Demand side resource hierarchical classification method based on interactive behavior driving factors

Abstract

The invention provides a demand side resource hierarchical classification method based on interactive behavior driving factors. The method comprises the following steps: acquiring a current clustering label set of a current demand side resource set by adopting a clustering total model constructed based on a convolutional neural network, an auto-encoder and a Gaussian mixture model; obtaining a current behavior feature set and a current interaction influence index set of the current demand side resource set by adopting a motivation analysis model constructed based on a graph neural network; adopting a hierarchical classification model to obtain a hierarchical classification result of the current demand side resource set according to the current clustering label, the current behavior feature set and the current interaction influence index set; the hierarchical classification model is constructed based on a historical demand side resource set, a clustering general model, a motivation analysis model, an adaptive depth cascade forest and multi-task reinforcement learning. According to the method, the data processing efficiency and precision are high, the prediction accuracy and decision support are deeply improved, the adaptability in a complex environment is high, and the intelligence and precision of resource scheduling are improved.

Description

Technical Field

[0001] This application relates to the field of power system optimization technology, specifically to a demand-side resource hierarchical classification method based on interactive behavior driving factors. Background Technology

[0002] In the context of smart grids and energy management, efficient management of demand-side resources has become crucial for driving energy transformation and optimizing power system operation. With the rapid development of big data and artificial intelligence, demand-side resource management (DSR) has demonstrated enormous potential in improving energy efficiency, optimizing power system dispatch, and reducing environmental impact. However, traditional DSR classification methods often rely on simple statistical models or manual rules, making it difficult to cope with complex user behaviors, resource characteristics, and dynamic changes in the market environment. Summary of the Invention

[0003] This application addresses the problems existing in the prior art by providing a demand-side resource hierarchical classification method based on interactive behavior-driven factors. This method can significantly improve the efficiency and accuracy of data processing, enhance prediction accuracy and decision support depth, and strengthen adaptability in complex environments, thereby effectively improving the intelligence and accuracy of resource scheduling.

[0004] To achieve the above objectives, the technical solution adopted in this application is as follows: A demand-side resource hierarchical classification method based on interactive behavior-driven factors, comprising: Obtain the current demand-side resource set; The current clustering label set of the current demand-side resource set is obtained using the overall clustering model; The current behavioral feature set and current interaction influence index set of the current demand-side resource set are obtained by using a motivation analysis model; A hierarchical classification model is adopted to obtain the hierarchical classification result of the current demand-side resource set based on the current cluster label set, the current behavior feature set, and the current interaction influence index set; The overall clustering model is obtained based on convolutional neural networks, autoencoders, Gaussian mixture models, and historical demand-side resource sets. The driving force analysis model is constructed based on graph neural networks and historical demand-side resource sets; The hierarchical classification model is constructed based on the historical demand-side resource set, the overall clustering model, the motivation analysis model, adaptive deep cascaded forest, and multi-task reinforcement learning.

[0005] In some embodiments, the demand-side resource set includes a resource load dataset, which is a collection of resource load data for each user; the historical demand-side resource set is a demand-side resource set for a historical period, which includes a historical resource load dataset; the historical resource load dataset is a resource load dataset for a historical period; the overall clustering model is obtained based on a convolutional neural network, an autoencoder, a Gaussian mixture model, and the historical demand-side resource set, including: An initial clustering model is constructed based on the convolutional neural network, the autoencoder, and the Gaussian mixture model. The initial clustering model is trained based on the historical resource load dataset to obtain intermediate clustering models and intermediate clustering results; Determine whether the intermediate clustering model has converged based on the intermediate clustering results; When the intermediate clustering model does not converge, the intermediate clustering model is set to the initial clustering model, and the process of training the initial clustering model based on the historical resource load dataset is repeated to obtain the intermediate clustering model and intermediate clustering results. When the intermediate clustering model converges, a historical clustering label set is obtained based on the intermediate clustering results, and the overall clustering model is obtained based on the intermediate clustering model.

[0006] In some embodiments, the initial clustering model includes a first initial model, a second initial model, and a third initial model, wherein the first initial model is constructed based on the convolutional neural network, the second initial model is constructed based on the autoencoder, and the third initial model is constructed based on the Gaussian mixture model; The initial clustering model is trained based on the historical resource load dataset to obtain intermediate clustering models and intermediate clustering results, including: The first initial model is trained using the historical resource load dataset to obtain the first intermediate model and the first intermediate training dataset; The second initial model is trained using the first intermediate training set to obtain the second intermediate model and the second intermediate training set; The third initial model is trained using the second intermediate training set to obtain the third intermediate model and intermediate clustering results; The intermediate clustering model includes the first intermediate model, the second intermediate model, and the third intermediate model.

[0007] In some embodiments, when the intermediate clustering model does not converge, the first intermediate model is set as the first initial model, the second intermediate model is set as the second initial model, the third intermediate model is set as the third initial model, and the process of repeatedly training the initial clustering model based on the historical resource load dataset to obtain intermediate clustering models and intermediate clustering results is repeated.

[0008] In some embodiments, when the third intermediate model converges, a historical clustering label set is obtained based on the intermediate clustering results, a feature extraction model is obtained based on the first intermediate model, a data compression model is obtained based on the second intermediate model, and a hybrid clustering model is obtained based on the third intermediate model. The overall clustering model includes the feature extraction model, the data compression model, and the hybrid clustering model.

[0009] In some embodiments, the current demand-side resource set includes a current resource load dataset, which is the resource load dataset for the current stage; obtaining the current cluster label set of the current demand-side resource set using a clustering overall model includes: The feature extraction model is used to obtain an output feature map based on the current resource load dataset; The data compression model described above is used to obtain a latent space representation set based on the output feature map; The hybrid clustering model is used to obtain the current clustering label set based on the latent space representation set.

[0010] In some embodiments, the feature extraction model includes convolutional layers and pooling layers; The feature extraction model described above is used to obtain an output feature map based on the current resource load dataset, including: The convolutional layer is used to obtain intermediate feature maps based on the current resource load dataset; The pooling layer is used to obtain the output feature map based on the intermediate feature map.

