Power system flexibility resource classification model construction method based on machine learning

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a method for constructing a flexible resource classification model of a power system based on machine learning, which comprises the following steps of: utilizing data from a plurality of sources such as power utilization records of a power grid company, a load acquisition system, historical power utilization data, weather environment data and power equipment operation data; and training is carried out through a random forest algorithm, and accurate classification of flexible resources is realized. Data are marked into different types manually, and the depth and splitting conditions of a decision tree are optimized through a dynamic constraint mechanism, so that the classification precision and robustness are improved. The importance of each feature in the decision tree generation process is quantified, the optimal feature is selected for node splitting, and a complete decision tree model is constructed step by step; and finally, through integrating prediction results of a plurality of decision trees, determining final classification output by utilizing a voting mechanism, completing effective classification of the flexible resources of the power system, and providing support for power dispatching and management.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method for constructing a power system flexibility resource classification model based on machine learning. Background Art

[0002] As the world pays more and more attention to the use of renewable energy, the operation mode of the power system is undergoing significant changes. Especially at a time when intermittent and fluctuating energy sources such as wind and solar energy are connected to the power grid on a large scale, the power system has an increasingly urgent need for flexible and intelligent dispatching strategies. These strategies are essential to ensure the stable operation of the power grid and the reliability of power supply. Flexible resources such as adjustable loads and energy storage equipment play a key role in the power system. They can be flexibly adjusted according to changes in electricity demand, which helps to alleviate the imbalance between supply and demand caused by fluctuations in renewable energy.

[0003] However, the current power system faces many challenges in the management of flexible resources. Existing management methods mainly rely on manual experience and static rules, but they are unable to cope with large-scale, multi-dimensional data. Specifically, multi-source data such as user electricity consumption records, load fluctuations, weather changes, and equipment status are often processed in isolation, lacking effective integration and deep mining. This processing method limits the accuracy of predicting fluctuations in electricity demand and makes it difficult to respond quickly to dynamic changes in the system.

[0004] In addition, existing management methods are limited by their characteristics based on static models or fixed parameters and lack the ability of adaptive optimization. The load characteristics of the power system will change with multiple factors such as time, weather, and equipment operating status, but traditional methods are difficult to flexibly adapt to these changes, thus affecting the efficiency and stability of power grid dispatching.

[0005] The information disclosed in this background technology section is only intended to deepen the understanding of the overall background technology of the present disclosure, and should not be regarded as acknowledging or suggesting in any form that the information constitutes the prior art known to those skilled in the art. Summary of the invention

[0006] The present invention provides a method for classifying power system flexibility resources based on machine learning, which can effectively solve the problems in the background technology.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is:

[0008] The method for constructing a power system flexibility resource classification model based on machine learning includes the following steps:

[0009] Collect multi-source power data and manually annotate it;

[0010] Feature extraction and data preprocessing based on multi-source power data;

[0011] Build a classification model based on the random forest algorithm and perform model training;

[0012] A dynamic constraint mechanism is used to optimize the generation process of decision trees in the classification model, and the classification model is obtained by integrating the outputs of multiple decision trees.

[0013] Furthermore, the collection of multi-source power data is based on power consumption records of power grid companies, load collection systems, historical power consumption data, weather and environmental data, and power equipment operation data;

[0014] The collected multi-source power data are manually labeled as high-flexibility resources, medium-flexibility resources, non-flexible resources, and non-adjustable loads.

[0015] Furthermore, feature extraction based on multi-source data includes load value, timestamp, electricity consumption type and load volatility, and the extracted feature samples and their corresponding sample labels are used as training data for the classification model.

[0016] Furthermore, the training process of the classification model includes the following steps:

[0017] Initialize the classification model and set the parameters of the random forest algorithm;

[0018] Dynamically adjust the generation constraints of decision trees in the random forest algorithm;

[0019] In the process of decision tree generation, node splitting features and splitting points are adaptively adjusted;

[0020] During the feature selection process at each node, dynamic evaluation of feature importance is performed;

[0021] The output results of all decision trees are integrated, and a voting mechanism is used to determine the classification result.

