CN116596105A - Charging station load prediction method considering power distribution network development - Google Patents
Charging station load prediction method considering power distribution network development Download PDFInfo
- Publication number
- CN116596105A CN116596105A CN202310235422.6A CN202310235422A CN116596105A CN 116596105 A CN116596105 A CN 116596105A CN 202310235422 A CN202310235422 A CN 202310235422A CN 116596105 A CN116596105 A CN 116596105A
- Authority
- CN
- China
- Prior art keywords
- node
- power distribution
- load prediction
- distribution network
- lstm
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000009826 distribution Methods 0.000 title claims abstract description 105
- 238000000034 method Methods 0.000 title claims abstract description 80
- 238000011161 development Methods 0.000 title claims abstract description 21
- 230000008569 process Effects 0.000 claims abstract description 34
- 230000008859 change Effects 0.000 claims abstract description 25
- 238000013528 artificial neural network Methods 0.000 claims abstract description 12
- 238000013135 deep learning Methods 0.000 claims abstract description 7
- 238000000605 extraction Methods 0.000 claims description 31
- 238000012549 training Methods 0.000 claims description 31
- 230000009466 transformation Effects 0.000 claims description 24
- 230000002776 aggregation Effects 0.000 claims description 18
- 238000004220 aggregation Methods 0.000 claims description 18
- 239000013598 vector Substances 0.000 claims description 18
- 238000004364 calculation method Methods 0.000 claims description 15
- 239000011159 matrix material Substances 0.000 claims description 12
- 230000006870 function Effects 0.000 claims description 9
- 230000004913 activation Effects 0.000 claims description 8
- 238000010276 construction Methods 0.000 claims description 6
- 238000010606 normalization Methods 0.000 claims description 3
- 230000004931 aggregating effect Effects 0.000 claims description 2
- 230000002708 enhancing effect Effects 0.000 claims description 2
- 238000007781 pre-processing Methods 0.000 claims description 2
- 230000001131 transforming effect Effects 0.000 claims description 2
- 238000004088 simulation Methods 0.000 abstract description 22
- 125000004122 cyclic group Chemical group 0.000 abstract description 3
- 230000000737 periodic effect Effects 0.000 abstract 1
- 230000001965 increasing effect Effects 0.000 description 18
- 230000007246 mechanism Effects 0.000 description 13
- 238000011156 evaluation Methods 0.000 description 12
- 238000010586 diagram Methods 0.000 description 11
- 230000000694 effects Effects 0.000 description 6
- 239000000203 mixture Substances 0.000 description 4
- 101001095088 Homo sapiens Melanoma antigen preferentially expressed in tumors Proteins 0.000 description 3
- 102100037020 Melanoma antigen preferentially expressed in tumors Human genes 0.000 description 3
- 230000005540 biological transmission Effects 0.000 description 3
- 239000006185 dispersion Substances 0.000 description 3
- 230000006872 improvement Effects 0.000 description 3
- 238000002407 reforming Methods 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 230000008030 elimination Effects 0.000 description 2
- 238000003379 elimination reaction Methods 0.000 description 2
- 238000013178 mathematical model Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 241000965478 Darksidea epsilon Species 0.000 description 1
- 238000000342 Monte Carlo simulation Methods 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000010835 comparative analysis Methods 0.000 description 1
- 230000000295 complement effect Effects 0.000 description 1
- 238000003912 environmental pollution Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 239000002360 explosive Substances 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 239000000446 fuel Substances 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 230000001788 irregular Effects 0.000 description 1
- 239000012633 leachable Substances 0.000 description 1
- 230000007774 longterm Effects 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000010248 power generation Methods 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000005096 rolling process Methods 0.000 description 1
- 238000012216 screening Methods 0.000 description 1
- 238000003860 storage Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
- G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
- G06Q10/063—Operations research, analysis or management
- G06Q10/0639—Performance analysis of employees; Performance analysis of enterprise or organisation operations
- G06Q10/06393—Score-carding, benchmarking or key performance indicator [KPI] analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
- G06Q50/06—Electricity, gas or water supply
Abstract
The application discloses a charging station load prediction method considering power distribution network development, and relates to the field of deep learning and load prediction. According to the method, a side graph annotation force network in a graph neural network is applied, key topological features including side features in the power distribution network topology are distributed with higher attention scores, and the spatial correlation of the predicted charging load is extracted efficiently; and capturing the periodic law of the load by using a long-period memory network and a short-period memory network in the cyclic neural network, and extracting the time correlation of the predicted charging load. In the process of power distribution network development, the prediction method provided by the application can consider the space-time correlation of the predicted charging station, has the capability of analyzing the change of the line parameters, strengthens the regression relationship between key topological features and labels, and ensures the efficiency and the precision of load prediction. The application uses Pytorch software to build and solve the depth network, and the simulation result verifies the rationality and effectiveness of the method.
Description
Technical Field
The application relates to the field of load prediction, in particular to a charging station load prediction method considering power distribution network development.
Background
In various countries of the world, along with the improvement of the living standard of people, fuel automobiles become indispensable tools in daily life of people, which brings great pressure to urban traffic networks and also causes serious damage to the environment, such as traffic jam, tail gas pollution and other problems. Aiming at the problem of environmental pollution, the traffic travel structure transformation policy is continuously implemented. The Electric Vehicles (EVs) in the traffic network are continuously enlarged in scale, and the load of the charging station is continuously increased, which brings great challenges to the safe and stable operation of the power network. Therefore, the power distribution network needs to conduct real-time and accurate scheduling strategies on charging loads, peak shaving pressure of the power grid is reduced as much as possible, and charging station load prediction needs to be conducted in advance.
Considering that EV users select charging stations, they are affected by time and space factors such as commute time and traffic congestion, and thus charging station loads exhibit correlations in time and space dimensions. In the deep learning method, the graph neural network captures the topological spatial attribute of the network, and the cyclic neural network captures the time dimension rule, so that the space-time correlation of the predicted object is effectively extracted. However, most of the existing load prediction researches of charging stations do not consider that as the load is continuously increased, the power distribution network is also subjected to upgrading and reconstruction work, wherein the work mainly comprises topology change and prediction situations of line reconstruction. When the topology changes, nodes and edges contained in the topology are increased, and efficient and accurate regression learning of the prediction model becomes more and more difficult; when the line is transformed, the method is equivalent to the change of the edge impedance characteristics in the topology, and if only the node characteristics are analyzed, the change information of the edge characteristics is lost, so that the fitting effect is affected. In summary, the conditions of topology scale increase and line upgrading and transformation in the upgrading process of the power distribution network may affect the accuracy of the prediction model, and bring practical problems to the processes of generating, scheduling and the like of the subsequent units.
