CN114282654A - An EV charging load calculation method based on conditional generative adversarial network - Google Patents

Abstract

The invention discloses an EV charging load calculation method based on a conditional generative confrontation network. The present invention establishes an accurate EV load prediction model considering air temperature, traffic congestion and user wishes, and uses the prediction result as real data. The load influencing factors such as temperature, traffic congestion, and user willingness are input into the CGAN generation model as conditions and noises to obtain the prediction data, and then the prediction data and the real data are respectively input into the CGAN discriminant model, and through the game training of CGAN, the generation model can be Load Influencers generate predicted load data for conditions. The method based on CGAN proposed by the present invention can change the condition data to quickly obtain EV charging data after the network training is completed, lay the foundation for online real-time scheduling, and can also quickly predict the load under some unusual conditions, which serves as a reference effect.

Description

An EV charging load calculation method based on conditional generative adversarial network

technical field

The invention relates to an electric vehicle charging service technology, in particular to an EV charging load calculation method based on a conditional generation confrontation network.

Background technique

In the existing EV charging load research, some researchers have used BP neural network, long short-term memory (LSTM), etc. to predict the charging load and achieved good results. However, neural networks rely on a large amount of historical data, and EVs are in the early stages of development, and their charging load data are often difficult to obtain. Therefore, using neural networks and deep learning to directly predict the charging load in a data-driven manner is difficult to generalize at this stage. Many researchers have carried out load forecasting in a model-driven manner by modeling related influencing factors such as EV travel paths, EV power consumption, EV charging, and traffic congestion. Model-driven methods do not require historical EV charging data, but the model after integrating “vehicle-road-network-human” is usually complex and computationally slow. In view of the above problems, the present invention considers combining the data-driven method with the model-driven method, obtains accurate load data through the model-driven method, and uses it as the training data of the deep learning network, although the network training process requires The network that has been trained for a long time can quickly predict the load and lay the foundation for online real-time scheduling.

SUMMARY OF THE INVENTION

Aiming at the deficiencies of the prior art, the present invention discloses an EV charging load prediction method based on a conditional generative adversarial network (CGAN), which combines a data-driven method and a model-driven method. The present invention establishes an accurate EV load prediction model considering air temperature, traffic congestion and user's will, and uses the prediction result as real data. Load factors such as temperature, traffic congestion, and user willingness are input into the CGAN generative model as conditions and noise to obtain prediction data, and then the predicted data and the real data are respectively input into the CGAN discriminant model, and through the game training of CGAN, the generative model is Load Influencers generate predicted load data for conditions. The present invention assumes that necessary information such as air temperature data, road congestion coefficient, user travel time, user travel destination, user destination stay time and the like can be obtained. The method is specifically implemented according to the following steps:

Step 1. Obtain the user's daily charging habits and preferences according to the questionnaire; fit the probability distribution of the power when the user has charging needs and the power at the end of each charge, and fit the user's preference for the shortest distance, the lowest charging cost, and the least time spent. The probability density of the ordering of the three options;

Step 2. Build a user travel model

EVs are divided into two types of private car operating vehicles according to their nature of use; most private cars are idle most of the day, while operating vehicles are running most of the day. In addition, the travel destination of private cars is more single and fixed than that of operating vehicles. Considering the large operational differences between private cars and operational vehicles, travel chains and OD (Origin-Destination) probability matrices are used to describe the travel behaviors of private cars and socially-operated vehicles, respectively.

A user will go to one or more places for activities in a day, and the set of destination nodes of his travel is a travel chain.

D={d ₁ ,d ₂ ,...,d _n ,...} (1)

In the formula, D is the destination set corresponding to the travel chain; n is the destination serial number; d ₁ is the departure point of the user's travel; d _n is the stop point in the travel process. The path set corresponding to the travel chain can be expressed by equation (2):

P={p(d ₁ ,d ₂ ),p(d ₂ ,d ₃ ),...p(d _n-1 ,d _n ),...} (2)

In the formula, P represents the path set corresponding to the travel chain, and p(d _n-1 ,d _n ) represents the path from the n-1th destination to the nth destination.

The OD probability matrix in each time period area can be regarded as the probability distribution of the travel destination of the operating vehicle, and its specific expression is shown in formula (3).

