An electric vehicle charging station short-term charging load prediction method and product

Abstract

This invention discloses a method and product for short-term charging load prediction of electric vehicle charging stations, relating to the field of electric vehicle charging station operation. The invention provides a method for short-term charging load prediction of electric vehicle charging stations, which measures nodes in an initial graph structure converted from historical operating data and considers multiple factors such as time direction, charging scenarios, load ramp-up propagation, capacity margin, and periodic recurrence. This constructs a comprehensive node influence model to evaluate network nodes, and then combines various topological attributes between nodes to perform probabilistic connectivity network modeling. The constructed probabilistic adjacency matrix can intuitively characterize the strength and uncertainty of potential correlations between historical charging load states, thus providing a more reliable data source for identifying key historical load nodes and ensuring the accuracy of subsequent predictions.

Description

Technical Field

[0001] This invention relates to the field of electric vehicle charging station operation, specifically to: 1. a method for predicting short-term charging load of electric vehicle charging stations by integrating improved node metrics and probabilistic connectivity networks; 2. a finished calculator program. Background Technology

[0002] With the continuous increase in the number of electric vehicles, the charging load of public charging stations, park charging stations, and highway service area charging stations exhibits significant peak-valley differences and random fluctuations. The short-term load of charging stations is not only affected by historical load levels but also by factors such as vehicle arrival time, charging demand, charging pile occupancy rate, queuing status, time-of-use pricing, weekday / holiday characteristics, surrounding traffic flow, weather conditions, and the capacity of the station's transformers. These factors result in nonlinearity, non-stationarity, sudden surges, and uncertainty in the charging load time series, posing a significant challenge to accurate short-term forecasting. If the short-term charging load for the next 15 minutes (or increments of 15 minutes) cannot be accurately predicted, charging stations may experience problems such as transformer overload, unreasonable power allocation of charging piles, increased queuing times, lagging charging and discharging strategies for energy storage devices, and exacerbated peak-valley fluctuations in the power distribution network.

[0003] Existing methods for predicting charging load mainly include statistical models, machine learning models, deep learning models, and complex network methods (such as spatiotemporal graph models). Statistical models have a certain ability to model stationary sequences, but they are difficult to characterize sudden increases and nonlinear changes in charging load. Single machine learning models tend to rely on finite lag features, making it difficult to fully utilize similar load states over long historical periods. While deep learning models have strong fitting capabilities, they are prone to overfitting or insufficient interpretability in scenarios with small to medium sample charging stations. Complex network methods can transform time series into a topological structure of nodes and edges to uncover implicit relationships between historical states, but their traditional network construction and K-shell node measurement methods consider relatively few factors, resulting in insufficient or inaccurate identification of key historical load nodes, leading to inaccurate prediction results that are difficult to directly serve charging station scheduling. Summary of the Invention

[0004] Therefore, it is necessary to provide a method and product for predicting the short-term charging load of electric vehicle charging stations, which addresses the problem that existing traditional network construction and K-shell node measurement methods consider only one factor and thus lead to inaccurate prediction results.

[0005] This invention is achieved using the following technical solution: In a first aspect, the present invention discloses a method for predicting short-term charging load at electric vehicle charging stations, comprising: S1, Construct an initial graph structure based on the historical operation data (History) of the target electric vehicle charging station; S2, decompose the initial graph structure to extract the basic topological attributes of each network node in the initial graph structure and construct the corresponding comprehensive influence of the nodes; The basic topology attributes include: time-directed neighborhood propagation term, load ramping propagation consistency term, charging scenario constraint shell diversity term, capacity margin correction term, and periodic recurrence stability term. S3. Combine the comprehensive influence of each network node in the initial graph structure with the mutual topological attributes between nodes to perform probabilistic connectivity network modeling, so as to obtain the probabilistic adjacency matrix P. The topological attributes between nodes include: edge connection probability between nodes, load state similarity, charging scenario similarity, and time decay term. S4, combined with P, find the node v corresponding to the target prediction node from the initial graph structure. h Several network nodes with high similarity are used to construct the prediction input sample matrix X. pred ; S5, select M trained machine learning models based on X respectively. pred Perform corresponding to v h The predictions are then weighted and fused to obtain a combined prediction value. .

[0006] This method for predicting short-term charging load at electric vehicle charging stations implements the method or process according to embodiments of this disclosure.

[0007] Secondly, the present invention discloses a computer program product, comprising a computer program. When executed by a processor, the computer program implements the steps of the short-term charging load prediction method for electric vehicle charging stations disclosed in the first aspect.

[0008] This type of computer program product implements the methods or processes according to embodiments of the present disclosure.

