Urban power lifeline dynamic risk assessment method based on machine learning

Abstract

The invention belongs to the technical field of power system dynamic assessment, and particularly relates to a machine learning-based urban power lifeline dynamic risk assessment method. The risk assessment method comprises the following steps: step S100, defining a power system risk assessment problem; step S200, the data input layer performs multi-source heterogeneous data fusion, and collects and integrates multi-source heterogeneous data; s300, a feature extraction layer extracts risk features capable of reflecting risks of the power system from the integrated multi-source heterogeneous data, and optimizes the risk features through an automatic feature selection mechanism to form feature vectors; step S400, the model training and evaluation layer uses the optimized risk features to train and debug a risk evaluation model, and evaluates the performance of the model; and step S500, dynamically displaying a risk assessment result by a result output layer. The risk assessment method has enough training samples, can overcome the defects of training data scarcity, difficult feature optimization and insufficient model generalization ability, and realizes accurate identification and dynamic assessment of urban power system risks.

Description

Technical Field

[0001] The present invention belongs to the technical field of dynamic assessment of power systems, and in particular relates to a dynamic risk assessment method for urban power lifelines based on machine learning. Background Art

[0002] As an important infrastructure for the operation of modern society, the safe and stable operation of urban power systems is directly related to the national economy, people's livelihood and the comprehensive development of the economy. However, with the intensification of global climate change and the frequent occurrence of extreme events, urban power systems are facing unprecedented multiple risk challenges. These risks come not only from the impact of natural disasters such as earthquakes, floods, ice and snow, but also from internal factors such as equipment failures and operational errors, and even involve new security threats such as cyber attacks and terrorist attacks. Traditional power system risk assessment methods, such as analysis based on reliability theory and qualitative judgment based on expert experience, can no longer effectively cope with the increasingly complex risk environment. Therefore, it is urgent to develop new theories and methods to dynamically, accurately and efficiently identify and warn of various risks, so as to provide a scientific basis for improving the disaster resistance and security defense of power systems.

[0003] As the core technology of artificial intelligence, machine learning provides a new perspective for solving the problem of power system risk assessment with its powerful data mining and complex modeling capabilities. Domestic and foreign scholars have conducted extensive explorations on the application of machine learning in the fields of intelligent dispatching, load forecasting, fault diagnosis, state estimation, etc. of power systems, and have achieved fruitful results. Advanced algorithms such as deep learning, support vector machines, and random forests show unique advantages in processing high-dimensional nonlinear data, extracting implicit features, and establishing robust models. Introducing machine learning into the dynamic assessment of multiple risk sources in urban power systems is expected to break through the limitations of traditional methods, realize intelligent perception and accurate early warning of risks, and thus achieve active defense and rapid recovery of power systems.

[0004] Although machine learning has broad application prospects in power system risk assessment, its implementation still faces many technical challenges. First, the acquisition and annotation of high-quality risk data is costly, and the lack of sufficient training samples will limit the performance of the model. Second, the fusion and feature engineering of massive heterogeneous data are the basis of machine learning models, but the diversity and correlation of risk factors make this task extremely complex. In addition, the interpretability and generalization ability of the model need to be improved. Black box models are difficult to gain the trust of experts, and data distribution drift will lead to performance degradation.

[0005] Therefore, whether a dynamic risk assessment method for urban power lifeline can be provided with sufficient training samples, which can overcome the scarcity of training data, the difficulty of feature optimization and the insufficient generalization ability of the model, and realize the accurate identification and dynamic assessment of urban power system risks is a technical problem to be urgently solved in the present invention. Summary of the invention

