Charging pile power prediction method and device, electronic equipment and storage medium

Abstract

The invention provides a charging pile power prediction method and device, equipment and a storage medium, and relates to the technical field of electric vehicle charging. The method comprises the following steps: acquiring real-time charging power data of a to-be-detected charging pile; inputting the real-time charging power data into a pre-trained charging power prediction model to obtain a charging power prediction result of the to-be-detected charging pile; wherein the charging power prediction model is obtained by training a causal perception dynamic graph neural network through a historical charging power data sample of a charging network where the charging pile is located. The charging power prediction model is obtained based on the causal perception dynamic graph neural network, the causal relationship and the dynamic state between key nodes in charging network load prediction can be clearly simulated, compared with a traditional charging power prediction method, the method provided by the invention can generate an accurate charging power prediction result, and the method is suitable for large-scale popularization and application. And a reliable basis is provided for efficient management and resource optimization of the power grid.

Description

Technical Field

[0001] The present application relates to the technical field of electric vehicle charging, and in particular to a charging pile power prediction method, device, electronic device and storage medium. Background Art

[0002] The widespread adoption of electric vehicles (EVs) is rapidly changing the transportation system, however, the increased demand for EV charging poses a major challenge to the stability and efficient power management of the power grid. The EV charging network (hereafter referred to as the charging network) plays a key role in facilitating the transition to electric transportation, requiring accurate and powerful load forecasting technology to predict future power demand and ensure a reliable and sustainable charging infrastructure.

[0003] Accurate charging network load forecasting is beneficial for grid operators to relieve peak demand pressure and optimize generation and distribution resources to effectively meet the charging needs of electric vehicles. Reliable forecasting is a key input for infrastructure expansion and upgrades in planning and investment decisions, ensuring sufficient capacity and flexibility to support the growing electric vehicle fleet. In addition, charging network load forecasting becomes the basis for providing consumers with personalized charging cost insights and encouraging behavioral adjustments to flatten the load curve, thereby improving grid efficiency and minimizing charging costs.

[0004] Most existing methods uniformly treat all nodes and interconnections in the charging network during representation learning, ignoring two key aspects that limit their prediction accuracy. First, existing methods do not distinguish between correlation and true causality, which may lead to less robust and interpretable predictions, especially as network dynamics evolve. Second, existing methods do not recognize that certain charging network nodes may have a disproportionately greater impact on the load behavior of the entire network. Because the unique roles of these influential "critical" nodes are not explicitly identified and utilized in the prediction process, key information may not be fully utilized, resulting in degraded performance and insufficient information, making it difficult for traditional prediction methods to achieve accurate predictions. Summary of the invention

[0005] The purpose of this application is to provide a charging pile power prediction method, device, electronic device and storage medium, which can improve the accuracy and reliability of charging power prediction results.

[0006] In a first aspect, an embodiment of the present application provides a charging pile power prediction method, comprising:

[0007] Obtain the real-time charging power data of the charging pile to be tested;

[0008] The real-time charging power data is input into a pre-trained charging power prediction model to obtain the charging power prediction result of the charging pile to be detected; wherein the charging power prediction model is obtained by training the causal-aware dynamic graph neural network through historical charging power data samples of the charging network where the charging pile is located.

[0009] In some embodiments of the present application, the charging power prediction model is pre-trained in the following manner:

[0010] Acquire historical charging power data of each charging pile in the charging network, and obtain historical charging power data samples after preprocessing the historical charging power data;

[0011] The historical charging power data samples are input into the causal-aware dynamic graph neural network for training, and the charging power prediction model is obtained after reaching a preset training cutoff condition.

[0012] In some embodiments of the present application, the causal-aware dynamic graph neural network includes:

[0013] A causal relationship graph learning module, used to construct a causal relationship graph between charging piles according to the historical charging power data samples, and perform a dynamic graph convolution operation on the causal relationship graph to obtain a causal relationship representation;

[0014] A key relationship graph learning module is used to identify influential key nodes in the charging network, construct a key relationship graph with key nodes as constituent elements, and perform a dynamic graph convolution operation on the key relationship graph to obtain a key relationship representation; wherein the node represents a charging pile in the charging network;

[0015] A time representation fusion module, used for fusing the causal relationship representation and the key relationship representation in time representation to obtain a refined time representation;

[0016] An output layer is used to generate a prediction result according to the refined time representation.