[0011] In some embodiments, the demand-side resource set further includes a behavior dataset, which is a collection of behavior data for each user; the historical demand-side resource set further includes a historical behavior dataset, which is a dataset of behavior data from historical periods; the hierarchical classification model is constructed based on the historical demand-side resource set, the overall clustering model, the motivation analysis model, the adaptive deep cascaded forest, and multi-task reinforcement learning, including: The historical clustering label set of the historical demand-side resource set is obtained based on the overall clustering model. Based on the aforementioned motivational analysis model, the historical behavioral characteristic set and historical interaction influence index set of the historical demand-side resource set are obtained; An initial hierarchical classification model is constructed based on the adaptive deep cascaded forest and the multi-task reinforcement learning; The hierarchical classification model is obtained by training the initial hierarchical classification model based on the historical clustering label set, the historical behavior feature set, and the historical interaction influence index set.

[0012] In some embodiments, the behavioral data includes individual characteristic data and social network relationship data.

[0013] In some embodiments, the behavioral data also includes market environment change data.

[0014] Compared with the prior art, this application has the following advantages: 1. This application proposes a deep clustering algorithm combining convolutional neural networks, autoencoders, and Gaussian mixture models for processing large-scale and complex unstructured data, particularly in identifying resource behaviors, drivers, and their interrelationships. Through this deep clustering algorithm, the application can discover latent patterns in high-dimensional data, overcoming the limitations of traditional clustering methods such as K-means in processing unstructured data. This method significantly improves the efficiency and accuracy of data processing by automatically identifying hidden structures and patterns in the data, thus providing more accurate and high-quality input data for subsequent analysis and decision-making.

[0015] 2. This application addresses the shortcomings of existing technologies in deeply analyzing the multi-layered driving factors of user interaction behavior by introducing graph neural networks to model the interactive relationships and social influences between users. Through the application of graph neural networks, the interrelationships between user behaviors can be accurately captured, and multi-dimensional external factors influencing electricity consumption behavior can be identified. Compared to traditional behavioral analysis methods, this approach provides a deeper understanding of the complex motivations behind user behavior, significantly improving the accuracy of electricity consumption behavior prediction and the depth of decision support.

[0016] 3. Addressing the static nature and lack of real-time adaptability of traditional classification methods, this invention proposes a method combining adaptive deep cascaded forest (daForest) and multi-task reinforcement learning (MT-RL) to achieve adaptive hierarchical classification of demand-side resources. The daForest algorithm, combined with adaptive reinforcement algorithms and hyperparameter optimization layers, improves the model's performance and stability under data-scarce conditions. Meanwhile, MT-RL, by sharing representation and policy networks across multiple related tasks, enables the classification engine to dynamically adjust its classification strategy based on real-time feedback, improving classification accuracy, stability, and flexibility, thus effectively adapting to the rapid changes in the electricity market and user demands. The introduction of this method significantly enhances the system's adaptability in complex environments and improves the intelligence and accuracy of resource scheduling. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the demand-side resource hierarchical classification method based on interactive behavior-driven factors in this application. Figure 2 The daily typical load curve is shown for the demand-side resource clustering label of the schedulable cluster in this application example. Figure 3 This is a typical daily load curve for the demand-side resource clustering label of the stable cluster in this application example; Figure 4 This is a typical daily load curve for the demand-side resource cluster labeled as a high-fluctuation cluster in the example of this application. Figure 5 This is a typical daily load curve for the demand-side resource clustering of the example in this application, labeled as an intermittent cluster. Figure 6 This is a graph showing the influence index of each user in the demand-side resource pool for this application example. Figure 7 This is a diagram showing the hierarchical classification results of the demand-side resource set in this application example. Detailed Implementation

[0018] Addressing the problems in the background technology, the maturity of technologies such as deep learning and graph neural networks (GNNs) has made it possible to leverage these advanced techniques for dynamic analysis and optimization of demand-side resources, becoming an effective way to overcome the bottlenecks of traditional methods. The behavior of demand-side resources is not only influenced by individual characteristics but also driven by complex factors such as social relationships between users and market changes. Especially in processing high-dimensional and unstructured data, traditional clustering methods struggle to deeply uncover these multi-layered and multi-dimensional potential driving factors. Therefore, how to accurately capture these complex driving factors and combine them with modern algorithms to optimize classification accuracy has become a current research hotspot.

[0019] Research Hotspot 1: In theoretical research, scholar Liu Xing, in his article "Classification of Demand-Side Adjustable Resources and Exploration of Load Aggregator Function," proposed a classification method for demand-side resources to improve the power system's regulation capacity and energy efficiency. The article analyzes in detail the characteristics of demand-side resources, particularly their discreteness and wide distribution. Based on the different characteristics of various industries, the importance of loads, and response time scales, Liu Xing proposed a multi-level classification method. First, based on the industry's adjustability potential, resources are divided into different levels, including industrial, commercial, and other user sectors. For the industrial sector, based on the type of electrical equipment and the potential impact of power outages, loads are divided into four importance levels: safety loads, primary production loads, auxiliary production loads, and non-production loads.

[0020] Furthermore, Liu Xing categorized demand-side resources based on time scales. According to the time characteristics of load response, loads were divided into Level I loads with long response times, Level II loads with response times in the hour, Level III loads with response times in the minute, and Level IV loads with response times in the second. This classification method can stratify resources according to response time and importance based on the characteristics of different industries and loads, thus providing a more accurate and effective scheduling basis for demand response implementation. This classification method provides a theoretical foundation for load aggregators, helping to better integrate and schedule demand-side resources and improve the operational stability of the power system during high-load periods.

[0021] However, the classification method proposed in the first research hotspot has two main drawbacks in demand-side resource classification. Firstly, this method typically relies on static criteria, such as industry, load importance level, and response time, neglecting the diversity and dynamic changes of demand-side resources. The differences in actual load characteristics and regulation potential among different users and equipment are ignored, resulting in classification results that fail to accurately reflect the actual regulation capacity of resources, affecting the accuracy and efficiency of demand response. Secondly, this classification method is too simplistic in its division of response time and load importance level, failing to fully consider the specific characteristics and response requirements of various load types. For example, loads of similar levels in different industries may have different regulation capabilities and timeliness; traditional methods fail to effectively distinguish these subtle differences, thus limiting the flexibility of demand response and the rapid response capability of the power system.

[0022] Research Hotspot Two: In theoretical research, scholar Zhang Tianmi, in his article "Research on Demand-Side Management Mechanism and Development Strategy under the New Power System," conducted an in-depth discussion on the classification of demand-side resources. Zhang Tianmi proposed that with the increasing proportion of new energy power generation, the load demand of the power system exhibits more complex diversity and dynamism. Therefore, traditional single classifications based on industry or load type can no longer meet the needs of modern power systems. To better adapt to this change, Zhang Tianmi proposed classifying demand-side resources into different response times and regulation potential levels, including load classifications for industrial, commercial, and residential users. Furthermore, he proposed a more refined hierarchical division based on load adjustability and response speed. This classification method can more accurately identify the regulation potential of various resources, thereby providing the power system with more efficient and flexible resource dispatching schemes.