Disclosure of Invention
The graph attention mechanism (Graph Attention Network, GAT) method can extract topological node attributes as distribution network side characteristic information, a topological structure is used as an adjacent matrix, attention mechanism learning is performed, higher attention scores are given to key nodes, regression characteristics of a prediction model are learned more pertinently in complex topology, and prediction precision and efficiency are improved. However, the GAT can only acquire node characteristics and cannot analyze side information, under the condition of upgrading and reforming a circuit, a side drawing attention network (Edge Graph Attention Network, EGAT) is introduced into the circuit, an EGAT-LSTM charging station load prediction framework is provided by combining a Long-short-term memory network (Long-short term memory, LSTM), the edge characteristics are extracted by utilizing the EGAT characteristics, the edge characteristics are aggregated to the node characteristics, the aggregated characteristics are input into the LSTM charging load prediction layer, each module parameter is learned by utilizing a back propagation method according to a historical load quantity label, and the influence of the change of the edge characteristics on the non-regression characteristics of the node characteristics is corrected while the time correlation of the charging load of the electric automobile is excavated.
In order to achieve the above purpose, the following technical scheme is adopted:
The invention relates to a charging station load prediction method considering power distribution network development, which comprises the following steps:
step 1, performing optimal power flow calculation on the load increase and the power distribution network upgrading process of an example power distribution network through a Matpower tool box in MATLAB software to obtain training data (characteristics and labels), wherein the characteristics comprise nodes and edge characteristics) of the neural network;
step 2, preprocessing the node and edge feature training data obtained in the step 1, inputting the data into an EGAT feature extraction layer in an EGAT-LSTM charging station load prediction framework, fusing line parameter change information into the node in the EGAT feature extraction layer, and outputting the data as new node features fused with edge features;
step 3, inputting the new node characteristics obtained in the step 2 into an LSTM charging load prediction layer in an EGTA-LSTM charging load prediction framework to obtain a predicted charging station load prediction result;
and 4, carrying out back propagation learning in an EGAT-LSTM charging station load prediction framework by matching the prediction result obtained in the step 3 with a load label to obtain global optimal model parameters considering node and edge characteristic changes.
The further technical scheme of the invention is that the specific process of the step 1 is as follows: step 1-1, simulating and generating daily load curves conforming to charging load trend of electric vehicles through a Matlab platform, taking the daily load curves as base lines, and floating up and down according to a certain random number to generate charging station load labels in a training time scale. Generating other load quantities of the power distribution network with the same time scale in a similar way;
Step 1-2, line transformation information and a generation method thereof related in the upgrading process of a power distribution network, wherein the line transformation information comprises the steps of modifying line parameters (resistance and reactance values) under different construction periods in an IEEE standard power distribution network topology under a Matpower tool kit in Matlab so as to simulate line transformation conditions under different power supply periods;
step 1-3, generating training data of an EGAT-LSTM charging station load prediction framework, which comprises the steps 1-1 and 1-2, using an own optimal power flow calculation module in a Matpower tool box to perform optimal power flow calculation of a training time scale on a selected standard IEEE power distribution network topology, and storing node voltage, active and reactive power and resistance and reactance of a power supply period where a line is located in an obtained optimal power flow result as characteristics of a training data set.
The further technical scheme of the invention is that the optimal power flow calculation is shown as follows:
step 1, setting the iteration number k=0 and selecting a proper initial valueWherein k is a counter value; />For iterating k times node voltage and phase angle, +.>Active and reactive power output of the unit is iterated k times;
step 2, solving the constraint equation of the equationGet solution->Wherein g is the equality constraint satisfied by the grid;
Step 3, inLinearizing the problem nearby, and carrying out linear programming solution on the variation delta x of x, wherein the step size limit delta is less than or equal to delta x and less than or equal to delta is required to be met;
step 4, taking k=k+1 and updating the current solution x k ＝x k -1+Δx, if x k Stopping when the boundary is exceeded, otherwise, entering the step 5;
step 5, adjusting the step size limit delta to return to step 2.
The invention further adopts the technical scheme that the specific process of the step 2 is as follows:
in order to improve training efficiency and match with the input dimension of a subsequent LSTM charging load prediction layer, after the training data obtained in the step 1 are packaged according to the sequence length of the LSTM charging load prediction layer, a plurality of subgraphs are constructed into a large graph, and the large graph is input to an EGAT feature extraction layer in parallel; aggregating the line parameter information to the nodes by utilizing the EGAT feature extraction layer, extracting the node and the side feature information at the same time, and describing the change condition of the topology attribute more comprehensively; the aggregation process of line parameter aggregation to nodes is mainly divided into attention score calculation and node aggregation, which comprises the following steps:
(1) The attention score was calculated as follows:
the attention score after the side characteristic information is aggregated is:
in the above, gamma' ji,p Is the attention score after the side characteristic information is aggregated, h i Is the feature vector of the node i, Is node j epsilon N i Feature vectors of (i.e. node voltage, active power)Rate and reactive power; r is (r) ji Is the edge eigenvector between connection nodes i and j, namely line resistance and reactance; />Is the next learnable parameter of the multi-head attention p head and can realize the characteristic vector from f η Direction of the dimension->Transforming; />Is subjected to a Softmax normalization operation; the g I G is to splice the feature vector; />To learn parameters under the multi-head attention p head;to learn parameters under the multi-head attention p head, the edge feature dimension is realized by f e To->Is a variation of (2); it is noted that all nodes of the same layer in the EGAT feature extraction layer share the same parameter matrix +.>And->N i A set of contiguous nodes that is node i; the LeakyReLU is a nonlinear activation function that is more commonly used in deep learning, and its expression is as follows:
wherein a is a coefficient, and x represents a part of the input nonlinear activation function;
(2) The node and polygon properties are aggregated as follows:
in the above, h i The node i is a new node i feature representation after aggregation of neighbor node features and edge features therebetween; p is the number of multi-head attention heads;is a learnable parameter that causes an edge feature vector r ji Linear transformation is carried out before node aggregation, and the dimension of edge characteristics is realized by f e Direction f η Variation of a; thus, the training data is obtained with new node characteristics containing side information, and the dimension of the new node characteristics can be set in a certain range.
The further technical scheme of the invention is that the specific process of the step 3 is as follows:
the new node characteristics obtained in the step 2 are flattened and subjected to dimensional transformation to obtain a dimensional form suitable for inputting the LSTM charging load prediction layer, and the number of hidden layers of the LSTM charging load prediction layer is set according to the characteristic number set by the new node characteristics; and inputting the new node characteristics after dimension transformation into an LSTM charging load prediction layer, and obtaining a predicted value of the predicted charging station load through parameter and bias calculation of a hidden layer.