[0017] In some embodiments of the present application, the causal relationship graph learning module is specifically used to:

[0018] Normalizing the input historical charging power data sample to obtain a normalized data sample;

[0019] Discovering the causal relationship in the normalized data sample based on the greedy equivalent search framework of the normalized data sample time segment, and constructing a causal relationship graph between the charging piles according to the causal relationship;

[0020] The normalized data sample time segment greedy equivalent search framework is constructed based on a generalized scoring function and a traditional greedy equivalent search algorithm.

[0021] In some implementations of the present application, the key relationship graph learning module is specifically used to:

[0022] The centrality measure based on historical load dependency is used to evaluate the centrality score of each node, and nodes whose centrality scores fall within a preset proportion of higher rankings are considered key nodes.

[0023] In some embodiments of the present application, the time representation fusion module is specifically used to:

[0024] The causal relationship representation and the key relationship representation are combined channel-wise to obtain a combined feature, and two temporal 1-dimensional convolutional layers are applied to the combined feature to obtain a refined temporal representation.

[0025] In some embodiments of the present application, the output layer includes a multilayer perceptron.

[0026] In a second aspect, an embodiment of the present application provides a charging pile power prediction device, comprising:

[0027] An acquisition module is used to obtain real-time charging power data of the charging pile to be detected;

[0028] A prediction module is used to input the real-time charging power data into a pre-trained charging power prediction model to obtain a charging power prediction result of the charging pile to be detected; wherein the charging power prediction model is obtained by training a causal-aware dynamic graph neural network through historical charging power data samples of the charging network where the charging pile is located.

[0029] In a third aspect, the present application provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described in the first aspect.

[0030] In a fourth aspect, the present application provides a computer-readable storage medium having computer-readable instructions stored thereon, wherein the computer-readable instructions can be executed by a processor to implement the method described in the first aspect.

[0031] Compared with the prior art, the charging pile power prediction method provided in the present application obtains the real-time charging power data of the charging pile to be detected; the real-time charging power data is input into a pre-trained charging power prediction model to obtain the charging power prediction result of the charging pile to be detected; wherein the charging power prediction model is obtained by training the causal-aware dynamic graph neural network through the historical charging power data samples of the charging network where the charging pile is located. The charging power prediction model of the present application is obtained based on the causal-aware dynamic graph neural network, which can clearly simulate the causal relationship and dynamics between key nodes in the charging network load prediction. Compared with the traditional charging power prediction method, the method proposed in the present application can generate accurate charging power prediction results, providing a reliable basis for efficient management of the power grid and resource optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present application. Also, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0033] Figure 1 A flow chart of a charging pile power prediction method provided in an embodiment of the present application is shown;

[0034] Figure 2 A schematic diagram of a prediction process of a charging power prediction model provided in an embodiment of the present application is shown;

[0035] Figure 3 A schematic diagram of a prediction process of another charging power prediction model provided in an embodiment of the present application is shown;

[0036] Figure 4 A schematic diagram of a charging pile power prediction device provided in an embodiment of the present application is shown;

[0037] Figure 5 A schematic diagram of an electronic device provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0038] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0039] It should be noted that, unless otherwise specified, the technical terms or scientific terms used in this application should have the common meanings understood by technicians in the field to which this application belongs.

[0040] In addition, the terms "first" and "second" etc. are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or devices.

[0041] Please refer to Figure 1 , Figure 1 1 is a flow chart of a charging pile power prediction method provided in an embodiment of the present application, comprising the following steps S101-S102:

[0042] S101. Obtain real-time charging power data of the charging pile to be detected.

[0043] S102. Input the real-time charging power data into a pre-trained charging power prediction model to obtain a charging power prediction result of the charging pile to be detected; wherein the charging power prediction model is obtained by training a causal-aware dynamic graph neural network through historical charging power data samples of the charging network where the charging pile is located.

[0044] The following describes how to pre-train a charging power prediction model. Specifically, the charging power prediction model may be pre-trained in the following manner, including the following steps:

[0045] S1. Obtain historical charging power data of each charging pile in the charging network, and pre-process the historical charging power data to obtain historical charging power data samples.

[0046] The historical charging power data refers to the historical data of the charging power of each charging pile in the charging network, and the preprocessing refers to data cleaning of the historical charging power data to remove incomplete, non-standard, and inaccurate data.

