Cloud energy storage user behavior multi-scale adaptive graph network prediction device

Abstract

The invention provides a cloud energy storage user behavior multi-scale adaptive graph network prediction device, which is characterized by comprising a storage module used for storing user characteristics and energy charging and discharging sequence data; the feature vector generation module is used for processing all user features of the specified user to obtain corresponding feature vectors; the multi-scale charge and discharge behavior module is used for processing charge and discharge sequence data of a specified user to obtain a plurality of subsequences; the adaptive graph learning module is used for obtaining an adjacent matrix shared by all the subsequences according to all the subsequences; the multi-scale time sequence diagram module is used for capturing time sequence modes of an adjacent matrix under different scales to obtain a multi-scale embedded vector fusing the time sequence modes; and the multi-scale fusion module is used for fusing the feature vector and the multi-scale embedded vector to obtain a charge and discharge behavior prediction result. In a word, the method can accurately predict the charging and discharging behaviors of the cloud energy storage user.

Description

Technical Field

[0001] The present invention relates to the field of cloud energy storage, and particularly to a multi-scale adaptive graph network prediction device for cloud energy storage user behavior. Background Art

[0002] Cloud energy storage has received wide attention due to its functions such as improving energy storage utilization rate and reducing total cost. Cloud energy storage centralizes the energy storage devices distributed on the user side to the cloud for sharing, and uses the virtual energy storage capacity in the cloud to replace the physical energy storage on the user side. However, due to the flexibility and randomness of the charging and discharging of cloud energy storage, it is difficult to predict the charging and load demands of users, which brings challenges to energy storage scheduling.

[0003] The patent application with the publication number CN116720631A discloses a method, system and storage medium for analyzing and predicting the power generation of distributed photovoltaic power generation. Specifically, it discloses establishing a prediction model based on a data analysis algorithm. After the prediction module analyzes the feature data through the prediction model, it predicts the power generation of the enterprise. The management decision-making module formulates a management strategy based on the predicted power generation of the enterprise to reasonably arrange the energy use plan and avoid waste of energy caused by the power generation exceeding the power consumption demand. However, this patent application only makes predictions based on user power generation data and has not modeled the load behavior of cloud energy storage users, making it difficult to comprehensively optimize the energy storage scheduling strategy.

[0004] The patent application with the publication number CN116663871A discloses a method and system for predicting power consumption demand. Specifically, it discloses clustering and analyzing historical power consumption information according to historical power consumption information and power consumption influencing factors through a clustering algorithm to determine the categories of historical power consumption information, constructing a load difference sequence, respectively extracting the power consumption characteristics of different categories of historical power consumption information, modeling the power consumption demand prediction, and outputting the predicted power consumption demand information of the target area at the target time period. However, this model only considers the load demand of users and does not analyze the charging demand of energy storage users, and is not applicable to the charge and discharge prediction of cloud energy storage users.

[0005] It can be seen that there is a lack of a method for accurately predicting the charge and discharge behavior of cloud energy storage users in the prior art. Summary of the Invention

[0006] The present invention is made to solve the above problems, and aims to provide a multi-scale adaptive graph network prediction device for cloud energy storage user behavior.

[0007] The present invention provides a multi-scale adaptive graph network prediction device for cloud energy storage user behavior, which is used to obtain the prediction results of the charging and discharging behavior of a specified user in a cloud energy storage platform, and has the following characteristics: including a storage module for storing multiple user characteristics of each user in the cloud energy storage platform and the charging and discharging energy sequence data of each user; a feature vector generation module for extracting all user characteristics of the specified user from the storage module and processing them to obtain the corresponding feature vector; a multi-scale charging and discharging behavior module including a multi-scale pyramid network for extracting the charging and discharging energy sequence data of the specified user from the storage module and processing them to obtain multiple subsequences of different scales; an adaptive graph learning module for performing adaptive graph learning based on matrix factorization according to all subsequences to obtain the adjacency matrix shared by all subsequences; a multi-scale temporal graph module including multiple temporal graph neural networks for capturing the temporal patterns of the adjacency matrix at different scales to obtain multi-scale embedding vectors that fuse temporal patterns; and a multi-scale fusion module for fusing the feature vector and the multi-scale embedding vectors to obtain the prediction results of the charging and discharging behavior.

