Charging pile charging power prediction method based on double-layer clustering

Abstract

The invention discloses a charging pile charging power prediction method based on double-layer clustering. The method comprises the following steps: acquiring private charging pile order data; performing data cleaning and feature processing on the order data; a density clustering algorithm (DBSCAN) is adopted to cluster the charging features of the order data dimensions, and key features of the orders are extracted; clustering the charging behavior time sequence of the user by adopting a K-Shape clustering algorithm, extracting the charging behavior time sequence of the user, and standardizing the charging behavior time sequence; integrating the clustering results of the order and the user charging behavior time sequence to generate a comprehensive user-order behavior group, and obtaining a clustering label of global new-phase model input data; and taking the clustering labels as input data, and inputting the input data into a global new-phase model to predict the charging power of the private charging pile. The problem that power prediction is not accurate due to the fact that an existing power prediction method cannot cope with uncertainty caused by user behavior differences and personalized features in a private charging pile scene is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of new energy vehicle charging, and specifically relates to a method for predicting the charging power of charging piles based on double-layer clustering. Background Art

[0002] In recent years, the market scale of private charging piles has grown rapidly. Mainly due to the continuous increase in the ownership of new energy vehicles, private charging piles have gradually become an important part of the charging infrastructure. In the field of private charging piles, users' requirements for charging efficiency, cost, and user experience are increasing day by day. Existing private charging piles only provide basic charging services, lacking the ability of power prediction and intelligent scheduling, and there are problems: the utilization rate of charging power is low, and it is difficult to meet the high-efficiency charging needs of users. The charging time is uncontrollable, and the user experience is poor. Characteristics such as user behavior differences and external grid fluctuations make it difficult for traditional power prediction models to adapt, and the accuracy and intelligence level of prediction are insufficient.

[0003] Although existing power prediction technologies have been applied in public charging piles and power dispatching systems to estimate future power demands through historical data and prediction models. However, due to the particularity of private charging piles (such as charging pile power limitations, user behavior differences, external grid fluctuations, etc.), existing power prediction methods cannot handle the uncertainties brought about by user behavior differences and personalized characteristics in the private charging pile scenario, resulting in relatively low accuracy and intelligence levels of power prediction. Summary of the Invention

[0004] Aiming at the problem that existing power prediction methods cannot handle the uncertainties brought about by user behavior differences and personalized characteristics in the private charging pile scenario, resulting in low power prediction accuracy, the present invention provides a method for predicting the charging power of charging piles based on double-layer clustering.

[0005] To achieve the above technical objectives, the technical solution adopted by the present invention is as follows:

[0006] A method for predicting the charging power of charging piles based on double-layer clustering, comprising the steps of:

[0007] Collect private charging pile order data;

[0008] Perform data cleaning and feature processing on the order data;

[0009] Use the density clustering algorithm (DBSCAN) to cluster the charging characteristics of the order data dimensions, and extract the key features of the orders;

[0010] Use the K-Shape clustering algorithm to cluster the charging behavior time series of users, extract the charging behavior time series of users, and perform standardization processing on the charging behavior time series;

[0011] Integrate the clustering results of the order and the user's charging behavior time series to generate a comprehensive user-order behavior grouping, and obtain the clustering labels of the input data for the global neural-prophet model.

[0012] Use the clustering labels as input data and input them into the global neural-prophet model to predict the charging power of private charging piles.

[0013] Furthermore, the order data includes the device number, order start time, order end time, and charging amount.

[0014] Furthermore, the detailed steps of using the density clustering algorithm to cluster the charging characteristics in the order data dimension are as follows:

[0015] Randomly select an unvisited point from the charging characteristic dataset in the order data dimension;

[0016] Check whether the point is a core point:

[0017] Calculate how many points are in the eps neighborhood of this point;

[0018] If the number of points in the neighborhood >= min samples: This point is a core point. Use it as the starting point of the cluster and start expanding the cluster;

[0019] If the number of points in the neighborhood < min samples: This point is not a core point;

[0020] Expand the cluster:

[0021] For each core point:

[0022] Find all the points in its neighborhood;

[0023] If there are other core points among the points in the neighborhood, add the neighborhood points of these core points to the cluster and recursively expand the cluster;

[0024] If the point in the neighborhood is a border point, only add it to the cluster but do not continue to expand;

[0025] Keep repeating this process until all the core points in the cluster have been processed;

[0026] Mark noise: If a point is neither a core point nor belongs to the neighborhood of any cluster, mark it as a noise point.