[0047] S2. Input the historical charging power data samples into the causal-aware dynamic graph neural network for training, and obtain the charging power prediction model after reaching a preset training cutoff condition.

[0048] The preset training cutoff condition may be that the number of training cycles reaches a set number, or the prediction error is less than a preset threshold.

[0049] Specifically, the above-mentioned causal-aware dynamic graph neural network includes: a causal relationship graph learning module, a key relationship graph learning module, a time representation fusion module and an output layer.

[0050] A causal relationship graph learning module, used to construct a causal relationship graph between charging piles according to the historical charging power data samples, and perform a dynamic graph convolution operation on the causal relationship graph to obtain a causal relationship representation;

[0051] The causal relationship graph learning module constructs a causal relationship graph between charging piles based on the historical charging power data samples, specifically including: normalizing the input historical charging power data samples to obtain normalized data samples; discovering the causal relationship in the normalized data samples based on the normalized data sample time segment greedy equivalent search framework, and constructing a causal relationship graph between charging piles based on the causal relationship; wherein the normalized data sample time segment greedy equivalent search framework is constructed based on a generalized scoring function and a traditional greedy equivalent search algorithm.

[0052] A key relationship graph learning module is used to identify influential key nodes in the charging network, construct a key relationship graph with key nodes as constituent elements, and perform a dynamic graph convolution operation on the key relationship graph to obtain a key relationship representation; wherein the node represents a charging pile in the charging network;

[0053] The key relationship graph learning module identifies influential key nodes in the charging network, specifically including: using a centrality measure based on historical load dependencies to evaluate the centrality score of each node, and considering nodes whose centrality scores fall within a preset proportion of higher rankings as key nodes.

[0054] A time representation fusion module, used for fusing the causal relationship representation and the key relationship representation in time representation to obtain a refined time representation;

[0055] The time representation fusion module is specifically used to: combine the causal relationship representation and the key relationship representation channel-wise to obtain a combined feature, and apply two time 1D convolutional layers to the combined feature to obtain a refined time representation.

[0056] An output layer is used to generate a prediction result according to the refined time representation.

[0057] The charging pile power prediction method provided in this application introduces a causal-aware dynamic graph neural network (CADGN), which can explicitly simulate the causal relationships and dynamics between key nodes in the charging network load forecasting. CADGN utilizes two key modules: the causal-aware graph learning module (CAGLM) and the critical relationship graph learning module (CRGLM). CAGLM infers a causal relationship graph from historical load data to represent the causal relationship between charging piles, and performs dynamic graph convolution on the causal graph to learn causal dependency representations. CRGLM focuses on modeling the complex dynamics between the most influential charging piles in the network, by identifying "critical" nodes and constructing a dynamic critical relationship graph to represent their interconnectivity. By dynamically capturing and representing these relationships, CRGLM extracts key information that is often overlooked by traditional graph-based models. The representations learned in these modules are then fused and processed by the temporal representation fusion module (TRFM) and the output layer, thereby improving the accuracy of the prediction, especially in the short-term and medium-term predictions.

[0058] Please refer to Figure 2 ,The charging power prediction model is introduced in detail as follows.

[0059] a) Constructing a causal graph through a causal graph learning module (CAGLM): By constructing a causal graph reflecting the causal relationship and then performing causal dynamic graph learning, the key causal representations between charging piles are extracted. First, the input electric vehicle charging network (EVCS) historical charging power data samples are normalized, and a time-fragment greedy equivalent search (TFGES) framework is designed using the generalized scoring function (GSF) and the traditional greedy equivalent search algorithm, focusing on causal relationship discovery in a shorter period, so that it can handle larger data sets and more complex causal structures.

[0060] The above step a) specifically includes the following 7 steps:

[0061] a1. Construct a causal graph that reflects these causal relationships and perform causal dynamic graph learning based on this causal structure to extract meaningful representations. CAGLM first normalizes the input EVCS historical charging power data samples to ensure that all features contribute equally to the model training process and improve convergence during the optimization process. The mathematical expression of the normalization operation is:

[0062]

[0063] in, represents the normalized data matrix, represents the average value of X in the time dimension, N T represents the number of historical time steps in the look-back window, X (t)represents the load vector at time step t, where t represents the look-back window N T The tth time step in .