[0008] In the multi-scale adaptive graph network prediction device for cloud energy storage user behavior provided by the present invention, it may also have the following characteristics: among them, the user characteristics are constructed from the collected user information and user charging and discharging behavior data, including basic information characteristics, charging and discharging behavior characteristics, and price demand response characteristics. The feature vector generation module includes: a basic information embedding and merging unit for embedding and merging all basic information characteristics of the specified user to obtain a basic information embedding vector; a charging and discharging behavior embedding and merging unit for embedding and merging all charging and discharging behavior characteristics of the specified user to obtain a charging and discharging behavior embedding vector; a price demand response embedding and merging unit for embedding and merging all price demand response characteristics of the specified user to obtain a price demand response embedding vector; an activation weight generation unit including a fully connected layer of a multi-layer perceptron for processing the charging and discharging behavior embedding vector and the price demand response embedding vector to obtain the corresponding activation weights; a weighted pooling unit for performing weighted sum and pooling on the activation weights and the corresponding price demand response embedding vectors to obtain a weighted pooling vector; and a merging and compression unit for sequentially merging and compressing the basic information embedding vector, the charging and discharging behavior embedding vector, and the weighted pooling vector to obtain a feature vector.

[0009] In the multi-scale adaptive graph network prediction device for cloud energy storage user behavior provided by the present invention, it may also have the following characteristics: among them, the basic information characteristics include user type, user business model, business hours, and user scale. The charging and discharging behavior characteristics include the charging and discharging times of the user on a daily, weekly, monthly, and quarterly basis, the user's horizontal time period characteristics, and the user's peak-valley time period characteristics. The price demand response characteristics include the change ratio of the load rate, the entropy of the demand response potential, the peak-valley difference ratio, and the change ratio of the peak-valley power consumption.

[0010] In the multi-scale adaptive graph network prediction device for cloud energy storage user behavior provided by the present invention, it may further have the following features: Among them, the multi-scale pyramid network in the multi-scale charge and discharge behavior module is composed of multiple pyramid layers stacked in sequence. The input of the first pyramid layer is the charge and discharge energy sequence data, and the input of each subsequent pyramid layer is the output of the previous pyramid layer. Each pyramid layer is used to process the corresponding input to obtain subsequences with a larger scale.

[0011] In the multi-scale adaptive graph network prediction device for cloud energy storage user behavior provided by the present invention, it may further have the following features: Among them, in the adaptive graph learning module, the initialized parameters include node embeddings shared among all scales and embeddings for each scale.

[0012] In the multi-scale adaptive graph network prediction device for cloud energy storage user behavior provided by the present invention, it may further have the following features: Among them, in the multi-scale time series graph module, the time series graph neural network includes a graph neural network and a temporal convolutional network.

[0013] In the multi-scale adaptive graph network prediction device for cloud energy storage user behavior provided by the present invention, it may further have the following features: Among them, the multi-scale fusion module includes: a merging pooling unit for merging and pooling multi-scale embedding vectors to obtain a first vector; an attention unit including an attention mechanism network for processing the first vector to obtain multi-scale important features; a fusion unit for performing a dot product operation on the feature vector and the multi-scale important features to obtain the charge and discharge behavior prediction result.

[0014] Functions and effects of the invention

[0015] According to the multi-scale adaptive graph network prediction device for cloud energy storage user behavior involved in the present invention, because, first, user portrait features, that is, user features, are characterized by basic information features, charge and discharge behavior features, and price demand response features, and feature vectors are constructed based on user features for predicting the charge and discharge behavior of a specified user; second, the charge and discharge energy sequence data is processed by a multi-scale pyramid network, adaptive graph learning based on matrix factorization, and multiple time series graph neural networks including a graph neural network and a temporal convolutional network, so as to retain potential temporal patterns at different time scales, capture temporal patterns specific to various scales, and further obtain multi-scale embedding vectors; finally, the feature vector and the multi-scale embedding vectors are fused through the multi-scale fusion module, thus considering the importance of user portrait features and the representations of each scale and capturing the correlations therein. Therefore, the multi-scale adaptive graph network prediction device for cloud energy storage user behavior of the present invention can accurately predict the charge and discharge behavior of cloud energy storage users. Brief description of the drawings

[0016] Figure 1It is a block diagram of the cloud energy storage user behavior multi-scale adaptive graph network prediction device in the embodiments of the present invention;

[0017] Figure 2 It is a schematic diagram of the structure and working principle of the multi-scale fusion module in the embodiments of the present invention;

[0018] Figure 3 It is a schematic diagram of the working process of the cloud energy storage user behavior multi-scale adaptive graph network prediction device in the embodiments of the present invention. Detailed implementation manners

[0019] In order to make the technical means, creative features, achieved purposes and effects of the present invention easy to understand, the following embodiments will specifically describe the cloud energy storage user behavior multi-scale adaptive graph network prediction device of the present invention in conjunction with the accompanying drawings.