[0027] Repeat the step of checking whether the point is a core point, select the next unvisited point, and repeat the above process until all points have been visited;

[0028] End: Divide all points into multiple clusters (each cluster consists of core points and border points), and a set of noise points.

[0030] Calculate the DBI index to obtain the eps and min_samples corresponding to the minimum DBI;

[0031] Update the optimal parameters.

[0032] Furthermore, the DBI index is used to tune and optimize the eps neighborhood and min_samples. The formula for the DBI index is:

[0033]

[0034] Furthermore, the detailed steps for clustering the user's charging behavior time series using the K-Shape clustering algorithm are as follows:

[0035] Initialize the cluster centers:

[0036] Randomly select K time series as the initial cluster centers (centroids);

[0037] The cluster center in K-Shape is defined as a shape center, representing the main trend or shape of the time series in the cluster.

[0038] Assign clusters: Calculate the shape similarity (using Shape-based Distance, SBD) between each user time series and all cluster centers;

[0039] Update the cluster centers: Recalculate the cluster centers based on the time series in the current clusters.

[0040] Iterate until convergence: Repeat the process of assigning clusters and updating the cluster centers until the cluster centers are stable or the maximum number of iterations is reached.

[0041] Furthermore, the calculation of shape similarity includes the following steps:

[0042] Align the time series: Align two time series through normalized cross-correlation (NCC) to maximize their shape similarity.

[0043] Calculate the similarity score: Measure the similarity using the maximum value of the normalized cross-correlation and use this to determine which cluster the time series belongs to.

[0044] Furthermore, the global neural-prophet model decomposes the order and user charging behavior time series into multiple components, and each component captures a specific attribute of the time series:

[0045] Trend captures the long-term trend changes in the time series; supports two types of trends: linear trend: suitable for data with stable growth or decline. Nonlinear trend: models the scenario of saturated growth through logistic regression.

[0046] Seasonality captures the periodic fluctuations in the time series; the Fourier expansion function is used to model seasonality, supporting multiple periods (such as daily, weekly, yearly, etc.).

[0047] Autoregressive (AR) The autoregressive module captures short-term autocorrelation; suitable for time series with strong fluctuations

[0048] Events Support modeling the impact of external events (such as holidays, promotional activities, etc.); events can be added through custom functions to model their positive or negative impacts on the time series.

[0049] Covariates Support external variables as inputs; improve the interpretability of the model.

[0050] Residuals Model the parts that the trend, seasonality, and AR components fail to capture. Ensure more accurate prediction results.

[0051] Furthermore, during the training process of the global neural-prophet model: impose additional penalties on overestimations (cases where the predicted value is greater than the true value);

[0052] Through the penalty term of the custom loss function, increase the penalty intensity for overestimated prediction results; thereby reducing the impact of overestimated prediction results on the grid operation.

[0053] The loss function is defined as follows:

[0054]

[0055] where N represents the number of samples, is the predicted value, y i is the true value, λ is the penalty parameter, MSE measures the overall deviation between the predicted value and the true value, and the hyperparameter λ is set to adjust the penalty intensity when the predicted value is greater than the true value.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] Through double-layer clustering, accurately identify the behavior patterns of orders and users, and provide high-quality grouping labels for the power prediction model.

[0058] The improved Neural-Prophet model can adapt to the complex requirements of the private charging pile scenario and has stronger modeling ability for multi-dimensional features.

[0059] Compared with the existing charging power prediction methods for charging piles, this application has the following advantages:

[0060] Achieve fine-grained grouping of user behavior based on double-layer clustering to capture multi-level characteristics.

[0061] Optimize the applicability of the power grid scheduling scenario by customizing the loss function to reduce resource waste. Brief Description of the Drawings

[0062] Figure 1 It is the overall flowchart of a charging power prediction method for charging piles based on double-layer clustering in an embodiment of the present invention. Detailed Implementation Manner

[0063] For the convenience of those skilled in the art, the present invention will be further described below in conjunction with the embodiments and the drawings. The content mentioned in the implementation manner does not limit the present invention.