[0064] a2. Subsequently, based on the generalized scoring function (GSF) and the traditional greedy equivalent search algorithm, a time segment greedy equivalent search (TFGES) framework was designed for causal discovery. The TFGES framework focuses on normalizing data segments. where t m ∈[1,t e ), t b ,t e ∈[1,N T ] represent the start time and end time of the time segment, and tx <t e By analyzing shorter cycles, such as N T With a historical time step range of ∈[12,72] hours, the TFGES framework can identify causal relationships that may be hidden in longer and more complex time series, enabling it to handle larger datasets and more complex causal structures.

[0065] a3. While balancing model complexity, GSF is used to evaluate candidate graphs based on their fit to the observed data. This method uses RKHS (Reproducing Kernel Hilbert Space) to capture dependencies in a non-parametric way, which can handle nonlinear causality and different data distributions. Cross-validation is used in the TFGES framework by Divide into N Q Training set and test set to reduce overfitting. For a given vertex V ca , Edge E ca Causal directed acyclic graph of Define cross validation as:

[0066]

[0067] in, represents the entire dataset, partitioned for cross validation. Represents the regression function learned from the qth training set represents the number found on the qth test set, is the qth test set in the cross-validation process, F represents the charging power prediction mapping function, N Q represents the number of cross-validation folds, q∈[1,N Q ] represents the qth cross-validation fold data, N S Indicates the number of electric vehicle charging piles, Indicates that for the graph and input data The cross-validation scoring function.

[0068] a4. For each node v in the charging network (i) , the regression function in RKHS is expressed as:

[0069]

[0070] In the formula, v (i) is the i-th individual variable representing the charging network, Z (i) For the picture Medium (i) The parent node set of . Function F tf (·) indicates that RKHS H v The parent node Z (i) Map to v (i) The real underlying function, U (i) represents the noise or error term. The maximum log-likelihood of this regression is given by:

[0071]

[0072] Among them, K X and K Z is the kernel matrix representing the pairwise similarity between time steps in the feature space, λ is the regularization parameter, and t m represents the time step, represents the maximum log-likelihood scoring function. I represents the identity matrix. The optimization problem is formulated as a DAG that minimizes the combined objective. :

[0073]

[0074] Among them, N s yes The number of variables (or nodes), In a given data segment and parameter Φ to evaluate the model fit for node i. For calculation provides the basis for The log-likelihood scores from multiple cross-validation folds are aggregated. Φ contains the specific parameter configuration and optimization criteria of the TFGES framework, and Ω is the regularization graph Complexity penalty.

[0075] a5. Establishing the causal graph structure After that, CAGLM uses a stacked graph convolutional network (SGCN), which has the same architecture as The inferred causal relationship of the encoding remains consistent. The traditional GCN layer adopts the propagation rule:

[0076] Guided by the adjacency structure of the graph, information is aggregated from neighboring nodes:

[0077] H (l+1) =σ(D -1 / 2 AD -1 / 2 H (l) W (l) );

[0078] in, is the node feature matrix of all EVCS nodes in the first layer, d l is the number of features in the first layer of GCN; D is the corresponding diagonal matrix, which ensures proper normalization during message passing, W (l) represents the trainable weight parameter associated with the lth layer, and σ(·) represents the hyperbolic tangent (tanh) activation function.

[0079] In order to enhance the generalization ability of the model, CAGLM constructs a temporarily masked data representation X m , as the initial input feature H (0) This is achieved by applying a shielding ratio R d The random masking operator M ca (R d ), which helps the model learn robust representations that are less sensitive to noise and missing information in the data.

[0080] a6. The implementation of GCN in CAGLM includes batch normalization (BN) and dropout layers distributed between convolutional layers. The following is a detailed description of SGCN using a two-layer GCN architecture in CAGLM:

[0081]

[0082] in, represents the intermediate feature representation generated at the output of the i-th SGCN block in CAGLM, d g is the feature dimension of the last layer of GCN output, and The trainable parameters of the first and second GCN layers are encapsulated in the i-th SGCN block, respectively, emphasizing that each layer and each block learns a different set of convolutional filters. In addition, the operator *G ca Represents the graph convolution operation performed, using A ca as a guiding adjacency structure. Finally, D(η) represents the application of a probability η to prevent overfitting and encourage the sparsity of learned representations.

