Electric vehicle (EV) charging load is greatly affected by many traffic factors, such as road congestion. Accurate ultra short-term load forecasting (STLF) results for regional EV charging load are important to the scheduling plan of regional charging load, which can be derived to realize the optimal vehicle to grid benefit. In this paper, a regional-level EV ultra STLF method is proposed and discussed. The usage degree of all charging piles is firstly defined by us based on the usage frequency of charging piles, and then constructed by our collected EV charging transaction data in the field. Secondly, these usage degrees are combined with historical charging load values to form the input matrix for the deep learning based load prediction model. Finally, long short-term memory (LSTM) neural network is used to construct EV charging load forecasting model, which is trained by the formed input matrix. The comparison experiment proves that the proposed method in this paper has higher prediction accuracy compared with traditional methods. In addition, load characteristic index for the fluctuation of adjacent day load and adjacent week load are proposed by us, and these fluctuation factors are used to assess the prediction accuracy of the EV charging load, together with the mean absolute percentage error (MAPE).
Environment-friendly EVs are being developed vigorously to deal with environmental challenges as the fast growth of global economy [
Traditional electric load forecasting methods are mainly used to predict the system-level loads based on the periodic characteristics and stochastic of the predicted load [
The focus of the charging load forecasting area is mainly on improving the accuracy of load forecasting. EV charging loads have similar characteristics to traditional loads, so the load forecasting methods applied to traditional loads are also widely used for EV charging loads. The charging load forecasting methods for EVs are also mainly divided into data-driven and model-driven methods, but load characteristics and influencing factors of EV charging loads have many specialties. The traditional electricity load generally shows a peak during the day and a low at night. As for the EV charging load, the EV charging load is less regular due to the spatial and temporal randomness of EV charging [
Due to the lack of real data for EV charging load, regional EV load forecasting methods mainly focus on model-driven methods. Specifically, combined with the charging characteristics of EVs, the spatial and temporal distribution characteristics of EVs are simulated to establish a corresponding load forecasting model [
Among the data-driven short-term load forecasting (STLF) methods, there are usually two stages, which include the data preprocessing stage and the load forecasting stage. In reference [
In fact, each charging pile in the region has its own unique power consumption characteristics during the day. During the day, the peak and trough charging periods of charging pile are different, and the frequency of use is also very different. The average load of the charging pile at the same time point on different days reflects the frequency of use of the charging pile at this time point, and on the contrary, the higher the frequency of use of the charging pile, the probability that the charging pile will be used at the same time point in the future will also increase, so there is a large probability of charging the charging pile at the same time point in the future. The average load at the same point in time on different days of the charging pile load forecasting. To improve the accuracy of the ultra short-term regional EV charging load forecasting results, a hybrid model and data-driven forecasting method is proposed and realized in this paper. The charging pile usage degree is defined and obtained by analyzing the daily usage curve of each charging pile. The charging pile usage degree would be used as the indicator for the traffic condition. Firstly, the abnormal EV charging daily load is removed by the Density-Based Spatial Clustering of Application with Noise (DBSCAN) clustering algorithm. Secondly, the load value of the charging pile is encoded to obtain the charging pile usage degree at each moment. Thirdly, the long-short term memory (LSTM) load forecasting model is constructed to deal with the charging pile usage degree and the historical charging load data. Fourthly, a novel ultra short-term forecasting method of EV charging loads is realized and verified by the field data. Finally, the load forecasting accuracy is directly influenced by the load fluctuation, this paper presents a load characteristic index for the fluctuation of adjacent day load and adjacent week load, which can be used to determine the mean absolute percentage error (MAPE) size of the forecasting result before forecasting and description of load fluctuation characteristics.
