Electric vehicle load optimization method based on graph convolution and deep belief network

Abstract

The present invention discloses an electric vehicle load optimization method based on graph convolution and deep belief network, which uses graph convolution network to extract a large amount of uncertainty feature data of the road network, and then trains the deep belief network through a data-driven method to output the optimal charging load distribution of the electric vehicle, optimize the charging load distribution of the electric vehicle while coping with the uncertainty of the road network, and quantitatively analyze the impact of the optimization results on the road network and the power grid. The load optimization method. The electric vehicle load optimization method of the present invention is based on the data-driven optimization method of graph convolution and DBN, which can generalize the uncertainty factors affecting the charging load of the electric vehicle and accurately and efficiently optimize the charging load of the electric vehicle; compared with solving the nonlinear optimization model, the present invention has low calculation difficulty, fast solution speed, and stronger generalization performance for uncertainty factors; compared with existing deep neural network and other technologies, the optimization results of the present invention are more accurate, and the effect of reducing the user's charging cost is more significant.

Description

Technical Field

[0001] The present invention relates to the field of power system planning and operation, and in particular to an electric vehicle load optimization method based on graph convolution and deep belief network. Background Art

[0002] As the global fossil energy depletion and environmental pollution problems become increasingly serious, electric vehicles have attracted much attention due to their good energy-saving and emission-reduction effects, and the rapid growth of their ownership has become an inevitable trend. Electric vehicles have the dual attributes of transportation and electricity consumption, and the charging load of electric vehicles has gradually become an important part of the power grid load. However, in the future, when the number of electric vehicles is very high, large-scale disordered charging will bring a certain burden to the power grid. Large power shocks will cause problems such as excessive voltage deviation and reduced power quality, and in severe cases, it will even affect the safe and stable operation of the power system. Therefore, solving the problem of large-scale disordered charging of electric vehicles is an indispensable part of power system planning and operation.

[0003] Existing technologies divide charging loads into two categories when regulating electric vehicle loads. One is slow charging at specific locations (such as homes and parking lots), which is not affected by uncertain factors in the road network. The second type mainly studies the charging load of fast charging stations in cities. The charging power is relatively high and is affected by multiple uncertain factors such as charging prices, traffic flow, and charging queue time. It has strong randomness and uncertainty. How to accurately evaluate and optimize the charging load of this type of electric vehicles is the key and difficult point of the research.

[0004] At present, most related studies model the driving and charging behaviors of a large number of electric vehicles as Nash equilibrium (User equilibrium, UE). There are a large number of nonlinear constraints in the model, which is difficult to solve and lacks generalization performance for the uncertainty of road and traffic information. As the scene changes, a large number of optimizations need to be re-solved, which increases the computational burden. Summary of the invention

[0005] The purpose of the present invention is to provide an electric vehicle load optimization method based on graph convolution and deep belief network, which fits the mapping relationship between environmental uncertainty and charging load through data-driven model-free deep learning, so as to optimize the electric vehicle charging load under arbitrary uncertainty scenarios efficiently and accurately.

[0006] The purpose of the present invention can be achieved through the following technical solutions:

[0007] An electric vehicle load optimization method based on graph convolution and deep belief network, the load optimization method comprising the following steps:

[0008] S1: Uncertainty feature extraction from road network features based on graph convolutional networks;

[0009] S2: Data-driven deep belief network training to obtain the optimal charging load distribution of electric vehicles after optimization;

[0010] S3: Quantitatively evaluate the benefits of the road network and power grid after charging load optimization.

[0011] Furthermore, in S1, the road network features are divided into point features and edge features, and then input into the edge feature convolutional network;

[0012] The road network point features include charging price, charging power, charging method, charging time, charging station queue time and charging station capacity. The point feature matrix is ​​represented by F N The road network edge features include road traffic volume, road grade, road capacity and road zero-flow travel time. The edge feature matrix is ​​represented by F A express;

[0013] Point features of the input layer and edge features is the unprocessed road network feature information; the node n feature of each layer is are all features of the previous layer of its adjacent nodes The weighted sum of

[0014] The convolution operation expression of ECCN is:

[0015]

[0016] In the formula, B(n) represents the set of node n and its adjacent nodes. is the trainable network weight, b (l) is the bias for the trainable network.

