CN114465256A - Multi-node electric vehicle charging load joint confrontation generation interval prediction method - Google Patents

Abstract

The invention discloses a multi-node electric vehicle charging load joint countermeasure generation interval prediction method, which comprises the steps of analyzing the time-space correlation among multi-node charging loads in a combined charging scene of a day to be predicted and a historical day, and determining an original multi-node multi-correlation day combined charging scene set for describing the charging behavior of a multi-node electric vehicle; generating strong randomness of space-time distribution of the confrontation network depicting charging load by utilizing the gradient punishment Wasserstein, and generating a large number of combined charging scenes which have similar probability distribution with the original scene set but have difference in time sequence distribution in confrontation; screening out a combined scene set strongly related to the days to be predicted by adopting a weighted 2-D correlation coefficient according to the generated multi-node multi-correlation day combined charging scene set; obtaining a multi-node charging load interval prediction result according to the combination scene set related to the intensity of day to be predicted; the method can effectively predict the charging load time-space distribution of the electric automobile in the distribution network space, and is more favorable for improving the stability and the economical efficiency of the operation of the distribution network.

Description

Prediction method of multi-node electric vehicle charging load joint confrontation generation interval

technical field

The invention belongs to the technical field of time-space distribution prediction of electric vehicle charging load, and particularly relates to a multi-node electric vehicle charging load joint confrontation generation interval prediction method.

Background technique

In the scenario of high penetration rate of electric vehicles (EV), a large number of charging loads with strong spatiotemporal uncertainty are connected to the power grid, which may bring problems such as low node voltage and line congestion to the power grid. Safe and stable operation brings great challenges. Therefore, after the EV is connected to the power grid, it is necessary to accurately predict the charging load of each node in the distribution network space, and describe the spatio-temporal distribution of the charging load. In this way, an important reference is provided for arranging a reasonable dispatch plan and reducing the negative impact of the charging load on the operation of the distribution network.

At present, there are two main methods for predicting the spatio-temporal distribution of EV charging load: one is to obtain the spatio-temporal distribution of charging load by establishing a physical model of EV charging load (ie, the physical model-driven method); the other is to use historical charging Load data-driven artificial intelligence algorithms predict charging loads (i.e., a data-driven approach). Existing EV charging load spatiotemporal distribution predictions are mostly driven by physical models. In some studies, after analyzing the EV travel time and daily mileage data, the Monte Carlo method is used to calculate the charging load. In addition, some studies take into account the randomness of different types of EV movement in different regions and time periods, the driver's willingness to drive, and the impact of electricity prices on the spatiotemporal distribution of charging load to predict, but the location and charging cycle of EVs are fixed, and the driving process of EVs is not considered. . In response to this problem, the influence of the transportation network and the distribution network on the spatiotemporal distribution of the charging load can be considered. However, this process does not consider the influence of traffic conditions on the EV driving path. In this regard, some studies have used traffic network modeling and travel chain theory to simulate the dynamic driving process of electric vehicles. A large number of research results have been obtained in the above researches, however, there are many variables involved in modeling, and many assumptions need to be set subjectively, which makes the time-space distribution of EV charging load in the model less objective.

Compared with physical models, the advantages of data-driven charging load prediction methods include: comprehensive utilization of historical charging load data and no need to set a large number of model parameters. A predictive model was built using historical traffic data and weather data to predict EV charging needs. Based on the actual electric vehicle load, a probabilistic prediction method of electric vehicle charging load for different geographical regions can also be proposed. In addition, data-driven deep learning methods have also achieved good results in the field of EV charging load prediction. However, these studies did not consider the spatial correlation of EV charging loads among multiple nodes in the distribution network space. And from the existing research, it is found that the deterministic prediction results of EV charging load are difficult to effectively reflect the risks brought by the strong spatiotemporal uncertainty of the charging load to the distribution network; compared with the deterministic prediction results, the charging load interval prediction results can More effectively characterize the strong randomness of the charging load.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a multi-node electric vehicle charging load joint confrontation generation interval prediction method, which solves the problem that the existing electric vehicle charging load spatiotemporal distribution prediction model has poor objectivity and does not consider the electric vehicle charging load in the distribution network space. The problem of spatial correlation of vehicle charging load.

The technical solution adopted in the present invention is that the multi-node electric vehicle charging load joint confrontation generation interval prediction method is specifically implemented according to the following steps:

Step 1. Map the historical data of electric vehicle charging load into the IEEE33 node distribution network system, and construct the original multi-node multi-correlation daily joint charging scene set based on the historical data of electric vehicle charging load;

Step 2, constructing a multi-node multi-correlation day joint charging scene generation model by using the original multi-node multi-correlation day joint charging scene set, and obtaining and generating a multi-node multi-correlation day joint charging scene set by using the multi-node multi-correlation day joint charging scene generation model;

Step 3, analyzing the correlation between the multi-node multi-correlation day joint charging scenario and the extremely strongly correlated historical day charging scenario used in the prediction, and selecting the one with a high degree of correlation as the to-be-predicted day-related joint scenario set;

Step 4: Obtain a multi-node charging load interval prediction result and a deterministic prediction result according to the data of the last day of the relevant joint scene set on the day to be predicted.