[0009] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention provides a method for predicting short-term charging load at electric vehicle charging stations. It measures nodes in an initial graph structure converted from historical operating data and considers multiple factors such as time direction, charging scenario, load ramp-up propagation, capacity margin, and periodic recurrence to construct a comprehensive node influence for evaluating network nodes. Furthermore, it combines various topological attributes between nodes to perform probabilistic connectivity network modeling, enabling the constructed probabilistic adjacency matrix to intuitively characterize the strength and uncertainty of potential correlations between historical charging load states. This provides a more reliable data source for identifying key historical load nodes and ensures the accuracy of subsequent predictions.

[0010] 2. This invention processes historical operating data, providing charging load time series, and also considers charging scenarios, taking the number of charging sessions and time attributes as features of charging scenarios, providing multi-level factor data sources for subsequent feature processing.

[0011] 3. The present invention uses the CEEMDAN method to decompose and reconstruct the charging load time series, which can effectively reduce the non-stationarity and complexity of the charging load time series and improve the stability of subsequent network modeling and prediction models.

[0012] 4. The present invention uses the finite penetration horizontal visualization method to convert the reconstructed charging load sequence into an initial graph structure, which can preserve the timing sequence information while allowing a small number of intermediate blocking points to be penetrated, thereby enhancing the ability to characterize the potential correlation between sudden fluctuations in charging load and peak-valley states.

[0013] 5. This invention combines a random walk mechanism with similarity score filtering to accurately identify key historical load nodes from the initial graph structure. Combined with subsequent multi-model combination prediction, it can achieve accurate prediction of charging load at the target prediction time. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart of the short-term charging load prediction method for electric vehicle charging stations provided in Embodiment 1 of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] It should be noted that when a component is said to be "installed on" another component, it can be directly on the other component or it may be in a component that is centered on it. When a component is said to be "set on" another component, it can be directly set on the other component or it may also be in a component that is centered on it. When a component is said to be "fixed to" another component, it can be directly fixed to the other component or it may also be in a component that is centered on it.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "or / and" as used herein includes any and all combinations of one or more of the associated listed items.

[0019] Example 1

[0020] This embodiment 1 provides a short-term charging load forecasting method for electric vehicle charging stations (referred to as CEEMDAN_PCNet_LRWR), which is used to forecast the short-term charging load of a target electric vehicle charging station. (Reference) Figure 1 It shows the flowchart of this method, which mainly includes the following steps: Step 1: Historical data preprocessing.

[0021] Step one mainly involves preprocessing the historical operating data (History) of the target electric vehicle charging station to improve data quality and obtain corrected operating data (E-History).

[0022] Specifically, step one includes: I. Obtain the historical operation data (History) of the target electric vehicle charging station—its total time span t. total Calculated monthly (minimum 3 months), with the time scale Δt calculated in 15-minute increments.

[0023] History includes the following data types: total charging load at the station level, number of active charging sessions (i.e., number of active charging vehicles), and time attributes (including: hourly attributes, weekday attributes, weekend attributes, etc.).

[0024] Among them, the total charging load at the station level is the total charging load of the target electric vehicle charging station at each historical moment; the number of active charging sessions represents the level of activity of the target electric vehicle charging station; and the time attribute includes information from both the time and scenario dimensions.

[0025] II. Preprocess the History data to obtain the corrected runtime data E-History.

[0026] Generally, the preprocessing is designed to perform missing value completion, outlier correction, and normalization in sequence.

[0027] Therefore, all types of data within E-History are strictly aligned on the time scale, and the total number of times is uniformly N=t. total / Δt, and it is data with strong temporal coherence.

[0028] Among them, the total station-level charging load in E-History constitutes the charging load time series X(t), and the number of active charging sessions and time attributes serve as charging scenario characteristics.

[0029] Step 2: Decomposition and reconstruction of the charging load sequence.

[0030] In simple terms, step two uses the CEEMDAN method to decompose and reconstruct the charging load time series X(t) to obtain the reconstructed charging load series X. r (t); where t represents the t-th historical moment (the interval between adjacent moments is the same as Δt).

[0031] To effectively reduce the non-stationarity and complexity of the charging load time series X(t) and improve the stability of subsequent network modeling and prediction models, the CEEMDAN method is used to first decompose the charging load time series X(t) into K intrinsic mode function (IMF) components IMF1(t)~IMF1(t) with different frequency characteristics. K (t) and one residual term R(t), namely: Then, based on IMF1(t)~IMF K The sample entropy, energy proportion, and correlation coefficient with X(t) are used to eliminate or merge IMF1(t)~IMF. K The high-frequency noise components in (t) are removed or merged if a certain IMF component simultaneously satisfies: its sample entropy is greater than the sample entropy threshold, its energy proportion is less than the energy threshold, and its correlation coefficient is less than the correlation threshold. The remaining IMF components are then combined to form a set of retained components Ω. Finally, the set of retained components Ω and R(t) are reconstructed into X. r (t), that is: .