[0020] The cloud energy storage user behavior multi-scale adaptive graph network prediction device of this embodiment is used to obtain the charge and discharge behavior prediction results of a specified user in the cloud energy storage platform.

[0021] Figure 1 It is a block diagram of the cloud energy storage user behavior multi-scale adaptive graph network prediction device in the embodiments of the present invention.

[0022] As Figure 1 shown, the cloud energy storage user behavior multi-scale adaptive graph network prediction device 1 includes a storage module 10, a feature vector generation module 20, a multi-scale charge and discharge behavior module 30, an adaptive graph learning module 40, a multi-scale time series graph module 50, and a multi-scale fusion module 60.

[0023] The storage module 10 is used to store multiple user features of each user in the cloud energy storage platform and the charge and discharge energy sequence data of each user.

[0024] Among them, the user features are constructed from the collected user information and user charge and discharge behavior data, and include basic information features, charge and discharge behavior features, and price demand response features.

[0025] Among them, the basic information features include user type, user business model, business hours, and user scale. The charge and discharge behavior features include the number of charge and discharge times of the user on a daily, weekly, monthly, and quarterly basis, the user's horizontal time period characteristics, and the user's peak-valley time period characteristics. The price demand response features include the change ratio of the load rate, the entropy of the demand response potential, the peak-valley difference ratio, and the change ratio of the peak-valley power consumption.

[0026] The feature vector generation module 20 is used to extract all user features of a specified user from the storage module 10 and process them to obtain corresponding feature vectors, including a basic information embedding and merging unit 201, a charge and discharge behavior embedding and merging unit 202, a price demand response embedding and merging unit 203, an activation weight generation unit 204, a weighted pooling unit 205, and a merging and compression unit 206.

[0027] The basic information embedding and merging unit 201 is used to embed and merge all basic information features of a specified user to obtain a basic information embedding vector.

[0028] The charge and discharge behavior embedding and merging unit 202 is used to embed and merge all charge and discharge behavior features of a specified user to obtain a charge and discharge behavior embedding vector.

[0029] The price demand response embedding and merging unit 203 is used to embed and merge all price demand response features of a specified user to obtain a price demand response embedding vector.

[0030] The activation weight generation unit 204 includes a multi-layer perceptron fully connected layer, which is used to process the charge and discharge behavior embedding vector and the price demand response embedding vector to obtain corresponding activation weights.

[0031] The weighted pooling unit 205 is used to perform weighted sum and pooling on the activation weights and the corresponding price demand response embedding vectors to obtain a weighted pooling vector.

[0032] The merging and compression unit 206 is used to sequentially merge and compress the basic information embedding vector, the charge and discharge behavior embedding vector, and the weighted pooling vector to obtain a feature vector.

[0033] The multi-scale charge and discharge behavior module 30 includes a multi-scale pyramid network, which is used to extract charge and discharge energy sequence data of a specified user from the storage module 10 and process it to obtain multiple subsequences of different scales.

[0034] Among them, the multi-scale pyramid network is composed of multiple pyramid layers stacked in sequence. The input of the first pyramid layer is the charge and discharge energy sequence data, and the input of each subsequent pyramid layer is the output of the previous pyramid layer. Each pyramid layer is used to process the corresponding input to obtain a subsequence of a larger scale.

[0035] In this embodiment, the subsequences of smaller scales can retain more fine-grained details, and the subsequences of larger scales can capture slow-changing trends.

[0036] The adaptive graph learning module 40 is used to perform adaptive graph learning based on matrix factorization according to all subsequences to obtain an adjacency matrix shared by all subsequences.

[0037] Among them, in the adaptive graph learning module 40, the initialized parameters include node embeddings shared among all scales and embeddings for each scale.

[0038] Existing learning-based methods only learn a shared adjacency matrix to model various temporal patterns. In many problems, the shared adjacency matrix helps to learn the most significant temporal patterns in multivariate time series, can significantly reduce the number of parameters, and avoid overfitting problems. However, when modeling multi-scale temporal patterns, this will lead to suboptimal results. The dynamics of short-term patterns are often affected by neighboring nodes, while long-term patterns are often reflected by nodes with similar inherent patterns. Therefore, it is very necessary to learn multiple adjacency matrices. However, directly learning a specific adjacency matrix for each scale will introduce too many parameters, making the model difficult to train, especially when the number of nodes is large. Therefore, in this embodiment, the adaptive graph learning module 40 uses matrix factorization-based adaptive graph learning to process all subsequences. From the perspective of a single scale, scale-related variable correlations can be extracted, and this way can capture both shared information and scale-related information at the same time.