[0017] Furthermore, in S2, the features extracted by ECCN in S1 are used as DBN input, and the optimal charging load distribution is used as DBN output; specifically, the following steps are included:

[0018] S21: Pre-training of restricted Boltzmann machines; DBN is composed of RBMs stacked layer by layer, each RBM contains a visible layer and a hidden layer, and the hidden layer of the previous layer serves as the visible layer of the next layer;

[0019] The energy function is:

[0020] E (l) (v,h|φ (l) )=-(a T v+b T h+v T wh)

[0021] Where v and h are the visible layer and hidden layer units respectively; w is the weight connecting the visible layer and the hidden layer; a and b are the biases of the visible layer and the hidden layer respectively; φ (l) =[a v ,b h ,w vh ] is the parameter that needs to be trained for the l-th layer RBM; the joint distribution function of v and h is:

[0022]

[0023] In the formula, Used to constrain the sum of joint distribution probabilities to be 1;

[0024] Therefore, the loss function of RBM training is:

[0025]

[0026] In the formula, S is the training sample set;

[0027] S22: Fine-tune the DBN network for back propagation, adjust the parameters obtained by pre-training to improve the prediction accuracy; use the mean square error between the prediction result and the label as the loss function for fine-tuning; the training data and labels are obtained by solving the following optimization problem:

[0028]

[0029] Where: ω is the time cost coefficient; s a 、s z are the traffic volume on road a and the number of charging vehicles at charging station z respectively; t a ,t z are the travel time on road a and the charging time at charging station z respectively; z , P z are the charging price and charging power of charging station z respectively; i,j,a , π i,j,z is a 0-1 variable, used to indicate whether road a or charging station z passes through the i-th path of the j-th starting and ending points; are the zero-flow travel time on road a and the average queuing time at charging station z, respectively, where μ is the queuing coefficient; c a 、c z are the capacity of road a and charging station z respectively; f i,j represents the traffic flow of the i-th path from the j-th starting point to the j-th ending point; d j is the total number of cars at the jth starting and ending point, is the total charging power of the zth charging station;

[0030] The objective function in the above formula is to minimize the total travel and charging cost of all electric vehicle users under the Nash equilibrium of the road network, which also includes the time cost of driving and charging queues and the charging price cost.

[0031] Furthermore, in S3, the optimal charging power of each charging station obtained in S2 is Substitute into the power grid flow calculation; the following optimization is required:

[0032]

[0033] In the formula, a x 、b x are the coefficients of the quadratic and linear terms of the cost function of generator x, respectively; are respectively the active power generation, reactive power generation and net active power generation of generator x; is the active transmission power, reactive transmission power and square of current amplitude of line l between nodes xy; r∈X y Indicates that the r node is a child node of y; are the resistance, reactance and impedance of line l respectively; is the non-electric vehicle load at node y; U x is the node voltage; is the square of the rated current of line l; are the lower and upper limits of the active output of the generator, U is the lower and upper limits of the reactive power output of the generator; min , U max is the square of the node voltage lower limit and upper limit; CE x 、CC x is the carbon emission and carbon capture of node x, which is positive if there is a carbon capture system and zero if there is no carbon capture system installed; is the power consumed by the carbon capture system when capturing carbon, β and ε are the carbon emission intensity coefficient and the carbon capture coefficient, respectively.

[0034] Beneficial effects of the present invention:

[0035] 1. The data-driven optimization method based on graph convolution and DBN can generalize the uncertainty factors affecting the charging load of electric vehicles and optimize the charging load of electric vehicles accurately and efficiently;

[0036] 2. Compared with solving nonlinear optimization models, the present invention has low computational difficulty, fast solution speed, and stronger generalization performance for uncertain factors;

[0037] 3. Compared with existing deep neural network (DNN) and other technologies, the optimization results of the present invention are more accurate and the effect of reducing user charging costs is more significant. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The present invention will be further described below in conjunction with the accompanying drawings.