[0032] This achieves effective extraction of non-stationary, multi-scale load characteristics, resulting in X. r (t) is used for subsequent initial graph structure and probabilistic connected network modeling.

[0033] Step 3: Construct the initial graph structure.

[0034] Step 3 uses the finite penetration horizontal visibility method to visualize X. r(t) is transformed into the initial graph structure, which allows a small number of intermediate occlusion points to be penetrated while preserving the timing sequence information, thereby enhancing the ability to characterize the potential correlation between sudden fluctuations in charging load and peak-valley states.

[0035] Specifically, step three includes: X r Each historical moment within (t) corresponds to a network node, and a visual judgment is made between any two network nodes; If two network nodes v i v j Based on visual judgment, then for v i v j Establish visual connection edges and perform domain direction labeling to delineate the corresponding domain sets. Where, v i Corresponding to X r The i-th historical moment t within (t) i Its value is X r (t)in t i Total charging load at station level x i ;v j Corresponding to X r The j-th historical moment t within (t) j Its value is X r (t)in t j Total charging load at the station level x j .

[0036] ① The condition for passing the visual judgment is: v i v j The number of intermediate points between them does not exceed L and is higher than or equal to v. i v j The smaller value in the range; L is the finite penetration parameter (L is generally taken as 1).

[0037] In other words, if v i v j The values ​​between are higher than or equal to v i v j The number of intermediate points of the smaller values ​​in the range does not exceed L, indicating that v i v j Visual judgment, and on v i v j Establish a visible connection edge; otherwise, indicate that v i v j The visual judgment was not passed.

[0038] So let v i v j The edge connection probability is A ij If v i vj If a visible edge is established between them, then the corresponding edge connection probability A is... ij =1; otherwise A ij =0 (Note that A) ii (Always 0).

[0039] ② The criteria for domain direction marking are: in v i v j If visible connection edges are established, then for v i In other words, v j As its neighboring node; if j < i, then v i v j The visible connection edges between them are marked as historical neighborhood edges (used to represent v). i (refer to similar historical load conditions), v j It can be further divided into historical neighbor nodes; if j > i, then v i v j The visible connection edges between them are marked as forward propagation edges (used to represent v). i (potential subsequent load evolution relationships), v j It can be further divided into forward neighbor nodes.

[0040] So for v i In other words, there are several visible connecting edges (the neighboring nodes on these edges form a directed neighborhood set N). i These visible connecting edges are further subdivided into: several historical neighborhood edges (the historical neighbor nodes on which form the historical neighborhood set N). i - Several forward propagation edges (the forward neighbor nodes on which form the forward neighborhood set N) i + ).

[0041] In other words, it will satisfy A ij Network node v with =1 and j<i j Classified into v i Historical neighborhood set N i - ; will satisfy A ij Network node v with =1 and j>i j Classified into v i forward neighborhood set N i + The historical neighborhood set N i - Forward neighborhood set N i + Together they form v i The directed neighborhood set N i This provides directional guidance for subsequent node metrics.

[0042] This iterates through X. r Within all historical moments (t), several visible connecting edges are obtained, forming an initial graph structure with all network nodes. Therefore, the adjacency matrix of the initial graph structure is A = [A...]. ij ] N×N .

[0043] It should be noted that steps one, two, and three are equivalent to constructing an initial graph structure based on the historical operation data (History) of the target electric vehicle charging station. These steps can be replaced with other equivalent operations, as long as the initial graph structure matches the History in terms of data and the domain orientation labeling of all network nodes within the initial graph structure is completed.

[0044] Step 4: Measure the nodes of the initial graph structure.

[0045] Since the initial graph structure only describes whether there are visible connections between network nodes, it can only reflect the topological associations between charging load states and cannot distinguish the importance of different network nodes in the overall network. Therefore, it is necessary to perform appropriate node measurement on the initial graph structure.

[0046] In simple terms, step four involves performing node metrics on the initial graph structure to extract the basic topological attributes of each network node within the initial graph structure and to construct the corresponding comprehensive influence of the nodes.

[0047] It's important to note that while traditional K-shell node metric methods, using K-shell decomposition, can measure a node's position within the network core from a global perspective, they typically overlook the order in which nodes are stripped from the network, the temporal direction, load ramp-up status, differences in charging scenarios, and capacity safety constraints. For electric vehicle charging stations, nodes at the same K-shell level may correspond to different operating scenarios, such as weekday morning peak hours, weekend nighttime off-peak hours, sudden concentrated charging during rainy days, or high-speed replenishment during holidays. If nodes are sorted solely based on topological shell values, nodes with different operating mechanisms can easily be mistakenly considered similar nodes.