[0022] Furthermore, the decision tree is trained based on the feature samples and their corresponding sample labels. After initializing the classification model, the number of decision trees N required for random forest algorithm training is set. u , specifically expressed as:

[0023]

[0024] Among them, Θ u Represents the set of all decision trees in the random forest algorithm, T u represents the u-th decision tree;

[0025] In the initial stage, no constraints are imposed on any decision tree, and the depth of any decision tree is dynamically adjustable.

[0026] Furthermore, according to the distribution characteristics of the training data, the generation constraints of the decision tree are dynamically adjusted. Specifically, the depth or splitting condition of the decision tree is adjusted by defining the constraint factor δu, which is specifically expressed as:

[0027]

[0028] Among them, f() is the activation function, which indicates the calculation method of the constraint factor; m c Represents the number of characteristic samples of training data; Ⅱ() is the indicator function, when Indicates that the sample labels of the i-th feature sample and the i'th feature sample are different.

[0029] Furthermore, the node splitting features and the corresponding splitting points are determined based on the information gain characteristics of the training data, and the splitting metric is defined as information gain, which is specifically expressed as:

[0030]

[0031] Among them, IG() is the information gain function; H(Y c ) represents the Shannon entropy of the target variable, H(Y c |X ck = r) indicates that in feature X ck When r is the condition, Y c Conditional entropy; m r Indicated in feature X ck = the number of samples under r; m c Represents the total number of training samples.

[0032] Furthermore, the optimal feature is selected based on the information gain metric And calculate the maximum value of information gain for feature division;

[0033] Based on the dynamic constraint mechanism, the decision tree structure is optimized by increasing the node depth or refining the splitting conditions. The specific adjustment method is expressed as follows:

[0034]

[0035] Among them, Δ Tu Represented as a decision tree T u The adjustment amount of the midpoint split; f() is the activation function used to smoothly adjust the split condition; η u Indicates the preset threshold.

[0036] Furthermore, a weighted average strategy is used to dynamically evaluate the importance of features, which is specifically expressed as follows:

[0037]

[0038] Among them, W cirepresents the weight of the i-th feature sample; Ⅱ(X ck =x ci ) is the indicator function, indicating that in the i-th sample, feature X ck Is it equal to its value x? ci .

[0039] Furthermore, any decision tree for input sample X u Classify and output a classification result The classification output of the classification model is determined by the voting results of all decision trees, which is specifically expressed as:

[0040]

[0041] in, is the final prediction output of the classification model; Ⅱ() is the indicator function, indicating whether the decision tree predicts category y.

[0042] The beneficial effects of the present invention are:

[0043] During the random forest training process, a dynamic constraint mechanism is used to optimize the generation process of the decision tree. According to the distribution characteristics of the samples, the depth of the tree and the node splitting conditions will be adaptively adjusted, which can improve the adaptability of the model to complex samples and effectively improve the classification accuracy and robustness.

[0044] The method of dynamically adjusting the depth of the decision tree is adopted. When the data samples are difficult to distinguish, the decision tree is allowed to generate a deeper structure. Compared with the traditional decision tree generation mechanism, it has better flexibility and can flexibly adjust the classification strategy according to the sample characteristics.

[0045] During feature selection, the importance of features is dynamically evaluated through a weighted average strategy to ensure that each decision tree can make split decisions based on the most relevant features. This is smarter than traditional static feature selection methods and can be optimized according to changes in the data set. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0047] Figure 1 It is a schematic diagram of the framework of the classification model construction method in the present invention. DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0049] It should be noted that when an element is referred to as being "fixed to" another element, it may be directly on the other element or there may be a central element. When an element is considered to be "connected to" another element, it may be directly connected to the other element or there may be a central element at the same time. The terms "vertical", "horizontal", "left", "right" and similar expressions used herein are for illustrative purposes only and do not represent the only implementation method.

[0050] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items.

[0051] In view of the supply-demand imbalance problem caused by the large-scale access of renewable energy to the power grid, the present invention proposes a strategy of adjusting flexibility resources to adapt to changes in electricity demand. Specifically, a power system flexibility resource classification model based on machine learning is used to classify flexibility resources. Therefore, a corresponding method for constructing a power system flexibility resource classification model based on machine learning is proposed. Figure 1 As shown, the following steps are included:

[0052] Collect multi-source power data and perform manual labeling; perform feature extraction and data preprocessing based on multi-source power data; build a classification model based on the random forest algorithm and perform model training; use a dynamic constraint mechanism to optimize the generation process of the decision tree in the classification model, and obtain the classification model by integrating the outputs of multiple decision trees.