[0083] To further enrich the feature representation, CAGLM includes an additional component that focuses on capturing short-term local temporal patterns from the raw charging power data. A multi-layer perceptron (MLP) that directly processes the masked data segments linearly. Produces output representation This representation captures the shallow and local temporal dynamics that emerge throughout the epoch. The intermediate representation obtained from GCN Carrying global causal information and local time representation Adding these two elements together yields This is a more comprehensive concatenated feature representation, and this fusion process enriches the expressiveness of the learned representation and contributes to the ability of the Causal-Aware Dynamic Graph Neural Network (CADGN) to model complex interactions in charging networks.

[0084] a7. Finally, to refine these feature maps, two additional standard one-dimensional (temporal) convolutional layers are applied in succession. These layers process Generate H (o) A refined representation of the expression:

[0085]

[0086] in, and represents two temporal convolution operations, each parameterized with learnable kernel weights and combined with nonlinear activations to model complex patterns in the temporal dimension of the input features. By stacking these convolutional layers, CAGLM performs hierarchical feature extraction, where higher-level features are extracted from From the perspective of CAGLM, this multi-level causal relationship representation can capture more abstract and complex patterns by integrating causal and temporal information. As a refined summary of the load dynamics, the causal structure of the network and the local temporal trends within the input data segment are captured.

[0087] b) Critical Relationship Graph Learning Module (CRGLM) performs dynamic learning of key node relationships: Determine and identify key nodes with significant influence in the charging network. The centrality score of each node is evaluated by centrality measurement based on historical load dependencies. Then, a dynamic critical relationship graph is used to represent the relationships between these key nodes, which further enhances the accuracy of prediction by dynamically capturing and representing these relationships.

[0088] The above step b) specifically includes the following 5 steps:

[0089] b1. Determine a subset of key nodes (influential charging piles) and then dynamically learn from the graph representing the relationships between them. To identify key nodes, a centrality measure based on historical load dependencies is used. First, the pairwise relationship score e (i,j) Indicates the charging network v (i) With v (j) The degree of influence between them is calculated as follows:

[0090]

[0091] In the formula, X (i,k) is the input feature vector of charging pile i at time step k; t s represents a time shift to account for potential lagged relationships; N T represents the historical time window considered for calculating the score, w (i,j) The terms serve as learnable weights, allowing the model to dynamically emphasize or de-emphasize specific pairwise charging pile relationships during training.

[0092] b2. The CRGLM module calculates the centrality score s(i) for each charging station i. This score aggregates the relationship scores of the charging station with all other charging stations in the network, aiming to identify charging stations with significant overall influence:

[0093]

[0094] b3, CRGLM uses the threshold parameter R k The Top-K selection mechanism guided by the centrality score identifies the most influential charging piles. This process formally defines the key node set C cr as follows:

[0095] C cr =Top-K(S,R k )={i|rank(s (i) ,S)≤R k};

[0096] In the formula, rank(s i ,S) represents the centrality score of charging pile i in the entire centrality score s i Ranking in the collection This model only considers the centrality scores with the highest R k The charging stations within the scale are considered as key nodes, thus focusing attention on the most important interactions.

[0097] b4. These determined key nodes become the components of the key relationship diagram. Inherits the same set of nodes as the complete charging network, namely V cr = V. However, the adjacency matrix A cr Different from the complete network diagram, A cr Specifically used to represent the interaction between key charging piles. The time-dependent adjacency tensor A at time step t cr Each entry in The definition is as follows:

[0098]

[0099] Among them, i and j represent different charging stations. cr A dynamic graph structure is encoded, in which a network of connections is established between key nodes and other nodes. The time-varying nature of this adjacency structure enables CRGLM to capture the changing influences between these important nodes throughout the charging power prediction process.

[0100] b5. Identify key nodes and form an initial graph Finally, first, CRGLM uses dynamic graphs to learn the structure and cross N cr Layer SGCN to process masked data X m Each layer generates an intermediate representation Specifically capture the patterns around key nodes. The adjacency matrix A of the key graph is used in the convolution operation. cr . Similar to CAGLM, The representations of are aggregated and then concatenated with separate, locally focused temporal representations. The fused feature representation of CRGLM is further refined through a two-layer convolution process similar to CAGLM, resulting in the key relation representation of the final output of CRGLM.