[0023] Compared to traditional static classification methods, Zhang Tianmi's research introduces multi-dimensional and dynamic factors to redefine demand-side resources, especially given the increasingly important role of these resources in the context of high-proportion renewable energy integration. By combining user electricity consumption characteristics, load response capabilities, and market mechanisms, Zhang Tianmi proposes a more flexible classification framework capable of responding to real-time changes in the electricity market. This framework considers not only the importance and response time of loads but also the enthusiasm and incentive mechanisms for user participation in demand response, thereby making resource scheduling more precise, response more timely, and improving the integration efficiency of demand-side resources in the power system. This classification method provides theoretical support for subsequent design and marketization paths of electricity demand response mechanisms and helps improve the system's economy and sustainability.

[0024] In the second research hotspot, namely the demand-side resource classification method proposed by Zhang Tianmi, on the one hand, this classification method enhances the accuracy of demand-side resource classification by refining the response time and regulation potential of resources, thus better adapting to the regulation needs of the power system. However, this classification method still relies too heavily on static electricity consumption characteristics and fails to fully consider the dynamic changes of demand-side resources and the complexity of user behavior. In practical applications, external factors such as market fluctuations and seasonal changes may cause changes in load characteristics, and the existing classification framework is difficult to adapt to these changes in real time, affecting its flexibility and accuracy. On the other hand, although the classification framework of this method is innovative in terms of refined resource scheduling, how to accurately identify and effectively schedule different categories of resources in actual operation still faces challenges. The diversity and dispersion of demand-side resources make it difficult to accurately classify and schedule certain user groups with personalized needs (such as specific industry or commercial users), affecting the implementation effect of the classification method. In addition, when considering user participation in demand response, the classification method does not fully consider economic costs and operational complexity, which may lead to some users having a low willingness to respond, thereby limiting the system's regulation capacity.

[0025] In summary, in the context of smart grids and energy management, there is an urgent need for a more intelligent, precise, and adaptable approach to address the challenges of stratified and categorized demand-side resources in order to efficiently manage demand-side resources, promote energy transition, and achieve optimized power system operation.

[0026] To clearly illustrate the technical features of this solution, the implementation methods of this application will be described in detail below with reference to the accompanying drawings and embodiments. This will allow for a full understanding and implementation of how this application uses technical means to solve technical problems and achieve corresponding technical effects. The embodiments of this application and the various features within them can be combined with each other without conflict, and the resulting technical solutions are all within the protection scope of this application.

[0027] See Figure 1This application proposes a demand-side resource hierarchical classification method based on interactive behavior-driven factors, including: Get the current demand-side resource set: The demand-side resource set includes a resource load dataset and a behavior dataset. The resource load dataset is a collection of resource load data for each user, and the behavior dataset is a collection of behavior data for each user. The historical demand-side resource set is the demand-side resource set for a historical period. The historical demand-side resource set includes the historical resource load dataset and the historical behavior dataset. The historical resource load dataset is the resource load dataset for a historical period, and the historical behavior dataset is the behavior dataset for a historical period. The current demand-side resource set is the demand-side resource set for the current stage. The current demand-side resource set includes the current resource load dataset and the current behavior dataset. The current resource load dataset is the resource load dataset for the current stage, and the current behavior dataset is the behavior dataset for the current stage. It is worth noting that the demand-side resource hierarchical classification method based on interactive behavior driving factors proposed in this application can be used for hierarchical classification of demand-side resources at a single point in time, which is suitable for scenarios that require rapid response, such as hierarchical classification of real-time demand-side resources, and can also be used for hierarchical classification of demand-side resource sets in stages, which is suitable for non-real-time or low real-time scenarios, such as hierarchical classification of demand-side resource sets in the past 24 hours. In addition, the same demand-side resource set samples can be divided into training set, validation set and test set. The training set and validation set are used as historical demand-side resource sets for model training, and the test set is used as the current demand-side resource set for application, thereby achieving hierarchical classification of the demand-side resource set samples.

[0028] The clustering model is used to obtain the current clustering label set of the current demand-side resource set based on the current resource load dataset; The motivation analysis model is used to obtain the current behavior feature set and the current interaction influence index set of the current demand-side resource set based on the current behavior dataset; A hierarchical classification model is adopted to obtain the hierarchical classification results of the current demand-side resource set based on the current cluster label set, the current behavioral feature set, and the current interaction influence index set. The overall clustering model was obtained based on a convolutional neural network, an autoencoder, a Gaussian mixture model, and a historical resource load dataset. The motivation analysis model is built based on graph neural networks and historical behavior datasets; The hierarchical classification model is constructed based on historical demand-side resource sets, clustering overall models, motivation analysis models, adaptive deep cascaded forests, and multi-task reinforcement learning.

[0029] This application proposes a deep clustering algorithm combining convolutional neural networks, autoencoders, and Gaussian mixture models for processing large-scale and complex unstructured data, particularly in identifying resource behaviors, drivers, and their interrelationships. Through this deep clustering algorithm, the application can discover latent patterns in high-dimensional data, overcoming the limitations of traditional clustering methods such as K-means in processing unstructured data. This method significantly improves the efficiency and accuracy of data processing by automatically identifying hidden structures and patterns in the data, thus providing more accurate and high-quality input data for subsequent analysis and decision-making.

[0030] This application addresses the shortcomings of existing technologies in deeply analyzing the multi-layered driving factors of user interaction behavior by introducing graph neural networks to model the interactive relationships and social influences between users. Through the application of graph neural networks, the interrelationships between user behaviors can be accurately captured, and multi-dimensional external factors influencing electricity consumption behavior can be identified. Compared to traditional behavioral analysis methods, this approach provides a deeper understanding of the complex motivations behind user behavior, significantly improving the accuracy of electricity consumption behavior prediction and the depth of decision support.

[0031] Addressing the static nature and lack of real-time adaptability of traditional classification methods, this invention proposes a method combining adaptive deep cascaded forest (daForest) and multi-task reinforcement learning (MT-RL) to achieve adaptive hierarchical classification of demand-side resources. The daForest algorithm, combined with adaptive reinforcement algorithms and hyperparameter optimization layers, improves the model's performance and stability under data-scarce conditions. Meanwhile, MT-RL, by sharing representation and policy networks across multiple related tasks, enables the classification engine to dynamically adjust its classification strategy based on real-time feedback, improving classification accuracy, stability, and flexibility, thus effectively adapting to rapid changes in the electricity market and user demands. The introduction of this method significantly enhances the system's adaptability in complex environments and improves the intelligence and accuracy of resource scheduling.