In
In
In
The charging conditions of EVs applied by our method is shown in
In fact, each charging pile in the region has its own unique power consumption characteristics during the day. During the day, the peak and trough charging periods of charging pile are different, and the frequency of use is also very different. The average load of the charging pile at the same time point on different days reflects the frequency of use of the charging pile at this time point, and on the contrary, the higher the frequency of use of the charging pile, the probability that the charging pile will be used at the same time point in the future will also increase, so there is a large probability of charging the charging pile at the same time point in the future. The average load at the same point in time on different days of the charging pile load forecasting. However, due to the complexity of the load, directly inputting the average load at the same point in time on different days into the data-driven load forecasting model will increase the complexity of the network, which is prone to overfitting, and instead does not yield better forecasting results. In this paper, the average load of charging piles at the same time point on different days is further coded to reduce the complexity of the data. This paper integrates the charging load mechanism of charging piles in a data-driven approach, and proposes a load forecasting method based on the usage degree of charging piles, which integrates the advantages of data-driven and model-driven and improves the forecasting accuracy.
The key influencing factor of EV charging load is the EV traffic information. The concept of the average daily usage degree of charging piles,
The average load of the EV charging pile in each small period can be used to reflect the usage degree of the charging pile at each moment. It is determined by the arrival times of EVs at charging stations at each moment, which indirectly reflects the nearby traffic condition. Therefore, the important information of charging pile usage degree is extracted as a key input variable for charging load prediction.
The steps of regional charging pile usage degree extraction work are given as follows:
There is certain regularity in the daily EV charging load at the regional level. The abnormal daily load is removed by the DBSCAN clustering algorithm. If cluster analysis results show that
DBSCAN is a density-based clustering technique, and its advantage is that it does not require a preset number of clusters and can be used to identify outliers from a set of daily load curves [
Calculate the average load values
There are
The load of each charging pile varies in a wide range, and it cannot be directly used as the input parameter of the deep learning-based forecasting model. These load data would be encoded as a new time series. The maximum element in the obtained matrix
Therefore, the average daily load distribution
The encoded value curve of average usage degree of the charging piles in the whole region in workdays can be obtained by the encoded values of each charging pile, which is calculated by
The encoded value curve of average usage degree of the charging piles in the whole region in holidays can be obtained by repeating the steps from 1 to 5. And then the encoded value curve of average usage degree of the charging piles in the whole region for the historical data can be obtained as a complete time series
LSTM is a kind of recurrent neural network (RNN) and is realized based on the general recurrent neural network. The internal structure of LSTM is improved so that it can maintain the long-range dependence of time series and effectively avoid gradient disappearance and gradient explosion. In this paper, LSTM is selected to construct the EV charging load forecasting model.
The historical EV charging load data
The EV charging load data and charging pile usage degrees need to be normalized to simplify the computation during training and speed up the network convergence before feeding the dataset into the model. The normalization formula is given as follows:
The input variables are fed into the LSTM for training after normalization. The constructed load prediction LSTM network consists of 1 input layer, 2 hidden layers, and 1 output layer. The input layer contains 2 cells and each hidden layer contains 20 memory cells.
The LSTM learning rate is used to control the learning progress of the model. A small learning rate will increase the learning time of the network, and a large learning rate may make the network to have difficulties in finding the optimal value and the network cannot converge. In this paper, the learning rate is selected empirically.
In the field, the load forecasting accuracy at the system level is obviously higher than that of power distribution networks. Indeed, the load forecasting accuracy is directly influenced by the load fluctuation, and strong periodicity would improve the prediction accuracy. For different type loads, the fluctuant components proportion is different. In our paper, regional EV charging load has high randomness and strong volatility. In the same area, the forecasting error of EV charging load is generally bigger than that of regular distribution electric load. Therefore, when traditional load forecasting methods are used to predict EV charging load, the MAPE would change. In this paper, in addition to MAPE being used to assess load forecasting errors, some metrics to characterize the adjacent day and adjacent week volatility of the historical data of the load to be forecast are proposed. It is convenient to confirm whether the load data can achieve satisfactory results when they are input into the forecasting model before the forecasting.