Among them, G(i) represents the OD probability matrix in the i-th time period; r, w, and b represent the residential area, work area, and commercial area, respectively; _{gw, r} represent the probability of the user going from the work area to the residential area. In order to calculate the spatiotemporal distribution of EV battery power, necessary information such as travel date type, initial charge, first travel time, travel destination, and destination stay time are also required, which can be obtained in reference [1].

Step 3. Build an EV energy consumption model.

When the road is clear, the vehicle can travel at a constant speed, and it takes the time t _i,j for the vehicle to travel from node i to node j, as shown in formula (4):

Among them, v is the road limit driving speed, and l _{i, j} represent the distance from node i to node j; when the road is congested, the higher the degree of congestion, the slower the vehicle travels, and the travel time is corrected by introducing a time-consuming coefficient, as shown in the formula (5).

为修正后的时间；δ_i为耗时系数；其值和交通拥堵程度相关，可参考文献[2]中取值。in,

is the corrected time; δ _i is the time-consuming coefficient; its value is related to the degree of traffic congestion, and can be obtained in reference [2].

The energy consumption of EV is mainly composed of two parts: air conditioning load and power consumption. Among them, the air-conditioning load is mainly affected by temperature, and the power consumption is mainly affected by factors such as EV driving speed, vehicle weight, vehicle acceleration, etc. The EV energy consumption model is specifically shown in formula (6):

为根据式(3)求得的行驶时间，E为EV电池容量。Among them, e _i,j is the amount of electricity consumed by EV traveling from node i to node j; e _cf is the power consumption per unit distance, l _i,j is the distance between node i and node j; P _a is the EV air conditioner power;

is the travel time obtained from equation (3), and E is the EV battery capacity.

Step 4. Build an EV charging model.

The EV charging time is specifically expressed by equation (7):

Among them, soc _e is the power when the user finishes charging; soc _s is the power when the user starts charging; p is the charging power of the charging pile; η is the charging efficiency of the charging pile; _te is the extra stay time after the EV is fully charged. At the end of charging, the electric quantity is less than the maximum electric quantity, then t _e is 0.

Step 5. Build an EV user charging scheme selection model.

When the user needs to charge, it is assumed that the user can receive the three charging schemes recommended by the dispatch center, including the scheme with the shortest time spent by the user, the scheme with the charging station closest to the user, and the scheme with the lowest charging cost for the user. After receiving the recommendation, the user chooses one of the schemes to go to the charging station for charging, in which the user's willingness to choose the charging scheme is obtained in the form of a questionnaire.

Step 6. Model-driven EV charging load prediction

According to the user travel model in step 2, calculate the temporal and spatial changes of EV battery power, in which the energy consumption of EV during operation is obtained in step 3; every time the user arrives at a destination, determine whether it needs to be charged, and if so, according to the step 5Choose a charging scheme; if no charging is required, the user will go to the next destination after staying for a while. Finally, the number of vehicles charged at the required time is counted, and the sum of the charging power is the charging load at the corresponding time, as shown in formula (8).

Among them, load _i is the charging load at time i; n is the total number of EVs being charged at the current time; p _j is the charging power of the j-th EV.

By changing the influencing factor data such as temperature, traffic congestion factor, date type, etc., the EV charging load value under various conditions can be obtained and used as the "real" data in the subsequent CGAN.

Step 7. Build the CGAN generator model

Because LSTM has good time series information processing ability, it is selected to build the CGAN generator model. The generator consists of a deep LSTM layer and a fully connected layer. The deep LSTM layer has 4 hidden layers, each with 200 LSTM units, and the hidden layer uses the ReLU function as the activation function. The input of the model is the influence factor data and random noise. The influencing factor data is the same influencing factor data as in the “real” data acquisition process in step 6, including the date type (weekdays or holidays), the lowest temperature value of the day, the highest temperature value of the day, traffic congestion factor, and user-generated charging. The probability distribution of EV power when demanded, the probability distribution of EV power when the user finishes charging, and the probability density of the user's charging scheme selection ranking. The influence factor data needs to be normalized before input. Random noise is set as a random variable obeying a Gaussian distribution. The output data of the model is the predicted load data.