The feature of the present invention also lies in:

计算公式为：The specific process of step 1 is: map the historical data of electric vehicle charging load to the IEEE33 node distribution network system, and number the charging scene space nodes in the IEEE33 node distribution network system 1, . . . , 32 to obtain the electric vehicle corresponding to each node. The historical data of the vehicle charging load is defined as a matrix D _nt for the multi-node combined charging scenario on the day to be predicted, and the multi-node combined charging scenario for the historical day is represented as a matrix (Di) _nt . According to all the historical data of the electric vehicle charging load, D _nt and ( Di) _nt time-space correlation between charging loads in two matrices

The calculation formula is:

In formula (1), n represents the spatial node number in the joint charging scenario, t represents the sampling time point of the spatial charging load in the joint charging scenario, and its ranges are n=1, 2,..., 32 and t=1, 2,... ,24; and

时，表示待预测日多节点联合充电场景与历史日多节点联合充电场景极强相关，将与待预测日多节点联合充电场景极强相关的历史日作为极强相关日；

When is , it means that the multi-node combined charging scenario on the day to be predicted is strongly correlated with the multi-node combined charging scenario on the historical day, and the historical day that is strongly related to the multi-node combined charging scenario on the to-be-forecasted day is regarded as a very strongly correlated day;

According to the correlation analysis, the multi-node joint charging scenarios on the extremely strongly correlated days and the days to be predicted are obtained, and the multi-node joint charging scenarios on the extremely strongly correlated days and the to-be-predicted days are arranged in time series to construct the original multi-node multi-correlation day joint charging scenario sets.

The specific process of step 2 is:

Step 2.1. Construct a gradient penalty Wasserstein generative adversarial network based on the original multi-node multi-correlation daily joint charging scene set, optimize the generator and discriminator in the adversarial network, and use the optimized generated network as a multi-node multi-correlation daily joint charging scene generate models;

Step 2.2. Input the centralized data of the original multi-node multi-correlation daily joint charging scene into the multi-node multi-correlation daily joint charging scene generation model, and generate a large number of multi-node multi-correlation of the same dimension with similar probability distribution to the original joint charging scene data but with different time series distribution. Daily joint charging scenarios, the generated massive multi-node multi-correlation daily joint charging scenarios constitute a multi-node multi-correlation daily joint charging scenario set.

Step 2.1 The specific process of optimizing the generator and discriminator in the adversarial network is as follows:

The difference between the generated data and the real data distribution is described by the Wasserstein distance instead of the JS divergence, and the Wasserstein distance is applied to the generative adversarial network, which is expressed as:

为期望；

为生成样本；

表示由判别器获得的结果；z 为生成器输入的噪声向量，且概率分布为Z～P_Z(z)；x为原始多节点多相关日联合充电场景集中样本特征向量，且X～P_X(x)；in,

for expectation;

to generate samples;

represents the result obtained by the discriminator; z is the noise vector input by the generator, and the probability distribution is Z ~ P _Z (z); x is the sample feature vector of the original multi-node multi-correlation daily joint charging scene set, and X ~ P _X (x);

The gradient penalty term is added to the discriminator loss function, and the objective function of the multi-node multi-correlation daily joint charging scene generation model is:

为电动汽车充电负荷历史数据和生成多节点多相关日联合充电场景数据概率分布间线性采样值。where λ is the gradient penalty coefficient,

A linear sampling value between the historical data of electric vehicle charging load and the probability distribution of multi-node multi-correlation daily joint charging scenario data.

The specific process of step 3 is:

Calculate the weighted 2-D correlation coefficient R _j of the j-th scene in the set of multi-node multi-node combined charging scenarios on the historical day that is strongly related to the multi-node combined charging scene on the day to be tested and the j-th scene in the generated multi-node multi-correlation daily combined charging scene set, and the expression is:

表示与待测日多节点联合充电场景极强相关的历史日多节点联合充电场景D-i与生成多节点多相关日联合充电场景集中第j个场景的 2-D相关系数；where:

Represents the 2-D correlation coefficient of the historical multi-node combined charging scenario Di strongly correlated with the multi-node multi-node combined charging scenario on the day to be tested and the j-th scenario in the generated multi-node multi-correlated daily combined charging scenario set;

The obtained correlation coefficients select the top M joint charging scenarios in order from high to low to form a set of relevant joint scenarios on the day to be predicted.