[0006] In view of this, the present invention is based on the cutting-edge theories and technologies of machine learning, aims to overcome the technical difficulties of scarce training data, difficult feature optimization, and insufficient model generalization ability, and realizes accurate prediction and dynamic assessment of urban power system risks.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions: A method for dynamic risk assessment of urban power lifeline based on machine learning, the risk assessment method comprising the following steps: Step S100, defining the power system risk assessment problem, clarifying the assessment objectives and scope, and determining the multi-source heterogeneous data and quantitative standards that affect the security of the power system; Step S200, the data input layer performs multi-source heterogeneous data fusion, collects and integrates multi-source heterogeneous data from different sources, and the multi-source heterogeneous data includes meteorological data, geographic data and equipment status; Step S300, the feature extraction layer extracts risk features that can reflect the risk of the power system from the integrated multi-source heterogeneous data, optimizes the risk features through an automatic feature selection mechanism, and forms a feature vector; Step S400, model training and evaluation layer, using the optimized risk features and integrated superposition generalization method to train and debug the risk assessment model, and evaluate the performance of the risk assessment model to improve the robustness of risk prediction, wherein the risk assessment model is a stacked ensemble learning model; Step S500: The result output layer dynamically displays the risk assessment results to provide an intuitive reference for decision-making.

[0008] Furthermore, in step S100, the power system risk assessment gives the multi-source input characteristics of the system and predicts the future risk level of the system, which is expressed as: , in represents the input feature vector, represents the predicted risk level, represents the risk assessment model, are model parameters; Quantification and measurement of risk levels The risk level is discretized into ordered levels, defined as: in Represents the risk from level 1 to k and introduces the learning ranking metric.

[0009] Furthermore, in step S200, the multi-source heterogeneous data fusion includes a data fusion framework based on an attention mechanism to assign importance weights to different multi-source heterogeneous data: in express The feature vector of the data source, is the corresponding attention weight, is the function of weight calculation.

[0010] Furthermore, in step S300, the automatic feature selection mechanism includes an embedded feature selection mechanism for automatically selecting a feature subset during model training: in represents the original loss function, represents the regularization coefficient, Indicates the feature dimension; in this step, L1 regularization technology is applied to the model training process to reduce the weight coefficients of some features to zero, thereby automatically screening out key features that have a significant impact on the model prediction target.

[0011] Furthermore, in step S400, the steps of training and debugging the risk assessment model in the model training and evaluation layer, and evaluating the performance of the risk assessment model include: Step S410, data preparation, the risk feature data set after input by the data input layer and data preprocessing by the feature extraction layer is divided by a stratified k-fold cross-validation method, and the data set is divided into k mutually exclusive subsets according to the risk level. In each cross-validation process, one subset is selected as the test set, and the remaining k-1 subsets are used as the training set. This is repeated k times and the average value is taken. By controlling the k value, the training data can be fully utilized while the test results are stabilized; Step S420, model training, each time the training set data divided by stratified k-fold cross validation enters the model training for stacked ensemble learning, for the first-level basic learner, the training set data is used for training to obtain the model of the first-level basic learner; for the basic learners of the subsequent levels, the output of the previous level basic learner and the features of the original training set data are used as input to train the basic learners of the current level; Step S430, model performance evaluation, each time the test set data divided by the stratified K-fold cross-validation enters the model evaluation to evaluate the performance of the trained multi-level base learner, and calculates the performance evaluation index value of the multi-level base learner on the current test set; Step S440, the performance evaluation index results of the multi-level base learner obtained in the k-times stratified k-fold cross-validation are summarized through the result output layer output, and the mean value and standard deviation statistics of the performance evaluation index results are calculated to obtain a comprehensive evaluation of the model performance, and the optimal model is selected according to the summarized performance evaluation results.

[0012] Furthermore, in step S420, model training is to improve the prediction ability of the model by optimizing model parameters, wherein the objective function and solution algorithm of model training include: The objective function includes an objective function based on ranking learning: , in Indicator function, is a regularization term, and the objective function considers the relative order relationship between samples to optimize the sorting index; The solution algorithm includes a gradient boosting based model solution method: in represents the t regression tree, is the corresponding weight coefficient. The algorithm alternately performs tree structure and weight coefficient optimization and continuously approaches the optimal solution.

[0013] Furthermore, in step S420, stacked ensemble learning in model training improves the robustness of risk prediction, including the following: firstly, multiple heterogeneous base learners are trained, and then a meta-learner is used for combined prediction: in represents the basic learner, represents the meta-learner.