[0048] Therefore, to improve the ability to identify critical load state nodes, step four proposes an improved K-shell node measurement method (referred to as CRP-KPNS) that takes into account multiple factors (such as time orientation, charging scenario, load ramp-up propagation, capacity margin, and periodic recurrence). Step four is the specific application of CRP-KPNS.

[0049] The main improvement of CRP-KPNS over traditional K-shell node metrics lies in the design of the following five basic topology attributes: First, a time-directed neighborhood propagation term is introduced to distinguish between historical neighborhoods and forward neighborhoods; Second, a load ramping propagation consistency term was introduced to highlight the indicative role of rapid rise or rapid fall in the future peak value. Third, a charging scenario-constrained shell diversity term is introduced to weight shell diversity based on charging scenario characteristics. Fourth, a capacity margin correction term was introduced to give higher safety weight to nodes that are close to the transformer capacity or the upper limit of the charging pile power. Fifth, a periodic recurrence stabilization term was introduced to enhance the identification of daily and weekly repetitive charging behaviors.

[0050] Specifically, step four includes: The K-shell value and stripping order of a network node within the initial graph structure are calculated to determine the basic topological attributes of that network node, which are then integrated to construct the overall node influence of that network node.

[0051] For ease of understanding, let's take network node v as an example. i For example, its K-shell value is ks i Its peeling sequence within its respective shell is O i Its forward neighborhood set is N i + Its historical neighborhood set is N i - .

[0052] Among them, ks i The initial graph structure is obtained through iterative stripping: first, the degree of each network node is calculated; then, starting from the nodes with the lowest degree, nodes are deleted layer by layer, and the degree of the remaining nodes is updated after each deletion; where, network node v i The shell layer number corresponding to the peeling is ks. i .

[0053] O i This indicates that network node v i The order in which nodes are deleted within their respective shells is used to further distinguish the relative coreness of different network nodes within the same shell.

[0054] Then, calculate network node v i CRP-KPNS node comprehensive influence i The process is as follows: Ⅰ. Calculate network node v i The time has a propagation of items to neighboring areas : ; In the formula, α and β are weighting parameters; Represents network node v i Maximum stripping order within the shell; deg j Represents network node v j The value of c; i Represents network node v i The corresponding charging scenario characteristics are one or more of the following: the number of active charging sessions and time attributes; D(c i ,c j ) represents network node v i v j The scene difference distance; τ represents the time decay parameter.

[0055] It not only considers the shell to which a network node belongs and the order of stripping, but also emphasizes the high-impact neighbors that a network node can connect to in subsequent time directions, thereby increasing the weight of historical nodes that guide future load evolution.

[0056] II. Calculate network node v i Load ramp propagation consistency term R i : ; In the formula, r i Represents network node v i The load ramp rate, ε is a constant to prevent the denominator from being zero; r mean This represents the mean of the absolute values ​​of the load ramp-up rates of all network nodes within the initial graph structure.

[0057] R i It describes the propagation characteristics during the rapid rise or fall of charging load: when network node v i When R and its forward neighbor nodes have the same load ramping direction and a large ramping amplitude, i An increase in the value indicates that network node v i It is more likely to be at a critical position during the rising edge, falling edge, or peak formation process of the charging load.

[0058] III. Calculate network node v i Charging scenario constraints on shell diversity : ; In the formula, This indicates that at network node v i The neighbor set N i In the context of charging scenarios, the probability of falling into the s-th shell is weighted according to the similarity of the charging scenarios. ; s represents the shell number.

[0059] Instead of simply grouping network nodes with significant differences in time location and operational activity into the same shell for diversity statistics, we prioritize counting neighboring nodes that are in similar time scenarios and operational states to the current network node.

[0060] IV. Calculate network node v i Capacity margin correction term M i : ; In the formula, C sta This indicates the available charging capacity at the station level (which can be determined from the design specifications of the target electric vehicle charging station, or by X). r (t) Capacity constraint parameters obtained by estimation); u i This indicates the relationship between network node v and network node v. i The corresponding number of active charging sessions is normalized to obtain the busyness correction term; η represents the capacity safety threshold coefficient; δ represents the smoothing coefficient; σ(.) represents the Sigmoid function; μ represents the weight coefficient of the busyness correction term.