[0101] c) Temporal feature fusion by the temporal representation fusion module (TRFM): The temporal representation fusion module (TRFM) acts as a bridge to fuse the temporal patterns extracted from the causal relationship graph learning module (CAGLM) and the key relationship graph learning module (CRGLM) to generate the final prediction output. The TRFM module can stitch together the dynamic relationship information of causal relationships and key nodes, thereby effectively improving the accuracy of prediction.

[0102] The above step c) specifically includes the following steps:

[0103] The TRFM module acts as a bridge and receives outputs from the CAGLM module and the CRGLM module respectively. and First, by channel-wise and Subsequently, two temporal 1D convolutional layers are applied to the combined features, resulting in a temporally rich output representation:

[0104]

[0105] in, It represents the refined time representation of the final output of the TRFM module; and represents the consecutive temporal convolution operations within the module, each of which is associated with its own set of learnable parameters. ψ(·) represents the rectified linear unit (ReLU) activation function, which adds complexity and nonlinearity. By processing the concatenated representation in this way, the TRFM module aims to discover and The complex interactions between temporal patterns present in the network ultimately provide the downstream output layer with richer information and a stronger basis for generating prediction results.

[0106] d) Output layer generates prediction results: the refined time representation from the TRFM module The input is fed to the output layer to generate the predicted result of the charging power of the charging pile. With the time step t b To time step t e Finally, this concatenated input is processed by an MLP, which consists of a series of fully connected layers that learn to map the fused spatiotemporal representation into target load forecasts, effectively capturing the various relationships learned in the various modules of CADGN. Mathematically, this output generation process can be expressed as:

[0107]

[0108] in, Indicates that from t e +1 to t e +N P The final charging pile load prediction is output by the causal-aware dynamic graph neural network with a time step of .

[0109] The performance of the causal-aware dynamic graph neural network during training is evaluated using the mean squared error (MSE) loss function, which compares the generated predictions and the corresponding basic truth value To calculate:

[0110]

[0111] in, Represents the loss of the model parameters θ, which contains parameters from all constituent modules, such as SGCN layers, traditional convolutional layers, MLPs, and other learnable components. Minimized by the optimization algorithm The driving model generates charging pile load forecasts.

[0112] During model training, the model parameters are adjusted according to the comparison results of the loss function, and the optimal charging power prediction model is obtained after reaching the preset training cutoff condition.

[0113] After obtaining the charging power prediction model, the real-time charging power data of the charging pile to be tested can be input into the charging power prediction model to predict the charging power of the charging pile in a future set time period, and finally the prediction result can be sent to the power grid operator.

[0114] like Figure 3 As shown, the charging pile load data (i.e., charging power data) is acquired in real time, the data is preprocessed, and then model training and real-time prediction are performed. The prediction results are sent to the grid operator, who can adjust the power supply of the charging network, providing a reliable basis for the efficient management and resource optimization of the grid, and effectively improving the stability and management efficiency of the grid.

[0115] The charging pile power prediction method provided in the embodiment of the present application has the following beneficial effects:

[0116] a) Improve the robustness and interpretability of predictions

[0117] The CADGN framework infers and represents the causal relationship between charging piles, rather than just the correlation relationship, through the Causal Relationship Graph Learning Module (CAGLM), which enables the model to capture the true causal dependency. The Critical Relationship Graph Learning Module (CRGLM) focuses on the dynamic relationship of the most influential key nodes in the modeling network. These nodes have a great influence on the overall load behavior of the network. By dynamically capturing and representing these relationships, the model can utilize important information that is ignored by traditional graph models, thereby improving the robustness and interpretability of the prediction.

[0118] b) Better prediction performance

[0119] CADGN uses a Temporal Representation Fusion Module (TRFM) to fuse the temporal patterns extracted from the Causal Relationship Graph Learning Module (CAGLM) and the Critical Relationship Graph Learning Module (CRGLM). This module helps to integrate the information of causal and critical node relationships and further improve the prediction accuracy.