Step 8. Build the CGAN discriminator model

The discriminator consists of a deep LSTM layer and a fully connected layer. The deep LSTM layer has 4 hidden layers, each with 200 LSTM units, and the hidden layer uses the ReLU function as the activation function. The sigmoid activation function is used in the fully connected layer to make true and false judgments. The load data obtained based on the model-driven and the predicted load data generated by the generator are respectively integrated with the influencing factors into the discriminator, and the discriminator outputs the probability that the input data is the real data.

Step 9: The generator model and the discriminator model are used for game training, and the trained generator model is used for load prediction.

When the CGAN fully learns the relationship between the data and reaches a balance, the load data under different conditions can be obtained by adjusting the conditional data input to the generator.

Preferably, the user's daily charging habits and preferences are counted in the form of a questionnaire. First, the Cronbach's coefficient method was used to test the reliability of the returned questionnaires, and valid samples were selected. The specific calculation of the Cronbach coefficient is shown in formula (9).

为第i题调查结果方差；

为全部调查结果方差。Among them, α is the reliability coefficient, the larger the value, the higher the reliability of the questionnaire; K is the number of questions in the questionnaire;

is the variance of the survey result of item i;

is the variance of all survey results.

Preferably, the influence factor data needs to be normalized before input.

Specifically, the maximum-minimum normalization method is adopted, which is shown in formula (10).

Among them, X _norm is the normalized data; X is the current data; X _min is the minimum value in the data; X _max is the maximum value in the data.

The advantages and beneficial results that the method of the present invention has are:

1. The present invention establishes a more accurate model-driven EV charging load calculation model in consideration of temperature, traffic, and user wishes.

2. The method of combining model-based driving and data-based driving proposed by the present invention can effectively utilize the respective advantages of the two methods, and is suitable for the situation that the current EV charging load is difficult to obtain.

3. The CGAN-based method proposed by the present invention can change the condition data to quickly obtain EV charging data after the network training is completed, lay the foundation for online real-time scheduling, and can also quickly predict the load under some uncommon conditions. Must be a reference.

Description of drawings

Fig. 1 Flow chart of model-driven EV charging load prediction;

Figure 2 CGAN generator model structure diagram;

Fig. 3 Flow chart of EV charging load prediction based on CGAN;

Detailed ways

Step 1. Obtain the user's daily charging habits and preferences according to the questionnaire; fit the probability distribution of the power when the user has charging needs and the power at the end of each charge, and fit the user's preference for the shortest distance, the lowest charging cost, and the least time spent. The probability density of the ordering of the three options;

The user's daily charging habits and preferences are counted in the form of questionnaires. First, the Cronbach's coefficient method was used to test the reliability of the returned questionnaires, and valid samples were selected. The specific calculation of the Cronbach coefficient is shown in formula (1).

为第i题调查结果方差；

为全部调查结果方差。Among them, α is the reliability coefficient, the larger the value, the higher the reliability of the questionnaire; K is the number of questions in the questionnaire;

is the variance of the survey result of item i;

is the variance of all survey results.

Step 2. Build a user travel model

EVs are divided into two types of private car operating vehicles according to their nature of use; most private cars are idle most of the day, while operating vehicles are running most of the day. In addition, the travel destination of private cars is more single and fixed than that of operating vehicles. Considering the large operational differences between private cars and operational vehicles, travel chains and OD (Origin-Destination) probability matrices are used to describe the travel behaviors of private cars and socially-operated vehicles, respectively.

The user will go to one or more places for activities in one day, and the set of destination nodes of his travel is the travel chain.

D={d ₁ ,d ₂ ,...,d _n ,...} (2)

In the formula, D is the destination set corresponding to the travel chain; n is the destination serial number; d ₁ is the departure point of the user's travel; d _n is the stop point in the travel process. The path set corresponding to the travel chain can be expressed by equation (3):

P={p(d ₁ ,d ₂ ),p(d ₂ ,d ₃ ),...p(d _n-1 ,d _n ),...} (3)

In the formula, P represents the path set corresponding to the travel chain, and p(d _n-1 ,d _n ) represents the path from the n-1th destination to the nth destination.