[0039] Figure 1 is a flow chart of the electric vehicle load optimization method of the present invention;

[0040] Figure 2 It is the structural principle diagram of ECCN of the present invention;

[0041] Figure 3 This is a schematic diagram of the DBN training principle of the present invention;

[0042] Figure 4 It is a topological diagram of a 6-node test system of the present invention;

[0043] Figure 5 It is a training and testing error curve diagram under different algorithms of the present invention;

[0044] Figure 6 It is a 20-way network node topology diagram of the present invention;

[0045] Figure 7 It is a topological diagram of the power grid node test system of the present invention 21;

[0046] Figure 8 is a comparison chart of the queuing time of each charging station before and after the optimization of the present invention;

[0047] Fig. 9 It is a schematic diagram of the changes in power generation output and carbon capture amount at each node of the power grid before and after the optimization of the present invention. DETAILED DESCRIPTION

[0048] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0049] An electric vehicle load optimization method based on graph convolution and deep belief networks (DBN), the electric vehicle load optimization method comprises the following steps:

[0050] S1: To extract features using graph convolutional networks, the road network features need to be divided into point features and edge features, and then input into ECCN.

[0051] The road network point features include charging price, charging power, charging method, charging time, charging station queue time, and charging station capacity. The point feature matrix is ​​represented by F N express.

[0052] The road network edge features include road traffic volume, road grade, road capacity, and road zero-flow travel time. The edge feature matrix is ​​represented by F A express.

[0053] The structure and principle of ECCN are as follows Figure 1 As shown. Point features of the input layer and edge features is the unprocessed road network feature information. The node n feature of each layer are all features of the previous layer of its adjacent nodes The weighted sum of; the convolution operation expression of ECCN is as follows:

[0054]

[0055] In the formula, B(n) represents the set of node n and its adjacent nodes, is the trainable network weight, b (l) is the bias for the trainable network.

[0056] S2: Using data-driven DBN, the features extracted by the previous ECCN are used as DBN input, and the optimal charging load distribution is used as the output of DBN.

[0057] The training principle of DBN is as follows Figure 2 As shown, it includes the following two steps:

[0058] S21: Pre-training of restricted Boltzmann machines (RBM). DBN is composed of RBM stacked layer by layer. Each RBM contains a visible layer and a hidden layer. The hidden layer of the previous layer serves as the visible layer of the next layer. The energy function is:

[0059] E (l) (v,h|φ (l) )=-(a T v+b T h+v T wh)

[0060] Where v and h are the visible layer and hidden layer units, respectively. w is the weight connecting the visible layer and the hidden layer. a and b are the biases of the visible layer and the hidden layer, respectively. φ (l) =[a v ,b h ,w vh] are the parameters that need to be trained for the l-th layer RBM.

[0061] The joint distribution function of v and h is:

[0062]

[0063] In the formula, Used to constrain the sum of joint distribution probabilities to be 1.

[0064] Therefore, the loss function of RBM training is:

[0065]

[0066] In the formula, S is the training sample set.

[0067] S22: Fine-tune the DBN network for back propagation and adjust the parameters obtained by pre-training to improve the prediction accuracy. The mean square error between the prediction result and the label is used as the loss function for fine-tuning. The training data and labels can be obtained by solving the following optimization problem:

[0068]

[0069] Where: ω is the time cost coefficient. sa and sz are the traffic volume on road a and the number of charging vehicles at charging station z, respectively. a ,t z are the travel time on road a and the charging time at charging station z respectively. z , P z are the charging price and charging power of charging station z respectively. i,j,a , π i,j,z It is a 0-1 variable used to indicate whether road a or charging station z passes through the i-th path of the j-th origin and destination (because each origin and destination can be connected by multiple paths and there are a large number of different origins and destinations in the road network). are the zero-flow travel time on road a and the average queuing time at charging station z, respectively, where μ is the queuing coefficient. a 、c z are the capacity of road a and the capacity of charging station z respectively. i,j represents the traffic flow of the i-th path from the j-th origin to the j-th destination. j is the total number of cars at the jth starting and ending point, is the total charging power of the zth charging station.

[0070] The objective function in the above formula is to minimize the total travel and charging cost of all electric vehicle (EV) users under the Nash equilibrium of the road network, which also includes the time cost of driving and charging queue and the charging price cost.

[0071] S3: Quantitative analysis of the optimization results can be used to calculate the optimal charging power of each charging station obtained in the previous step. Substitute it into the power grid flow calculation. The following optimization problem needs to be solved:

[0072]

[0073] In the formula, a x 、b x are the quadratic and linear coefficients of the cost function of generator x, respectively. They are respectively the active power generation, reactive power generation and net active power generation of generator x. is the active transmission power, reactive transmission power and square of current amplitude of line l between nodes xy. y Indicates that the r node is a child of the y node. are the resistance, reactance and impedance of line l respectively. is the non-electric vehicle load at node y. x is the node voltage. It is the square of the rated current of line l. are the lower and upper limits of the active output of the generator, U is the lower and upper limits of the reactive power output of the generator. min , U max is the square of the node voltage lower limit and upper limit. x 、CC x is the carbon emission and carbon capture amount of node x (positive if there is a carbon capture system, and zero if there is no carbon capture system installed). is the power consumed by the carbon capture system when capturing carbon, β and ε are the carbon emission intensity coefficient and the carbon capture coefficient, respectively.

