A Data-driven Approach for Charging Characteristic Parameter Identification Method of Electric Vehicles

In order to accurately identify new energy electric vehicles charging behaviour characteristic parameter, the theory of consistency is put forward based on the k-means clustering method. The complex coupling network including consistency control is introduced into the data clustering analysis to accurately describe the consistency characteristics of the electric vehicles charging load data in different periods, and the dissimilarity measure with the state update of the adjacent cluster data is used to quickly calculate the initial cluster center. The k-means method is used to quickly identify the expected value of EV initial charging time in typical scenes, and to accurately extract the probability distribution function of EV charging probability and charging initial time. Combined with a practical case, it is verified that the proposed method has the advantages of simple calculation, fast clustering.


Introduction
New energy electric vehicles have attracted increasing attention and application due to their advantages of superior handling performance, low pollution emission and high energy security [1]. However, with the electric car mass connected to the electricity grid charging load random impact load, the safe and stable operation of power grids caused adverse effects. For electric vehicle charging probability and charging time probability distribution function can comprehensively reflect the driving habits, mileage, holidays and seasonal factors on the electric car battery load model [2]. In order to achieve rapid and accurate modelling of EV charging load, the k-means method can be used to complete EV charging load data analysis and extract charging behaviour characteristic parameters [3]. However, in order to improve the accuracy of k-means clustering, the number of clustering centers needs to be accurately judged in advance. M.G.Quiles [4] et al. proposed a region detection method based on particle competition. Each particle can control as many data points as possible according to the set competition mechanism, and finally form a particle population with different characteristics to achieve the accurate calculation of clustering center. Using the idea of coupling resonance synchronization, L. Zhao [5] et al. cluster the data points with synchronization characteristics into a data group in a similar period of time. When all the data points reach a common state, the consistency control is realized [6]. Therefore, the consensus control method can make the coupled resonant network have the characteristics of consistency and synchronization. By applying the consistency control method, T. Chen [7] et al. proved that the complex coupling network could realize the expected group solution of data with different characteristics. Therefore, the key problem to be 2 solved in data-driven EV charging load modelling is to use the consensus control method to solve the selection problem of the initial clustering center of the k-means method, while maintaining the advantage of small computation amount of the k-means method, and realize the accurate online identification of EV charging behaviour characteristic parameters. In order to solve the technical difficulties of the existing data-driven EV charging parameter identification methods, which are complicated in operation and difficult to achieve engineering application, a new EV charging parameter identification method based on consensus k-means clustering was proposed [8], which automatically completed the initial clustering center selection without human intervention and improved the clustering accuracy. The probability distribution function of EV charging period was quickly and accurately calculated to establish EV charging time distribution under typical scenes to improve the accuracy of the model.

Consensus control method of containment
2.1.1 Initial cluster center calculation. A constraint consensus control method was designed, and each data updated its own clustering state according to the dissimilarity measure with neighboring data. When all data completed the updating process, the preliminary selection of clustering center was completed. Based on the consensus control theory, the dissimilarity measure i d between data point i and data point j in the charging load data network is defined as: However, the clustering boundary of collective decision mean α is not reasonable in practical application. Therefore, the consensus control method is proposed, and the dissimilarity measurement i d between data point i and data point j is designed to better meet the requirements of engineering application: By comparing Equations (1) and (2), it can be seen that by introducing the consensus control constraint item, only a small number of local feedback controllers can be injected into the data network to achieve reasonable setting of the clustering boundary of the data network.  (2) analysis shows that contain consensus control ( ) i s t boils down to is expect clustering boundary restrain the consistency of the data points, to simplify the operation, both contain consensus, quickening the response of the control law in the electric vehicles charging load data network will only a small number of data points is set to contain, do not break general, choose one every p point data contain points, set up internal coupling function: (4) Equation (2) can be written as After completing the calculation of the dissimilarity measure i d between data point i and data point j according to Equation (6), gradually achieve the ideal state of divisions of expected clustering boundary ( ) i s t scales, with close to i d data points to form the initial clustering fast, determine the initial clustering center number k-means algorithm for k-means clustering accurately calculate the probability of electric vehicles charging and charging time probability distribution function preparation conditions.

K-means clustering algorithm based on consensus control.
According to the initial number of clustering centers k obtained by the consensus control method, the weekly charging load data set V of EV charging stations under typical scenes is divided into k groups ( k n < ), and each group represents a class. On the basis of the initial clustering completed by the constraint consistency control, the kmeans algorithm is adopted to calculate the similarity between the data points and construct the adjacent matrix. Through repeated iteration, the clustering center is updated and the grouping is refined until the clustering results no longer change, which is the optimal clustering result. The computational complexity of the k-means clustering method has a linear relationship with the data set size, and the calculation is fast. The constraint consistency control method completes the calculation of the initial number of clustering centers, which ensures the accuracy of the clustering results of the kmeans algorithm. The adjacent matrix was constructed by similarity measurement, and the two closest groups were found out and represented by G1 and G2. The two closest element points between the two groups were connected, and the average dissimilarity between the vertices in G1 and G2 of each group was calculated. If its dissimilarity is less than the defined threshold, merge G1 and G2 into a larger group. When the number of groups reaches the pre-defined group number k, the grouping is ended, and the clustering center of each group is calculated. This process is iterated repeatedly until every k clustering center value does not change, that is, the k-means clustering result is obtained, and the mathematical expected value of the k clustering center value is obtained. Figure 1 shows the clustering results of a random data set. The proposed k-means clustering method based on constraint consistency control is applied to perform clustering analysis on the data set. Information is exchanged between each data point and at least the nearest two data points. As can be seen from the analysis of Figure 1, the proposed constraint consistency control method rapidly forms the initial clustering by measuring the dissimilarity between adjacent data points, and determines the number of initial clustering centers of the k-means algorithm, which has good convergence. The proposed k-means clustering method can quickly complete the accurate calculation of each grouping clustering center, and has a better clustering effect.

Case studies
Based on the charging load data of an EV charging station in a financial center in Hangzhou in 2020, clustering analysis was conducted on the charging load data of EV charging stations in each season to test the correctness and feasibility of the proposed EV charging load modelling method based on consensus k-means clustering.

Consensus k-means clustering performance test
In 2020, spring, summer, autumn and winter in financial center electric vehicles charging station a week were charging load data: January 6-January 12 (324 points), April 13-April 19th(368 points), August 10-August 16(352 points),November 9-November 15(307 points), validate the proposed consistency k-means clustering method is feasible. Figure 3(a)-(d) shows the clustering analysis results of charging load data in four natural weeks. In the figure, the abscissa is the starting time of charging, and the ordinate is the charging load of electric vehicles. From the analysis of figure 3, the use of check consensus control method, through charging load data set between adjacent data points not similarity measure computation, fast accurate finished the initial clustering charging load data, through the k-means clustering method of realization of electric vehicles charging feature difference between different set of classes is clear, 6-13 financial center is relatively high demand for electric vehicles charging, at around 9 in the peak load. From 0 to 6 o'clock, 13-16