[0061] M i This is to ensure that the node measurement results serve the safe operation of the charging station: if network node v i When the corresponding charging load is close to the capacity safety threshold or the charging pile occupancy rate is high, M i Increase the size of the model to make it more sensitive to critical nodes that may cause overload, power limiting, or queuing.

[0062] V. Calculate network node v i The periodic recurrence stable term Q i : ; In the formula, k represents the periodic recurrence enhancement coefficient; Ω i Indicates the relationship with network node v i A set of historical periodic nodes that are in the same or similar time period; q represents Ω. i The number of any node within; Indicates network node v i The local load window centered on the σ q This represents the window similarity scale parameter corresponding to q.

[0063] Q i It can characterize the daily and weekly recurrence patterns in charging behavior: when network node v i When the nearby load pattern repeats multiple times in the daily or weekly cycles of historical time, Q i An increase indicates that network node v i The load state it represents has a high degree of regularity and can serve as a stable reference node in short-term forecasting.

[0064] VI, based on R i , M i Q i A fusion process is performed to calculate CRP-KPNS. i : ; In the formula, Norm(.) represents the normalization function; ω1, ω2, and ω3 are weight coefficients, and ω1+ω2+ω3=1.

[0065] It should be noted that ω1, ω2, and ω3 can be determined using preset weights, entropy weighting, the inverse of the validation set error, or an adaptive update method based on prediction error feedback.

[0066] CRP-KPNS i It can further distinguish nodes with different contributions to future charging load peaks, ramp changes, and overload risks within the same shell.

[0067] Step 5: Construct a probabilistic connected network.

[0068] To overcome the limitations of traditional deterministic networks in depicting the strength and uncertainty of potential correlations between historical charging load states, step five introduces a probabilistic connected network to model the charging load state nodes. This probabilistic connected network not only preserves the topological connections in the initial graph structure but also provides a probabilistic description of the connections between nodes. In other words, step five combines the comprehensive influence of each network node within the initial graph structure with the topological attributes between nodes (including edge connection probabilities, load state similarity, charging scenario similarity, and time decay terms) to model the probabilistic connected network, resulting in the probabilistic adjacency matrix P.

[0069] Specifically, step five includes: I. CRP-KPNS i After normalization, the normalized node comprehensive influence Z is obtained. i .

[0070] Right now, ; In the formula, min(CRP-KPNS) represents the minimum value of the total influence of all network nodes; max(CRP-KPNS) represents the maximum value of the total influence of all network nodes.

[0071] II. Calculate the values ​​of any two network nodes v i v j Load state similarity between Similarity of charging scenarios and time decay term T ij .

[0072] in, , , ; In the formula, σ x σ is the load condition difference scale parameter. c Parameters for different charging scenarios.

[0073] As mentioned above. , T ij With A ij These all belong to the topological attributes between nodes.

[0074] III, combined with Z i Z j , , T ij A ij Calculate any two network nodes v i v j The connection probability EP between ij .

[0075] .

[0076] Thus, EP ij This also reflects v i v j The topological correlations, consistency of operating scenarios, and uncertainty of load fluctuations between historical charging load states are used to intuitively depict the strength and uncertainty of potential correlations between them.

[0077] IV. Based on connection probability EP ij Construct the probabilistic adjacency matrix P.

[0078] In fact, the probability adjacency matrix P is: P = [p ij ] N×N ;p ij Let v be the element in the i-th row and j-th column of the probabilistic adjacency matrix P, representing a network node v. i With network node v j The probabilistic connection strength between them.

[0079] Specifically, The resulting probability adjacency matrix P is the matrix representation of the probabilistic connected network, characterizing X. r The potential correlation strength within (t).

[0080] Step 6: Search for similar nodes.

[0081] Step 6: Using the probabilistic adjacency matrix P, find the node v that is related to the target prediction node from the initial graph structure. h (It corresponds to the target prediction time t) h Several network nodes with high similarity (also known as key historical load nodes) are selected to construct the prediction input sample matrix X. pred .

[0082] Specifically, step six includes: I. Based on the random walk mechanism, the network nodes of the initial graph structure and the target prediction node v h The node similarity scores are iteratively calculated until convergence.

[0083] It is recommended to use a personalized PageRank algorithm with restart or a local random walk restart method to apply the random walk mechanism. This can characterize the potential relationships between nodes and iteratively calculate the relationship between network nodes and the target prediction node v. h The node similarity score—its iterative form is: ; In the formula, W represents the probability transition matrix obtained by normalizing the probability adjacency matrix P; ρ is the restart probability; e q For the target prediction node v h The initial vector; r (l) is the node similarity score at the l-th iteration; l represents the number of iterations in the random walk mechanism.

[0084] II. Select the top m network nodes with the highest node similarity scores to form a set S of similar nodes. m .