其中n表示节点编号，且n∈[1,32]，The specific process of step 4 is: according to the relevant joint scenarios on the day to be forecasted, the charging load of each node in the scene on the last day is taken as the charging load of each node on the day to be forecasted:

where n represents the node number, and n∈[1,32],

Equation (6) is used to calculate the charging load interval prediction results and deterministic prediction results of each node:

分别表示t时刻节点n充电负荷区间预测结果的上下限；

表示t时刻节点n的电动汽车充电负荷确定性预测结果。in

respectively represent the upper and lower limits of the prediction result of the charging load interval of node n at time t;

Represents the deterministic prediction result of electric vehicle charging load at node n at time t.

The beneficial effects of the present invention are:

The multi-node electric vehicle charging load combined confrontation generation interval prediction method of the present invention considers the interval prediction of the spatial correlation between the multi-node electric vehicle charging loads, each prediction index is better, and can more effectively predict the electric vehicle charging load in the distribution network space- Empty distribution is more conducive to improving the stability and economy of the operation of the distribution network.

Description of drawings

Fig. 1 is the correlation analysis diagram between the charging load of the to-be-forecasted day and the multi-history day multi-node combined charging scenario in the implementation of the present invention;

2 is a structural diagram of a multi-node multi-correlation day joint charging scenario in the implementation of the present invention;

FIG. 3 is a flow chart for generating a multi-node multi-correlated day joint charging scenario in the implementation of the present invention;

4 is an analysis diagram of probability distribution characteristics of the centralized data generated in the multi-node multi-correlation day combined charging scene and the original multi-node multi-correlation day combined charging scene centralized data in the implementation of the present invention;

FIG. 5 is an analysis diagram of statistical characteristics of a multi-node multi-correlation daily joint charging scenario sample in the implementation of the present invention;

6 is an analysis diagram of spatial correlation between electric vehicle charging loads in a joint charging scenario in the implementation of the present invention;

Fig. 7 is the time sequence distribution analysis diagram of the charging load of each node in the original multi-node multi-correlation day joint charging scene set and the multi-node multi-correlation day joint charging scene set in the implementation of the present invention;

FIG. 8 is a flow chart of the charging load prediction of a multi-node electric vehicle in the implementation of the present invention;

FIG. 9 is a graph showing the statistical results of the evaluation indicators of each prediction method in the implementation of the multi-node electric vehicle charging load joint confrontation generation interval prediction method of the present invention;

Fig. 10 is a graph of the prediction result of EV charging load interval in each season in the implementation of the multi-node electric vehicle charging load joint confrontation generation interval prediction method of the present invention;

FIG. 11 is an evaluation index diagram of the charging load interval prediction result in each season in the implementation of the present invention.

Detailed ways

The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

The multi-node electric vehicle charging load joint confrontation generation interval prediction method of the present invention is specifically implemented according to the following steps:

Step 1. Construct the original multi-node multi-correlation daily joint charging scene set;

There is a strong random access of electric vehicles with high penetration rate in the distribution network space. After reckoning the charging load to different distribution network nodes, the distribution of the charging load is different, and the impact on the voltage level of each node of the distribution network is different. Accordingly, when the power grid reasonably arranges the electric vehicle dispatching plan, it is necessary to consider the influence of the spatial distribution of electric vehicle charging load on the operation of the distribution network. In order to meet the needs of the electric vehicle scheduling plan arranged by the power grid, the charging load of 32 charging stations (including 229 charging piles) in a certain area is mapped to each node of the IEEE33 node distribution network system. Prediction of charging load interval. The electric vehicle charging load data required for the study was collected from the measured data in a certain area in 2019. The data is the daily electric vehicle charging load data record in this area with a sampling interval of 1 hour.

The choice of charging location for electric vehicle users has a certain inertia, and there is a correlation between the charging behavior of users during different days. However, there are differences in the vehicle behavior and charging frequency of electric vehicle users in the corresponding geographical areas of each node in the distribution network. Therefore, it is necessary to select a historical day with strong correlation with the day to be predicted, and construct a multi-node multi-correlation day joint charging scenario. This type of scenario set is used to describe the impact of the charging load of the multi-node joint charging scenario on the to-be-forecasted day for multiple historical days. In order to construct a multi-node and multi-correlation day joint charging scenario, a 2-D correlation coefficient is used to comprehensively analyze the day-to-day similarity of the multi-node joint charging scenario on the day to be predicted and the historical day and the multi-node charging in the scene from the two dimensions of time and space. The similarity between loads overcomes the limitation of traditional correlation analysis methods for single-dimensional analysis of one-dimensional data.