[0053] During the random forest training process, a dynamic constraint mechanism is used to optimize the generation process of the decision tree. According to the distribution characteristics of the sample, the depth of the tree and the node splitting conditions will be adaptively adjusted, which can improve the adaptability of the classification model to complex samples and effectively improve the classification accuracy and robustness. The method of dynamically adjusting the depth of the decision tree allows the decision tree to generate a deeper structure when the data samples are difficult to distinguish. Compared with the traditional decision tree generation mechanism, it has better flexibility and can flexibly adjust the classification strategy according to the sample characteristics.

[0054] In this embodiment, the collection of multi-source power data is based on the power consumption records of the power grid company, the load collection system, historical power consumption data, weather and environmental data, and power equipment operation data; the collected multi-source power data is manually labeled as high-flexibility resources, medium-flexibility resources, non-flexible resources, and non-adjustable loads.

[0055] Specifically, the power grid company's electricity consumption record data comes from the power company's monitoring and recording of the electricity consumption of users, regions or equipment. The power grid company collects information such as electricity consumption, load changes, power factor, etc. of each user or region through regular or real-time monitoring data acquisition systems (for example, smart meters), reflecting the dynamic changes in power load.

[0056] The load collection system data realizes real-time load monitoring of different power users, equipment or areas through the load collection system (such as smart meters, distribution automation systems, etc.). The load collection system can collect data such as load volatility, peak values, valley values ​​and duration to form time series data.

[0057] Historical electricity consumption record data is obtained by analyzing historical electricity consumption data to obtain users' electricity consumption trends, fluctuations, and periodic patterns over long time periods, reflecting the long-term changing characteristics of user demand and its impact on the power system.

[0058] Weather and environmental data This data comes from the meteorological monitoring system and includes environmental factors such as temperature, humidity, wind speed, and air pressure. Environmental factors will directly affect the fluctuation of electricity demand. For example, high temperature weather will lead to increased use of air conditioning, thereby putting greater pressure on the power system.

[0059] The status and operation data of power equipment include the operation data of power system equipment such as generators, substations, and transmission lines. The load status, voltage, current, frequency and other information of the equipment are also important reference data for the classification of flexibility resources.

[0060] In one embodiment, the attributes of the data include:

[0061] Ra is the load value (in kilowatts, representing the load of the user or area at a certain moment);

[0062] Da is the timestamp (in hours, indicating the specific time of data collection);

[0063] Fa is the type of electricity consumption (such as household, commercial, industrial, etc., used to indicate the type of user);

[0064] La is the load volatility (indicating the magnitude of load fluctuation, such as standard deviation).

[0065] It should be noted that this embodiment is only used to illustrate one data format and type of the present invention. In actual applications, the attributes of data are usually more than 10 attributes, and the number of attributes of data may reach dozens or even hundreds.

[0066] Furthermore, the collected data is annotated. The annotating method of the present invention is manual annotating. In one embodiment, the annotated categories include:

[0067] Type A: Flexible resources (high flexibility);

[0068] Type B: Flexible resources (medium flexibility);

[0069] Type C: Inflexible resources (low flexibility);

[0070] Type D: Non-adjustable load.

[0071] Furthermore, useful features are extracted from the original data. Feature extraction based on multi-source data includes load value, timestamp, electricity consumption type and load volatility. The extracted features are preprocessed, such as data cleaning (noise removal, missing value processing), feature scaling (making different features on the same scale), etc., to ensure the effectiveness and accuracy of model training. The extracted feature samples and their corresponding sample labels are used as training data for the classification model.

[0072] The training process of the classification model includes the following steps:

[0073] Initialize the classification model and set the parameters of the random forest algorithm; dynamically adjust the generation constraints of the decision tree in the random forest algorithm; adaptively adjust the node splitting features and splitting points during the generation of the decision tree; dynamically evaluate the importance of features during the feature selection process of each node; integrate the output results of all decision trees and use a voting mechanism to determine the classification results.