[0085] In this case, m is usually taken as 30, but m can also be preset according to the actual situation.

[0086] III. Extract the set S of similar nodes m The multidimensional feature vectors corresponding to each network node are used to form the prediction input sample matrix X. pred .

[0087] Among them, X pred ∈R m×d d represents S m The total dimension of the multidimensional feature vector of a single network node.

[0088] For ease of explanation, let S m Any network node v g That is, v g ∈S m For network node v g For example, the corresponding historical moment is t. g Its multidimensional feature vector is cxg —Including: t g The corresponding total charging load at the station level x g t g-1 The corresponding total charging load at the station level x g-1 (Also known as station-level lagging load), CEEMDAN decomposition characteristic group (containing x g The corresponding K IMF components and 1 residual term, and the time feature group (including t) g Corresponding hour attributes, day of the week attributes, weekend attributes), and running status characteristics (for t). g Corresponding number of active charging sessions), network feature groups (including v g The node degree, K-shell value, stripping order, overall node influence, and node similarity score.

[0089] Additionally, it can be based on X pred Further construct the prediction sample set D pred : Where h represents the prediction step size; y g+h For t g The actual charging load value corresponding to the predicted step size h is used as the actual tag.

[0090] Step 7: Multi-model combination prediction.

[0091] To eliminate the error of a single model prediction, step eight chose a multi-model combined prediction approach: Select M trained machine learning models to predict the input sample matrix X. pred Based on X respectively pred Perform corresponding to v h The prediction (i.e., the corresponding target prediction time t) h M prediction results ), and obtain the combined predicted value through weighted fusion. .

[0092] All M machine learning models are trained—trained on a training set to adjust model parameters and validated on a validation set to examine error. The training and validation sets can be constructed based on commonly known datasets, but it is recommended to use a prediction sample set D. pred Divide into sections.

[0093] The prediction weights for each machine learning model can be taken from empirical values. However, it is highly recommended to use the validation set error adaptive weighting method to determine the prediction weights for each machine learning model. This involves constructing the corresponding prediction weights based on the error of each model on the validation set, resulting in: ; ; In the formula, ω represents the prediction result of the m-th machine learning model; m Represents the prediction weights of the m-th machine learning model; e m Let m represent the error of the m-th machine learning model on the validation set; m∈[1,M].

[0094] In this embodiment 1, three trained models were selected: Support Vector Regression (SVR), LightGBM Regression, and Random Forest Regression. SVR uses kernel function mapping to handle nonlinear relationships and is suitable for small-to-medium sample charging station scenarios. LightGBM Regression is characterized by high training efficiency and strong feature interaction modeling capabilities, making it suitable for handling multi-source operational state features. Random Forest Regression can improve stability and anti-overfitting ability through the ensemble of multiple decision trees. Of course, other types or numbers of machine learning models can also be selected, as long as their predictive performance is satisfactory.

[0095] Finally, based on It can then provide subsequent overload warnings, charging power allocation, energy storage charging and discharging, and queuing scheduling suggestions.

[0096] To illustrate the effect of CEEMDAN_PCNet_LRWR, this Example 1 introduces an ablation experiment and performs performance tests on the same dataset (the publicly available electric vehicle charging dataset ACN-Data).

[0097] The ablation experiment is set up as follows: Use CEEMDAN_PCNet_LRWR as the benchmark method; Remove steps five and six from CEEMDAN_PCNet_LRWR to form the comparison method CEEMDAN_CRP_KPNS; Replace step four in CEEMDAN_CRP_KPNS with traditional K-shell node metrics to form the comparison method CEEMDAN_Base.

[0098] The predictive performance metrics of the above three methods (including: MAE, RMSE, MAPE_gt1, R) 2 As shown in Table 1 below: Table 1 Comparison of Predictive Performance Indicators

[0099] As shown in Table 1, the prediction errors of the three methods listed above, in the order of CEEMDAN_Base, CEEMDAN_CRP_KPNS, and CEEMDAN_PCNet_LRWR, continuously decreased while the R² coefficient of determination continuously increased.

[0100] Specifically, compared to CEEMDAN_Base, CEEMDAN_PCNet_LRWR reduces RMSE by approximately 31.81%, MAE by approximately 37.06%, and MAPE_gt1 by approximately 34.87%, indicating that CEEMDAN_PCNet_LRWR has a significant improvement in prediction accuracy.

[0101] Example 2

[0102] This embodiment 2 discloses a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the short-term charging load prediction method for electric vehicle charging stations disclosed in embodiment 1.