[0014] Furthermore, the risk assessment method also includes model assessment and dynamic update, including: Step S501, offline evaluation based on backtesting, uses the backtesting method to evaluate the risk assessment model offline, and the historical data is divided into a training set, a validation set, and a test set to simulate the prediction performance of the model in different time periods: in, , , They represent the training set, validation set, and test set of m respectively, and comprehensively examine the historical performance of the model by combining the evaluation indicators of different periods; Step S502: Real-time update of the risk assessment model is based on an adaptive dynamic update mechanism: in represents the new data t collected at the first moment, Represents the learning speed.

[0015] The beneficial effects of the present invention are: Urban power system risk assessment is a systematic project involving multi-source heterogeneous data fusion and complex correlation modeling. Traditional single models are difficult to handle effectively. This application designs a modular, multi-level dynamic risk assessment model and assessment method, including data input layer, feature extraction layer, model training and evaluation layer, and result output layer. The data input layer is responsible for integrating multi-source data such as meteorology, geography, and equipment status. The feature extraction layer optimizes risk characteristics through feature selection, dimensionality reduction, and other technologies. The model training and evaluation layer adopts a comprehensive learning strategy, integrates multiple machine learning algorithms, and introduces cross-validation, grid search, and other parameter adjustment and performance evaluation methods. The result output layer dynamically displays the risk assessment results to provide an intuitive reference for decision-making. The standard interface design between modules not only ensures the relative independence of the functions, but also takes into account the coordinated integration of the system, providing convenience for the iterative optimization of the model.

[0016] The dynamic risk assessment method described in the present invention has conducted in-depth research on key technologies such as data preprocessing, feature selection and optimization, model training and evaluation. This application aims to overcome the "bottleneck" problem of machine learning implementation, such as feature engineering and model generalization, and the research results directly serve the needs of engineering practice. In addition, this application establishes a dynamic assessment model for multiple risk sources in urban power systems based on machine learning, providing intelligent authorization for power dispatching, emergency management, etc., and has broad application prospects.

[0017] This application also takes into account the limitations of a single model and uses ensemble learning methods such as stacking and mixing of stacked ensemble learning models to further improve the robustness and accuracy of the model by combining the prediction results of multiple base learners. When selecting a model, we not only focus on the inherent mechanism of the theoretical analysis algorithm, but also compare the actual performance of different models in specific risk assessment tasks through empirical comparison, striving to select the optimal modeling solution. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is the overall architecture diagram of the present invention.

[0019] Figure 2 It is a schematic diagram of the training and debugging process of the stacked ensemble learning model of the present invention.

[0020] Figure 3 It is the average value and standard deviation of the performance indicators of different machine learning modes on the test set in the embodiments of the present invention.

[0021] Figure 4It is a Nemenyi test result diagram based on the F1 value in an embodiment of the present invention.

[0022] Figure 5 This is an example of the risk assessment results of the model for the first week of August 2022. DETAILED DESCRIPTION

[0023] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0024] like Figure 1-Figure 2 As shown, a dynamic risk assessment method for urban power lifeline based on machine learning includes steps S100-S500.

[0025] Step S100, define the power system risk assessment problem, clarify the assessment objectives and scope, and determine the multi-source heterogeneous data and quantitative standards that affect the security of the power system. The definition of the assessment problem can help guide the process of multi-source heterogeneous data collection and analysis. Clarifying which data is important and which data is secondary can improve the efficiency of data collection and ensure the correct direction of data analysis.

[0026] Power system risk assessment is a supervised learning problem, which gives the multi-source input characteristics of the system and predicts the future risk level of the system. The expression is: , in represents the input feature vector, represents the predicted risk level, represents the risk assessment model, are model parameters.

[0027] Quantification and measurement of risk levels The risk level is discretized into ordered levels, defined as: in Indicates the risk from level 1 to k, and introduces the learning ranking indicator. For example, NDCG@K: , in express The relevance score of the sample. $NDCG@K$ is able to consider the sequential quality of the predicted risk level.

[0028] Step S200: The data input layer performs multi-source heterogeneous data fusion according to the defined power system risk assessment problem, and collects and integrates multi-source heterogeneous data from different sources. The multi-source heterogeneous data includes meteorological data, geographic data, equipment status data and other multi-source data.