The further technical scheme of the invention is that the specific process of the step 4 is as follows:
carrying out counter propagation operation on the predicted result value obtained in the step 3 in combination with a truth value tag in an EGAT-LSTM charging station load prediction framework, updating the coefficient of an attention score calculation module in an EGAT feature extraction layer, searching key features in topology, and enhancing the regression of the key features and the tag; and meanwhile, the opening degrees of the input door, the forgetting door and the output door in the LSTM charging load prediction layer are updated, so that the proportion of the current time information to the historical information can be controlled, and the prediction precision is improved.
By adopting the technical scheme, the invention has the following technical progress:
1. the model considering the attention mechanism can extract key topological features in the topological structure, and when the topological scale becomes larger and more complex, the EGAT feature extraction layer can improve the accuracy and efficiency of model regression learning and is more suitable for the power grid environment which is continuously developed at the present stage.
2. The EGAT feature extraction layer can expand attention from node features to node features and edge features, change conditions of line parameters are extracted in the process of line upgrading and reconstruction, and the influence of the line upgrading and reconstruction process on a prediction model is reduced by optimizing feature parameters and attention scores through model learning.
3. Through a comparison experiment, the proposed EGAT-LSTM charging station load prediction framework not only can capture key topological attributes in more complex topologies and improve training efficiency, but also can learn the influence of line characteristic changes on a prediction model when a line is upgraded and transformed, and shows good EV charging station load prediction performance in different situations of power distribution network upgrading.
Drawings
FIG. 1 is a schematic diagram of an upgrade process of a power distribution network according to the method of the present invention;
FIG. 2 is a schematic diagram of an EGAT-LSTM EV charging station load prediction method of the invention;
FIG. 3 is a schematic diagram of the LSTM charge load prediction layer of the method of the present invention;
FIG. 4 is a schematic diagram of an EGAT feature extraction layer of the method of the invention;
FIG. 5 is a schematic diagram of an EGAT-LSTM charging station load prediction framework of the method of the invention;
FIG. 6 is a schematic diagram of the load increase of the power distribution network of the method of the present invention;
FIG. 7 is a schematic diagram of line resistance and reactance for different power cycles during line upgrade and retrofit of the method of the present invention;
FIG. 8 is a schematic diagram of simulation results of an IEEE33 node power distribution system in accordance with the method of the present invention;
FIG. 9 is a schematic diagram of the simulation fitting effect of the IEEE33 node power distribution system of the method of the present invention;
FIG. 10 is a schematic diagram of simulation results of an IEEE69 node power distribution system according to the method of the present invention;
FIG. 11 is a schematic diagram of the simulation fit effect of the IEEE69 node power distribution system of the method of the present invention.
Detailed Description
The invention is further described below with reference to the accompanying drawings:
referring to fig. 1 to 11, the method of the present invention comprises the following steps:
step 1, determining the working condition of a power distribution network accompanied by power distribution network upgrading and the influence on load prediction, wherein the working condition is specifically as follows:
step 1-1, determining the background of power distribution network upgrading and the importance of EV charging station load prediction
With the continuous implementation of clean energy policies in countries around the world, the operation process in each field has changed greatly. In the power distribution network, new energy power generation devices such as wind energy and solar energy are continuously increased, and the power distribution network is continuously updated due to the fact that the power consumption is continuously increased. In a traffic network, the storage quantity of the EV is continuously increased, and the load of the EV charging station is also increased in an explosive manner, so that huge pressure is brought to power distribution network dispatching and peak shaving, and accurate and efficient EV charging load prediction becomes more important to a power grid.
Step 1-2, influence of power distribution network development on EV charging load prediction
Load prediction is essentially a complex time series prediction problem, which takes into account the subjectivity of EV user charging station selection, resulting in correlation in both time and space dimensions. In predicting TF in a traffic network, most students use the traffic network ratio as a graph structure to describe spatial correlation between TF and different nodes in the traffic network (e.g., TF in the traffic network typically has a larger flow in the vicinity of business circles, schools, and hospitals, etc.). The inverted distribution network is also a standard graph structure, and topological attributes in the graph can be extracted by using a method in a graph neural network (Graph Neural Network, GNN), so that the spatial correlation of the predicted charging stations can be learned and described.
The upgrading of the distribution network mainly comprises two conditions of topology change and distribution line upgrading and transformation, wherein the number of nodes and edges is continuously increased in a topological graph, and meanwhile, the characteristics of the nodes and the edges are changed. The traditional prediction model lacks the pertinence of key topology information and the analysis capability of line characteristic change, can not accurately describe the influence of node and line attribute change on the load time-space characteristics of the EV charging station, and reduces the precision and efficiency of the prediction model. The upgrade schematic is shown in fig. 1, and in general, the upgrade and transformation process of the distribution line is divided into a plurality of construction periods, and is orderly carried out according to a certain line arrangement, for example, 3-period engineering is taken as an example, 4 line working periods 1-4 are formed in a conformal manner, the size of a legend in the drawing represents the output or load of a unit, and the thickness of the line represents whether the line has been transformed or not. When the topology scale of the power distribution network becomes large, nodes and edges in the topology are increased, the correlation between a predicted object and other topology objects becomes more complex and difficult to extract, and when a line is transformed, the resistance and reactance of the line change, so that the line loss value can be influenced, the larger the power distribution network scale is, the higher the line loss value is, the power flow can be redistributed under the regulation and control of the optimal power flow, the characteristic information of a power supply and other nodes can change in non-regression characteristics, and the neural network learning can consider that the characteristic change is caused by load change, so that the fitting precision of a prediction model is influenced.
Step 2, aiming at the influence of power distribution network upgrading on a prediction model, an EGAT-LSTM load prediction framework is provided
With the continuous development of society, the electric power load is continuously increased, the topology structure of the power distribution network is increased and is huge and complex due to the upgrading of the power distribution network, and the topology attribute is also enriched and enriched, for example, in the figure 2, the topology of the power distribution network changes situation. By the aid of the GAT method, topological node attributes are extracted to serve as power distribution network side characteristic information, a topological structure serves as an adjacent matrix, attention mechanism learning is conducted, higher attention scores are given to key nodes, prediction model regression learning is conducted more efficiently in complex topology, and model prediction accuracy is improved. However, the GAT can only acquire node characteristics and cannot distribute attention scores to the side information, and as in the case of the line transformation period 1-n of fig. 2, the application introduces an EGAT characteristic extraction layer to aggregate the node characteristics and the transformed side characteristics for different line working periods, and more comprehensively describes the change condition of the power distribution network topology. And then inputting the aggregated characteristics into an LSTM load prediction layer, and learning the EGAT-LSTM charging station load prediction framework parameters according to the historical load quantity, so as to ensure the global optimality of the prediction framework parameter learning. In fig. 2, the color of the node in the EGAT feature extraction layer represents the importance degree of the topological node to the predicted node after the edge feature is aggregated.