Load characteristic indexes include load fluctuation rate, load rate, peak-to-valley difference, etc. These load characteristic indexes describe the characteristics of the intra-day load or the load characteristics of the total historical data amount in the load history
This paper presents a load characteristic index for the fluctuation of adjacent day load and adjacent week load, considering the difference between adjacent day load and adjacent week load. Accumulate the ratio of the load difference between different adjacent days and adjacent weeks at the same time to the total daily load and total weekly load, respectively. Therefore, it is convenient to confirm whether the load data can achieve satisfactory results when they are input into the forecasting model before the forecasting. The adjacent day load fluctuation rate and the adjacent week load fluctuation rate are defined as
In summary,
MAPE as the cost function in this paper, the training process of the load forecasting model reduces the MAPE of the forecasting results. The specific formulas for RMSE and MAE are listed below:
The overall forecasting process is shown in
The transaction data of all EV charging stations in some region in the field is collected by us in Hubei province. The EV charging stations can be divided into two categories based on their operating locations, including urban charging stations and highways circumjacent charging stations. In addition, the EV charging transaction data includes the start time of the charging process, power consumption of the charging process, the charging cost, the charging pile location, and the end time of the charging process.
Since the raw data is recorded in transaction order format, it is necessary to perform the data preprocessing work to obtain the charging load in time series firstly.
Specific preprocessing efforts for the EV charging transaction data include the following steps:
Reorder each transaction order into a column by the start and end time, and divide the transaction power consumption by the total charging time to express the average charging power for that period.
The start and end times of these charging transaction orders are relatively random and the charging time is inconsistent, which is not conducive to conducting cluster and prediction studies. Hence interpolation process should be done to form the charging load in time series. Due to many charging times lasting only about two minutes, the expanded time scale is 1 min. Power data is interpolated with zero power supplement between two order times, and the charging power in one order is regarded as the same value.
We selected two-day daily load curves of urban charging stations and highways circumjacent charging stations in a region, and it is shown in
The regional EV charging load can be obtained after the preprocessing steps, and it is shown in
There is a certain difference in the regional EV load between the workday and the holiday. Overall, the overall change of 24 h between different days has a strong regularity, which also shows that the load elements corresponding to the time points of the day are encoded as an input variable into the data-driven load forecasting model is reasonable. It also shows that it is necessary to take the day as the unit to extract the daily distribution of the use degree of EV charging piles in the region.
The EV load data can be divided into different regions according to the geographical location information. In this paper, the data of a small and a large region are used for analysis and research. The small filed contains 10 charging stations, and the large region contains 69 charging stations. Among them, the data of January in a small region are used as the basic data of simulation, and other data are used for comparative analysis and research.
The EV load box plot of holiday and workday in small region and large region is shown in
The DBSCAN clustering algorithm is used to analyze the total charging load of EVs in January. The clustering results are as shown in
The load data of January 05 and 06 are removed from the load data of 88 EV charging piles in a certain area. And then the daily average load value is calculated in these regular days. The number of charging piles within the same range of charging load power is counted to identify the proper encoded value of the charging pile usage degree. The charging load power is firstly divided into several ranges with the interval of 0.05 kW, and then the boundary values of these ranges are adjusted by the statistical charging load power. The final divided ranges and the corresponding encoded values for the charging piles are given in
Load/kW | Number | Encoded value |
---|---|---|
0–0.049 | 2873 | 1 |
0.050–0.149 | 333 | 2.5 |
0.150–0.249 | 191 | 4.5 |
0.250–0.449 | 156 | 7.5 |
0.450–0.949 | 209 | 14.5 |
0.950–1.699 | 222 | 27 |
1.700–7.699 | 240 | 94.5 |
The specific method to obtain the usage degree of regional charging piles is in
The load value of different intervals is encoded by the multiple of the median value of the interval, which indicates the daily use of different charging piles.
The abnormal load in the daily load clustering results is removed from the daily load in January. Considering the different charging modes of holidays and working days, the daily average values of the charging pile load are calculated and plotted for holidays and weekdays. The sum of the encoded values of all charging piles in the selected region is obtained and used in the regional charging load forecasting. The results are shown in
It can be seen that the usage degree of regional daily charging piles in January is expressed as that the charging peak time is around 5 pm, and the usage degree of charging piles on workdays is significantly higher than that on holidays.
The regional average daily charging pile usage degree obtained above is replaced repeatedly according to the distribution of holidays and workdays to obtain a complete regional charging pile usage degree time series, which is consistent with the length of the regional charging load sequence in January. Then correlation analysis and load forecasting are studied.