[0120] In the above embodiment, a charging pile power prediction method is provided. Correspondingly, the present application also provides a charging pile power prediction device. The charging pile power prediction device provided in the embodiment of the present application can implement the above charging pile power prediction method. The charging pile power prediction device can be implemented by software, hardware, or a combination of software and hardware. For example, the charging pile power prediction device can include integrated or separate functional modules or units to perform the corresponding steps in the above methods. Please refer to Figure 4 As shown, the charging pile power prediction device 10 includes:

[0121] The acquisition module 101 is used to acquire the real-time charging power data of the charging pile to be detected;

[0122] The prediction module 102 is used to input the real-time charging power data into a pre-trained charging power prediction model to obtain the charging power prediction result of the charging pile to be detected; wherein the charging power prediction model is obtained by training the causal-aware dynamic graph neural network through historical charging power data samples of the charging network where the charging pile is located.

[0123] In some embodiments of the present application, the device further includes: a training module, configured to pre-train the charging power prediction model in the following manner:

[0124] Acquire historical charging power data of each charging pile in the charging network, and obtain historical charging power data samples after preprocessing the historical charging power data;

[0125] The historical charging power data samples are input into the causal-aware dynamic graph neural network for training, and the charging power prediction model is obtained after reaching a preset training cutoff condition.

[0126] In some embodiments of the present application, the causal-aware dynamic graph neural network includes:

[0127] A causal relationship graph learning module, used to construct a causal relationship graph between charging piles according to the historical charging power data samples, and perform a dynamic graph convolution operation on the causal relationship graph to obtain a causal relationship representation;

[0128] A key relationship graph learning module is used to identify influential key nodes in the charging network, construct a key relationship graph with key nodes as constituent elements, and perform a dynamic graph convolution operation on the key relationship graph to obtain a key relationship representation; wherein the node represents a charging pile in the charging network;

[0129] A time representation fusion module, used for fusing the causal relationship representation and the key relationship representation in time representation to obtain a refined time representation;

[0130] An output layer is used to generate a prediction result according to the refined time representation.

[0131] In some embodiments of the present application, the causal relationship graph learning module is specifically used to:

[0132] Normalizing the input historical charging power data sample to obtain a normalized data sample;

[0133] Discovering the causal relationship in the normalized data sample based on the greedy equivalent search framework of the normalized data sample time segment, and constructing a causal relationship graph between the charging piles according to the causal relationship;

[0134] The normalized data sample time segment greedy equivalent search framework is constructed based on a generalized scoring function and a traditional greedy equivalent search algorithm.

[0135] In some implementations of the present application, the key relationship graph learning module is specifically used to:

[0136] The centrality measure based on historical load dependency is used to evaluate the centrality score of each node, and nodes whose centrality scores fall within a preset proportion of higher rankings are considered key nodes.

[0137] In some embodiments of the present application, the time representation fusion module is specifically used to:

[0138] The causal relationship representation and the key relationship representation are combined channel-wise to obtain a combined feature, and two temporal 1-dimensional convolutional layers are applied to the combined feature to obtain a refined temporal representation.

[0139] In some embodiments of the present application, the output layer includes a multilayer perceptron.

[0140] An embodiment of the present application also provides an electronic device corresponding to the method provided in the aforementioned embodiment, and the electronic device may be an electronic device for a client, such as a mobile phone, a laptop computer, a tablet computer, a desktop computer, etc., to execute the above-mentioned charging pile power prediction method.

[0141] Please refer to Figure 5 , which shows a schematic diagram of an electronic device provided by some embodiments of the present application. Figure 5 As shown, the electronic device 20 includes: a processor 200, a memory 201, a bus 202 and a communication interface 203, and the processor 200, the communication interface 203 and the memory 201 are connected via the bus 202; the memory 201 stores a computer program that can be run on the processor 200, and when the processor 200 runs the computer program, it executes the charging pile power prediction method provided in any of the aforementioned embodiments of the present application.

[0142] The memory 201 may include a high-speed random access memory (RAM), and may also include a non-volatile memory, such as at least one disk memory. The communication connection between the system network element and at least one other network element is realized through at least one communication interface 203 (which may be wired or wireless), and the Internet, wide area network, local area network, metropolitan area network, etc. may be used.

[0143] The bus 202 may be an ISA bus, a PCI bus, or an EISA bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc. Among them, the memory 201 is used to store programs, and the processor 200 executes the program after receiving the execution instruction. The charging pile power prediction method disclosed in any implementation of the embodiment of the present application may be applied to the processor 200, or implemented by the processor 200.