[0032] The overall clustering model combines convolutional neural networks (CNNs), autoencoders, and Gaussian mixture models (GMMs) to complete the clustering task through multi-level data processing and feature extraction. Specifically, firstly, meaningful spatial or temporal features are extracted from the data using CNNs; then, the feature representations are reduced in dimensionality and compressed using autoencoders; finally, clustering analysis is performed using a Gaussian mixture model; and finally, the loss functions of these steps are combined for joint optimization.

[0033] In some embodiments, a clustering overall model is obtained based on a convolutional neural network, an autoencoder, a Gaussian mixture model, and a historical resource load dataset, including: Step S11: Construct an initial clustering model. The initial clustering model is constructed based on a convolutional neural network, an autoencoder, and a Gaussian mixture model. In some embodiments, the initial clustering model includes a first initial model, a second initial model, and a third initial model. The first initial model is constructed based on a convolutional neural network, the second initial model is constructed based on an autoencoder, and the third initial model is constructed based on a Gaussian mixture model. Step S12: Train an initial clustering model based on the historical resource load dataset to obtain an intermediate clustering model and intermediate clustering results, including: The first initial model and the first intermediate model and training set were obtained by training the historical resource load dataset. The second initial model is obtained by training the first intermediate training set; The third initial model is trained using the second intermediate training set to obtain the third intermediate model and intermediate clustering results; The intermediate clustering model includes the first intermediate model, the second intermediate model, and the third intermediate model.

[0034] Step S13: Determine whether the intermediate clustering model has converged based on the intermediate clustering results; When the intermediate clustering model fails to converge, the intermediate clustering model is set as the initial clustering model, and the training of the initial clustering model is repeated to obtain the intermediate clustering model and intermediate clustering results. In some embodiments, when the intermediate clustering model fails to converge, the first intermediate model is set as the first initial model, the second intermediate model is set as the second initial model, the third intermediate model is set as the third initial model, and the training of the initial clustering model based on the historical resource load dataset is repeated to obtain the intermediate clustering model and intermediate clustering results.

[0035] When the intermediate clustering model converges, a historical clustering label set is obtained based on the intermediate clustering results, and a total clustering model is obtained based on the intermediate clustering model. In some embodiments, when the third intermediate model converges, a historical clustering label set is obtained based on the intermediate clustering results, a feature extraction model is obtained based on the first intermediate model, a data compression model is obtained based on the second intermediate model, and a hybrid clustering model is obtained based on the third intermediate model. The total clustering model includes the feature extraction model, the data compression model, and the hybrid clustering model.

[0036] In some embodiments, a clustering overall model is used to obtain the current clustering label set of the current demand-side resource set based on the current resource load data, including: Step S21: Obtain an output feature map based on the current resource load dataset using a feature extraction model; in some embodiments, the feature extraction model includes convolutional layers and pooling layers; obtaining the output feature map based on the current resource load dataset using the feature extraction model includes: obtaining an intermediate feature map based on the current resource load dataset using a convolutional layer; the calculation formula for the convolutional layer is: ; in, The output feature map of the convolutional layer is generated when the input image is based on the current resource load dataset, i.e., during application. This is an intermediate feature map. The activation function of the convolutional layer. The weight matrix of the convolution kernel (filter) , The spatial size of the convolution kernel. For the number of channels, The number of convolution kernels, The input image is generated based on the historical resource load dataset during training, and generated based on the current resource load dataset during application. , For the height of the input image, The width of the input image, The bias term for the convolution kernel is an additional parameter for the convolution operation, used to increase the flexibility of the feature extraction model. , , All were generated through training.

[0037] Pooling layers are used to obtain output feature maps from intermediate feature maps, which are then used to reduce... The pooling layer reduces the spatial size of the feature map while retaining key information; the calculation formula for the pooling layer is: ; in, The output feature map of the pooling layer is generated when the input image is based on the current resource load dataset, i.e., during application. To output the feature map, For pooling operations, the pooling operation can be max pooling or average pooling. Max pooling takes the maximum value in the local region, while average pooling calculates the average value in the local region. This is the window size for pooling operations. , All were generated through training.

[0038] Step S22: Obtain the latent space representation set based on the output feature map using a data compression model. The data compression model is used to reduce the dimensionality of the output feature map. The model includes an encoder and a decoder. The encoder compresses the high-dimensional output feature map into the latent space using its weight matrix to generate the latent space representation set. The decoder reconstructs the latent space representation set back to the original space. The encoder's calculation formula is: ; in, The output of the encoder is the latent space representation set when applied. , Let the first data in the latent space representation set be... The second data point in the latent space representation set. For the first in the latent space representation set One data point, The potential space represents the sequence number of the data in the set. It also represents the user sequence number in the resource load dataset. Indicates the first Potential spatial representation of resource load data for each user For the first in the latent space representation set One data point, The potential space represents the total number of centralized data, and the total number of centralized data represented by the potential space is the same as the total number of users represented by the resource load dataset. Here, "Encoder Function" represents the process by which the encoder maps features from the input output feature map to the latent space. Here is the parameter set for the encoder. include , and , Generated through training, This is the activation function of the encoder. Here is the weight matrix of the encoder. This is the encoder bias term, used to adjust the encoder output.

[0039] The decoder calculation formula is: ; in, The output of the decoder is the reconstructed data. For decoder functions, it represents the process by which the decoder reconstructs the latent space representation set back to the original data space. Here is the parameter set for the decoder. include , and , Generated through training, For the activation function of the decoder, Here is the weight matrix of the decoder. This is the bias term for the decoder, used to adjust the decoder's output.

[0040] In training the second initial model into a data compression model, a key objective is to minimize the difference between the input image and the reconstructed data. This difference is typically measured by a reconstruction loss function, which measures the error between the input image and the reconstructed data obtained through encoding and decoding processes. The reconstruction loss function is: ; in, To reconstruct the loss function.

[0041] Step S23: Use a hybrid clustering model to obtain a clustering label set based on the latent space representation set.