[0074] In this embodiment, random forest can capture different patterns in the data by training multiple decision trees, each of which is trained on different sub-samples and feature subsets, thereby increasing the diversity of the model; integrating multiple decision trees helps to reduce the dependence of each tree on the noise in the training data, reduce the risk of overfitting, and improve the generalization ability of the model.

[0075] Specifically, the decision tree is trained based on the feature samples and their corresponding sample labels. After the classification model is initialized, the number of decision trees N required for random forest algorithm training is set. u , specifically expressed as:

[0076]

[0077] Among them, Θ uRepresents the set of all decision trees in the random forest algorithm, T u represents the u-th decision tree;

[0078] In the initial stage, no constraints are imposed on any decision tree, and the depth of any decision tree is dynamically adjustable.

[0079] The number of decision trees is a hyperparameter that can be adjusted according to the complexity of the data and computing resources. Generally, the more decision trees there are, the more stable the performance of the model. In the initial stage, the depth of each decision tree is dynamically adjustable, which means that during the training process, the model will automatically adjust the depth of the tree according to the characteristics of the data and the training effect.

[0080] In the specific implementation process, by combining the prediction results of multiple trees, random forests are able to make more accurate predictions and usually perform better than a single decision tree; random forests are insensitive to slight changes in the training data, and even if some trees are overfitting, other trees can compensate, thereby improving the robustness of the model; the depth of each tree is dynamically adjustable in the initial stage, which means that the model can automatically adjust according to the complexity and noise level of the data to optimize performance.

[0081] In this embodiment, considering the different difficulty levels of data samples, for samples that are easy to distinguish, a shallow tree can achieve good classification results; for samples that are difficult to distinguish, a deeper tree is required to capture complex patterns. Therefore, by dynamically adjusting the depth of the tree, the adaptability of the model to complex samples can be improved, unnecessary computational overhead can be avoided, and the accuracy of the model can be maintained.

[0082] Specifically, during the growth of the decision tree, the generation constraints of the decision tree are dynamically adjusted according to the distribution characteristics of the training data. When the data samples are easy to distinguish, the growth of the tree is limited and the tree depth is shallow; when the data samples are difficult to distinguish or highly overlapping, the growth of the tree has no obvious restrictions, allowing a deeper structure to be generated, thereby improving the adaptability of the model to complex samples. Specifically, the constraint factor δ is defined u Adjust the depth or splitting condition of the decision tree, specifically expressed as:

[0083]

[0084] Among them, f() is the activation function, which indicates the calculation method of the constraint factor; m c Represents the number of characteristic samples of training data; Ⅱ() is the indicator function, when Indicates that the sample labels of the i-th feature sample and the i'th feature sample are different; the constraint factor δ u It is calculated based on the proportion of inconsistent sample labels. A higher proportion means more complex splitting conditions for the tree, allowing the tree to grow deeper.

[0085] In this embodiment, different features and split points have a significant impact on the performance of the model. Adaptive adjustment can ensure the selection of the optimal features and split points, and improve the accuracy of the model. Selecting the optimal features through metrics such as information gain can ensure that the model captures the key information in the data. At the same time, combined with a dynamic constraint mechanism, overfitting can be avoided and the generalization ability of the model can be improved.

[0086] Specifically, in the process of generating each tree, adaptive adjustments are made. First, the node splitting features and splitting points are selected based on the statistical characteristics of the data (such as information gain, Gini index, etc.). Then, in the process of tree growth, the tree structure is adaptively adjusted according to the aforementioned dynamic constraint mechanism. Assume that each time a split occurs, feature X is selected. ck As the optimal splitting point, the node splitting features and the corresponding splitting points are determined based on the information gain characteristics of the training data, and the splitting metric is defined as information gain, which is specifically expressed as:

[0087]

[0088] Among them, IG() is the information gain function; H(Y c ) represents the Shannon entropy of the target variable, H(Y c |X ck = r) indicates that in feature X ck When r is the condition, Y c Conditional entropy; m r Indicated in feature X ck = the number of samples under r; m c Represents the total number of training samples.