[0064] As Figure 1 shown, this embodiment provides a charging power prediction method for charging piles based on double-layer clustering, including the steps:

[0065] Collect private charging pile order data;

[0066] Perform data cleaning and feature processing on the order data;

[0067] Use the density clustering algorithm (DBSCAN) to cluster the charging characteristics of the order data dimension and extract the key features of the order;

[0068] Use the K-Shape clustering algorithm to cluster the charging behavior time series of users, extract the charging behavior time series of users, and perform standardization processing on the charging behavior time series;

[0069] Integrate the clustering results of the order and the user's charging behavior time series to generate a comprehensive user-order behavior grouping and obtain the clustering labels of the input data of the global neural-prophet model;

[0070] Use the clustering labels as input data and input them into the global neural-prophet model to predict the charging power of private charging piles.

[0071] The order data includes the device number, order start time, order end time, and charging amount.

[0072] The detailed steps of using the density clustering algorithm to cluster the charging characteristics of the order data dimension:

[0073] Randomly select an unvisited point from the charging characteristic data set of the order data dimension;

[0074] Check whether the point is a core point:

[0075] Calculate how many points are in the eps neighborhood of this point;

[0076] If the number of points in the neighborhood >= min_samples: This point is a core point. Use it as the starting point of a cluster and start expanding the cluster.

[0077] If the number of points in the neighborhood < minsamples: This point is not a core point.

[0078] Expand the cluster:

[0079] For each core point:

[0080] Find all the points in its neighborhood;

[0081] If there are other core points among the points in the neighborhood, add the neighborhood points of these core points to the cluster as well and recursively expand the cluster;

[0082] If the points in the neighborhood are border points, only add them to the cluster but do not continue to expand;

[0083] Keep repeating this process until all the core points in this cluster have been processed;

[0084] Mark noise: If a point is neither a core point nor belongs to the neighborhood of any cluster, mark it as a noise point.

[0085] Repeat the step of checking whether a point is a core point, select the next unvisited point, and repeat the above process until all points have been visited;

[0086] End: Divide all points into multiple clusters (each cluster consists of core points and border points), and a set of noise points.

[0088] Calculate the DBI index, and obtain the eps and min_samples corresponding to the minimum DBI;

[0089] Update the optimal parameters.

[0090] Use the DBI index to tune and optimize eps neighborhood and min_samples. The DBI index calculation formula is:

[0091]

[0092] Detailed steps for clustering the user's charging behavior time series using the K-Shape clustering algorithm:

[0093] Initialize the class center:

[0094] Randomly select K time series as the initial class centers (centroids);

[0095] The class center in K-Shape is defined as a shape center, representing the main trend or shape of the time series in the cluster.

[0096] Assign clusters: Calculate the shape similarity (using Shape-based Distance, SBD) between each user time series and all class centers.

[0097] Update class centers: Recalculate the class centers based on the time series in the current clusters.

[0098] Iterate until convergence: Repeat the process of assigning clusters and updating class centers until the class centers are stable or the maximum number of iterations is reached.

[0099] The calculation of shape similarity includes the following steps:

[0100] Align time series: Align two time series through Normalized Cross-Correlation (NCC) to maximize their shape similarity.

[0101] Calculate similarity score: Measure the similarity with the maximum value of the normalized cross-correlation and use it to determine which cluster the time series belongs to.

[0102] The global neural-prophet model decomposes the order and user charging behavior time series into multiple components, each component capturing specific attributes of the time series:

[0103] Trend captures the long-term trend changes in the time series; supports both linear and non-linear trends: Linear trend: Suitable for data with stable growth or decline. Non-linear trend: Models saturated growth scenarios through logistic regression.

[0104] Seasonality captures the periodic fluctuations in the time series; uses Fourier expansion functions to model seasonality and supports multiple periods (such as daily, weekly, annual, etc.).

[0105] Auto-Regression (AR) The auto-regression module captures short-term autocorrelations; suitable for time series with strong fluctuations

[0106] Events supports modeling the impact of external events (such as holidays, promotions, etc.); events can be added through custom functions to model their positive or negative impacts on the time series.

[0107] Covariates supports external variables as inputs; improves the interpretability of the model.

[0108] Residuals models the parts that the trend, seasonality, and AR components fail to capture. Ensures more accurate prediction results.

[0109] During the training process of the global neural-prophet model: impose additional penalties on overestimations (cases where the predicted value is greater than the true value);

[0110] Through the penalty term of the custom loss function, increase the penalty intensity for the overestimated prediction results; thereby reducing the impact of the overestimated prediction results on the power grid operation.

[0111] The loss function is defined as follows:

[0112]

[0113] where N represents the number of samples, is the predicted value, y i is the true value, λ is the penalty parameter, MSE measures the overall deviation between the predicted value and the true value, and the hyperparameter λ is set to adjust the penalty intensity when the predicted value is greater than the true value.