[0042] Gaussian Mixture Models (GMMs) are generated by combining multiple Gaussian distributions, with each data point having a probability of belonging to each Gaussian component. In this way, GMMs can divide a dataset into clusters of Gaussian components generated by multiple Gaussian distributions; that is, when applied, they generate a set of cluster labels and calculate the probability of each data point belonging to a different Gaussian component within that cluster. ;in, for The probability density, This represents the index of the Gaussian component in the Gaussian component cluster, and in application, it indicates the index of the cluster label in the cluster label set. This represents the total number of Gaussian components in the Gaussian component cluster. In practical applications, it represents the total number of cluster labels within a cluster label set. The first in the Gaussian component cluster The mixing coefficient of the _th Gaussian component, representing the _th The weights of each Gaussian component in the Gaussian component cluster. Let be the probability density function of a Gaussian distribution. express In the Probability density under Gaussian components For the first The arithmetic mean of the probability densities of the Gaussian components. For the first The covariance matrix of the nth Gaussian component represents the... The degree of dispersion of each Gaussian component, i.e. and The distribution range between them , ,and All are generated during training.

[0043] During training, the parameters of the third initial model are updated through the responsibility level. , ,and This yields a hybrid clustering model. The formula for calculating responsibility is: ; in, for Belongs to the The responsibility degree of each Gaussian component indicates In the The probability of belonging to each Gaussian component Let be the index of each possible Gaussian distribution in the Gaussian component cluster. The range of values ​​and k They are exactly the same, both from 1 to K .

[0044] ; ; ; Among them, superscript T Indicates transpose. for, and The difference vector between them.

[0045] The alternating iterations of responsibility calculation (E-step) and parameter update (M-step) in Gaussian Mixture Models (GMMs) play a crucial role. The E-step calculates the responsibility of each data point belonging to each Gaussian component, providing the GMM with probabilistic information about data assignment. These responsibility values ​​reflect the degree to which data points belong to different Gaussian components, forming the basis for effective clustering in GMMs. The M-step, on the other hand, uses the responsibility values ​​calculated in the E-step to update the model's parameters (including the mean, covariance matrix, and mixing coefficients), continuously optimizing the clustering results and ensuring that the GMM better fits the data. This alternating optimization process ultimately helps the model achieve optimal data grouping, thus enabling accurate clustering. Through repeated iterations of the E-step and M-step, Gaussian Mixture Models can effectively perform clustering analysis on complex datasets and handle fuzziness and uncertainty in the data.

[0046] A motivational analysis model is employed to obtain a set of current interaction influence indices based on the current user behavior dataset. The goal of this model is to deeply analyze the motivations influencing user interaction behavior and explore how these motivations affect final electricity consumption behavior through factors such as social network relationships. To achieve this goal, a Graph Neural Network (GNN) is used to model the user behavior data. Motivational analysis is the core of the entire model, encompassing, but not limited to, various factors such as individual user characteristics, social network relationships, and changes in the market environment. In other words, the collected user behavior dataset includes, but is not limited to, individual user characteristics, social network relationships, and changes in the market environment. These factors are modeled and analyzed in the GNN through the characteristics of nodes and edges. Identifying the motivations is one of the key tasks of the motivational analysis model.

[0047] Individual characteristics include factors such as age, gender, consumption habits, and electricity demand, which directly influence user behavior patterns. Individual characteristic data is a data-driven representation of these characteristics. Social network relationships include social connections between users, such as recommendations from friends and family, and information dissemination, which affect user decisions and behavior. Social network relationship data is a data-driven representation of social network relationships. Market environment changes include external factors such as electricity price fluctuations and policy changes, which significantly impact user electricity consumption behavior. Market environment change data is a data-driven representation of changes in the market environment.

[0048] Using a behavioral dataset that includes users' individual characteristics and social network relationship data as an example, construct a motivational input graph.

[0049] In the causal analysis model, the causal input diagram Used to represent the interaction relationship between users. For user sets, , For the first Each user is a node in the input graph, representing a driving force. For social network relationship datasets, , For the first The user and the first The social connections between users are represented by the edges in the input graph, which are the driving forces. For users, a user serial number is used. Each user has a feature vector. , Indicates the first Each user's feature vector contains individual characteristics such as age, gender, and electricity usage history.

[0050] Adjacency matrix of the driver input graph It represents the social relationships between users. Indicates the total number of users. ,in, For the first The user and the first Social interaction relationships between users Indicates the first The user and the first There is social interaction between users, and vice versa. Indicates the first The user and the first There is no social interaction between users.

[0051] In addition, external factors, such as changes in the market environment, can be considered. These factors can be modeled by introducing additional features to each edge or node of the driver input graph.

[0052] During training, in a graph neural network, the update of the node representation in each layer is obtained by propagating information from neighboring nodes. The update formula is: ; in, For the first The node feature matrix of the layer, for This refers to the layer number in the graph neural network. This is the first nonlinear activation function, its purpose is to introduce nonlinearity, enabling the model to learn more complex patterns. The normalized adjacency matrix of the input graph is used to balance the influence of each node. For the first The node feature matrix of the layer is used for each user in the first layer. Layer feature representation, For the first The layer's weight matrix is ​​used for feature transformation. For the first The characteristics of each edge in a layer represent the strength of social relationships between users or the influence of external factors; The above update formula means that the new representation of each node, i.e., the user, is a weighted sum of the user's information and the information of the neighboring users, which combines the social network relationship between the user and the neighboring users, as well as the user node's own feature information and the neighboring user node's own feature information, or combines the user's social network relationship with the neighboring users, external factors, and the user node's own feature information and the neighboring user node's own feature information.

[0053] In multi-layer graph convolutional networks, each layer updates the representation of its nodes and passes these representations to the next layer, ultimately learning deeper levels of user behavior patterns. The update formula for multi-layer propagation is: ; in, For the first multi-layer graph convolutional network The node representation matrix of the layer, These are the layer numbers in a multi-layer graph convolutional network. Through multi-layer propagation, the node representations are gradually updated, ultimately learning representations that reflect deep-seated patterns of user behavior. For the second nonlinear activation function, The feature vector matrix of the input graph represents the set of initial features for each node. This is the weight matrix for layer 1, used to map the node features to the representation space of the next layer. The weight matrix for each layer is learned and optimized through training. The feature matrix of each edge in the graph is used as the driving force, representing the strength of the relationship between each pair of users in the social network, or the influence of external factors (such as market environment or policy changes). In this way, the features of the edges can influence the final representation of the nodes. For the first The layer weight matrix is ​​used to map edge features (such as external factors) to the next layer.