计算公式为：For the charging scene space nodes in the _IEEE33 node distribution network system, number 1, . Di) _nt , according to all the historical data of electric vehicle charging load, calculate the time-space correlation between charging loads in the two matrices D _nt and (Di) _nt

The calculation formula is:

In formula (1), n represents the spatial node number in the joint charging scenario, t represents the sampling time point of the spatial charging load in the joint charging scenario, and its ranges are n=1, 2,..., 32 and t=1, 2,... ,24; and

According to the correlation analysis, the historical multi-node combined charging scenarios with strong correlation on the to-be-forecasted day are obtained, and the historical multi-node combined charging scenarios that are strongly related to the multi-node combined charging scenarios on the to-be-forecasted day are regarded as the multi-node combined charging on the highly correlated day. Scenarios, the multi-node joint charging scenarios on extremely strongly correlated days and the multi-node joint charging scenarios on days to be predicted are arranged in time series to construct the original multi-node multi-correlation day joint charging scenario sets.

According to all the historical data of electric vehicle charging load, formula (1) is applied to calculate the time-space correlation between the multi-node joint charging scenario on the day to be predicted and the multi-node charging load in the joint charging scenario on the previous ten days. And when |R|∈(0.8,1], it means that the multi-node joint charging scenario on the day to be predicted is strongly correlated with the multi-node joint charging scenario on the historical day. The method of the present invention combines the highly correlated day and the multi-node joint charging on the day to be predicted. The scenarios are arranged in time series to construct a multi-node multi-correlation daily joint charging scenario. Based on the measured data of electric vehicle charging load, the first joint charging scenario is constructed from January 9, 2019. Each multi-node multi-correlation daily joint charging scenario includes The 6-day electric vehicle charging load data of 32 nodes. When carrying out the multi-node charging load interval prediction, the original multi-node multi-correlation daily combined charging scene set is divided into training set and test set according to the ratio of 4:1.

Step 2. Multi-node multi-correlation daily joint charging scenario set generation

In order to carry out the research on the spatiotemporal distribution prediction of the charging load in the distribution network space, a multi-node multi-correlation daily joint charging scenario generation model is constructed based on the original multi-node multi-correlation daily joint charging scene set, and the multi-node multi-correlation daily joint charging scene generation model is used to generate the model. Obtain and generate a multi-node multi-correlation daily joint charging scenario set.

Step 2.1. Multi-node multi-correlation daily joint charging scenario generation model

The gradient penalty Wasserstein generative adversarial network is constructed based on the original multi-node multi-correlation daily joint charging scene set. The gradient penalty Wasserstein generative adversarial network needs to determine the generator and discriminator. The generator learns the multi-node multi-correlation daily joint charging scene sample data probability Distribution, mining the potential spatiotemporal distribution of electric vehicle charging load in the distribution network space, the discriminator is responsible for supervising the quality of the multi-node multi-correlation daily joint charging scene sample data to ensure that the generated joint charging scene data and the historical joint charging scene data have a similar probability distribution .

When training the gradient penalty Wasserstein generative adversarial network, the discriminator and the generator alternately optimize the mutual game, and the model training is completed when the Nash equilibrium point is reached. The Wasserstein distance is used instead of the JS divergence to describe the difference between the generated data and the real data distribution. The Wasserstein distance is applied to Generative Adversarial Networks and is expressed as:

为期望；

为生成样本；

表示由判别器获得的结果；z 为生成器输入的噪声向量，且概率分布为Z～P_Z(z)；x为原始多节点多相关日联合充电场景集中样本特征向量，且X～P_X(x)；in,

for expectation;

to generate samples;

represents the result obtained by the discriminator; z is the noise vector input by the generator, and the probability distribution is Z ~ P _Z (z); x is the sample feature vector of the original multi-node multi-correlation daily joint charging scene set, and X ~ P _X (x);

In order to solve the problems of difficult training and mode collapse of the WGAN model, a gradient penalty term is added to the discriminator loss function, and the Lipschitz limit is optimized to ensure that the gradient penalty is within the limited threshold. The objective function of the multi-node multi-correlation daily joint charging scene generation model is:

为电动汽车充电负荷历史数据和生成多节点多相关日联合充电场景数据概率分布间线性采样值。where λ is the gradient penalty coefficient,

A linear sampling value between the historical data of electric vehicle charging load and the probability distribution of multi-node multi-correlation daily joint charging scenario data.

Step 2.2. Generate a multi-node multi-correlation daily joint charging scene set

Input the centralized data of the original multi-node multi-correlation day joint charging scenario into the multi-node multi-correlation day joint charging scenario generation model, and generate a large number of multi-node multi-correlation day joint charging with the same dimension and the original joint charging scene data with similar probability distribution but different time series distribution. The generated massive multi-node multi-correlation daily joint charging scenarios constitute the generated multi-node multi-correlation daily joint charging scene set.