[0114] Compared with the prior art, the present invention has the following beneficial effects:

[0115] Through double-layer clustering, accurately identify the behavior patterns of orders and users, and provide high-quality grouping labels for the power prediction model.

[0116] The improved Neural-Prophet model can adapt to the complex requirements of the private charging pile scenario and has stronger modeling ability for multi-dimensional features.

[0117] Compared with the existing charging power prediction methods for charging piles, the present application has the following advantages:

[0118] Based on double-layer clustering, achieve fine grouping of user behaviors and capture multi-level characteristics.

[0119] The custom loss function optimizes the applicability of the power grid scheduling scenario and reduces resource waste.

[0120] The above has introduced in detail a method for predicting the charging power of charging piles based on double-layer clustering. The description of specific embodiments is only used to help understand the method and its core idea of the present application. It should be noted that for those of ordinary skill in the art of this technology, without departing from the principle of the present application, several improvements and modifications can still be made to the present application, and these improvements and modifications also fall within the protection scope of the claims of the present application.

Claims

1. A charging pile charging power prediction method based on double-layer clustering, characterized in that: Includes steps: Collect private charging pile order data; Perform data cleaning and feature processing on order data; The density clustering algorithm is used to cluster the charging features of the order data dimension and extract the key features of the order; The K-Shape clustering algorithm is used to cluster the user's charging behavior time series, extract the user's charging behavior time series, and standardize the charging behavior time series; Integrate the clustering results of order and user charging behavior time series to generate comprehensive user-order behavior groupings and obtain cluster labels for the input data of the global neural-prophet model; The cluster labels are used as input data and fed into the global neural-prophet model to predict the charging power of private charging piles.

2. A charging pile charging power prediction method based on double-layer clustering according to claim 1, characterized in that: The order data includes the device number, order start time, order end time and charge amount.

3. A charging pile charging power prediction method based on double-layer clustering according to claim 2, characterized in that: Detailed steps for clustering charging features of order data dimension using density clustering algorithm: Randomly select an unvisited point from the charging feature dataset of the order data dimension; Check whether the point is a core point; Expand a cluster: find all points in its neighborhood; Repeat the above process until all core points in the cluster are processed; Mark noise: If a point is neither a core point nor belongs to the neighborhood of any cluster, it is marked as a noise point; Repeat the steps to check if the point is a core point, select the next unvisited point, and repeat the above process until all points have been visited; End: Divide all points into multiple clusters, as well as a set of noise points.

4. A charging pile charging power prediction method based on double-layer clustering according to claim 3, characterized in that: The DBI index is used to adjust and optimize the eps neighborhood and min_samples. The DBI index calculation formula is: Among them, k represents the total number of clusters in the data set, S i and S j Represent the average distance within cluster i and cluster j, respectively, and measure the compactness within each cluster. ij The centroid distance between cluster i and cluster j.

5. A charging pile charging power prediction method based on double-layer clustering according to claim 4, characterized in that: Detailed steps for clustering user charging behavior time series using K-Shape clustering algorithm: Initialize the class center: randomly select K time series as the initial class center; Assign clusters: Calculate the shape similarity of each user’s time series with all cluster centers; Update cluster center: recalculate the cluster center based on the time series in the current cluster; Iterate until convergence: Repeat the process of assigning clusters and updating cluster centers until the cluster centers stabilize or the maximum number of iterations is reached.

6. A charging pile charging power prediction method based on double-layer clustering according to claim 5, characterized in that: The calculation of shape similarity includes the following steps: Align Time Series: Align two time series via normalized cross-correlation to maximize their shape similarity. Calculate the similarity score: Use the maximum value of the normalized cross-correlation to measure the similarity and use it to determine which cluster the time series belongs to.

7. A charging pile charging power prediction method based on double-layer clustering according to claim 6, characterized in that: During the training of the global neural-prophet model: additional penalties are imposed on high estimates; By customizing the penalty term of the loss function, the penalty for overestimation of prediction results can be increased, thereby reducing the impact of overestimation of prediction results on power grid operation. The loss function is defined as follows: Where N represents the number of samples, is the predicted value, y i is the true value, λ is the penalty parameter, MSE measures the overall deviation between the predicted value and the true value, and sets the hyperparameter λ. When the predicted value is greater than the true value, the penalty intensity is adjusted.

Images