[0039] The multi-scale temporal graph module 50 includes multiple temporal graph neural networks for capturing the temporal patterns of the adjacency matrix at different scales and obtaining multi-scale embedding vectors that fuse the temporal patterns.

[0040] Among them, the temporal graph neural network includes a graph neural network and a temporal convolutional network.

[0041] The multi-scale fusion module 60 is used to fuse the feature vector and the multi-scale embedding vector to obtain the charge and discharge behavior prediction result.

[0042] Figure 2 It is a schematic diagram of the structure and working principle of the multi-scale fusion module 60 in the embodiment of the present invention.

[0043] As Figure 2 shown, the multi-scale fusion module 60 includes a merging pooling unit 601, an attention unit 602, and a fusion unit 603. The working process of the multi-scale fusion module 60 is as follows: First, the merging pooling unit 601 merges and pools the multi-scale embedding vectors to obtain a first vector; secondly, the attention unit 602 including an attention mechanism network processes the first vector to obtain multi-scale important features; finally, the fusion unit 603 performs a dot product operation on the feature vector and the multi-scale important features to obtain the charge and discharge behavior prediction result as the output of the multi-scale fusion module 60.

[0044] In this embodiment, the multi-scale fusion module 60 not only considers the portrait features of a specified user, that is, the feature vector, but also can consider the importance of scale-specific temporal patterns, that is, the multi-scale embedding vector, and capture the cross-scale correlations therein.

[0045] Figure 3 It is a schematic diagram of the working process of the multi-scale adaptive graph network prediction device for cloud energy storage user behavior in the embodiments of the present invention.

[0046] As Figure 3 shown, the working process of the multi-scale adaptive graph network prediction device 1 for cloud energy storage user behavior includes the following steps:

[0047] Step S1, the basic information embedding and merging unit 201 embeds and merges all the basic information features of the specified user to obtain a basic information embedding vector.

[0048] Step S2, the charge and discharge behavior embedding and merging unit 202 embeds and merges all the charge and discharge behavior features of the specified user to obtain a charge and discharge behavior embedding vector.

[0049] Step S3, the price demand response embedding and merging unit 203 embeds and merges all the price demand response features of the specified user to obtain a price demand response embedding vector.

[0050] Step S4, the activation weight generation unit 204 processes the charge and discharge behavior embedding vector and the price demand response embedding vector to obtain the corresponding activation weights.

[0051] Step S5, the weighted pooling unit 205 performs weighted sum and pooling on the activation weights and the corresponding price demand response embedding vectors to obtain a weighted pooling vector.

[0052] Step S6, the merging and compression unit 206 sequentially merges and compresses the basic information embedding vector, the charge and discharge behavior embedding vector, and the weighted pooling vector to obtain a feature vector.

[0053] Step S7, the multi-scale charge and discharge behavior module 30 extracts the charge and discharge energy sequence data of the specified user from the storage module 10 and processes it to obtain multiple subsequences of different scales.

[0054] Step S8, the adaptive graph learning module 40 performs adaptive graph learning based on matrix decomposition according to all the subsequences to obtain an adjacency matrix shared by all the subsequences.

[0055] Step S9, the multi-scale temporal graph module 50 captures the temporal patterns of the adjacency matrix at different scales to obtain a multi-scale embedding vector integrating the temporal patterns.

[0056] Step S10, the merging and pooling unit 601 merges and pools the multi-scale embedding vectors to obtain a first vector.

[0057] Step S11, the attention unit 602 processes the first vector to obtain multi-scale important features.

[0058] In step S12, the fusion unit 603 performs a dot product operation on the feature vector and the multi-scale important features to obtain the charge and discharge behavior prediction result.

[0059] Functions and effects of the embodiment

[0060] According to the cloud energy storage user behavior multi-scale adaptive graph network prediction device involved in this embodiment, first, the user portrait features, i.e., user features, are characterized by basic information features, charge and discharge behavior features, and price demand response features, and a feature vector is constructed according to the user features for predicting the charge and discharge behavior of a specified user; secondly, the charge and discharge energy sequence data is processed by a multi-scale pyramid network, an adaptive graph learning based on matrix factorization, and multiple temporal graph neural networks including a graph neural network and a temporal convolutional network, so as to retain potential temporal patterns at different time scales, capture various scale-specific temporal patterns, and then obtain multi-scale embedding vectors; finally, the feature vector and the multi-scale embedding vectors are fused through a multi-scale fusion module, thereby considering the importance of the user portrait features and each scale representation and capturing the correlation therein. In short, this method can accurately predict the charge and discharge behavior of cloud energy storage users.