In order to ensure the effectiveness of generating multi-node and multi-correlation daily joint charging scenarios, the quality of the generated multi-node and multi-correlation daily joint charging scenarios is evaluated.

1) Analysis of the probability distribution characteristics of data generated in joint charging scenarios

In order to evaluate the data generation capability of the joint charging scenario generation model, without considering the coupling relationship between the electric vehicle charging loads of each node in the distribution network space, the multi-node multi-correlation day joint charging scenario and the multi-node multi-correlation day joint charging scenario were generated respectively. Probability distribution characteristic analysis is carried out on all data of all nodes in the scene set collection. Using the probability density function and the empirical cumulative distribution function, the data probability distribution characteristics of the original multi-node multi-correlation day joint charging scene set and the multi-node multi-correlation day joint charging scene set are analyzed. Furthermore, three statistics of mean value, variance and maximum value are used to analyze the probabilistic and statistical characteristics of the multi-node multi-correlation daily joint charging scene set samples and the original multi-node multi-correlation daily joint charging scene set samples.

2) Spatial correlation analysis of charging load between nodes in multi-node multi-correlation daily combined charging scenario

The spatial correlation of charging loads between multiple nodes in a multi-node multi-correlation daily joint charging scenario must conform to the correlation law of historical joint charging scenarios. In order to evaluate the spatial correlation of charging loads among multi-nodes in the multi-node multi-correlation day joint charging scenario, the original multi-node multi-correlation day joint charging scenario set and the generated multi-node multi-correlation day joint charging scenario set are respectively collected for all the charging loads of each node. The data is reshaped into one row (that is, 32 rows of charging load data are obtained from the original multi-node multi-correlation daily joint charging scene set and the generated multi-node multi-correlation daily joint charging scene set respectively). After the data is reshaped, the correlation between the charging load data of each row is calculated separately. Thereby, the original multi-node multi-correlation daily joint charging scene set and the generated multi-node multi-correlation daily joint charging scene set and the multi-node charging load spatial correlation (ie, matrix row data correlation) are respectively obtained. To further quantify the analysis results, structural similarity and feature similarity evaluation indicators are used. From the perspective of the structure and characteristics of the spatial correlation between multi-node charging loads in joint charging scenarios, it is verified that the original multi-node multi-correlation daily joint charging scenario set and the generated multi-node multi-correlation daily joint charging scenario set have similar spatial correlations between multi-node charging loads. sex. When the evaluation index value of structural similarity and feature similarity is larger, it indicates that the degree of similarity of spatial correlation is higher.

3) Analysis of time series distribution characteristics of generated joint charging scenarios

In order to analyze the time series distribution characteristics of the generated joint charging scenarios, boxplots are used to analyze the original multi-node multi-correlation daily joint charging scene set and the multi-node multi-correlation daily joint charging scene set based on the gradient penalty Wasserstein generative adversarial network. Charging load data. It is verified that the charging load data of each node in the multi-node multi-correlation daily combined charging scenario set is in line with the time-series distribution characteristics of the charging load data of each node in the original multi-node multi-correlated daily combined charging scenario set.

According to the analysis of the above three characteristics, the validity of the generated multi-node multi-correlation daily joint charging scenario can be obtained.

Step 3. Analyze the correlation between the multi-node multi-correlation day joint charging scenario and the extremely strong correlation day joint charging scenario used for prediction, and select the one with a high degree of correlation as the to-be-predicted day-related joint scenario set; the specific process is as follows:

Calculate the combined charging scenario of the extremely strongly correlated day on the day to be predicted and the j-th scenario weighted 2-D correlation coefficient R _j in the set of multi-node multi-correlated day combined charging scenarios, and the expression is:

表示待预测日极强相关日联合充电场景和生成多节点多相关日联合充电场景集D-i与生成多节点多相关日联合充电场景集中第j个场景的2-D相关系数；where:

Represents the 2-D correlation coefficient between the extremely strongly correlated day joint charging scenario on the day to be predicted, the generated multi-node multi-correlation day joint charging scenario set Di and the j-th scenario in the multi-node multi-correlation day joint charging scenario set;

The obtained correlation coefficients select the top M joint charging scenarios in order from high to low to form a set of relevant joint scenarios on the day to be predicted.