Step 3, in the EGAT-LSTM load prediction framework, the LSTM prediction module is used for indicating
Load prediction is essentially a complex time series prediction problem with time dimension characteristics. For traditional model driven methods, predictive models mostly assume a known statistical distribution, starting with the possibility of distribution variation, to obtain the predicted result. It is very difficult to achieve accurate EV charging station load prediction. The number of hidden layers of the network in deep learning is usually large, and under the condition that the training sample data is sufficient, enough parameters are provided for learning complex load characteristics.
The cyclic neural network (Recurrent Neural Network, RNN) is used as one of the networks in deep learning, and can describe the relation between the current time output and the previous information in one sequence data, and fully extract the time correlation of the time sequence data set. Taking single-step rolling prediction as an example, the prediction problem can be defined as the following mathematical model.
F t ＝f(T 1 ,,,T t-1 )
F t+1 ＝f(T 2 ,,,F t )
...(1)
Wherein F is a prediction result, F is a trained prediction model, T is a feature specified by the model, and T-1 is a seq_len value.
Because the RNN has the problems of gradient elimination and long-term dependence, the LSTM is selected as an execution part of a prediction model in an EGAT-LSTM charging load prediction framework, the execution part comprises a forgetting gate opening value which can be controlled between 0 and 1, history gradient information can be freely controlled, so that the gradient cannot be dominated by a short-distance small gradient, the problem of gradient elimination and the like can be well solved, and the brief principle is shown in the figure 3.
Fig. 3 shows the transformation structure of LSTM neural network, the external dotted line is the input at time t, unlike conventional RNNs, LSTM incorporates control gate concepts including input gates, forget gates and output gates. The input gate is responsible for controlling input information; forget gate is responsible for controlling previous time history information C t-1 The influence on the present time is due to C t-1 Can describe the correlation of the sequence data in the time dimension; the output gate is responsible for controlling the output information at the moment. Output h of last moment t-1 And information input x at this time t Together as the real input of the moment, after passing through the parameter equation f, the variable I with the value between 0 and 1 is obtained by entering the activation function sigma t ，F t ，O t Controlling the opening degree of the input door, the forget door and the output door, and carrying out information screening to finally obtain the output h at the moment t . The ratio of the current time information to the historical information can be controlled by controlling the opening degree of the 3 doors through the network learning, and the output result is optimized. The switch control variable and LSTM output are shown in the following equation.
h t ＝O t ×tanh(C t ) (3)
In the formula, the function f is composed of a leachable parameter W i And b i ，W f And b f ，W O And b O Composition, C t To be stored in the memory unit.
Step 4, in the EGAT-LSTM load prediction method, an EGAT aggregation module is used for schematic representation
Because the EV charging station load prediction only depends on the RNN method, spatial information can be lost, and the accuracy of a prediction result is reduced. Whether the traffic network or the power distribution network is well-suited to be described as a graph structure, in the traffic network, buildings with different properties can be regarded as nodes, and roads can be regarded as edges; in the power distribution network, different types of loads and power sources can be used as nodes, and a power transmission line can be used as an edge. Traffic flow on roads can be predicted by the traffic network through various attributes on the roads and building coordinate attributes, and the anti-observation power distribution network can also utilize collected topology information, such as node characteristics (node voltage, active power and reactive power are used in the application), edge characteristics (line resistance and reactance are used in the application) and the like, so that spatial correlation between a predicted charging station and other topology nodes or edges is mined, and the space-time characteristics of charging loads are more comprehensively described.
First, a graph model of a power distribution network can be defined as g= (N, E, D), where N represents all nodes and features in the power distribution network, and is defined by vectorsComposition, E.epsilon.R g×r Representing all transmission lines and characteristics in the system, and the transmission lines and characteristics are represented by vectors +.>Composition, D.epsilon.R m×m To describe an adjacency matrix of connection relations between nodes, the connection relation between two nodes is represented by element d ij Composition is prepared.
The adjacency between nodes is in the form of a matrix:
wherein m is the number of nodes in the topology, N i Representing the neighbor nodes and feature sets of node i.
The nodes and features are described as a matrix form as follows:
wherein b is the characteristic number of nodes, h i Is the node feature vector of node i.
The edges and features are described in the form of a matrix:
wherein g is the number of edges in the topological graph, r is the number of edge features, e ij Is the edge feature between node i and node j.
The further technical proposal is that the EGAT mathematical model is as follows:
the attention mechanism of the EGAT method sets attention scores for aggregation of the nodes, adjacent nodes and side information between the adjacent nodes, so that the mutual connection between the nodes is differentiated, and the model can pay attention to important node information better. The EGAT process is largely divided into attention score computation and node aggregation.
(1) The attention score was calculated as follows:
the attention score after the side characteristic information is aggregated is:
in the above, gamma' ji,p Is the attention score after the side characteristic information is aggregated, h i Is the feature vector of the node i,is node j epsilon N i I.e. node voltage, active power and reactive power; r is (r) ji Is the edge eigenvector between connection nodes i and j, namely line resistance and reactance; / >Is the next learnable parameter of the multi-head attention p head and can realize the characteristic vector from f η Dimension f η Conversion; />Is subjected to a Softmax normalization operation; the g I G is to splice the feature vector; />To learn parameters under the multi-head attention p head;to learn parameters under the multi-head attention p head, the edge feature dimension is realized by f e Direction f e Variation of a; it is noted that all nodes of the same layer in the EGAT feature extraction layer share the same parameter matrix +.>And->N i A set of contiguous nodes that is node i; the LeakyReLU is a nonlinear activation function that is more commonly used in deep learning, and its expression is as follows:
wherein a is a coefficient, and x represents a part of the input nonlinear activation function;
(2) The node and polygon properties are aggregated as follows:
in the above, h i The node i is a new node i feature representation after aggregation of neighbor node features and edge features therebetween; p is the number of multi-head attention heads;is a learnable parameter that causes an edge feature vector r ji Linear transformation is carried out before node aggregation, and the dimension of edge characteristics is realized by f e Direction f η Variation of a; thus, the training data is obtained with new node characteristics containing side information, and the dimension of the new node characteristics can be set in a certain range.
The further technical scheme is characterized in that the EGAT-LSTM charging load prediction framework is briefly described as follows:
The EGAT-LSTM charging station load prediction framework splices the EGAT characteristic extraction layer and the LSTM load prediction layer together for learning and training, ensures that the LSTM layer parameters learn the charging station load time correlation, and simultaneously endows the key topological features with higher attention weight when the nodes and the edge features are fused, and learns the spatial correlation of the predicted object and other topological features more effectively. Compared with the simple EGAT as the feature extractor, the spliced prediction model parameter learning ensures the global optimal characteristic.