[0061] The above embodiments are preferred cases of the present invention and are not used to limit the protection scope of the present invention.

Claims

1. A multi-scale adaptive graph network prediction device for cloud energy storage user behavior, which is used to obtain the charge and discharge behavior prediction results of a specified user in a cloud energy storage platform. Characterized in that: It includes: A storage module, which is used to store multiple user characteristics of each user in the cloud energy storage platform and the charge and discharge energy sequence data of each user; A feature vector generation module, which is used to extract all the user characteristics of the specified user from the storage module and process them to obtain corresponding feature vectors; A multi-scale charge and discharge behavior module, including a multi-scale pyramid network, which is used to extract the charge and discharge energy sequence data of the specified user from the storage module and process it to obtain multiple subsequences of different scales; An adaptive graph learning module, which is used to perform adaptive graph learning based on matrix factorization according to all the subsequences to obtain an adjacency matrix shared by all the subsequences; A multi-scale time series graph module, including multiple time series graph neural networks, which is used to capture the time series patterns of the adjacency matrix at different scales to obtain multi-scale embedding vectors that fuse time series patterns; A multi-scale fusion module, which is used to fuse the feature vectors and the multi-scale embedding vectors to obtain the charge and discharge behavior prediction results.

2. The multi-scale adaptive graph network prediction device for cloud energy storage user behavior according to claim 1 ， Characterized in that: Among them, The user characteristics are constructed from the collected user information and user charge and discharge behavior data, and include basic information characteristics, charge and discharge behavior characteristics, and price demand response characteristics. The feature vector generation module includes: A basic information embedding merging unit, which is used to embed and merge all the basic information characteristics of the specified user to obtain a basic information embedding vector; A charge and discharge behavior embedding merging unit, which is used to embed and merge all the charge and discharge behavior characteristics of the specified user to obtain a charge and discharge behavior embedding vector; A price demand response embedding merging unit, which is used to embed and merge all the price demand response characteristics of the specified user to obtain a price demand response embedding vector; An activation weight generation unit, including a fully connected layer of a multi-layer perceptron, which is used to process the charge and discharge behavior embedding vector and the price demand response embedding vector to obtain corresponding activation weights; A weighted pooling unit, which is used to perform weighted sum pooling on the activation weights and the corresponding price demand response embedding vectors to obtain a weighted pooling vector; A merging and compressing unit, which is used to merge and compress the basic information embedding vector, the charge and discharge behavior embedding vector, and the weighted pooling vector in sequence to obtain the feature vector.

3. The multi-scale adaptive graph network prediction device for cloud energy storage user behavior according to claim 2, characterized in that: Among them, The basic information characteristics include user type, user business model, business hours, and user scale. The charge and discharge behavior characteristics include the charge and discharge times of the user on a daily, weekly, monthly, and quarterly basis, the user's horizontal time period characteristics, and the user's peak-valley time period characteristics. The price demand response characteristics include the change ratio of the load rate, the entropy of the demand response potential, the peak-valley difference ratio, and the change ratio of the peak-valley power consumption.

4. The cloud energy storage user behavior multi-scale adaptive graph network prediction device according to claim 1, wherein: Among them, The multi-scale pyramid network in the multi-scale charge and discharge behavior module is composed of multiple pyramid layers stacked in sequence, The input of the first pyramid layer is the charge and energy sequence data, and the input of each subsequent pyramid layer is the output of the previous pyramid layer, Each pyramid layer is used to process the corresponding input to obtain a subsequence of a larger scale.

5. The cloud energy storage user behavior multi-scale adaptive graph network prediction device according to claim 1, wherein: Among them, In the adaptive graph learning module, the initialized parameters include node embeddings shared among all scales and embeddings for each scale.

6. The cloud energy storage user behavior multi-scale adaptive graph network prediction device according to claim 1, wherein: Among them, In the multi-scale temporal graph module, the temporal graph neural network includes a graph neural network and a temporal convolutional network.

7. The cloud energy storage user behavior multi-scale adaptive graph network prediction device according to claim 1 ， wherein: Among them, The multi-scale fusion module includes: A merging pooling unit for merging and pooling the multi-scale embedding vectors to obtain a first vector; An attention unit including an attention mechanism network for processing the first vector to obtain multi-scale important features; A fusion unit for performing a dot product operation on the feature vector and the multi-scale important features to obtain the charge and discharge behavior prediction result.

Images