The OD probability matrix in each time period area can be regarded as the probability distribution of the travel destination of the operating vehicle, and its specific expression is shown in formula (4).

where G(i) represents the OD probability matrix in the i-th time period; r, w, and b represent the residential area, work area, and commercial area, respectively; _{gw, r} represent the probability of the user going from the work area to the residential area. In order to calculate the spatiotemporal distribution of EV battery power, necessary information such as travel date type, initial power level, first travel time, travel destination, and destination stay time are also required, which can be obtained in reference [1].

Step 3. Build an EV energy consumption model.

When the road is clear, the vehicle can travel at a constant speed, and the time t _i,j for the vehicle to travel from node i to node j is shown in formula (5):

Among them, v is the road limit driving speed, and l _{i, j} represent the distance from node i to node j; when the road is congested, the higher the degree of congestion, the slower the vehicle travels, and the travel time is corrected by introducing a time-consuming coefficient, as shown in the formula (6).

为修正后的时间；δ_i为耗时系数；其值和交通拥堵程度相关，可参考文献[2]中取值。in,

is the corrected time; δ _i is the time-consuming coefficient; its value is related to the degree of traffic congestion, and can be obtained in reference [2].

The energy consumption of EV is mainly composed of two parts: air conditioning load and power consumption. Among them, the air-conditioning load is mainly affected by temperature, and the power consumption is mainly affected by factors such as EV driving speed, vehicle weight, vehicle acceleration, etc. The EV energy consumption model is specifically shown in formula (7):

为根据式(3)求得的行驶时间，E为EV电池容量。Among them, e _i,j is the amount of electricity consumed by EV traveling from node i to node j; e _cf is the power consumption per unit distance, l _i,j is the distance between node i and node j; P _a is the EV air conditioner power;

is the travel time obtained from equation (3), and E is the EV battery capacity.

Step 4. Build an EV charging model.

The EV charging time is specifically expressed by equation (8):

Among them, soc _e is the power when the user finishes charging; soc _s is the power when the user starts charging; p is the charging power of the charging pile; η is the charging efficiency of the charging pile; _te is the extra stay time after the EV is fully charged. At the end of charging, the electric quantity is less than the maximum electric quantity, then t _e is 0.

Step 5. Build an EV user charging scheme selection model.

When the user needs to charge, it is assumed that the user can receive the three charging schemes recommended by the dispatch center, including the scheme with the shortest time spent by the user, the scheme with the charging station closest to the user, and the scheme with the lowest charging cost for the user. After receiving the recommendation, the user chooses one of the schemes to go to the charging station for charging, in which the user's willingness to choose the charging scheme is obtained in the form of a questionnaire.

Step 6. Model-driven EV charging load prediction

The specific process of model-driven EV charging load prediction is shown in Figure 1. According to the user travel model in step 2, calculate the temporal and spatial changes of EV battery power, in which the energy consumption of EV during operation is obtained in step 3; whenever the user arrives at a destination, determine whether it needs to be charged, and if so, according to the step 5Choose a charging scheme; if no charging is required, the user will go to the next destination after staying for a while. Finally, the number of vehicles charged at the required time is counted, and the sum of the charging power is the charging load at the corresponding time, as shown in formula (9).

Among them, load _i is the charging load at time i; n is the total number of EVs being charged at the current time; p _j is the charging power of the j-th EV.

By changing the influencing factor data such as temperature, traffic congestion factor, date type, etc., the EV charging load value under various conditions can be obtained and used as the "real" data in the subsequent CGAN.

Step 7. Build the CGAN generator model

Because LSTM has good time series information processing ability, it is selected to build the CGAN generator model. Its specific structure is shown in Figure 2. The generator consists of a deep LSTM layer and a fully connected layer. The deep LSTM layer has 4 hidden layers, each with 200 LSTM units, and the hidden layer uses the ReLU function as the activation function. The input of the model is the influence factor data and random noise. The influencing factor data is the same influencing factor data as in the “real” data acquisition process in step 6, including the date type (weekdays or holidays), the lowest temperature value of the day, the highest temperature value of the day, traffic congestion factor, and user-generated charging. The probability distribution of EV power when demanded, the probability distribution of EV power when the user finishes charging, and the probability density of the user's charging scheme selection ranking. Random noise is set as a random variable obeying a Gaussian distribution. The output data of the model is the predicted load data.