[0029] In step S200, multi-source heterogeneous data fusion includes a data fusion framework based on an attention mechanism, which is used to assign importance weights to different multi-source heterogeneous data sources and integrate them into a unified data set: in express The feature vector of the data source, is the corresponding attention weight, is the function of weight calculation.

[0030] Step S300: The feature extraction layer extracts risk features that can reflect the risk of the power system from the integrated multi-source heterogeneous data, optimizes the risk features through an automatic feature selection mechanism, and forms a feature vector.

[0031] In step S300, the automatic feature selection mechanism is the key to risk assessment, including an embedded feature selection mechanism for automatic selection of feature subsets during model training: in represents the original loss function, represents the regularization coefficient, Indicates the feature dimension; in this step, L1 regularization technology is applied to the model training process to reduce the weight coefficients of some features to zero, thereby automatically screening out key features that have a significant impact on the model prediction target.

[0032] Step S400, model training and evaluation layer, uses the optimized risk features and integrated superposition generalization method to train and debug the risk assessment model, and evaluates the performance of the risk assessment model to improve the robustness of risk prediction, wherein the risk assessment model is a stacked ensemble learning model.

[0033] Step S500: The result output layer dynamically displays the risk assessment results to provide an intuitive reference for decision-making.

[0034] In step S400, the steps of training and debugging the risk assessment model in the model training and evaluation layer, and evaluating the performance of the risk assessment model include S410-S440.

[0035] Step S410, data preparation, the risk feature data set after input by the data input layer and data preprocessing by the feature extraction layer is divided by a stratified k-fold cross-validation method, and the data set is divided into k mutually exclusive subsets according to the risk level. In each cross-validation process, one subset is selected as the test set, and the remaining k-1 subsets are used as training sets. This is repeated k times and the average value is taken. By controlling the k value, the training data can be fully utilized while the test results are stabilized.

[0036] Step S420, model training, the training set data divided by each stratified k-fold cross-validation enters the model training for stacked ensemble learning. For the first-level basic learner, the training set data is used for training to obtain the model of the first-level basic learner; for the basic learners of subsequent levels, the output of the previous level basic learner and the features of the original training set data are used as input to train the basic learners of the current level.

[0037] Step S430, model performance evaluation, each time the test set data divided by stratified K-fold cross-validation enters the model evaluation to perform performance evaluation on the trained multi-level base learner, and calculates the performance evaluation index value of the multi-level base learner on the current test set.

[0038] Step S440, the performance evaluation index results of the multi-level base learner obtained in the k-times stratified k-fold cross-validation are summarized through the result output layer output, and the mean value and standard deviation statistics of the performance evaluation index results are calculated to obtain a comprehensive evaluation of the model performance, and the optimal model is selected according to the summarized performance evaluation results.

[0039] In step S420, model training is to improve the prediction ability of the model by optimizing model parameters, wherein the objective function and solution algorithm of model training include: The objective function includes an objective function based on ranking learning: , in Indicator function, is a regularization term, and the objective function considers the relative order relationship between samples to optimize the sorting index.

[0040] The solution algorithm includes a gradient boosting based model solution method: in represents the t regression tree, is the corresponding weight coefficient. The algorithm alternately performs tree structure and weight coefficient optimization and continuously approaches the optimal solution.

[0041] In step S420, stacked ensemble learning in model training improves the robustness of risk prediction, including the following: first, multiple heterogeneous base learners are trained, and then a meta-learner is used for combined prediction: in represents the basic learner, represents the meta-learner.

[0042] The risk assessment method also includes model assessment and dynamic update, including: Step S501, offline evaluation based on backtesting, uses the backtesting method to evaluate the risk assessment model offline, and the historical data is divided into a training set, a validation set, and a test set to simulate the prediction performance of the model in different time periods: in, , , They represent the training set, validation set, and test set of m, respectively, and comprehensively examine the historical performance of the model by combining the evaluation indicators of different periods; Step S502: Real-time update of the risk assessment model is based on an adaptive dynamic update mechanism: in represents the new data t collected at the first moment, Represents the learning speed.