[0054] The ultimate goal of the causal analysis model is to uncover the key factors influencing user electricity consumption behavior using a graph neural network (GNN). After model training, the final representation of each node integrates the node's original features with the structural information of its neighboring nodes. Therefore, the causal analysis model can identify which factors have the greatest impact on user behavior based on these representations. To identify which factors have a significant impact on user behavior, it is first necessary to define an influence metric that assesses the importance of each node in the model's predictions. The contribution of each node to the final decision can be calculated using the node representations. Taking a multi-layer graph convolutional network as an example, the prediction formula for the causal analysis model is as follows: ; in, for The predicted output, namely the behavioral feature set, represents... The corresponding factors' impact on users' electricity consumption behavior is classified into probabilities. This means transforming the output into a probability distribution, representing the predicted probability of each category. for In the The feature vector of a layer, after passing through multiple graph convolutions, contains all the information about that node and its neighbors; Based on the predictions, gradient calculation is used to quantify the influence of each node's feature on the final prediction. The gradient calculation formula is as follows: ; in, for The gradient value, or influence index value, represents... Influence metrics, used to measure Contribution to the final decision; a larger gradient indicates that the user has a greater influence on the prediction results, therefore the factors corresponding to this user may be key factors influencing user behavior.

[0055] This method enables Graph Neural Networks (GNNs) to effectively identify key factors influencing user electricity consumption behavior. By calculating the gradient of user characteristics with respect to the prediction results, the contribution of each user can be quantified, and the data can be sorted according to the magnitude of the gradient to ultimately extract factors that significantly impact user behavior. This process ensures that the causal analysis model can identify key factors from complex graph structures, providing an important basis for user behavior analysis and optimization.

[0056] In demand-side resource management, the diversity and dynamism of resources require a hierarchical classification model to efficiently manage the hierarchical and categorized management of demand-side resource sets. To achieve this goal, this application proposes a hierarchical classification model combining adaptive deep cascaded forest (daForest) and multi-task reinforcement learning (MT-RL). By dynamically adjusting the structure and strategy of the hierarchical classification model, it achieves fine-grained classification and hierarchical management of demand-side resources. In some embodiments, the hierarchical classification model is constructed based on historical demand-side resource sets, a clustering overall model, a motivation analysis model, adaptive deep cascaded forest, and multi-task reinforcement learning, including: Based on the overall clustering model, the historical clustering label set of historical demand-side resource sets is obtained from the historical resource load dataset; Based on the motivation analysis model, the historical behavioral feature set and historical interaction influence index set of historical demand-side resource sets are obtained from the historical behavioral dataset; An initial hierarchical classification model was constructed based on adaptive deep cascaded forest and multi-task reinforcement learning; The hierarchical classification model is obtained by training an initial hierarchical classification model based on the historical clustering label set, historical behavior feature set, and historical interaction influence index set.

[0057] Within the framework of multi-task learning, we first define a set of... Task set of tasks Each task has a different objective. Each task uses a shared strategy and value function for optimization, and knowledge can be transferred between tasks through shared underlying feature representations. The number of cluster labels in the cluster label set is equal to and corresponds one-to-one with the number of cluster labels in the cluster label set.

[0058] Input dataset for hierarchical classification model From the first input dataset Second input dataset Jointly generated. The first input dataset includes a resource load dataset that combines cluster label sets. , Represents the resource load data in the first... The feature vector of a user's resource load data includes information such as resource status, historical load data, and resource availability. , Indicates the first Cluster labels for each user, representing the different categories to which resource load data belongs, such as peak periods, off-peak periods, etc. This represents a cluster label set.

[0059] To improve the classification ability of hierarchical classification models, an adaptive deep cascaded forest (daForest) is introduced. This is a layer-by-layer optimized model structure, where each layer uses multiple classifiers (such as decision trees or random forests) to classify demand-side resources. The overall output of the adaptive deep cascaded forest is obtained by weighted summation of the outputs of each layer's classifiers. Specifically, given an input feature vector set... Input feature vector set For the input dataset The generated feature vector set, the output of the adaptive deep concatenated forest can be represented as: ; in, For input The output of a time-adaptive deep cascaded forest. The layer number in the adaptive deep cascaded forest. To adapt the number of layers in a deep cascaded forest, This represents the classifier index for the adaptive deep cascaded forest. The total number of classifiers in an adaptive deep cascaded forest. For the first Layer The parameters of each classifier Indicates that the input is And the weight is Time Layer The output of each classifier For the first Layer The weights of each classifier; Through the layer-by-layer structure of cascaded forests, adaptive deep cascaded forests can extract features step by step from low to high levels and optimize the performance of each classifier.

[0060] This application aims to enable the sharing of policies and value functions among multiple tasks in multi-task reinforcement learning (MT-RL) to achieve knowledge sharing and improve overall learning efficiency. The objective of each task is to maximize its cumulative reward. The objective function of MT-RL can be expressed as: ; in, For multi-task reinforcement learning, For the first The weighting factor for each task, Let be the mathematical expectation, representing the average value over all possible state-action paths. The time step index represents the time dimension of the sequence decision. For at any time The discount factor is used to control the impact on long-term returns. For the first The immediate reward for each task, based on the specific objective definition of the task. For at any time state, For a moment Actions such as demand response and load regulation. This is the initial state. For the first k The initial state of each task. This is a multi-task shared policy function that outputs the probability distribution of each action. By employing a sharing strategy, effective knowledge sharing can be achieved across multiple tasks, optimizing the learning objectives for each task.

[0061] An adaptive feature selection mechanism is also introduced, where the input features for each classifier layer are selected through optimization. The feature selection process measures the importance of features in classification using methods such as information gain (IG) and selects the optimal features for training. Specifically, the goal of feature selection is to maximize the classification accuracy of the task, and the optimization formula for feature selection is: ; in, For the first Layer The classifier selected is the first... The input feature vector of each user The variable that maximizes the objective function. for The information gain value measures the ability of a feature to distinguish between different output categories; a higher information gain indicates that the feature is more important.

[0062] Furthermore, to effectively share knowledge across tasks, a shared feature representation and a task-specific feature transformation layer are introduced. The shared feature representation captures the commonalities between tasks, while the task-specific features include: ; in, For the loss function of shared feature representation, For the first Shared feature representation of each task For the first The specific characteristics of a task are represented. The square of the Euclidean norm represents the degree of dissimilarity between two vectors; To further improve the efficiency of multi-task learning, an adaptive mechanism is adopted to automatically adjust the weighting factor of each task based on its learning progress. The weighting factor for the task is related to the magnitude of the gradient, and the specific update rule is as follows: ; in, For the first The learning rate hyperparameter for each task controls the adjustment range of the weighting factor. For; the Each task in the current model parameters The gradient reflects the learning rate of the task. The updated task weights are determined by the gradient: the larger the gradient (i.e., the learning has not converged), the smaller the weight; conversely, the smaller the gradient, the larger the weight, indicating that the task has been learned stably.