Normalization processing is also required before the input of the influencing factor data; specifically, the maximum-minimum normalization method is adopted, which is shown in formula (10).

Among them, X _norm is the normalized data; X is the current data; X _min is the minimum value in the data; X _max is the maximum value in the data.

Step 8. Build the CGAN discriminator model

The discriminator consists of a deep LSTM layer and a fully connected layer. The deep LSTM layer has 4 hidden layers, each with 200 LSTM units, and the hidden layer uses the ReLU function as the activation function. The sigmoid activation function is used in the fully connected layer to make true and false judgments. The load data obtained based on the model-driven and the predicted load data generated by the generator are respectively integrated with the influencing factors into the discriminator, and the discriminator outputs the probability that the input data is the real data.

Step 9: The generator model and the discriminator model are used for game training, and the trained generator model is used for load prediction.

When the CGAN fully learns the relationship between the data and reaches a balance, the load data under different conditions can be obtained by adjusting the conditional data input to the generator. The process of EV charging load prediction based on CGAN is shown in Figure 3.

references

[1] Long Xuemei, Yang Jun, Wu Fuzhang, et al. Electric vehicle charging load prediction considering road network-grid interaction and user psychology [J]. Automation of Electric Power Systems, 2020, 44(14): 86-93.

[2] Yan J, Zhang J, Liu Y, et al. EV charging load simulation and forecasting considering traffic jam and weather to support the integration of renewables and EVs [J]. Renewable Energy, 2020, 159: 623-641.

Claims (3)

1. An EV charging load calculation method for generating a countermeasure network based on conditions is characterized by specifically comprising the following steps of:

step 1, acquiring daily charging habits and preferences of a user according to a questionnaire; fitting the probability distribution of the electric quantity when the user has a charging demand and the electric quantity when each charging is finished, and fitting the probability density of the user for sequencing three schemes, namely, the shortest distance, the lowest charging cost and the least time spent;

step 2, constructing a user travel model

The EV is divided into two types of private car operation vehicles according to the use property; considering the larger operation difference of the private car and the operation vehicle, respectively using a travel chain and an OD probability matrix to describe the travel behaviors of the private car and the social operation vehicle;

the method comprises the following steps that a user moves to one or more places in a meeting in one day, a set formed by destination nodes of travel of the user is a travel chain, and the specific expression of representing the travel of a private car user by using the travel chain is shown as a formula (1):

D＝{d₁,d₂,...,d_n,...} (1)

d is a destination set corresponding to the trip chain; n is a destination serial number; d₁Is the starting point of the user trip;

d_na stopping point in the travel process; the path set corresponding to the trip chain can be represented by equation (2):

P＝{p(d₁,d₂),p(d₂,d₃),...p(d_n-1,d_n),...} (2)

wherein P represents a path set corresponding to the trip chain, P (d)_n-1,d_n) Representing a path from the (n-1) th destination to the nth destination;

the OD probability matrix in each time period region can be regarded as the probability distribution of the travel destination of the operating vehicle, and the specific expression is shown as a formula (3);

wherein G (i) represents an OD probability matrix in the ith time period; r, w, b denote a residential area, a work area, and a business area, respectively; g_w,rRepresenting the probability of a user traveling from a work area to a residential area;

step 3, constructing an EV energy consumption model;

when the road is smooth, the vehicle can run at a constant speed, and the time t is spent on the vehicle running from the node i to the node j_i,jAs shown in formula (4):

where v is the road speed limit,/_i,jRepresenting the distance from the node i to the node j; when the road is congested, the higher the congestion degree is, the slower the vehicle runs, and the running time is corrected by introducing a time-consuming coefficient, specifically shown as a formula (5);

wherein,

is the corrected time; delta_iIs a time consumption coefficient;

the EV energy consumption model is specifically shown as formula (6):

wherein e is_i,jThe amount of power consumed for EV to travel from node i to node j; e.g. of the type_cfRepresents the power consumption per unit distance,/_i,jExpressed as the distance between the inode and the j node; p_aIs EV air-conditioning power;

e is the EV battery capacity for the travel time obtained according to equation (3);