[0074] After this optimization is solved, the changes in data such as power generation and carbon emissions of the power system before and after the optimization can be analyzed.

[0075] This embodiment includes a 6-node road network small system to illustrate the superiority of the proposed method compared with the prior art, and includes a 20-node road network + 21-node power grid large system to illustrate the benefits of the optimization results of the proposed method to the road network and the power grid. It should be clear that the present invention is not limited to the given embodiments, and changing the system nodes, topology, simulation parameters, etc. does not change the working principle, so they are all within the scope of protection.

[0076] See the attached 6-node small system topology diagram Figure 3, including 3 fast charging stations (FCS), road sections L1-L9, as well as distributed edge generators (DG) and photovoltaic generators (PV). Assume that a total of 1000 EVs travel from node 1 to node 6, and the road section capacity is 400 veh / h. Solve 1000 groups of UE optimization as training data for ECCN-DBN, of which 100 groups are used as verification data. Figure 4 The proposed ECCN-DBN method is compared with the existing DNN, ECCN-DNN, and DBN methods in terms of the improvement in the accuracy of the training results. Table 1 shows the training and test mean square errors of each algorithm at convergence, as well as the total charging cost of all EVs (i.e., the objective function of the UE).

[0077] Table 1. Training and test mean square error and total EV charging cost of each algorithm at convergence

[0078] method DNN ECCN-DNN DBN ECCN-DBN Training mean square error 0.0550 0.0086 0.0105 0.0075 Test mean square error 0.0570 0.0094 0.0113 0.0076 Total cost (yuan) 62,250 41,532 42,570 40,968

[0079] It can be seen that the proposed ECCN-DBN method has higher accuracy than other deep learning methods, and has the lowest mean square error and EV charging cost at convergence. Its mean square error is reduced by 86.67%, 19.15%, and 32.74% compared with DNN, ECCN-DNN, and DBN, respectively, and the total charging cost is reduced by 34.19%, 1.36%, and 3.76%, respectively.

[0080] The system topology diagram of 20 network nodes + 21 power grid nodes is as attached. Figure 5 -Attached Figure 6 As shown. The road network contains 6 FCSs, charging power of 100kW, road capacity of 1200veh / h, charging station capacity of 2,800veh / h, and a total of 10,000 EVs. The starting point and destination are randomly sampled. The sampling probability of the central dotted box is 0.7, and the sampling probability of the outer dotted box is 0.3. In addition to ordinary nodes and FCS nodes, PV power generation, CCS-containing thermal power and non-CCS thermal power nodes are also set up in the power grid. Solving the optimization generates 10,000 sets of training samples, of which 9,000 sets are used as training data and 1,000 sets are used as test data. The distribution of charging load before and after optimization is shown in Table 2:

[0081] Table 2 Charging load distribution of each charging station before and after optimization

[0082] Charging Station Before optimization(MW) After optimization (MW) FCS1 64.80 134.31 FCS2 134.73 141.30 FCS3 131.49 139.95 FCS4 189.67 170.10 FCS5 228.02 175.14 FCS6 151.29 142.20

[0083] The optimized charging queue time is as follows Figure 7 As shown in the above table and the attached Figure 7It can be seen that the optimized distribution of electric vehicle charging conditions helps to shorten the charging queue time and alleviate charging congestion, and the total UE cost of the road network is reduced from 155,470 yuan to 123,972 yuan, a reduction of 20.26%.

[0084] The grid operation status (such as the power generation output and carbon capture changes of each node in the grid) is shown in the attached Figure 8 The specific data are shown in Table 3. It can be seen that the proposed optimization method is helpful to reduce the power generation cost of the power grid, reduce network losses and carbon emissions, and increase the photovoltaic absorption capacity and carbon dioxide capture.

[0085] Table 3 Changes in grid operation data before and after optimization

[0086]

[0087]

[0088] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0089] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements all fall within the scope of the present invention to be protected.