Based on the node characteristics and edge characteristics selected herein, namely node voltage, active power and reactive power, line resistance and reactance. An example of the aggregation of EGATs taking into account the topology properties of the distribution network side is shown in fig. 4. In the figure, circles represent node and edge features, including the active power (P), reactive power (Q) and voltage value (V) of the node, and line resistance (R), reactance (X). Sigma represents the activation function, x is the vector multiplication, and + is the vector addition. The portion enclosed by the gray solid line represents the module that calculates the attention score. The Concat-marked rectangle is characterized by that it is Linear layer, and contains many learning parameters for learning the correlation between adjacent nodes and between nodes and edges, in which the nodes share parameters And edge sharing parameter->This is a common means of feature enhancement, and +.>Mapping the spliced high-dimensional topological characteristic to a correlation coefficient e ji On top of that, and through the Softmax layer, it can be converted into a attention fraction gamma' ji . The portion enclosed by the gray dashed line is the module that applies the attention score to the node and edge features. The wavy line is polymerized by using a multi-head attention mechanism to obtain a custom specialNew node characteristics h characterizing dimensions C1-Cn i ”。
Step 5, the load prediction method of the EGAT-LSTM charging station load prediction framework is schematically as follows:
the EGAT-LSTM load prediction method is characterized in that an EGAT characteristic extraction layer and an LSTM charging load prediction layer are spliced together for learning and training, the parameter of the LSTM charging load prediction layer is ensured to learn EV charging load time correlation, and meanwhile, the EGAT characteristic extraction layer gives higher attention weight to key topological characteristics when fusing node and edge characteristics, and learns the spatial correlation of a predicted object and other topological characteristics. Compared with the method that the EGAT feature extraction layer is simply used as a feature extractor, the spliced prediction model can ensure the global optimality of model parameter learning aiming at the object to be predicted. The prediction framework is shown in fig. 5.
The figure is the whole structure of the charge prediction method of the EGAT-LSTM EV charging station, and mainly comprises an EGAT characteristic extraction layer and an LSTM charging load prediction layer. It should be noted that, the input features on the left side of the graph are all data packaged according to the seq_len time points, and when batch data are built to improve the model learning efficiency, the topological structures of all time point data in each packaged data need to be combined, that is, a plurality of sub-graphs are built into a large graph, the adjacent matrix of the large graph is a block diagonal matrix, and blocks on the diagonal line are respectively adjacent matrixes of all sub-graphs, so that adjacent nodes do not exist between different time point data, and model learning cannot be affected. In the method, when the EGAT-LSTM charging station load prediction framework model is constructed, the batch_size EGAT feature extraction layers are constructed, and therefore the constructed large graph cannot contain a large number of nodes to influence the training and prediction efficiency. And correspondingly inputting the packaged data of the batch_size into an EGAT feature extraction layer to extract features. For a rectangular EGAT feature extraction layer, starting from the topology features of the power distribution network in the 1 st group of Batch, the topology features are taken as input of EGAT, and the input is the new node features of the fusion node and the edge features. After the Batch_size data are spliced and dimension transformation is carried out, the data can be input into a cuboid LSTM charging load prediction layer, and a model prediction result is output. And taking the charging load value in each group of batch data as the label input of the LSTM charging load prediction layer, and after carrying out loss calculation on the charging load value and the prediction result, reversely transmitting and updating parameters of the EGAT characteristic extraction layer and the LSTM charging load prediction layer. It is noted that the features of the input LSTM charge load prediction layer are flattened features, which enables the prediction framework to learn the importance of each feature well, and to fuse the predicted object into the prediction framework better.
Step 6, load increase and line modification are schematically as follows:
with the development of society, the load in the power distribution network is continuously increased, and the predicted No. 2 hybrid node is taken as an example, and the load data increasing process of the power distribution network is shown in fig. 6. The abscissa is the EV charge load and other load increase process divided into 7 cycles, cycles I-VII respectively, each of which has a period in the data set as shown in the table on the left of the figure, which does not coincide with the line upgrade and retrofit cycle times mentioned above, given the randomness of the construction data set. The ordinate is the load active power value, and the fluctuation range of the load on each cycle is set to + -15%. As can be seen from the development of period ii, the predicted EV charge load exhibits a time dependence at a certain period.
The load increase is usually complementary with the development of the power distribution network, in order to improve the load capacity, the line upgrading condition is unavoidable, and it is noted that upgrading and reconstruction engineering belongs to construction projects and has a certain construction period and standard, so that the resistance and reactance upgrading and reconstruction information of all sides can be obtained without a signal acquisition device, and the line resistance and reactance of the whole graph can be obtained no matter where the middle node is located. An example of the line resistance and reactance for different power supply cycles during upgrade and retrofit is shown in fig. 7.
According to the topological structure line arrangement sequence of the IEEE33 or 69 node distribution network system, in the 8760h data set, line upgrading and reconstruction work is orderly carried out in 5-stage engineering, namely 6 different power supply periods 1-6 exist. The left-hand table of the figure shows the corresponding time period on the data set for each power cycle. The shaded portion in the legend represents the reduced resistance/reactance value of each phase of engineering after upgrading and reforming for the line corresponding to the label on the abscissa. The blue solid line represents the resistance/reactance values before the transformation of different lines, and the complete blue solid line also represents the resistance/reactance values of each line in the 1 st power supply period; the red solid line represents the resistance/reactance value after the transformation of different lines, and the complete red solid line also represents the resistance/reactance value of each line in the 6 th power supply period; it should be noted that all lines in the data set topology are modified, and each line is modified only once, and the red solid line and the blue solid line together form resistance/reactance values on different lines in the 2 nd-5 th power supply period.
Step 7, simulation results are shown as follows:
in the section, the change situation of evaluation indexes on different models under the conditions of upgrading and reconstruction of power distribution network lines and topology change which occur in the development of the power distribution network is mainly shown. Wherein, step 7-1 is to perform simulation analysis in the IEEE33 node power distribution system. Step 7-2 is an expansion scenario of the equivalent distribution network, namely, simulation analysis is performed in the IEEE69 node distribution system. Step 7-3 is to comprehensively compare the IEEE33 node system and the IEEE69 node system, and analyze the advantages of the EGAT-LSTM charging station load prediction framework model constructed herein in the development process of the power distribution network. It should be noted that in the initial stage of the simulation process, there is a high probability that the R2 evaluation index is less than 0, and such data has no meaning, so each model is recorded from the EPOCH when the R2 value is greater than 0, and a comparative analysis is performed. And the simulation results of all the models are subjected to filtering operation, so that the advantages and disadvantages of different models on different evaluation indexes are more intuitively displayed.
