warping algorithm with the time decay coefficient, and generate an extreme weather scenario set; Extract the load fluctuation data from the extreme weather scenario set, calculate the hierarchical indicators according to the system level, the regional level and the equipment level, assign the weight by mixing the analytic hierarchy process and the entropy weight method, and generate a resilience evaluation matrix; Combine the investment budget and the policy incentive factors to match the economic parameters, perform dynamic benefit analysis, and generate an investment priority scheme; Adjust the key parameters of the multi-dimensional correlation model weight, threshold and regularization coefficient based on the updated demand list, and generate a closed-loop update instruction.

2. The method of claim 1, wherein, The weather-power mapping relationship is called from the multi-dimensional correlation database, which specifically comprises: According to the time and space aligned new energy output and key weather parameter data stored in the multi-dimensional correlation database, the weather variables closely related to the new energy power output are automatically retrieved and screened out; The extraction of the weather-power mapping relationship adopts a feature selection algorithm, including a method based on mutual information and recursive feature elimination, combined with domain knowledge rules, to eliminate redundant and weakly correlated features and form a high-correlation weather variable set; For the high-correlation variables, the corresponding relationship between the input feature matrix and the output power sequence is constructed as the training sample of the machine learning model.

3. The method of claim 2, wherein, The process of the training sample as the training sample further comprises: The training sample is expanded by data enhancement technology before training, including time series slicing, random noise injection and weather condition simulation; The machine learning model training adopts a supervised machine learning algorithm, including gradient boosting decision tree, support vector machine and deep neural network parallel training, combined with cross-validation strategy to evaluate the model performance; The training process adjusts the hyperparameters, including learning rate, tree depth and regularization coefficient, by using grid search and Bayesian optimization method.

4. The method of claim 1, wherein, The cross-validation of the initial model parameters specifically comprises: The K-fold cross-validation method is used to divide the training data into K subsets, and K-1 subsets are used for model training and the remaining 1 subset is used for validation, which is repeated K times, and the performance indicators of each round of validation are calculated, including mean square error, determination coefficient and mean absolute error; For the model parameters that perform poorly in cross-validation, model pruning and regularization adjustment are performed; The model weight adjustment scheme is automatically generated based on the validation results.

5. The method of claim 4, wherein, In combination with the verification results, an automatic model weight adjustment scheme is generated, specifically including: Using model integration technology to weight and fuse the weights of multiple training models, using weighted average and stacking methods; The weight fusion process is optimized by an algorithm, including genetic algorithm or particle swarm optimization, to determine the optimal weight combination under the condition of meeting the preset performance threshold.

6. The method of claim 1, wherein, Call the wind speed fluctuation and radiation intensity gradient of the multi-dimensional correlation model, specifically including: Call the weather fluctuation parameters of the multi-dimensional correlation model, extract the key parameter set related to weather changes from the multi-dimensional correlation model, including wind speed fluctuation amplitude, solar radiation intensity change rate, temperature gradient, and humidity fluctuation range; The weather fluctuation parameters are automatically calculated by the time series analysis module and the spatial distribution module inside the model, combined with the statistical characteristics of historical data and the output results of the prediction model, to form a multi-dimensional parameter matrix.

7. The method of claim 1, wherein, Monte Carlo algorithm initialization is performed to generate a random seed set and an extreme weather scenario, specifically including: The Monte Carlo algorithm module uses the random seed set and sampling parameters to generate extreme weather scenarios covering different time periods and spatial regions through a large number of random simulation iterations; The generation process includes multi-stage simulation: Generate preliminary weather variable sequences and adjust the weather variables of each geographical unit through spatial coupling based on the spatial correlation model; During the simulation process, the system dynamically adjusts the sampling strategy to increase the probability of extreme value occurrence.