step 4, constructing an EV charging model;

the EV charging period is specifically represented by equation (7):

wherein, soc_eThe electric quantity when the user finishes charging; soc_sThe electric quantity when the user starts to charge; p is charging power of the charging pile; eta is charging efficiency of the charging pile; t is t_eFor the extra stay time after the EV is fully charged, if the electric quantity is less than the maximum electric quantity when the user finishes charging, t_eIs 0;

step 5, constructing an EV user charging scheme selection model;

when a user needs to charge, the user is supposed to receive three charging schemes recommended by a scheduling center, specifically, the three charging schemes include a scheme that the time spent by the user is shortest, a scheme that a charging station is closest to the user, and a scheme that the charging cost of the user is lowest; after receiving the recommendation, the user selects one scheme to go to a charging station for charging, wherein the user's will for selecting the charging scheme is obtained in the form of a questionnaire;

step 6, forecasting EV charging load based on model driving

Calculating the space-time change of the electric quantity of the EV battery according to the user travel model in the step 2, wherein the energy consumption of the EV in the running process is obtained in the step 3; every time the user arrives at a destination, judging whether the user needs to be charged, and if so, selecting a charging scheme according to the step 5; if the charging is not needed, the user stays for a period of time and then goes to the next destination; finally, counting the number of the vehicles charged at the required moment, wherein the sum of the charging power of the vehicles is the charging load at the corresponding moment, and the charging load is shown as a formula (8);

wherein, load_iThe charging load at the moment i; n is the total number of the EV's being charged at the current moment; p is a radical of_jCharging power for the jth EV;

the EV charging load values under various conditions can be obtained by changing the influence factor data, and the data are used as 'real' data in the subsequent CGAN;

step 7, building a CGAN generator model

The generator consists of a depth LSTM layer and a full-connection layer, wherein the depth LSTM layer is provided with 4 hidden layers, each layer has 200 LSTM units, and the hidden layers use a ReLU function as an activation function; the input of the model is influence factor data and random noise; the influence factor data is the same influence factor data as the influence factor data obtained in the real data obtaining process in the step 6, and specifically comprises a date type, a lowest value of the temperature of the day air, a highest value of the temperature of the day air, a traffic jam coefficient, probability distribution of electric quantity of the EV when the user generates a charging demand, probability distribution of electric quantity of the EV when the user finishes charging, and probability density of user charging scheme selection sequencing; setting random noise as a random variable subject to Gaussian distribution; the output data of the model is predicted load data;

normalization processing is needed before the influence factor data are input;

step 8, constructing a CGAN discriminator model

The discriminator consists of a depth LSTM layer and a full-connection layer, wherein the depth LSTM layer is provided with 4 hidden layers, each layer has 200 LSTM units, and the hidden layers use a ReLU function as an activation function; using a sigmoid activation function to judge whether the connection layer is true or false; load data obtained based on model driving and predicted load data generated by a generator are respectively integrated with influence factors and input into a discriminator, and the discriminator outputs the probability that the input data is real data;

step 9, carrying out game training on the generator model and the discriminator model, and carrying out load prediction by using the trained generator model;

when the relation between the CGAN fully-learned data is balanced, the load data under different conditions can be obtained by adjusting the condition data input to the generator.

2. An EV charging load calculation method for a conditional generation countermeasure network according to claim 1, characterized in that: counting the daily charging habits and preferences of the user in the form of questionnaire survey; firstly, carrying out reliability inspection on the recovered questionnaire survey by using a clone Bach coefficient method, and screening out effective samples; the calculation of the cloned Bach coefficient is specifically shown as a formula (9);

wherein, alpha is a reliability coefficient, and the higher the value is, the higher the reliability of the questionnaire is; k is the number of questionnaire questions; s_i ²The variance of the investigation result of the ith question is shown; s_x ²The variance of all the survey results.

3. An EV charging load calculation method for a conditional generation countermeasure network according to claim 1, characterized in that: before the influence factor data is input, normalization processing is required,

specifically, a maximum-minimum normalization method is adopted, which is specifically shown as a formula (10);

wherein, X_normThe normalized data is obtained; x is current data; x_minIs the minimum value in the data; x_maxIs the maximum value in the data.

Images