The simulation analysis uses standard IEEE33 and IEEE69 node power distribution systems to carry out simulation analysis, and the condition after the topology change of the distribution network is equivalently assumed by using the IEEE69 node power distribution system. The application divides the node types in the power distribution network topology into power supply nodes, charging load nodes, other load nodes and mixed load nodes, and signal acquisition devices are arranged on the nodes and the connected edges of the nodes. However, there are some nodes that have no signal acquisition device, cannot acquire node characteristics, and cannot acquire edge characteristics connected thereto, and these nodes are collectively called intermediate nodes. The node classification is shown in the table. Regarding the hybrid node, if the hybrid node is near the city center according to the position of the EV charging station, the hybrid node is combined with other load nodes into the hybrid node in the process of power grid topology planning; in suburban areas away from the urban center, EV charging stations may also appear solely as a node in the distribution network.
Regardless of the topology change, an IEEE33 node power distribution system includes 33 nodes and 32 edges, while an IEEE69 node power distribution system includes 69 nodes and 68 edges. The nodes and the edges comprise various characteristics as shown in the following table, and the characteristics in the table are selected to equivalent the topology information of the nodes and the edges of the power distribution network according to the problems to be solved. The charging station positions are mapped into individual nodes of the topological graph, and load changes in the power distribution network are reflected into node characteristics, including changes of node voltage, active power and reactive power. According to the Monte Carlo simulation method, 8760h initial distribution network load data including other loads and EV charging station loads are generated, and the two types of loads are different in time correlation and different in load distribution. And solving the optimal power flow of the IEEE33 and IEEE69 node distribution network by using a Matpower kit in Matlab to obtain a data set containing 8760h node characteristics, a charging load label and an adjacent matrix, wherein 70% of the data are used as training, 10% of the data are used as verification sets for adjusting super parameters in a model, 20% of the data are used as test sets, and the training effect of the model is tested.
The application selects the GCN-LSTM model as a load prediction basic method considering space-time correlation to study the performance of the proposed EGAT-LSTM charging station load prediction framework model prediction model. In the simulation process, LSTM charging load prediction layer parameters in the three models are set to be the same, number layer=3, hidden size=32, dropout=0.05, and the output dimension after feature aggregation is 8. Multi-heads=4 for GAT layer and erat feature extraction layer. The model presented herein was constructed using Pytorch software and trained using intel Xeon E-2224CPU and 32GB memory.
The simulation adopts average absolute percentage error (Mean Absolute Percentage Error, MAPE), average absolute error (Mean Absolute Error, MAE) and judgment coefficient R2 as evaluation indexes of the models to compare and analyze the performances of different models. The judgment coefficient R2 is the proportion of the square sum of the dispersion of the sample interpretation by using the regression model in the total square sum of the dispersion, namely the correlation degree percentage of the predicted value and the true value, if the fitting degree is good, the closer each sample observation point is to the regression line, the closer R2 is to 1, the equivalent of the R2 value is the curve fitting precision, and the mathematical expression of the evaluation index is as follows:
Wherein SSR is the sum of squares of regression, SSE is the sum of squares of residual errors, and SST is the sum of squares of total dispersion. yi is the true value of time i,for the model prediction value of time i, +.>Is the average value of the true valuesN is the total time series length.
Step 7-1, the simulation result of the IEEE33 node power distribution system is shown as follows:
for hybrid node (node No. 2) EV charging load prediction set for an IEEE33 node distribution system, fig. 8 shows an evaluation index case of the model when the EPOCH is set to 300. As can be seen from fig. 8 (a), after training is finished, the difference between the period that the GAT-LSTM model reaches the period that the EPOCH converged by R2 and the period that the load prediction frame model reaches the period that the epoch=110 converged by R2 is about 100 is not great, and the small difference is that the GAT does not gather the edge features, does not need to learn the attention score of the edge features, and has higher convergence rate; the EPOCH of the GCN-LSTM model without the attention mechanism reaching the upper limit of R2 is near the EPOCH=140 and is obviously larger than the model introducing the attention mechanism, and the R2 value of the GAT-LSTM is near 0.92 and is better than that of the GCN-LSTM, so that the improvement is about 4 percent. The reason for the phenomenon is that when the number of nodes or the characteristics contained in the graph structure is large, the attention mechanism can learn the attention scores of different topological characteristics in the training process, key topological characteristics are given to higher attention scores in complex topological characteristics, and the feedback to the model is the optimization of parameters and the efficient extraction of the characteristics. The R2 value of the EGAT-LSTM charging station load prediction framework model reaches 0.93, which is improved by about 1% compared with the R2 value of the GAT-LSTM model. As shown in FIGS. 8 (b) and (c), in the evaluation indexes of MAE and MAPE, the EGAT-LSTM charging station load prediction framework model is better than the GAT-LSTM and GCN-LSTM prediction models, and the error is smaller. The reason for this phenomenon is that in the process of upgrading and reforming a line, resistance and reactance values are reduced, line loss is changed, and under the scheduling strategy of optimal power flow, node characteristics are changed beyond regression characteristics, so that the prediction effect of a model is affected. Compared with GAT, the EGAT feature extraction layer integrates the edge features into the aggregation process, the model learning edge feature change condition can correct the non-regression feature change of the node features by optimizing parameters in the aggregation process, the influence of the line upgrading and transformation process on the prediction model is reduced, and the prediction precision is ensured.
The fitting of 100 time points during the test set test selected herein is shown in fig. 9. The ordinate is the normalized load value, the red line is the true value, and the blue line is the model predictive value. Since the data set is generated by considering the fluctuation of other loads, EV charging loads and the increase in the time dimension, the fluctuation amount is irregular, no external factor features are used for learning the models, and the fact that the two models are poorly fitted at the peak of the truth curve can be obviously seen.
The fitted curve of the EGAT-LSTM charging station load prediction framework model shows that the curve trend is similar to the true curve trend, points with insufficient fitting degree are mostly possibly smaller than the true value, the fluctuation range is small, and the phenomenon of burrs near 60 time points is seldom generated; however, the predicted value of GAT fluctuates greatly compared with the true value, for example, the predicted result near the time points of 20, 40, 70 and 95 is changed in a manner of floating up and down sharply on the true value, a lot of burrs appear, and the curve is not smooth enough, which causes difficulty in unit scheduling.
Step 7-2, the simulation result of the IEEE69 node power distribution system is shown as follows:
the IEEE33 node power distribution system is expanded to an IEEE69 node power distribution system, the number of nodes and edges is increased, the topological characteristics are more abundant, and the structure is more complex. For hybrid node (node No. 2) EV charging load prediction set for an IEEE69 node distribution system, fig. 10 shows an evaluation index case of the model when the EPOCH is set to 300. According to the simulation result in the IEEE33 node power distribution system shown in fig. 10 (a), the convergence of the GCN-LSTM model R2 is far greater than that of the GAT-LSTM model and the EGAT-LSTM charging station load prediction framework model which draw attention, and the results of the EGAT-LSTM charging station load prediction framework model in all evaluation indexes are obviously superior to those of other models, the R2 value reaches about 0.95, and the improvement is about 3% compared with the R2 value of the GAT-LSTM model of 0.92. Referring to FIGS. 10 (b) and (c), the EGAT-LSTM charging station load prediction framework model is superior to the GAT-LSTM and GCN-LSTM prediction models in both the MAE and MAPE evaluation metrics.