[0043] like Figure 2-4 As shown, this embodiment selects high-quality operating data provided by a large urban power company. The data set covers the entire operation of the power grid from 2015 to 2023, including multi-source heterogeneous data such as meteorological conditions, equipment conditions, fault information and load characteristics, with a total capacity of more than 10TB. This application adopts a systematic data cleaning and fusion strategy to remove missing values ​​and outliers, and constructs a structured sample set containing 53 feature dimensions and 5 levels of risk labels. The sample size exceeds 10 million, which meets the requirements for large-scale machine learning algorithm training and evaluation. The representativeness and richness of this data set provide a solid data foundation for the research of the embodiments of this application.

[0044] In order to strictly evaluate the generalization performance of the model and prevent the risk of overfitting, the stratified k-fold cross-validation method is used to divide the data set. Specifically, the data set is divided into k mutually exclusive subsets according to the risk level. Each time, one subset is selected as the test set, and the remaining k-1 subsets are used as training sets. The experiment is repeated k times and the average value is taken. By controlling the k value, a balance is achieved between making full use of the training data and stabilizing the test results. In this embodiment, k is 5, that is, 5-fold cross-validation is used. At the same time, in order to verify the robustness of the results, the division and experiment were repeated 10 times, and the average value and standard deviation of the 10 experiments were reported. This data partitioning strategy not only ensures the objectivity and fairness of the model evaluation, but also improves the credibility of the experimental results.

[0045] In this embodiment, systematic model training and adjustment are also performed on mainstream machine learning algorithms such as SVM, RF and LightGBM. Taking RF as an example, the Bayesian optimization method is used to automatically search for hyperparameters such as the number of trees, maximum depth and minimum number of node samples, and the optimization target is the cross-validated f1 value. In this process, the accuracy and complexity of the model are comprehensively considered, and an early stopping strategy is introduced to control the training time. Similarly, the penalty coefficient and kernel function type of the support vector machine are optimized, and key parameters such as the number of leaf nodes and learning rate of LightGBM are optimized. GPU acceleration and parallelization technology are used in the hyperparameter search process, which greatly improves the tuning efficiency.

[0046] like Figure 2 As shown, in order to further explore the potential of machine learning, the embodiment explores a variety of integrated learning paradigms, studies the superposition generalization method in detail, and designs a multi-level base learner combination strategy. Taking the first layer of 5 LightGBMs and the second layer of SVM as an example, the base learners use different feature subspaces and hyperparameters to enhance the diversity of the set, and the meta-learner uses SVR for fusion to improve the generalization ability. The training of the integrated method also uses k-fold cross-validation, and the Optuna framework is used to automatically optimize the hyperparameters of the base learners and the meta-learner weights. This automated and refined integrated model development process greatly improves model performance.

[0047] Figure 3 The mean and standard deviation of the performance indicators of different machine learning models on the test set are listed (in brackets). It can be seen that LightGBM is significantly better than SVM and RF in terms of precision, recall, F1 value and other indicators, reflecting its advantages in processing high-dimensional and nonlinear data. At the same time, the performance of the comprehensive model is generally better than that of a single model. The stacked ensemble learning model (LR+LGBM) has the best overall performance, with an F1 value of 0.915 ± 0.006, which is 2.3% higher than the best single-machine model LGBM. The McNemar test confirmed that the performance difference between the comprehensive model and other models was significant, with a significance level of 0.01, indicating the effectiveness of stacked integration. In addition, the embodiment also plotted the precision-recall curve of the model, as shown in Figure 4 As shown in Figure 2, the curves of the integrated model always surround the individual models, further demonstrating the superiority of the integrated approach. The results show that the integrated strategy can significantly improve the performance and robustness of the model, and superposition generalization shows the best improvement.

[0048] In order to more systematically compare the advantages and disadvantages of different types of algorithms, the statistical significance of the performance differences between the two algorithms was analyzed using Friedman test and Nemenyi follow-up test based on model ranking. Figure 4 The results of Nemenyi test based on F1 value are shown. The rankings of different algorithms are represented by points, and the horizontal lines represent the critical differences. It can be seen that the ranking of the integrated model is significantly better than that of the single model. Superposition (LR + LGBM) obtained the best average ranking of 1.6, followed by superposition (SVM + LGBM) and LGBM. The rankings of support vector machine (SVM) and radio frequency (RF) are relatively low, and the performance difference between them is not significant. The results show that considering multiple evaluation indicators comprehensively, the superposition integration model with LGBM as the base learner can obtain the best comprehensive performance, which is better than the single model and other integration schemes. It further verifies the effectiveness and superiority of the superposition generalization method in the power system risk assessment task.