[0063] The training process of multi-task reinforcement learning is accomplished by minimizing a comprehensive loss function. This comprehensive loss function includes classification error, regularization, and feature selection, and its goal is to simultaneously optimize the classifier's performance, the effectiveness of feature selection, and knowledge sharing between tasks. The comprehensive loss function can be expressed as: ; in, To minimize the function, Clustering labels are hour In the The loss function of the layer, For the first Layer Sparse feature selection vectors for each classifier As the first coefficient, As the second coefficient, The third coefficient, , , The values ​​of are 0 to 1, which are used to balance the importance of different parts of the loss function; L2 regularization is used to prevent overfitting. L1 sparsification is used to facilitate feature selection; After training, hierarchical classification predictions can be performed on new samples. Through a shared strategy, the hierarchical classification model can make decisions on each task and effectively classify demand-side resources. The hierarchical classification prediction process can be represented as: ; in, The optimal action selected at the end, i.e. the final hierarchical classification result, The motion space contains all feasible control operations. The current state of the new sample; The hierarchical classification model achieves efficient hierarchical classification of demand-side resources by combining adaptive deep cascaded forests and multi-task reinforcement learning. Through layer-by-layer classifier optimization, shared feature representations, and knowledge sharing between tasks, the model effectively handles multiple tasks, improves overall performance, and provides support for resource scheduling in practical applications.

[0064] Calculation example: In this example, the demand-side resource set is derived from the actual operation data of a provincial power grid in 2023. Through systematic analysis of the 96-dimensional time-series load curves of 16 typical demand-side resources (including air conditioning clusters, energy storage power stations, and industrial furnaces), a complete dataset containing 25,000 resource entities was constructed. The data system integrates cluster label sets extracted by CNN-AE-GMM clustering analysis (schedulable, stable, highly volatile, and intermittent), historical interaction influence index sets generated by GNN analysis (e.g., 0.82 for air conditioning clusters, 0.81 for energy storage power stations), and a 20-dimensional feature space formed by multi-dimensional feature fusion (including key parameters such as load characteristics, behavioral patterns, and value dimensions). All data has been verified through the provincial power grid data platform to ensure that the analysis results are both statistically significant and have engineering practical value, providing comprehensive and reliable data support for the hierarchical classification of demand-side resources.

[0065] Table 1. Detailed Cluster Analysis of Demand-Side Resource Sets Depend on Figure 2-5 From Table 1, we can see that, Figure 2-5Standardized load typically refers to load data that has undergone normalization, often used to eliminate the influence of dimensions and facilitate comparison between different load curves. In other words, standardized load is a dimensionless value. Through a clustering model, the resource load dataset of 16 users was clearly divided into four typical behavioral pattern clusters, i.e., four cluster label sets: schedulable resources, stable resources, highly volatile resources, and intermittent resources. Schedulable resources (orange), such as air conditioning clusters and energy storage power stations, exhibit significant bimodal characteristics, consistent with the characteristics of peak electricity consumption in the morning and evening during actual operation. Stable resources (blue), such as industrial furnaces and data centers, maintain consistently stable load curves, reflecting the operating characteristics of continuous production equipment. Highly volatile resources (red), such as EV charging stations and commercial refrigeration, exhibit multi-peak fluctuation characteristics. Intermittent resources (green), such as agricultural irrigation and distributed photovoltaic systems, exhibit random start-stop operation modes. The clustering results show that the proportion of each type of demand-side resource belonging to the main cluster exceeds 95%, indicating that the grouping effect is stable and reliable.

[0066] Cluster analysis revealed the essential differences among users in the demand-side resource set, with schedulable resources accounting for 31.25%, including high-value regulatory resources such as air conditioning clusters and energy storage power stations. This clustering result provides a solid foundation for subsequent behavioral motivation analysis and helps to understand the operational characteristics and regulatory potential of different types of resources.

[0067] Depend on Figure 6 Cluster analysis revealed the essential differences among users in the demand-side resource set, with schedulable resources accounting for 31.25%, including high-value regulation resources such as air conditioning clusters and energy storage power stations. This clustering result provides a solid foundation for subsequent behavioral motivation analysis and helps to understand the operational characteristics and regulation potential of different types of demand-side resources. The detailed influence values ​​of the 16 categories of demand-side resources are shown in Table 2. Table 2 Ranking of the interactive influence of demand-side resource sets Impact analysis not only validated the rationality of the clustering results but also provided a new dimension for demand-side resource stratification. High-impact demand-side resources should be the primary targets of demand response projects, while low-impact demand-side resources are suitable for providing ancillary services.

[0068] Based on the analysis results of the first two stages, a random forest classifier was used to divide the 16 types of demand-side resources into three management levels. For example... Figure 7As shown, core layer resources (orange) include air conditioning clusters, energy storage power stations, smart lighting, and emergency power supplies, with a classification accuracy of 96.45%; production layer resources (blue) include industrial furnaces, data centers, production lines, and communication base stations, with an accuracy of 91.40%; and basic layer resources (green) include the remaining 8 types of resources, with an accuracy of 87.16%. The classification results show that core layer resources are the easiest to accurately identify due to their distinct bimodal characteristics and high influence; while the basic layer resources have high similarity in characteristics, leading to some misclassification, which is consistent with the actual resource characteristics. The classification results for the 16 types of demand-side resources are shown in Table 3. Table 3. Hierarchical Classification Results of Demand-Side Resources The stratified results form a complete demand-side resource management system: core layer resources (4 types) undertake real-time frequency regulation tasks with a response time of <5 minutes; production layer resources (4 types) are responsible for planned adjustments such as load shifting with a response time of 15-30 minutes; and basic layer resources (8 types) provide auxiliary services such as energy efficiency management with a response time >60 minutes. Applying the demand-side resource stratification and classification method based on interactive behavior drivers proposed in this application to pilot areas improved the resource scheduling efficiency of demand response projects by 42%.

[0069] This example fully implements a scientific stratification of 16 types of demand-side resources: air conditioning clusters, energy storage power stations, smart lighting, and emergency power supplies are classified as the core layer, possessing minute-level response capabilities; industrial furnaces, data centers, production lines, and communication base stations are classified as the production layer, suitable for planned regulation; EV charging stations, commercial refrigeration, residential water heaters, distributed photovoltaics, agricultural irrigation, elevator systems, medical equipment, and wastewater treatment belong to the basic layer, providing ancillary services. Through a technical approach of "feature extraction-motivation analysis-stratified classification," a three-tiered management system precisely matched to resource characteristics was established, and practical applications show a significant improvement in scheduling efficiency. This stratification result provides a standardized solution for smart grid demand-side management; future research will optimize the classification adaptability under extreme scenarios.