Claims

1. Electric vehicle load optimization method based on graph convolution and deep belief network, characterized in that: The load optimization method comprises the following steps: S1: Uncertainty feature extraction from road network features based on graph convolutional networks; S2: Data-driven deep belief network training to obtain the optimal charging load distribution of electric vehicles after optimization; S3: Quantitatively evaluate the benefits of the road network after charging load optimization; In S2, the features extracted by ECCN in S1 are used as DBN input, and the optimal charging load distribution is used as DBN output; The specific steps include: S21: Pre-training of restricted Boltzmann machines; DBN is composed of RBMs stacked layer by layer, each RBM contains a visible layer and a hidden layer, and the hidden layer of the previous layer serves as the visible layer of the next layer; The energy function is: E (l) (v,h|φ (l) )=-(a T v+b T h+v T wh) Where v and h are the visible layer and hidden layer units respectively; w is the weight connecting the visible layer and the hidden layer; a and b are the biases of the visible layer and the hidden layer respectively; φ (l) =[a v ,b h ,w vh ] is the parameter that needs to be trained for the l-th layer RBM; the joint distribution function of v and h is: In the formula, Used to constrain the sum of joint distribution probabilities to be 1; Therefore, the loss function of RBM training is: In the formula, S is the training sample set; S22: Fine-tune the DBN network for back propagation, adjust the parameters obtained by pre-training to improve the prediction accuracy; use the mean square error between the prediction result and the label as the loss function for fine-tuning; the training data and labels are obtained by solving the following optimization problem: Where: ω is the time cost coefficient; s a 、s z are the traffic volume on road a and the number of charging vehicles at charging station z respectively; t a ,t z are the travel time on road a and the charging time at charging station z respectively; z , P z are the charging price and charging power of charging station z respectively; i,j,a , π i,j,z is a 0-1 variable, used to indicate whether road a or charging station z passes through the i-th path of the j-th starting and ending points; are the zero-flow travel time on road a and the average queuing time at charging station z, respectively, where μ is the queuing coefficient; c a 、c z are the capacity of road a and charging station z respectively; f i,j represents the traffic flow of the i-th path from the j-th starting point to the end point; d j is the total number of cars at the jth starting and ending point, is the total charging power of the zth charging station; The objective function in the above formula is to minimize the total travel and charging cost of all electric vehicle users under the Nash equilibrium of the road network, which also includes the time cost of driving and charging queues and the charging price cost.

2. The electric vehicle load optimization method based on graph convolution and deep belief network according to claim 1 is characterized in that: In S1, the road network features are divided into point features and edge features, and then input into the edge feature convolutional network; The road network point features include charging price, charging power, charging method, charging time, charging station queue time and charging station capacity. The point feature matrix is ​​represented by F N The road network edge features include road traffic volume, road grade, road capacity and road zero-flow travel time. The edge feature matrix is ​​represented by F A express; Point features of the input layer and edge features is the unprocessed road network feature information; the node n feature of each layer is are all features of the previous layer of its adjacent nodes The weighted sum of The convolution operation expression of ECCN is: In the formula, B(n) represents the set of node n and its adjacent nodes. is the trainable network weight, b (l) is the bias for the trainable network.

3. The electric vehicle load optimization method based on graph convolution and deep belief network according to claim 1 is characterized in that: In S3, the optimal charging power of each charging station obtained in S2 is Substitute into the power grid flow calculation; the following optimization is required: In the formula, a x , b x are the coefficients of the quadratic and linear terms of the cost function of generator x, respectively; are respectively the active power generation, reactive power generation and net active power generation of generator x; is the active transmission power, reactive transmission power and square of current amplitude of line l between nodes xy; r∈X y Indicates that the r node is a child node of y; are the resistance, reactance and impedance of line l respectively; is the non-electric vehicle load at node y; U x is the node voltage; is the square of the rated current of line l; is the lower and upper limits of the active output of the generator, U is the lower and upper limits of the reactive power output of the generator; min , U max is the square of the lower and upper limits of the node voltage; CE x 、CC x is the carbon emission and carbon capture of node x, which is positive if there is a carbon capture system and zero if there is no carbon capture system installed; is the power consumed by the carbon capture system when capturing carbon, β and ε are the carbon emission intensity coefficient and the carbon capture coefficient, respectively.

Images