[0103] The computer equipment can be either a mobile terminal or a fixed terminal. Examples of the former include mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), and in-vehicle terminals (such as in-vehicle navigation terminals); examples of the latter include digital TVs and desktop computers.

[0104] This embodiment 2 also discloses a readable storage medium that stores computer program instructions. When the computer program instructions are read and executed by a processor, the steps of the short-term charging load prediction method for electric vehicle charging stations disclosed in embodiment 1 are performed.

[0105] The readable storage medium may include, but is not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination of the above.

[0106] This embodiment 2 also discloses a computer program product, including a computer program. When executed by a processor, the computer program implements the steps of the short-term charging load prediction method for electric vehicle charging stations disclosed in embodiment 1.

[0107] It should be noted that the computer program used to execute the above can be written in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—as well as conventional procedural programming languages—such as C or similar languages. The computer program can be executed entirely on the user's computer, partially on the user's computer, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer through any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN).

[0108] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0109] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for predicting short-term charging load at electric vehicle charging stations, characterized in that, It includes: S1, Construct an initial graph structure based on the historical operation data (History) of the target electric vehicle charging station; S2, perform node measurement on the initial graph structure to extract the basic topological attributes of each network node in the initial graph structure and construct the corresponding comprehensive node influence; The basic topology attributes include: time-directed neighborhood propagation term, load ramping propagation consistency term, charging scenario constraint shell diversity term, capacity margin correction term, and periodic recurrence stability term. S2 includes: Calculate the K-shell value and stripping order of a network node within the initial graph structure to determine the basic topological attributes of that network node, and then integrate and construct the overall node influence of that network node. Let v be two network nodes. i v j ;v i The expression for the overall influence of a node is: ； In the formula, CRP-KPNS i Indicates v i The overall influence of nodes; Norm(.) represents the normalization function; ω1, ω2, and ω3 are weight coefficients; Q i Indicates v i The periodic recurrence of stable terms; M i Indicates v i The capacity margin correction term; Indicates v i The charging scenario constrains the shell diversity term; R i Indicates v i The load ramp propagation consistency term; Indicates v i The time period includes propagation of terms to neighboring regions; ； ； ； ； ； In the formula, ks i For v i K-shell value; α and β are weight parameters; Indicates v i Maximum stripping order within the corresponding shell; O i For v i The order of peeling within the respective shells; N i + For v i forward neighborhood set; deg i Indicates v i The value of c; i Indicates v i Corresponding charging scenario characteristics; D(c i ,c j ) represents v i With v j Scene difference distance; t i For X r The i-th historical moment within (t); X r (t) represents the reconstructed charging load sequence; t represents the t-th historical moment; τ represents the time decay parameter; r i Indicates v i The load ramp rate; ε is a constant to prevent the denominator from being zero; r mean This represents the mean of the absolute values ​​of the load ramp-up rates of all network nodes within the initial graph structure. Indicates that in N i In the context of charging scenario similarity, the probability of falling into the s-th shell is calculated by weighting the similarity of the charging scenarios; s represents the shell number; C sta Indicates the available charging capacity at the station level; u i Indicates to v i The corresponding number of active charging sessions is normalized to obtain the operational busyness correction term; η represents the capacity safety threshold coefficient; δ represents the smoothing coefficient; σ(.) represents the Sigmoid function; μ represents the weighting coefficient of the operational busyness correction term; k represents the periodic recurrence enhancement coefficient; Ω i Indicates with v i A set of historical periodic nodes that are in the same or similar time period; q represents Ω. i The number of any node within it; Indicates with v i The local load window centered on the σ q x represents the window similarity scale parameter corresponding to q; i For X r (t)in t i Total charging load at the station level at that time; S3. Combine the comprehensive influence of each network node in the initial graph structure with the mutual topological attributes between nodes to perform probabilistic connectivity network modeling, so as to obtain the probabilistic adjacency matrix P. The topological attributes between nodes include: edge connection probability between nodes, load state similarity, charging scenario similarity, and time decay term. S4, combined with P, find the node v corresponding to the target prediction node from the initial graph structure. h Several network nodes with high similarity are used to construct the predicted input sample matrix X. pred ; S5, select M trained machine learning models based on X respectively. pred Perform corresponding to v h The predictions are then weighted and fused to obtain a combined prediction value. .

2. The method for predicting short-term charging load at electric vehicle charging stations according to claim 1, characterized in that, S1 includes: Step 1: Preprocess the History data to improve its quality and obtain the corrected runtime data E-History; The total time span of History is t. total The time scale is Δt; the data types of History include: total station-level charging load, number of active charging sessions, and time attributes; the total station-level charging load in E-History constitutes the charging load time series X(t), and the number of active charging sessions and time attributes serve as charging scenario features; t represents the t-th historical moment; Step 2: The CEEMDAN method is used to decompose and reconstruct X(t) to obtain the reconstructed charging load sequence X. r (t); Step 3: Use the finite penetration horizontal visibility method to visualize X. r (t) is transformed into the initial graph structure.