[0070] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus (systems), or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0071] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce an instruction that executes via the processor of the computer or other programmable data processing apparatus to create an instruction for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0072] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0073] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0074] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0075] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

[0076] Finally, it should be noted that the above content is only used to illustrate the technical solution of this application, and is not intended to limit the scope of protection of this application. Simple modifications or equivalent substitutions made by those skilled in the art to the technical solution of this application shall not depart from the substance and scope of the technical solution of this application.

Claims

1. A demand-side resource hierarchical classification method based on interactive behavior-driven factors, characterized in that, include: Obtain the current demand-side resource set; The current clustering label set of the current demand-side resource set is obtained using the overall clustering model; The current behavioral feature set and current interaction influence index set of the current demand-side resource set are obtained by using a motivation analysis model; A hierarchical classification model is adopted to obtain the hierarchical classification result of the current demand-side resource set based on the current cluster label set, the current behavior feature set, and the current interaction influence index set; The overall clustering model is obtained based on convolutional neural networks, autoencoders, Gaussian mixture models, and historical demand-side resource sets. The driving force analysis model is constructed based on graph neural networks and historical demand-side resource sets; The hierarchical classification model is constructed based on the historical demand-side resource set, the overall clustering model, the motivation analysis model, adaptive deep cascaded forest, and multi-task reinforcement learning.

2. The demand-side resource hierarchical classification method based on interactive behavior-driven factors according to claim 1, characterized in that, The demand-side resource set includes a resource load dataset, which is a collection of resource load data for each user; the historical demand-side resource set is a demand-side resource set for a historical period, including a historical resource load dataset; the overall clustering model is obtained based on a convolutional neural network, an autoencoder, a Gaussian mixture model, and the historical demand-side resource set, including: An initial clustering model is constructed based on the convolutional neural network, the autoencoder, and the Gaussian mixture model. The initial clustering model is trained based on the historical resource load dataset to obtain intermediate clustering models and intermediate clustering results; Determine whether the intermediate clustering model has converged based on the intermediate clustering results; When the intermediate clustering model does not converge, the intermediate clustering model is set to the initial clustering model, and the process of training the initial clustering model based on the historical resource load dataset is repeated to obtain the intermediate clustering model and intermediate clustering results. When the intermediate clustering model converges, a historical clustering label set is obtained based on the intermediate clustering results, and the overall clustering model is obtained based on the intermediate clustering model.

3. The demand-side resource hierarchical classification method based on interactive behavior-driven factors according to claim 2, characterized in that, The initial clustering model includes a first initial model, a second initial model, and a third initial model. The first initial model is constructed based on the convolutional neural network, the second initial model is constructed based on the autoencoder, and the third initial model is constructed based on the Gaussian mixture model. The initial clustering model is trained based on the historical resource load dataset to obtain intermediate clustering models and intermediate clustering results, including: The first initial model is trained using the historical resource load dataset to obtain the first intermediate model and the first intermediate training dataset; The second initial model is trained using the first intermediate training set to obtain the second intermediate model and the second intermediate training set; The third initial model is trained using the second intermediate training set to obtain the third intermediate model and intermediate clustering results; The intermediate clustering model includes the first intermediate model, the second intermediate model, and the third intermediate model.

4. The demand-side resource hierarchical classification method based on interactive behavior-driven factors according to claim 3, characterized in that, When the intermediate clustering model does not converge, the first intermediate model is set as the first initial model, the second intermediate model is set as the second initial model, the third intermediate model is set as the third initial model, and the process of training the initial clustering model based on the historical resource load dataset is repeated to obtain intermediate clustering models and intermediate clustering results.

5. The demand-side resource hierarchical classification method based on interactive behavior-driven factors according to claim 3, characterized in that, When the third intermediate model converges, a historical clustering label set is obtained based on the intermediate clustering results, a feature extraction model is obtained based on the first intermediate model, a data compression model is obtained based on the second intermediate model, and a hybrid clustering model is obtained based on the third intermediate model. The overall clustering model includes the feature extraction model, the data compression model, and the hybrid clustering model.

6. The demand-side resource hierarchical classification method based on interactive behavior-driven factors according to claim 5, characterized in that, The current demand-side resource set includes the current resource load dataset, which is the resource load dataset for the current stage; The current clustering label set of the current demand-side resource set is obtained using a clustering overall model, including: The feature extraction model is used to obtain an output feature map based on the current resource load dataset; The data compression model described above is used to obtain a latent space representation set based on the output feature map; The hybrid clustering model is used to obtain the current clustering label set based on the latent space representation set.

7. The demand-side resource hierarchical classification method based on interactive behavior-driven factors according to claim 6, characterized in that, The feature extraction model includes convolutional layers and pooling layers; The feature extraction model described above is used to obtain an output feature map based on the current resource load dataset, including: The convolutional layer is used to obtain intermediate feature maps based on the current resource load dataset; The pooling layer is used to obtain the output feature map based on the intermediate feature map.

8. The demand-side resource hierarchical classification method based on interactive behavior-driven factors according to claim 5, characterized in that, The demand-side resource set also includes a behavior dataset, which is a collection of behavior data for each user; the historical demand-side resource set also includes a historical behavior dataset, which is a behavior dataset from a historical period. The hierarchical classification model is constructed based on the historical demand-side resource set, the overall clustering model, the motivation analysis model, adaptive deep cascaded forest, and multi-task reinforcement learning, including: The historical clustering label set of the historical demand-side resource set is obtained based on the overall clustering model. Based on the aforementioned motivational analysis model, the historical behavioral characteristic set and historical interaction influence index set of the historical demand-side resource set are obtained; An initial hierarchical classification model is constructed based on the adaptive deep cascaded forest and the multi-task reinforcement learning; The hierarchical classification model is obtained by training the initial hierarchical classification model based on the historical clustering label set, the historical behavior feature set, and the historical interaction influence index set.

9. The demand-side resource hierarchical classification method based on interactive behavior-driven factors according to claim 8, characterized in that, The behavioral data includes individual characteristic data and social network relationship data.

10. The demand-side resource hierarchical classification method based on interactive behavior-driven factors according to claim 9, characterized in that, The behavioral data also includes data on changes in the market environment.

Images