Similar to the fitted curve illustration in the IEEE33 node system simulation, the simulated fitted curve illustration in the IEEE69 node system is shown in FIG. 11. In the IEEE69 node system, the number of topological features is increased compared with that of IEEE33 nodes, the model is easier to mine the correlation between the predicted object and different features, so that the model prediction curve does not have the phenomenon of burrs opposite to the true curve in the IEEE33 node system, but the fitting degree of the EGAT-LSTM charging station load prediction framework prediction model is obviously better than that of the GAT-LSTM prediction model in most time points.
Step 7-3, the simulation comprehensive analysis of the distribution system based on the IEEE33 and the IEEE69 nodes is as follows:
with the development of a power distribution network, the topology structure is more and more complex, namely the topology is continuously expanded, the number of nodes and edges is continuously increased, the types of features are possibly more and more, and the extraction of key topology features by a model without a attention mechanism is more and more difficult. It is clear that training the GCN-LSTM model on an IEEE69 node distribution system with a larger and more complex topology achieves R2 converged EPOH around 160, while the GAT-LSTM model achieves R2 converged EPOH around 90, the R2 converged EPOH gap between GCN and GAT is larger than that of IEEE33 nodes. In the fitting precision, the R2 value of the GAT-LSTM is improved by about 5% compared with that of the GCN-LSTM model in the IEEE69 node distribution system, and in the simulation of the IEEE33 node distribution system, the R2 value of the GAT-LSTM is improved by about 4% compared with that of the GCN-LSTM model, and other evaluation indexes also show similar rules. Therefore, as the power distribution network is upgraded, the difference between the model without the attention mechanism and the prediction effect of the model with the attention mechanism is further enlarged. In the future, when there is a need for training an EPOCH in the context of more and more complex power distribution network topologies, we recommend that the use of a predictive model that includes an attention mechanism can guarantee a certain prediction accuracy with fewer training rounds.
The line upgrading and reconstruction work is also one of important performances of the development of the power distribution network. With the expansion of the topology of the power distribution network, the number of lines is increased continuously, the loss difference of the network lines is larger before and after upgrading and transformation, the optimal power flow is redistributed, so that the variation of non-regression characteristics on key characteristics is larger than that of small topology, and in more complex topology, the limitation of the GAT-LSTM model is amplified. Simulation results show that in the IEEE33 node power distribution system, the R2 value of the EGAT-LSTM charging station load prediction framework model is improved by about 1% compared with that of the GAT-LSTM model, while in the IEEE69 node power distribution system simulation, the R2 value of the EGAT-LSTM charging station load prediction framework model is improved by about 3% compared with that of the GAT-LSTM model, other evaluation indexes also show similar rules, and the difference of fitting precision of the two models is further increased along with topology change.
The foregoing describes specific embodiments of the present application. It is to be understood that the application is not limited to the particular embodiments described above, and that various changes or modifications may be made by those skilled in the art within the scope of the appended claims without affecting the spirit of the application. The embodiments of the application and the features of the embodiments may be combined with each other arbitrarily without conflict.
Claims (6)
1. A charging station load prediction method considering power distribution network development is characterized by comprising the following steps:
step 1, performing optimal power flow calculation on the load increase and the power distribution network upgrading process of an example power distribution network through a Matpower tool box in MATLAB software to obtain training data (characteristics and labels), wherein the characteristics comprise nodes and edge characteristics) of the neural network;
step 2, preprocessing the node and edge feature training data obtained in the step 1, inputting the data into an EGAT feature extraction layer in an EGAT-LSTM charging station load prediction framework, fusing line parameter change information into the node in the EGAT feature extraction layer, and outputting the data as new node features fused with edge features;
step 3, inputting the new node characteristics obtained in the step 2 into an LSTM charging load prediction layer in an EGTA-LSTM charging load prediction framework to obtain a predicted charging station load prediction result;
and 4, carrying out back propagation learning in an EGAT-LSTM charging station load prediction framework by matching the prediction result obtained in the step 3 with a load label to obtain global optimal model parameters considering node and edge characteristic changes.
2. The charging station load prediction method considering power distribution network development according to claim 1, wherein the specific process of step 1 is as follows:
Step 1-1, different types of power distribution network load generation methods comprise the steps of simulating and generating a daily load curve which accords with the charging load trend of an electric automobile through a Matlab platform, taking the daily load curve as a base line, and generating charging station load labels in a time scale for training according to up-and-down floating of a certain random number; generating other load quantities of the power distribution network with the same time scale in a similar way;
step 1-2, line transformation information and a generation method thereof related in the upgrading process of a power distribution network, wherein the line transformation information comprises the steps of modifying line parameters (resistance and reactance values) under different construction periods in an IEEE standard power distribution network topology under a Matpower tool kit in Matlab so as to simulate line transformation conditions under different power supply periods;
step 1-3, generating training data of an EGAT-LSTM charging station load prediction framework, which comprises the steps 1-1 and 1-2, using an own optimal power flow calculation module in a Matpower tool box to perform optimal power flow calculation of a training time scale on a selected standard IEEE power distribution network topology, and storing node voltage, active and reactive power and resistance and reactance of a power supply period where a line is located in an obtained optimal power flow result as characteristics of a training data set.
3. A charging station load prediction method considering the development of a power distribution network according to claim 2, wherein the optimal power flow calculation is shown as follows:
Step 1, setting the iteration number k=0 and selecting a proper initial valueWherein k is a counter value; />For iterating k times node voltage and phase angle, +.>Active and reactive power output of the unit is iterated k times;
step 2, solving the constraint equation of the equationGet solution->Wherein g is the equality constraint satisfied by the grid;
step 3, inLinearizing the problem nearby, and carrying out linear programming solution on the variation delta x of x, wherein the step size limit delta is less than or equal to delta x and less than or equal to delta is required to be met;
step 4, taking k=k+1 and updating the current solution x k ＝x k -1+Δx, if x k Stopping when the boundary is exceeded, otherwise, entering the step 5;
step 5, adjusting the step size limit delta to return to step 2.