The correlation between regional charging pile usage degree and regional EV charging load in January was 0.539. The Pearson correlation coefficient between 0.5 and 0.8 indicates a moderate correlation. There is a certain correlation between the extracted charging pile usage degree in January and the charging load of EVs. Therefore, it is feasible to input the charging pile usage degree as an input variable into the load forecasting model.
The historical load data in January were selected for simulation analysis. The first 29 days in the historical data is used in the data training model, the last two days of data is used for forecasting validation. This paper sets the learning rate as 0.01 according to historical experience. The forecasting results were obtained after 150 iterations. Traditional method for LSTM model with input load data only, the constructed load prediction LSTM network consists of 1 input layer, 2 hidden layers, and 1 output layer. The input layer contains 1 cell and each hidden layer contains 20 memory cells. This paper sets the learning rate as 0.01 according to historical experience. The forecasting results were obtained after 150 iterations.
For the LSTM-based forecasting model, the load forecasting results of multiple runs will be different. The reason is that the weights or parameters of the training layer in the artificial intelligence model are randomly initialized. Given the randomness of the forecasting results, this paper selects 30 times to run the LSTM-based EV charging load prediction model. MAPE for the prediction results is shown in
From the above analysis results, it can be seen that MAPE of the prediction results by our proposed method is smaller than those by the traditional method twenty-four times, accounting for four-fifths of the total operation times. In most cases, the forecasting effect of our proposed method is better than the traditional method.
The average value of the forecasting load curve of the 30 operation models on January 30 and 31 is shown in
To further verify the validity of the method proposed in this paper, the same method was applied to the regional EV load data for July and October. MAPE for calculating the average of predicted results of 30 times runs in January, July, and October. The results are given in
January | July | October | ||
---|---|---|---|---|
Traditional method | MAPE | 33.1% | 31.6% | 35.3% |
RMSE | 10.93 | 9.84 | 8.41 | |
MAE | 8.80 | 7.95 | 6.31 | |
BP | MAPE | 42.5% | 36.6% | 40.8% |
RMSE | 9.96 | 10.02 | 7.15 | |
MAE | 7.82 | 7.65 | 5.72 | |
SVR | MAPE | 33.2% | 34.5% | 31.0% |
RMSE | 10.93 | 9.79 | 7.24 | |
MAE | 8.26 | 7.92 | 5.75 | |
Our proposed method | MAPE | 28.9% | 25.3% | 29.9% |
RMSE | 9.83 | 8.40 | 8.07 | |
MAE | 7.36 | 6.64 | 6.10 |
The prediction results and MAPE for the prediction results in July and October are as shown follow.
The MAPE for the prediction results in July and October are as shown in
The average value of the forecasting load curve of the 30 operation models in July and October is shown in
For the charging load data of EVs in January, July, and October, the MAPE of the forecasting results of our proposed method is nearly 5% higher than that of the traditional method. The MAPE of the BP forecasting method forecasting results for the small areas in January, July and October are on average about 11% higher than the MAPE of the load forecasting method prediction results in this paper. The MAPE of the forecasting results of the SVR forecasting method for small areas in January, July and October is on average about 5% higher than the MAPE of the forecasting results of the load forecasting method in this paper. The RMSE of the forecasting results of the traditional load forecasting methods for small areas in January, July and October are on average about 1 higher than the RMSE of the forecasting results of the load forecasting methods in this paper. The MAE of the traditional load forecasting method prediction results for small areas in January, July and October are on average about 1.3 higher than the MAE of the load forecasting method forecasting results in this paper. which fully proves the effectiveness of our proposed method.
The data used above is the total charging load of 10 charging stations in the small region, the maximum charging load shall not exceed 100 kW. To further verify the validity of the method proposed in this paper, the same method was applied to the regional EV load data for the total charging load of 69 charging stations in the big region.
The prediction results and MAPE for the prediction results in January, July and October in the large region are as shown in
It can be seen that the MAPE of the prediction results by our proposed method is smaller than those by the traditional method every time in July and October, and there are only three exceptions in January. In most cases, the forecasting effect of our proposed method is better than the traditional method.