[0144] The processor 200 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the hardware integrated logic circuit or software instructions in the processor 200. The above processor 200 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a readily available programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The methods, steps and logic block diagrams disclosed in the embodiments of the present application can be implemented or executed. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. The steps of the method disclosed in the embodiments of the present application can be directly embodied as a hardware decoding processor to be executed, or the hardware and software modules in the decoding processor can be executed. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory 201, and the processor 200 reads the information in the memory 201 and completes the steps of the above method in combination with its hardware.

[0145] The electronic device provided in the embodiment of the present application and the charging pile power prediction method provided in the embodiment of the present application are based on the same inventive concept and have the same beneficial effects as the methods adopted, operated or implemented therein.

[0146] An embodiment of the present application also provides a computer-readable storage medium corresponding to the charging pile power prediction method provided in the aforementioned embodiment, on which a computer program (i.e., a program product) is stored. When the computer program is run by a processor, it will execute the charging pile power prediction method provided in any of the aforementioned embodiments.

[0147] It should be noted that examples of the computer-readable storage medium may also include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other optical or magnetic storage media, which are not listed here one by one.

[0148] The computer-readable storage medium provided in the above-mentioned embodiments of the present application and the charging pile power prediction method provided in the embodiments of the present application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the application programs stored therein.

[0149] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present application, and they should all be included in the scope of the claims and specification of the present application.

Claims

1. A charging pile power prediction method, characterized in that: include: Obtain the real-time charging power data of the charging pile to be tested; The real-time charging power data is input into a pre-trained charging power prediction model to obtain the charging power prediction result of the charging pile to be detected; wherein the charging power prediction model is obtained by training the causal-aware dynamic graph neural network through historical charging power data samples of the charging network where the charging pile is located.

2. The method according to claim 1, characterized in that The charging power prediction model is pre-trained in the following manner: Acquire historical charging power data of each charging pile in the charging network, and obtain historical charging power data samples after preprocessing the historical charging power data; The historical charging power data samples are input into the causal-aware dynamic graph neural network for training, and the charging power prediction model is obtained after a preset training cutoff condition is reached.

3. The method according to claim 2, characterized in that The causal-aware dynamic graph neural network comprises: A causal relationship graph learning module, used to construct a causal relationship graph between charging piles according to the historical charging power data samples, and perform a dynamic graph convolution operation on the causal relationship graph to obtain a causal relationship representation; A key relationship graph learning module is used to identify influential key nodes in the charging network, construct a key relationship graph with key nodes as constituent elements, and perform a dynamic graph convolution operation on the key relationship graph to obtain a key relationship representation; wherein the node represents a charging pile in the charging network; A time representation fusion module, used for fusing the causal relationship representation and the key relationship representation in time representation to obtain a refined time representation; An output layer is used to generate a prediction result according to the refined time representation.

4. The method according to claim 3, characterized in that The causal relationship graph learning module is specifically used for: Normalizing the input historical charging power data sample to obtain a normalized data sample; Discovering the causal relationship in the normalized data sample based on the greedy equivalent search framework of the normalized data sample time segment, and constructing a causal relationship graph between the charging piles according to the causal relationship; The normalized data sample time segment greedy equivalent search framework is constructed based on a generalized scoring function and a traditional greedy equivalent search algorithm.

5. The method according to claim 3, characterized in that: The key relationship graph learning module is specifically used for: The centrality measure based on historical load dependency is used to evaluate the centrality score of each node, and nodes whose centrality scores fall within a preset proportion of higher rankings are considered key nodes.

6. The method according to claim 3, characterized in that The time representation fusion module is specifically used for: The causal relationship representation and the key relationship representation are combined channel-wise to obtain a combined feature, and two temporal 1-dimensional convolutional layers are applied to the combined feature to obtain a refined temporal representation.

7. The method according to claim 3, characterized in that The output layer includes a multilayer perceptron.

8. A charging pile power prediction device, characterized in that: include: An acquisition module is used to obtain real-time charging power data of the charging pile to be detected; A prediction module is used to input the real-time charging power data into a pre-trained charging power prediction model to obtain a charging power prediction result of the charging pile to be detected; wherein the charging power prediction model is obtained by training a causal-aware dynamic graph neural network through historical charging power data samples of the charging network where the charging pile is located.

9. An electronic device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method according to any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that: Computer-readable instructions are stored thereon, and the computer-readable instructions can be executed by a processor to implement the method according to any one of claims 1 to 7.

Images