3. The method for predicting short-term charging load at electric vehicle charging stations according to claim 2, characterized in that, In step two, the CEEMDAN method is used to decompose X(t) into K eigenmode function components IMF1(t)~IMF with different frequency characteristics. K (t) and one residual term R(t), then based on IMF1(t)~IMF K The sample entropy, energy proportion, and correlation coefficient with X(t) are used to eliminate or merge IMF1(t)~IMF. K The high-frequency noise components in (t) are extracted, and the remaining intrinsic mode function components are used to form a retained component set Ω. Finally, the retained component set Ω and R(t) are reconstructed into X. r (t).

4. The method for predicting short-term charging load at electric vehicle charging stations according to claim 3, characterized in that, Step three includes: X r Each historical moment within (t) corresponds to a network node, and a visual judgment is made between any two network nodes; If two network nodes v i v j Based on visual judgment, then for v i v j Establish visual connection edges and perform domain direction labeling to delineate the corresponding domain sets; where v i Corresponding to X r The i-th historical moment t within (t) i ;v j Corresponding to X r The j-th historical moment t within (t) j ; The pass condition for visual judgment is: v i v j The number of intermediate points between them does not exceed L, which is higher than or equal to the value of v. i v j The smaller value in the range; L is the finite penetration parameter; If v i v j If a visible connection edge is established between them, then A ij =1; otherwise, A ij =0;A ij Represents network node v i v j The edge connection probability; The criteria for domain direction marking are: in v i v j If visible connection edges are established, and j < i, then v i v j The visible connection edges between them are marked as historical neighborhood edges, v j Classified into v i Historical neighborhood set N i - If j > i, then v i v j The visible connection edges between them are marked as forward propagation edges, v j Classified into v i forward neighborhood set N i + N i - N i + Composition v i The directed neighborhood set N i ; Traverse X r For all historical moments within (t), several visible connection edges are obtained and an initial graph structure is formed with all network nodes.

5. The method for predicting short-term charging load of electric vehicle charging stations according to claim 1, characterized in that, S3 include: S301, CRP-KPNS i After normalization, the normalized node comprehensive influence Z is obtained. i ; S302, Calculate v i v j Load state similarity between Similarity of charging scenarios and time decay term T ij ; in, , , ;σ x σ is the load condition difference scale parameter. c Scale parameters for differences in charging scenarios; S303, combined with Z i Z j , , T ij A ij Calculate v i v j The connection probability EP between ij ; in, A ij Represents network node v i v j The edge connection probability; S304, based on connection probability EP ij Construct the probabilistic adjacency matrix P=[p ij ] N×N ; Where, p ij Let be the element in row i and column j of P; .

6. The method for predicting short-term charging load of electric vehicle charging stations according to claim 1, characterized in that, S4 include: S401, Based on the random walk mechanism, the network nodes of the initial graph structure and the target prediction node v are compared. h The node similarity scores are iteratively calculated until convergence. S402, select the top m network nodes with the highest node similarity scores to form a similar node set S. m ; S403, Extract the set of similar nodes S m The multidimensional feature vectors corresponding to each network node are used to form X. pred ∈R m×d d represents S m The total dimension of the multidimensional feature vector of a single network node; Wherein, if network node v g ∈S m Then its multidimensional eigenvector cx g Includes: t g The corresponding total charging load at the station level x g t g-1 The corresponding total charging load at the station level x g-1 CEEMDAN decomposition feature groups, time feature groups, operational status features, and network feature groups; t g For v g The corresponding historical moment; The CEEMDAN decomposition feature set includes: x g The corresponding K IMF components and 1 residual term are generated; the time feature group includes: t g The corresponding hour attribute, day of the week attribute, and weekend attribute; the running status characteristics include: t g The corresponding number of active charging sessions, network feature groups include: v g The node degree, K-shell value, stripping order, overall node influence, and node similarity score.

7. The method for predicting short-term charging load at electric vehicle charging stations according to claim 1, characterized in that, In S5, the adaptive weighting method of validation set error is used to determine the prediction weights of each machine learning model; in, ; ; In the formula, ω represents the prediction result of the m-th machine learning model; m e represents the prediction weight of the m-th machine learning model; m Let m represent the error of the m-th machine learning model on the validation set; m∈[1,M].

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the short-term charging load prediction method for electric vehicle charging stations as described in any one of claims 1-7.

Images