8. The method of claim 1, wherein, The expression of the multi-dimensional correlation model includes: New energy output sequence P after space-time alignment is called from a multi-dimensional correlation database t with weather feature matrix W t , a feature selection model for constructing weather-output mapping relationship is constructed, and feature importance is calculated by improved Shannon entropy: where H s (w i ) is the characteristic importance indicator of weather variable w i ; w i is the i-th weather variable; i is the weather variable index; k is the time window index; n is the total number of time windows; p k (w i ) is the probability distribution value of weather variable w i in the k-th time window; μ i is the historical mean of weather variable w i ; σ i is the historical standard deviation of weather variable w i . Further, the improved PageRank algorithm is used to cross-validate the initial model parameters to generate optimized model weights, and the initial model parameters θ 0 K-fold cross-validation is performed, and the improved PageRank algorithm is used to fuse the model weights: where PR v (θ m ) is the PageRank weight with the output as the model parameter θ m ; θ m is the model parameter obtained by the m-fold cross-validation training; α is the damping factor; K is the number of K-fold cross-validation; θ n is the n-th model parameter in the neighborhood model set; θ is the neighborhood model set with similar performance to the model parameter θ m ; L(θ n ) is the loss function value of the model parameter θ n . Further, load the optimized model weight θ into the deep neural network to build a spatio-temporal attention mechanism: wherein A t is the output of the spatio-temporal attention weight matrix; softmax is the standard normalization function; Q t is the query matrix; W t is the weather feature matrix; θ is the optimized model weight; K t is the key matrix; P t is the new energy output sequence; d k is the scaling factor; denotes the transpose of the new energy output sequence vector P t .

9. The method of claim 1, wherein, The expression of the extreme weather scenario set includes: Calling weather fluctuation parameters Ψ in multidimensional association models t , including wind speed fluctuations δ v and radiation intensity gradients Define improved Monte Carlo initialization: where s j is the random seed for the jth weather parameter; is the fluctuation component for the jth weather parameter; is the historical mean of the fluctuation component ; k2 is the time step index; T is the total length of the time series; is the historical standard deviation of the fluctuation component ; k2 is the time step index; T is the total length of the time series; is the radiation intensity gradient at the k2th time step; is the wind speed fluctuation at the k2th time step; exp is the standard exponential function; based on the jth weather parameter random seed s j , construct the jth2 scene random seed s j2 : s j2 = SpatialInterp(s j , Loc j2 ) where SpatialInterp() is a geolocation-based interpolation function, Loc j2 is the spatial coordinate of the j2th scene, and j2 is the scene index. Further, based on the random seed s of the j2th scenario j2 , define the extreme scenario coverage verification: where C s is the extreme scenario coverage score; j2 is the scenario index; N is the total number of Monte Carlo generated scenarios; Φ(s j2 ) is the statistical feature vector of the j2th scenario; Φ hist is the statistical feature vector of the historical extreme event; ||·||1 is the L1 norm; I(·) is the indicator function; is the maximum wind speed fluctuation of the j2th scenario; τ v is the wind speed threshold; Further, use the dynamic time warping algorithm combined with the time decay coefficient to perform redundancy screening and define an improved similarity measure: where D s is the scene similarity distance; p, q are time point indices; is the optimization variable π, the objective function is minimized; is the value of the jth2scene at the pthtime point; is the value of the kth2scene at the qthtime point; γ is the time decay coefficient; t p is the timestamp of the pthtime point; t q is the timestamp of the qthtime point; λ is the path penalty term; len(π) is the length of the alignment path π.

10. A power system resilience quantitative evaluation system for high proportion of new energy, applied to the method of any one of claims 1-9, characterized in that, Including: The data acquisition module is used to obtain new energy unit installation parameters and historical weather data; The preprocessing module is used to perform data cleaning and outlier correction operations; The time series alignment module is used to generate a synchronized correlation data set; The strategy decision module is used to call the resilience evaluation strategy table to output control instructions; The control execution module is used to output force threshold adjustment instructions to the new energy station controller; The database management module is used to store and update the multi-dimensional correlation database parameters.

Images