[0049] Through rigorous empirical analysis, this experiment verifies the effectiveness of the power system risk assessment model based on machine learning, reveals the differences in the applicability of different types of algorithms, and proves the significant superiority of stacked ensemble learning. These findings provide new ideas and methods for the practice of power system risk management.

[0050] In order to verify the applicability and effectiveness of the model in the actual power system, this embodiment selects the power system of a city as the application demonstration object. The city is located in the central region. The power grid covers an area of ​​about 20,000 square kilometers. The power supply targets cover multiple fields such as industry, commerce and residents. The load types are diverse, and the power demand has obvious seasonal and cyclical characteristics. At the same time, the city is located in an area prone to earthquakes and floods, extreme weather events occur frequently, the problem of aging power equipment is prominent, and the network security situation is severe. The power system faces a variety of complex risk sources.

[0051] This example collects the city's power system operation data from 2018 to 2023, including meteorological information, equipment status, fault records, load curves, etc., and constructs a risk assessment sample set containing 48 features based on the data preprocessing process. On this basis, the trained stacked ensemble model is used to conduct an hourly risk assessment of the city's power system. Considering the temporal applicability of the model, the sliding time window method is used for prediction, that is, the data of the past three months is used each time to assess the risk of the next week, and the model is continuously updated over time.

[0052] Figure 5An example of the risk assessment results of the model in the first week of August 2022 is shown. It can be seen that the model can accurately reflect the changes in the risk level of the power system in different periods, and can identify the high-risk periods on August 3 and August 5. Further analysis shows that the high risk on August 3 was mainly due to the severe convective weather in City A, and the occurrence of disastrous weather such as heavy rain and hail in some areas, which caused damage to the transmission lines and substation equipment and caused large-scale power outages. The high risk on August 5 was related to a major equipment accident that occurred on that day. A main transformer failed due to insulation aging, causing a large area of ​​load to lose power.

[0053] In order to quantitatively evaluate the effect of the model in practical applications, this application uses indicators such as accuracy, false alarm rate, and false alarm rate as benchmarks to compare the risk level predicted by the model with actual risk events. The verification period is for the whole year of 2020, and the evaluation is carried out once a week. The results show that the risk prediction accuracy of the model reaches 87.5%, and the false alarm rate and false alarm rate are controlled within 9.6% and 7.8% respectively. Further use of the prediction results to guide power grid operation and emergency response can increase the risk prevention and control capabilities and rapid recovery capabilities of the power system by more than 15%. This fully demonstrates the reliability and practicality of the model in the actual power system.

[0054] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for dynamic risk assessment of urban power lifeline based on machine learning, characterized in that: The risk assessment method comprises the following steps: Step S100, defining the power system risk assessment problem, clarifying the assessment objectives and scope, and determining the multi-source heterogeneous data and quantitative standards that affect the security of the power system; Step S200, the data input layer performs multi-source heterogeneous data fusion, collects and integrates multi-source heterogeneous data from different sources, and the multi-source heterogeneous data includes meteorological data, geographic data and equipment status; Step S300, the feature extraction layer extracts risk features that can reflect the risk of the power system from the integrated multi-source heterogeneous data, optimizes the risk features through an automatic feature selection mechanism, and forms a feature vector; Step S400, model training and evaluation layer, using the optimized risk features and integrated superposition generalization method to train and debug the risk assessment model, and evaluate the performance of the risk assessment model to improve the robustness of risk prediction, wherein the risk assessment model is a stacked ensemble learning model; Step S500: The result output layer dynamically displays the risk assessment results to provide an intuitive reference for decision-making.