其中n表示节点编号，且n∈[1,32]，Step 4. Obtain the multi-node charging load interval prediction results and deterministic prediction results according to the data on the last day of the relevant joint scene set on the day to be predicted. The load is taken as the charging load of each node on the day to be forecasted as

where n represents the node number, and n∈[1,32],

Equation (6) is used to calculate the charging load interval prediction results and deterministic prediction results of each node:

分别表示t时刻节点n充电负荷区间预测结果的上下限；

表示t时刻节点n的电动汽车充电负荷确定性预测结果。in

respectively represent the upper and lower limits of the prediction result of the charging load interval of node n at time t;

Represents the deterministic prediction result of electric vehicle charging load at node n at time t.

Example

The electric vehicle charging load data required for the study was collected from the measured data in a certain area in 2019. The data is the daily electric vehicle charging load data record in this area with a sampling interval of 1 hour.

(1) Construction of multi-node multi-correlation daily joint charging scenario set

Figure 1 shows a schematic diagram of the correlation analysis between the charging load of the to-be-forecasted day and the multi-node multi-node combined charging scenario. It can be seen from Figure 1 that there are five extremely strongly correlated historical day charging scenarios on the to-be-predicted day. 1 day before, 2 days before, 6 days before, 7 days before, 8 days before. A multi-node multi-correlation day joint charging scenario is constructed by arranging the multi-node multi-node joint charging scenarios of the five extremely strong correlation days and the days to be predicted in time series. The scene structure is shown in Figure 2. Based on the measured data of electric vehicle charging load, the first joint charging scenario was constructed from January 9, 2019. Each multi-node multi-correlation daily joint charging scenario contains 32 nodes 6-day electric vehicle charging load data. Then the original multi-node multi-correlation daily joint charging scene set contains 357 joint charging scenarios. When carrying out the multi-node charging load interval prediction, the scene data of the original multi-node multi-correlated daily joint charging scene set are divided into training set and test set according to the ratio of 4:1.

(2) Multi-node multi-correlation daily joint charging scenario generation

Figure 3 shows the generation process of multi-node and multi-correlation daily joint charging scenarios, and obtains a large number of multi-node multi-correlation daily joint charging scenarios of the same dimension with similar probability distribution to the original joint charging scene data but with different timing distributions. Among them, when the scale of the multi-node multi-correlation daily joint charging scene set is set to 5000 groups through experiments, the effect of covering the historical joint charging scene is the best.

Figure 4 shows the analysis results of the probability distribution characteristics of all data of all nodes. It can be seen from the analysis results that, compared with the Wasserstein generative adversarial network, the gradient penalized Wasserstein generative adversarial network generates the PDF, ECDF curve and historical joint charging of the joint charging scene set The scene set corresponds to the curve height fitting. It shows that on the whole, the generated joint charging scene data has a probability distribution characteristic similar to the historical joint charging scene data. After the evaluation of the data probability distribution characteristics, the three statistics of mean, variance and maximum value shown in Table 1 are further used to analyze the original multi-node multi-correlation daily joint charging scenario set samples and the generated multi-node multi-correlation daily joint charging scenarios. Probabilistic and statistical properties of a set of samples. x represents the multi-node multi-correlation daily joint charging scenario sample, the average value reflects the distribution characteristics of the joint charging scenario charging load; the variance represents the charging load dispersion degree of the joint charging scenario sample; the maximum value represents the maximum charging load of the electric vehicle in the joint charging scenario sample.

Table 1

As shown in Figure 5, three statistical analysis results of the combined charging scene sample. From the analysis results shown in the figure, it can be seen that compared with the Wasserstein generative adversarial network, the multi-node and multi-correlation daily joint charging scene samples generated by the gradient penalized Wasserstein generative adversarial network are closer to the historical joint charging scene sample distribution, which is effective. It covers the scattered points of historical joint charging scene samples, and contains potential charging loads that conform to the charging behavior rules of electric vehicle users. It proves that the multi-node and multi-correlation daily joint charging scene set generated by the gradient penalty Wasserstein generative adversarial network can more effectively reflect the distribution network space. The fluctuation law of charging load of internal electric vehicles reflects the effectiveness of generating a centralized sample of multi-node and multi-correlation daily joint charging scenarios.

As shown in Figure 6, the spatial correlation analysis results between the original multi-node multi-correlation daily joint charging scenario set and the multi-node multi-correlation daily joint charging scenario set multi-node charging load. In order to further quantify the visual analysis results in Figure 6, the structural similarity and feature similarity evaluation indicators shown in Table 2 are used to verify the original multi-node multi-correlation daily joint charging scene set and the generated multi-node multi-correlation daily joint charging scene set multi-node. There is a similar spatial correlation for charging loads. From the analysis results shown in Figure 6 and Table 2, it can be seen that compared with the Wasserstein generative adversarial network model, the structure similarity and characteristics of the overall data of the multi-node multi-correlation daily joint charging scene set generated by the gradient penalty Wasserstein generative adversarial network are similar. The degree index values are 0.94 and 0.97, respectively, which effectively describe the spatial correlation between multi-node charging loads in historical joint charging scenarios. Demonstrate the effectiveness of joint joint scenes generated by gradient penalized Wasserstein generative adversarial nets.