[0089] Furthermore, the optimal feature is selected based on the above information gain metric And calculate the maximum value of information gain, which is used as the basis for feature division.

[0090] Based on the dynamic constraint mechanism, if the information gain of a split is less than the preset threshold η u , the split will not continue, but the depth of the node will be increased or more detailed split conditions will be added. The decision tree structure will be optimized by increasing the node depth or refining the split conditions. The specific adjustment method is expressed as:

[0091]

[0092] Among them, Δ Tu Represented as a decision tree T u The adjustment amount of the node split; f() is the activation function used to smoothly adjust the splitting conditions to ensure the continuity and adaptability of the splitting process; η uRepresents the preset threshold. Through the above steps, the decision tree can dynamically adjust the node splitting during the generation process, thereby improving the overall performance of the model.

[0093] In this embodiment, considering that different features have different importance to the model, dynamically evaluating the feature importance can ensure that the features that contribute most to the model are selected. Using a weighted average strategy to evaluate feature importance can fully consider the weights and feature values ​​of the samples, thereby improving the sensitivity and accuracy of the model to the data.

[0094] Specifically, in the process of generating each tree, feature selection is performed by dynamically evaluating the importance of the features of the training data. i When performing feature selection, a weighted average strategy is used to dynamically evaluate the importance of features, which is specifically expressed as:

[0095]

[0096] Among them, W ci represents the weight of the i-th feature sample; Ⅱ(X ck =x ci ) is the indicator function, indicating that in the i-th sample, feature X ck Is it equal to its value x? ci .

[0097] During the implementation, a weight W is first assigned to each sample ci , these weights may be based on sample rarity, accuracy, or other metrics; for each sample, check feature X ck Is it equal to its specific value x? ci , if they are equal, the indicator function value is 1, otherwise it is 0; the weight W of each sample is ci Multiply it with the corresponding indicator function value, and then sum these products to get the feature X ck Final rating of .

[0098] This embodiment quantifies the importance of each feature in the decision tree generation process, thereby selecting the optimal feature for node splitting and gradually building a complete decision tree model, ensuring that the model can make the most reasonable decision based on the feature distribution of the training data.

[0099] In this embodiment, integrating the outputs of multiple decision trees can reduce the errors that may exist in a single tree and improve the stability and accuracy of the model; adopting a voting mechanism to summarize the outputs of all trees can make full use of the information of each tree and improve the model's generalization ability and prediction performance for data.

[0100] Specifically, after all decision trees are trained, the outputs of all decision trees are summarized through the integration method.u Classify and output a classification result The classification output of the classification model is determined by the voting results of all decision trees, which is specifically expressed as:

[0101]

[0102] in, is the final prediction output of the classification model; Ⅱ() is the indicator function, indicating whether the decision tree predicts category y.

[0103] During the implementation process, multiple decision trees are first trained using algorithms such as random forests. Each decision tree will independently train the input sample X u Classify and output a classification result The final classification result is not determined by a single decision tree, but by the output of all decision trees. The voting mechanism adopts a "one-to-many" strategy, that is, each tree votes for each category; the total number of predictions for a certain category in all decision trees is calculated, and the category with the largest number of predictions is selected as the final classification result through the argmax function. This embodiment improves the accuracy of classification by integrating the prediction results of multiple decision trees and using a voting mechanism to determine the final classification output.

[0104] Based on the above scheme, the dynamic constraint mechanism and adaptive adjustment of tree depth and split conditions enable the random forest model to make more accurate classification decisions when faced with complex and difficult-to-distinguish samples. The integration of multiple decision trees enhances the robustness of the model and effectively avoids overfitting or underfitting problems.

[0105] Accurate classification of flexibility resources can help power systems better forecast, dispatch and manage loads, especially in the context of more renewable energy access. The system can balance supply and demand fluctuations and optimize the stability and operating efficiency of the power grid by rationally dispatching flexibility resources.

[0106] Those skilled in the art should understand that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A method for constructing a power system flexibility resource classification model based on machine learning, characterized in that: The following steps are involved: Collect multi-source power data and manually annotate it; Feature extraction and data preprocessing based on multi-source power data; Build a classification model based on the random forest algorithm and perform model training; A dynamic constraint mechanism is used to optimize the generation process of decision trees in the classification model, and the classification model is obtained by integrating the outputs of multiple decision trees.