The average value of the forecasting load curve of the 30 operation models in January, July and October is shown in
MAPE for calculating the average of predicted results of 30 times runs in January, July, and October. The results are given in
January | July | October | ||
---|---|---|---|---|
Traditional method | MAPE | 16.2% | 14.3% | 13.1% |
RMSE | 17.34 | 21.96 | 16.16 | |
MAE | 12.80 | 16.70 | 12.47 | |
BP | MAPE | 16.4% | 14.4% | 14.7% |
RMSE | 14.87 | 19.71 | 15.94 | |
MAE | 11.67 | 16.10 | 12.69 | |
SVR | MAPE | 14.7% | 14.1% | 12.6% |
RMSE | 14.97 | 20.56 | 14.28 | |
MAE | 11.21 | 15.82 | 11.24 | |
Our proposed method | MAPE | 14.4% | 12.0% | 11.3% |
RMSE | 14.35 | 18.17 | 13.41 | |
MAE | 11.09 | 14.39 | 10.45 |
For the charging load data of EVs in January, July, and October, the MAPE of the forecasting results of our proposed method is nearly 2% higher than that of the traditional method. The MAPE of the BP load forecasting method prediction results for large areas in January, July and October are on average about 2.6% higher than the MAPE of the load forecasting method forecasting results in this paper. The MAPE of the forecasting results of the SVR load forecasting method for large regions in January, July and October is on average about 1.2% higher than the MAPE of the forecasting results of the load forecasting method in this paper. The RMSE of the forecasting results of the traditional load forecasting method for large areas in January, July and October are on average about 3 higher than the RMSE of the forecasting results of the load forecasting method in this paper. The MAE of the traditional load forecasting method prediction results for large areas in January, July and October are on average about 1.8 higher than the MAE of the load forecasting method forecasting results in this paper. It can also prove that the forecasting accuracy of forecasting method in this paper is higher than the traditional load forecasting method. It can prove that the forecasting accuracy of forecasting method in this paper is higher than the traditional load forecasting method.
When the results in
Calculate the load characteristic indexes in
Evaluation indicators | January | July | October | |
---|---|---|---|---|
10 charging stations | Traditional method MAPE | 33.1% | 31.6% | 35.3% |
Our proposed method MAPE | 28.9% | 25.3% | 29.9% | |
14.164 | 9.207 | 10.352 | ||
1.163 | 0.830 | 1.079 | ||
0.696 | 0.423 | 0.504 | ||
69 charging stations | Traditional method MAPE | 16.2% | 14.3% | 13.1% |
Our proposed method MAPE | 14.4% | 12.0% | 11.3% | |
7.696 | 4.882 | 5.142 | ||
0.724 | 0.485 | 0.552 | ||
0.420 | 0.339 | 0.389 |
For the total load data of 10 charging stations in a small region,
In
The MAPE and load characteristic indexes of the forecasting results for different types as well as different scales of loads are shown in
Traditional method MAPE | ||||
---|---|---|---|---|
10 charging stations | 31.6% | 9.207 | 0.830 | 0.423 |
69 charging stations | 14.3% | 4.882 | 0.485 | 0.339 |
50 households | 15.6% | 6.034 | 0.631 | 0.320 |
100 households | 11.5% | 3.927 | 0.411 | 0.299 |
150 households | 9.1% | 3.444 | 0.357 | 0.314 |
200 households | 8.5% | 3.000 | 0.301 | 0.299 |
It can be found that for the same type of load, the larger the scales of the load, the smaller the MAPE of the forecasting result, and the smaller
The MAPE of load data forecasting results for different types of load data maintains a high positive correlation with
In summary, the ability of LSTM load forecasting model to obtain better forecasting results is highly correlated with the load data itself. Data with smaller
To improve the ultra short-term predication accuracy of EV charging load, EV charging pile usage degree is defined by us based on the usage frequency of charging piles. And then the EV charging pile usage degree data is merged with the historical charging load data as the input variables in the LSTM-based ultra short-term forecasting model.
Simulation results show that the larger the maximum power of the historical load data is, the higher the load forecasting accuracy is.
The simulation results show that when the charging load data
This work is supported by
The authors declare that they have no conflicts of interest to report regarding the present study.