Table 2

Fig. 7 shows the time-series distribution analysis results of the charging load of each node in the original multi-node multi-correlation daily joint charging scenario set and the generated multi-node multi-correlation daily joint charging scenario set. It can be seen from Fig. 7 that the change rule of the charging load of each node in the generated joint charging scenario conforms to the time-series distribution characteristics of the charging load in the historical joint charging scenario. At the same time, compared with the historical joint charging scenario, the upper edge of the charging load data distribution of each node in the generated joint charging scenario is higher, the lower edge is lower, and the data has more outliers. It shows that the multi-node multi-correlation daily joint charging scenario set can not only cover all historical charging scenarios of each node, but also contain potential charging loads that conform to the charging behavior rules of electric vehicle users.

(3) Multi-node electric vehicle charging load interval prediction

Figure 8 shows the process of forecasting the charging load interval of a multi-node electric vehicle based on a multi-node multi-correlation daily joint charging scenario set.

Figure 9 shows the statistical results of evaluation indicators for each prediction method. The input feature set of the GPR model in the comparative experiment includes all the charging loads on the 8th, 7th, 6th, and 2nd days before the prediction date of the node, and the charging load at time t 1 day before the prediction date, a total of 97-dimensional features, and the ratio of the training set to the test set. Consistent with the present invention, and run in the MatlabR2018b environment, the confidence is set to 95%, and the kernel function is an ardexponential function. The minimum values of the PICP indicators obtained by the GPR method and the method of the present invention are respectively 77.9% and 90.4%, and the minimum values of the average values are 80.9% and 92.7%, respectively. The maximum values of the PINAW indicators are 36.7% and 32.1%, respectively, and the maximum values of the average values are 34.2% and 29.2%, respectively. It can be seen from the analysis that the PICP index value of the prediction result of the method of the present invention is larger. It can be seen from this that the method of the present invention can make the electric vehicle charging load prediction interval of each node more reliable and has a higher degree of sophistication. The maximum values of MAPE index of the deterministic prediction results of the two methods are GPR22.7%, the method of the present invention is 17.7%, and the maximum value of the average value is GPR19.7%, and the method of the present invention is 15.8%, respectively. By comparison, it can be seen that the deterministic prediction result of the method of the present invention has a smaller MAPE index value and a higher accuracy of the prediction result.

FIG. 10 shows the prediction result, and FIG. 11 shows the evaluation index. In order to display the prediction effect of each method in the interval, the prediction results of node 3, node 26 and node 31 with relatively poor prediction effect are selected for display. By analyzing the prediction results of different nodes and different date types in Fig. 10 and Fig. When the spatio-temporal distribution of electric vehicle charging load in the network space changes drastically, the GPR method has limited ability to track the change of charging load in the prediction interval, resulting in poor interval prediction effect. And compared with the GPR method that predicts the charging load of electric vehicles at each node separately, the method of the present invention considering the spatial correlation of the charging load between nodes has a higher prediction interval PICP value, higher reliability of the prediction interval; and a lower PINAW value , the prediction interval is closer to the actual charging load. At the same time, the deterministic prediction result of the electric vehicle charging load at each node has a smaller MAPE value and a higher prediction accuracy. This proves that the present invention considers the charging load of the electric vehicle at each node. Validity of load interval forecasting methods.

Through the above method, the present invention discloses a multi-node electric vehicle charging load joint confrontation generation interval prediction method, and the interval prediction considering the spatial correlation between multi-node EV charging loads has better prediction indicators and can more effectively predict the distribution network space. The time-space distribution of the internal EV charging load is more conducive to improving the stability and economy of the operation of the distribution network.

Claims (6)

1. The multi-node electric vehicle charging load joint countermeasure generation interval prediction method is characterized by being implemented according to the following steps:

step 1, mapping historical charging load data of an electric vehicle into an IEEE33 node power distribution network system, and constructing an original multi-node multi-correlation day combined charging scene set based on the historical charging load data of the electric vehicle;

step 2, constructing a multi-node multi-correlation day combined charging scene generation model through the original multi-node multi-correlation day combined charging scene set, and obtaining a multi-node multi-correlation day combined charging scene set through the multi-node multi-correlation day combined charging scene generation model;

step 3, analyzing and generating the correlation between the multi-node multi-correlation-day combined charging scene and the extremely strong correlation historical-day charging scene used for prediction, and selecting the high correlation degree as a day-to-be-predicted correlation combined scene set;

and 4, obtaining a multi-node charging load interval prediction result and a certainty prediction result according to the last day data of the day-related combined scene set to be predicted.

2. The multi-node electric vehicle charging load joint countermeasure generation interval prediction method according to claim 1, wherein the specific process of the step 1 is as follows: mapping the historical charging load data of the electric automobile into an IEEE33 node power distribution network system, numbering space nodes of a charging scene in the IEEE33 node power distribution network system by 1, … and 32 to obtain the historical charging load data of the electric automobile corresponding to each node, and defining a multi-node combined charging scene to be predicted on day to be expressed as a matrix D_ntThe multi-node combined charging scene of the historical day is expressed as a matrix (D-i)_ntD is calculated according to all historical data of the charging load of the electric automobile_ntAnd (D-i)_ntTime-space correlation between charging loads in two matrices

The calculation formula is as follows:

in equation (1), n denotes a spatial node number in the joint charging scenario, t denotes a spatial charging load sampling time point in the joint charging scenario, and ranges of n ═ 1,2, …,32, and t ═ 1,2, …,24, respectively; and is

The multi-node combined charging scene of the day to be predicted is strongly related to the multi-node combined charging scene of the historical day, and the historical day which is strongly related to the multi-node combined charging scene of the day to be predicted is taken as a strongly related day;

and obtaining a multi-node combined charging scene of the extremely strong correlation day of the day to be predicted according to correlation analysis, and constructing an original multi-node multi-correlation day combined charging scene set by arranging the multi-node combined charging scene of the extremely strong correlation day and the multi-node combined charging scene of the day to be predicted according to a time sequence.

3. The multi-node electric vehicle charging load joint countermeasure generation interval prediction method according to claim 1, wherein the step 2 specifically comprises the following steps:

step 2.1, constructing a gradient penalty Wasserstein generation countermeasure network based on an original multi-node multi-correlation day combined charging scene set, optimizing a generator and a discriminator in the countermeasure network, and taking the generated network after optimization as a multi-node multi-correlation day combined charging scene generation model;

and 2.2, inputting the concentrated data of the original multi-node multi-correlation-day combined charging scene into a multi-node multi-correlation-day combined charging scene generation model, generating a same-dimension multi-node multi-correlation-day combined charging scene with the mass similar to the data of the original combined charging scene in probability distribution but different in time sequence distribution, and forming a multi-node multi-correlation-day combined charging scene set by the generated mass multi-node multi-correlation-day combined charging scenes.

4. The multi-node electric vehicle charging load joint countermeasure generation interval prediction method according to claim 3, wherein the specific process of optimizing the generator and the discriminator in the countermeasure network in step 2.1 is as follows:

using the Wasserstein distance instead of the JS divergence to describe the difference between the generated data and the true data distribution, applying the Wasserstein distance to the generation of the countermeasure network, expressed as:

wherein,

as desired;

to generate a sample;

representing the result obtained by the discriminator; z is the noise vector input by the generator, and the probability distribution is Z-P_Z(z); x is the characteristic vector of the original multi-node multi-correlation day combined charging scene concentrated sample, and X-P_X(x)；

Adding a gradient penalty item in a discriminator loss function, wherein an objective function of a multi-node multi-correlation day combined charging scene generation model is as follows:

in the formula, lambda is a gradient penalty coefficient,

and charging load historical data for the electric automobile and generating a multi-node multi-correlation day combined charging scene data probability distribution-based linear sampling value.

5. The multi-node electric vehicle charging load joint countermeasure generation interval prediction method according to claim 1, wherein the specific process of step 3 is as follows:

calculating a historical day multi-node combined charging scene extremely strongly related to a multi-node combined charging scene on a day to be measured and generating a weighted 2-D correlation coefficient R of the jth scene in a multi-node multi-related day combined charging scene set_jThe expression is:

in the formula:

2-D correlation coefficients of a history day multi-node combined charging scene D-i which is extremely strongly related to a multi-node combined charging scene on a day to be measured and a jth scene in a multi-node multi-related day combined charging scene set are represented;

and sequentially selecting the first M combined charging scenes from high to low to obtain the correlation coefficient to form a day-to-day associated combined scene set to be predicted.

6. The multi-node electric vehicle charging load joint countermeasure generation interval prediction method according to claim 1, characterized in that the specific process of step 4 is as follows: according to the charging load of each node in the last day scene in the day-related combined scene set to be predicted, the charging load of each node in the day to be predicted is

Where n represents the node number and n ∈ [1,32 ]]，

And (3) calculating the charging load interval prediction result and the certainty prediction result of each node by adopting an equation (6):

wherein

Respectively representing the upper limit and the lower limit of the prediction result of the node n charging load interval at the time t;

and (4) representing the electric vehicle charging load certainty prediction result of the node n at the time t.

Images