2. The method for constructing a power system flexibility resource classification model based on machine learning according to claim 1, characterized in that: The collection of multi-source power data is based on power consumption records of power grid companies, load collection systems, historical power consumption data, weather and environmental data, and power equipment operation data; The collected multi-source power data are manually labeled as high-flexibility resources, medium-flexibility resources, non-flexible resources, and non-adjustable loads.

3. The method for constructing a power system flexibility resource classification model based on machine learning according to claim 1, characterized in that: Feature extraction based on multi-source data includes load value, timestamp, electricity consumption type and load volatility, and the extracted feature samples and their corresponding sample labels are used as training data for the classification model.

4. The method for constructing a power system flexibility resource classification model based on machine learning according to claim 1 is characterized in that: The training process of the classification model includes the following steps: Initialize the classification model and set the parameters of the random forest algorithm; Dynamically adjust the generation constraints of decision trees in the random forest algorithm; In the process of decision tree generation, node splitting features and splitting points are adaptively adjusted; During the feature selection process at each node, dynamic evaluation of feature importance is performed; The output results of all decision trees are integrated, and a voting mechanism is used to determine the classification result.

5. The method for constructing a power system flexibility resource classification model based on machine learning according to claim 4 is characterized in that: The decision tree is trained based on the feature samples and their corresponding sample labels. After initializing the classification model, the number of decision trees N required for random forest algorithm training is set. u , specifically expressed as: Among them, Θ u Represents the set of all decision trees in the random forest algorithm, T u represents the u-th decision tree; In the initial stage, no constraints are imposed on any decision tree, and the depth of any decision tree is dynamically adjustable.

6. The method for constructing a power system flexibility resource classification model based on machine learning according to claim 4 is characterized in that: According to the distribution characteristics of the training data, the generation constraints of the decision tree are dynamically adjusted. Specifically, the depth or splitting conditions of the decision tree are adjusted by defining the constraint factor δu, which is specifically expressed as: Among them, f() is the activation function, which indicates the calculation method of the constraint factor; m c Represents the number of characteristic samples of training data; Ⅱ() is the indicator function, when Indicates that the sample labels of the i-th feature sample and the i'th feature sample are different.

7. The method for constructing a power system flexibility resource classification model based on machine learning according to claim 4 is characterized in that: The node splitting features and corresponding splitting points are determined based on the information gain characteristics of the training data, and the splitting metric is defined as information gain, which is specifically expressed as: Among them, IG() is the information gain function; H(Y c ) represents the Shannon entropy of the target variable, H(Y c |X ck = r) indicates that in feature X ck When r is the condition, Y c Conditional entropy; m r Indicated in feature X ck = the number of samples under r; m c Represents the total number of training samples.

8. The method for constructing a power system flexibility resource classification model based on machine learning according to claim 7 is characterized in that: Selecting the best features based on information gain metric And calculate the maximum value of information gain for feature division; Based on the dynamic constraint mechanism, the decision tree structure is optimized by increasing the node depth or refining the splitting conditions. The specific adjustment method is expressed as follows: Among them, Δ Tu Represented as a decision tree T u The adjustment amount of the midpoint split; f() is the activation function used to smoothly adjust the split condition; η u Indicates the preset threshold.

9. The method for constructing a power system flexibility resource classification model based on machine learning according to claim 4 is characterized in that: The weighted average strategy is used to dynamically evaluate the importance of features, which is specifically expressed as follows: Among them, W ci represents the weight of the i-th feature sample; Ⅱ(X ck =x ci ) is the indicator function, indicating that in the i-th sample, feature X ck Is it equal to its value x? ci .

10. The method for constructing a power system flexibility resource classification model based on machine learning according to claim 4, characterized in that: Any decision tree for input sample X u Classify and output a classification result The classification output of the classification model is determined by the voting results of all decision trees, which is specifically expressed as: in, is the final prediction output of the classification model; Ⅱ() is the indicator function, indicating whether the decision tree predicts category y.

Images