2. A method for dynamic risk assessment of urban power lifeline based on machine learning according to claim 1, characterized in that: In step S100, the power system risk assessment gives the multi-source input characteristics of the system and predicts the future risk level of the system, which is expressed as: ， in represents the input feature vector, represents the predicted risk level, represents the risk assessment model, are model parameters; Quantification and measurement of risk levels The risk level is discretized into ordered levels, defined as: in Represents the risk from level 1 to k and introduces the learning ranking metric.

3. A method for dynamic risk assessment of urban power lifeline based on machine learning according to claim 1, characterized in that: In step S200, the multi-source heterogeneous data fusion includes a data fusion framework based on an attention mechanism to assign importance weights to different multi-source heterogeneous data: in express The feature vector of the data source, is the corresponding attention weight, is the function of weight calculation.

4. A method for dynamic risk assessment of urban power lifeline based on machine learning according to claim 1, characterized in that: In step S300, the automatic feature selection mechanism includes an embedded feature selection mechanism for automatic selection of feature subsets during model training: in represents the original loss function, represents the regularization coefficient, represents the feature dimension; In this step, L1 regularization technology is applied to the model training process to reduce the weight coefficients of some features to zero, thereby automatically screening out key features that have a significant impact on the model prediction target.

5. A method for dynamic risk assessment of urban power lifeline based on machine learning according to claim 1, characterized in that: In step S400, the steps of training and debugging the risk assessment model in the model training and evaluation layer, and evaluating the performance of the risk assessment model include: Step S410, data preparation, the risk feature data set after input by the data input layer and preprocessed by the feature extraction layer is divided by a stratified k-fold cross-validation method, and the data set is divided into k mutually exclusive subsets according to the risk level. In each cross-validation process, one subset is selected as the test set, and the remaining k-1 subsets are used as the training set. This is repeated k times and the average value is taken. By controlling the k value, the training data can be fully utilized while the test results are stabilized; Step S420, model training, each time the training set data divided by stratified k-fold cross validation enters the model training for stacked ensemble learning, for the first-level basic learner, the training set data is used for training to obtain the model of the first-level basic learner; for the basic learners of the subsequent levels, the output of the previous level basic learner and the features of the original training set data are used as input to train the basic learners of the current level; Step S430, model performance evaluation, each time the test set data divided by the stratified K-fold cross-validation enters the model evaluation to evaluate the performance of the trained multi-level base learner, and calculates the performance evaluation index value of the multi-level base learner on the current test set; Step S440, the performance evaluation index results of the multi-level base learner obtained in the k-times stratified k-fold cross-validation are summarized through the result output layer output, and the mean value and standard deviation statistics of the performance evaluation index results are calculated to obtain a comprehensive evaluation of the model performance, and the optimal model is selected according to the summarized performance evaluation results.

6. A method for dynamic risk assessment of urban power lifeline based on machine learning according to claim 5, characterized in that: In step S420, model training is to improve the prediction ability of the model by optimizing model parameters, wherein the objective function and solution algorithm of model training include: The objective function includes an objective function based on ranking learning: ， in Indicator function, is a regularization term, and the objective function considers the relative order relationship between samples to optimize the sorting index; The solution algorithm includes a gradient boosting based model solution method: in represents the t regression tree, is the corresponding weight coefficient. The algorithm alternately performs tree structure and weight coefficient optimization and continuously approaches the optimal solution.

7. A method for dynamic risk assessment of urban power lifeline based on machine learning according to claim 5, characterized in that: In step S420, stacked ensemble learning in model training improves the robustness of risk prediction, including the following: first, multiple heterogeneous base learners are trained, and then a meta-learner is used for combined prediction: in represents the basic learner, represents the meta-learner.

8. A method for dynamic risk assessment of urban power lifeline based on machine learning according to claim 1, characterized in that: The risk assessment method also includes model assessment and dynamic update, including: Step S501, offline evaluation based on backtesting, uses the backtesting method to evaluate the risk assessment model offline, and the historical data is divided into a training set, a validation set, and a test set to simulate the prediction performance of the model in different time periods: in, , , They represent the training set, validation set, and test set of m respectively, and comprehensively examine the historical performance of the model by combining the evaluation indicators of different periods; Step S502: Real-time dynamic update of the risk assessment model is based on an adaptive dynamic update mechanism: in represents the new data t collected at the first moment, Represents the learning speed.

Images