4. The charging station load prediction method considering power distribution network development according to claim 1, wherein the specific process of step 2 is as follows:
in order to improve training efficiency and match with the input dimension of a subsequent LSTM charging load prediction layer, after the training data obtained in the step 1 are packaged according to the sequence length of the LSTM charging load prediction layer, a plurality of subgraphs are constructed into a large graph, and the large graph is input to an EGAT feature extraction layer in parallel; aggregating the line parameter information to the nodes by utilizing the EGAT feature extraction layer, extracting the node and the side feature information at the same time, and describing the change condition of the topology attribute more comprehensively; the aggregation process of line parameter aggregation to nodes is mainly divided into attention score calculation and node aggregation, which comprises the following steps:
(1) The attention score was calculated as follows:
the attention score after the side characteristic information is aggregated is:
in the above, gamma' ji,p Is the attention score after the side characteristic information is aggregated, h i Is the feature vector of the node i,is node j epsilon N i I.e. node voltage, active power and reactive power; r is (r) ji Is the edge eigenvector between connection nodes i and j, namely line resistance and reactance; />Is the next learnable parameter of the multi-head attention p head and can realize the characteristic vector from f η Direction of the dimension->Transforming; />Is subjected to a Softmax normalization operation; the g I G is to splice the feature vector; />To learn parameters under the multi-head attention p head; />To learn parameters under the multi-head attention p head, the edge feature dimension is realized by f e To->Is a variation of (2); it is noted that all nodes of the same layer in the EGAT feature extraction layer share the same parameter matrix +.>And->N i A set of contiguous nodes that is node i; the LeakyReLU is a nonlinear activation function that is more commonly used in deep learning, and its expression is as follows:
wherein a is a coefficient, and x represents a part of the input nonlinear activation function;
(2) The node and polygon properties are aggregated as follows:
in the above, h i The new characteristic representation of the node i after the node i aggregates the neighbor node characteristics and the edge characteristics between the neighbor node characteristics; p is the number of multi-head attention heads; Is a learnable parameter that causes an edge feature vector r ji Linear transformation is carried out before node aggregation, and the dimension of edge characteristics is realized by f e To->Is a variation of (2); thus, the training data is obtained with new node characteristics containing side information, and the dimension of the new node characteristics can be set in a certain range.
5. The charging station load prediction method considering the development of the power distribution network according to claim 1, wherein the specific process of the step 3 is as follows:
the new node characteristics obtained in the step 2 are flattened and subjected to dimensional transformation to obtain a dimensional form suitable for inputting the LSTM charging load prediction layer, and the number of hidden layers of the LSTM charging load prediction layer is set according to the characteristic number set by the new node characteristics; and inputting the new node characteristics after dimension transformation into an LSTM charging load prediction layer, and obtaining a predicted value of the predicted charging station load through parameter and bias calculation of a hidden layer.
6. The charging station load prediction method considering power distribution network development according to claim 1, wherein the specific process of step 4 is as follows:
carrying out counter propagation operation on the predicted result value obtained in the step 3 in combination with a truth value tag in an EGAT-LSTM charging station load prediction framework, updating the coefficient of an attention score calculation module in an EGAT feature extraction layer, searching key features in topology, and enhancing the regression of the key features and the tag; and meanwhile, the opening degrees of the input door, the forgetting door and the output door in the LSTM charging load prediction layer are updated, so that the proportion of the current time information to the historical information can be controlled, and the prediction precision is improved.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310235422.6A CN116596105A (en) | 2023-03-13 | 2023-03-13 | Charging station load prediction method considering power distribution network development |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310235422.6A CN116596105A (en) | 2023-03-13 | 2023-03-13 | Charging station load prediction method considering power distribution network development |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116596105A true CN116596105A (en) | 2023-08-15 |
Family
ID=87592592
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310235422.6A Pending CN116596105A (en) | 2023-03-13 | 2023-03-13 | Charging station load prediction method considering power distribution network development |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116596105A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116894163A (en) * | 2023-09-11 | 2023-10-17 | 国网信息通信产业集团有限公司 | Charging and discharging facility load prediction information generation method and device based on information security |
-
2023
- 2023-03-13 CN CN202310235422.6A patent/CN116596105A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116894163A (en) * | 2023-09-11 | 2023-10-17 | 国网信息通信产业集团有限公司 | Charging and discharging facility load prediction information generation method and device based on information security |
| CN116894163B (en) * | 2023-09-11 | 2024-01-16 | 国网信息通信产业集团有限公司 | Charging and discharging facility load prediction information generation method and device based on information security |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103293487B (en) | Based on the lithium ion battery life-span prediction method of integrated model | |
| CN110007235A (en) | A kind of accumulator of electric car SOC on-line prediction method | |
| Li et al. | A Kriging-based bi-objective constrained optimization method for fuel economy of hydrogen fuel cell vehicle | |
| CN112101684A (en) | Plug-in hybrid electric vehicle real-time energy management method and system | |
| CN113822481A (en) | Comprehensive energy load prediction method based on multi-task learning strategy and deep learning | |
| Shao et al. | The traffic flow prediction method using the incremental learning-based CNN-LTSM model: the solution of mobile application | |
| CN105809349A (en) | Scheduling method considering incoming water correlation cascade hydropower stations | |
| CN103698627A (en) | Transformer fault diagnostic method based on gray fuzzy firefly algorithm optimization | |
| Yang et al. | Market evolution modeling for electric vehicles based on system dynamics and multi-agents | |
| CN110826244A (en) | Conjugate gradient cellular automata method for simulating influence of rail transit on urban growth | |
| CN116596105A (en) | Charging station load prediction method considering power distribution network development | |
| CN109583588A (en) | A kind of short-term wind speed forecasting method and system | |
| CN113406503A (en) | Lithium battery SOH online estimation method based on deep neural network | |
| CN115907122A (en) | Regional electric vehicle charging load prediction method | |
| CN114707292A (en) | Voltage stability analysis method for power distribution network containing electric automobile | |
| CN104217296A (en) | Listed company performance comprehensive evaluation method | |
| Poczeta et al. | Application of fuzzy cognitive maps to multi-step ahead prediction of electricity consumption | |
| Mirjalili et al. | A comparative study of machine learning and deep learning methods for energy balance prediction in a hybrid building-renewable energy system | |
| Wang et al. | Improved hybrid fuzzy logic system for evaluating sustainable transportation systems in smart cities | |
| CN115577647B (en) | Power grid fault type identification method and intelligent agent construction method | |
| CN107491841A (en) | Nonlinear optimization method and storage medium | |
| Nutkiewicz et al. | Exploring the integration of simulation and deep learning models for urban building energy modeling and retrofit analysis | |
| CN115392143A (en) | Mobile energy storage charging and discharging space-time planning method based on deep reinforcement learning | |
| Li et al. | DiffCharge: Generating EV Charging Scenarios via a Denoising Diffusion Model | |
| Du et al. | AGRU and convex optimization based energy management for plug-in hybrid electric bus considering battery aging |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |