GNN-based'network-source-storage-vehicle 'cooperative energy supply system energy management and control method

Abstract

The invention discloses a GNN-based'network-source-storage-vehicle 'cooperative energy supply system energy management and control method, and the method comprises the steps: carrying out the working condition division of operation subjects, such as new energy, energy storage and electric locomotives, constructing a graph neural network model based on an optimization model, extracting the spatial-temporal feature information through a graph convolutional network and an attention mechanism, and carrying out the calculation of the spatial-temporal feature information. Real-time advanced energy management and control are realized, and online self-updating learning of the graph neural network model is realized through an online learning strategy; the energy management method is used in a network-source-storage-train cooperative function system of an electrified railway, and energy exchange is completed by controlling a power fusion device and a DC / DC converter. According to the method, the operation conditions of different main bodies can be accurately predicted, the local absorption capacity of a new energy system and the absorption efficiency of regenerative braking energy are effectively improved, the power impact of the peak power of an electric locomotive on a traction network is reduced, and technical support is provided for efficient and stable operation of a cooperative energy supply system.

Description

Technical Field

[0001] This invention belongs to the field of electrified railway technology, specifically relating to an energy management method for a "grid-source-storage-vehicle" collaborative energy supply system based on GNN. Background Technology

[0002] Electrified railways possess numerous significant advantages, including clean energy, large transport capacity, and low operating costs, making them a crucial component for enhancing national productivity and developing the national economy. However, the rapid expansion of railway networks has also generated greater demand for electricity, making electrified railway systems a major source of carbon emissions. In the context of global carbon neutrality, achieving energy conservation and emission reduction in railway systems, optimizing energy consumption, and lowering railway operating costs are of paramount importance. The traction power supply system is the sole source of energy for electric locomotives; however, traditional traction power supply systems have a simple structure, lack DC interfaces, and cannot directly interact with energy storage systems or new energy systems.

[0003] To address the above issues, integrating new energy sources and energy storage systems into the traction power supply system to construct a collaborative energy supply system for rail transit, encompassing the grid, energy source, storage, and vehicle, is a crucial pathway for the green and low-carbon development of rail transit. The primary method involves connecting new energy sources and energy storage to the existing traction power supply system via back-to-back railway power regulators, achieving interconnection and interoperability of energy from new energy sources, energy storage, and electric locomotives. This approach offers significant advantages such as minimal engineering modifications and low construction costs. However, with the integration of new energy sources and energy storage systems, the energy flow direction of the interactions among multiple entities (new energy sources, energy storage, and electric locomotives) becomes complex and variable. Furthermore, the load on electric locomotives exhibits impulsive and time-varying characteristics, and the output of new energy sources is subject to fluctuations and uncertainties, significantly increasing the difficulty of energy management in the collaborative energy supply system. Therefore, achieving energy management among these multiple entities is one of the core issues in realizing the efficient utilization of energy in a collaborative energy supply system.

[0004] Based on existing literature, energy management strategies for "grid-source-storage-vehicle" collaborative energy supply systems can be broadly categorized into two types: rule-based energy management strategies and optimization-based energy management strategies. Rule-based energy management strategies offer fast response times and simple energy control, making them suitable for practical engineering applications. However, these methods rely heavily on the designer's subjective experience, resulting in low energy regulation accuracy. Optimization-based energy management strategies, while offering high control accuracy and strong learning capabilities, suffer from slow response times, difficult solutions, and a lack of real-time online control capabilities, limiting their practical application. Comparing these two approaches, rule-based energy management strategies, with their fast response times, simple algorithms, and lack of complex model solutions, have greater practical engineering value. However, improving the accuracy of rule-based energy management while maintaining fast response times is of significant practical importance. Furthermore, neither rule-based nor optimization-based energy management strategies possess real-time energy control capabilities nor autonomous learning capabilities.

[0005] The patent "An Energy Management Method for Electrified Railway Energy Storage Traction Power Supply Based on Energy Router" (Publication No.: CN219513799U) proposes an energy management method based on an energy router. This method aims at negative sequence compensation and achieves energy management by setting a discharge threshold. However, this method requires customized adaptation for specific traction substations, lacking universality and portability, thus limiting its application in different scenarios. The patent "An Electrified Railway In-Phase Energy Storage Power Supply Device and Its Control Method" (Publication No.: CN109936135B) proposes an energy management method for electrified railways suitable for in-phase energy storage power supply devices. This method sets charging and discharging thresholds based on historical power demand curves and is applicable to in-phase energy storage power supply devices. Although the method is simple and intuitive, it lacks an adaptive control mechanism, resulting in insufficient flexibility and difficulty in coping with complex or dynamic operating conditions. The patent "An Energy Management Method for Electrified Railway Energy Storage System Based on Reinforcement Learning" (Publication No.: CN116316755B) proposes an energy management method based on reinforcement learning. This method uses reinforcement learning algorithms for energy management, which can improve control accuracy. However, as accuracy improves, the output power gradient of the energy storage system increases, leading to a rapid expansion of the action-state space, which prolongs training time and makes it difficult to meet real-time requirements. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention proposes an energy management method for a "grid-source-storage-vehicle" collaborative energy supply system based on GNN. This method first divides the operating conditions of the main entities (new energy, energy storage, and electric locomotives), and then constructs a graph neural network model based on an optimization model. It utilizes graph convolutional networks and attention mechanisms to extract spatiotemporal feature information, establishing a multi-level feature extraction network structure. By comparing real-time sample data and employing an online learning strategy to dynamically update the adjacency matrix and add characteristic information, it achieves real-time, efficient, and proactive energy management, as well as online self-updating learning of the graph neural network model. This accurately predicts the operating conditions of different entities, effectively improving the local absorption capacity of the new energy system and the absorption efficiency of regenerative braking energy, and reducing the power impact of electric locomotive peak power on the traction network.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] The energy management method for the "grid-source-storage-vehicle" coordinated energy supply system based on GNN is based on a graph neural network model. It utilizes graph convolutional networks and attention mechanisms to achieve real-time proactive energy management and achieves online self-updating learning of the graph neural network model through an online learning strategy. The energy management method is used in the "grid-source-storage-vehicle" coordinated functional system of electrified railways, and completes energy exchange by controlling the power fusion device and DC / DC converter.

[0009] The specific steps for constructing the graph neural network model include:

[0010] S111: Load power data from several time periods of the collaborative energy supply system, and divide the operating conditions of each operating entity from a data-driven perspective. The operating entities include at least new energy sources, energy storage, and electric locomotives.

[0011] S112: Based on the results of the division of the main operating conditions, different operating conditions of the main operating conditions are regarded as nodes, and the transfer relationship between different operating conditions is regarded as edges. A weighted and directed operating condition transfer network graph composed of nodes and edges is constructed.

[0012] S113: Construct an optimization model to optimize the working condition transfer relationship between the operating entities, obtain the working condition transfer network diagram of the operating entities, and output the theoretically optimal energy control result obtained through the optimization model;

[0013] S114: Construct a graph neural network model based on the operating condition transfer network diagram of the main operating entity;

[0014] The specific steps for achieving real-time advance energy management using graph convolutional networks and attention mechanisms include:

[0015] S121: Based on the graph neural network model, load the adjacency matrix and feature matrix;

[0016] S122: Extract local spatial feature information through the first graph convolutional layer and output the extracted feature matrix;

[0017] S123: The attention mechanism compensates for the high-dimensional feature weighting, calculates the attention weights between all node pairs, obtains global interaction information, and extracts temporal feature information. The node pair includes two nodes and one edge, reflecting the state transition relationship between the two nodes in the node pair.

[0018] S124: Extract comprehensive spatiotemporal features through the second graph convolutional layer;

[0019] S125: Extract global features through the third graph convolutional layer;

[0020] S126: The energy management layer predicts the operating conditions of each operating entity at the next moment based on the feature matrix output by the last graph convolutional layer of the graph convolutional network and attention mechanism, further predicts the power of the electric locomotive, and outputs the energy management results after the graph neural network model. The energy management results include at least the optimal output power of the new energy system and the charging and discharging power of the energy storage system at the next moment.

[0021] S127: If a stop command is received, the process ends; if no stop command is received, repeat S121 to S126.

[0022] The specific steps for implementing online self-updating learning of graph neural network models using online learning strategies include:

[0023] S131: Load the energy control results obtained through the graph neural network model, and the theoretically optimal energy control results obtained through the optimization model;

[0024] S132: After calculating the mean squared error loss function, determine the control accuracy of the graph neural network model;

[0025] S133: Compare whether the result of the loss function exceeds the set threshold;

[0026] S134: If the set threshold is not exceeded, no processing is performed; if the set threshold is exceeded, the data group is added to the data buffer.

[0027] S135: Update the adjacency matrix, add feature information, and wait for the next training update of the graph neural network model.

[0028] Furthermore, the optimization model uses maxf ％ To describe it, as in equation (A1):

[0029]

[0030] In formula (A1), enes％ For the energy utilization efficiency of new energy systems, To improve the absorption efficiency of regenerative braking energy in energy storage systems, Energy feedback efficiency of energy storage systems;

[0031] The energy utilization efficiency e of the new energy system nes％ The calculation is as shown in equation (A2):

[0032]

[0033] In formula (A2), Let be the power output of the renewable energy system to the external power grid at time t. Let be the power output from the new energy system to the energy storage system at time t. Let be the power output of the new energy system to the electric locomotive at time t. Let t be the maximum power that the renewable energy system can output at time t, and m be the length of the historical power data sequence.

[0034] The energy storage system's regenerative braking energy absorption efficiency The calculation is as shown in equation (A3):

[0035]

[0036] In formula (A3), The charging power of the energy storage system at time t. Let t be the regenerative braking power of the electric locomotive;

[0037] Energy feedback efficiency of the energy storage system The calculation is as shown in equation (A4):

[0038]

[0039] In formula (A4), Let t be the discharge power of the energy storage system at time t.

[0040] Furthermore, by stacking the first, second, and third graph convolutional layers, deep features of nodes are extracted step by step to capture local dependencies between the running entities.

[0041] The specific method for extracting aggregated spatial features using the first graph convolutional layer to obtain the first layer's spatial feature information and output feature matrix is ​​as follows: Aggregate the spatial features of each node at time t to generate a preliminary representation at time t+1, and then use a learnable weight matrix W. (l) A linear transformation is performed, and a nonlinearity is introduced by a nonlinear activation function σ to achieve aggregate spatial feature extraction of the graph convolutional layer, outputting the first graph convolutional layer feature matrix H. (l) As shown in equation (A5):

[0042]

[0043] In equation (A5), σ is the nonlinear activation function ReLU. Let A be the angle matrix after adding the self-loop, and let A be the characteristic matrix. D is the feature matrix of the previous layer. l-1 For feature dimension, D is a learnable weight matrix. l It is the output feature dimension of the current layer, where the initial input feature H (0) =X∈R N×F F = 2.

[0044] Furthermore, to extract temporal feature information at a deeper level, local spatial feature information is first extracted through a first-layer graph convolutional network, and the feature matrix H is output. (1) Next, the attention mechanism is based on H (1) By using high-dimensional feature weighting to compensate, the attention weights between all node pairs are calculated to obtain the global interaction information H. (2) Extracting temporal feature information. The specific calculation process of the attention mechanism includes:

[0045] S141: Through linear transformation, the feature matrix H of the first convolutional layer is transformed... (l) Projecting onto three different spaces generates corresponding query matrix Q, key matrix K, and value matrix V, as shown in equation (A6):

[0046]

[0047] In formula (A6), There are three learnable projection matrices, d k =D1 / h is the feature dimension of each attention head, and h is the number of attention heads. This is achieved by learning W... Q W K W V Three weight matrices yield three shapes. The new matrices Q, K, V.

[0048] S142: To measure the attention level of the query matrix Q to the key matrix K, click operations are performed on both the query matrix Q and the key matrix K, and the result is divided by a scaling factor. Calculate the attention score and obtain the attention weight matrix F. attn As shown in equation (A7):

[0049]

[0050] In equation (A7), the attention weight matrix F attnThe attention score is converted into a probability distribution, representing the degree of attention each node pays to other nodes.

[0051] S143: Obtain the attention weight matrix F attn Next, information from other nodes needs to be fused to generate a new feature representation for each node, and the single-head attention output h needs to be calculated. i As shown in equation (A8):

[0052] h i =F attn V(A8)

[0053] Concatenate all single-head attention outputs along the feature dimension to fuse the multi-head attention, resulting in a multi-head attention concatenation matrix. Simultaneously, the multi-head attention concatenation matrix H′ is mapped to the original dimension to obtain the output feature matrix H. attn As shown in equation (A9):

[0054]

[0055] In formula (A9), To output the projection matrix, the feature matrix H output by the multi-head attention mechanism attn This is the input to the second graph convolutional layer, H (2) =H attn Substitute into equation (A5) and continue the convolution operation.

[0056] Furthermore, spatial features are extracted through graph convolutional layers and temporal features are extracted through a multi-head attention mechanism. The energy control layer receives the feature information extracted by the graph convolutional layers and the attention mechanism, and outputs the final feature matrix H. (L) The specific process of predicting the operating condition of the main body at the next moment, using a fully connected architecture, is shown in equation (A10):

[0057] M = softmax(W) m H (L) +b m (A10)

[0058] In formula (A10), W m b m These are all parameters of the classification layer, M∈R N×K Let K be the probability distribution of working conditions for each node, where K is the number of working condition categories and N is the total number of working conditions.

[0059] Furthermore, based on the received feature matrix, the energy management layer predicts the power of the electric locomotive, the optimal output power of the photovoltaic system and the charging and discharging power of the energy storage system at the next moment, according to the prediction results of the operating conditions of the main body. The specific power output process is shown in equation (A11):

[0060]

[0061] In formula (A11), All are learnable parameters, P∈R N×1 This represents the output power value.

[0062] Furthermore, to measure the error between the test value and the true value and improve the energy management effect of the graph neural network, it is necessary to calculate the loss function of the graph neural network. Since the graph neural network needs to predict the operating conditions of the same entity and manage the energy flow of the coordinated energy supply system, in order to fully utilize the potential correlation between the operating conditions and power values ​​of different entities, a weighted combination of the operating condition prediction loss and the energy management result loss is used to describe the loss function of the graph neural network model, as shown in equation (A12):

[0063] L=λ(L NES +L ESS +L EL )+L MSE (A12)

[0064] In equation (A12), the predicted loss under the working condition is represented by λ(L). NES +L ESS +L EL ) to describe, L NES L ESS L EL The predicted losses for photovoltaic systems, energy storage systems, and electric locomotives are L, respectively. MSE λ represents the energy management result loss, and is a hyperparameter used to balance the loss from operating condition prediction and the loss from energy management results.

[0065] The energy management result loss L MSE The quantification is achieved by calculating the squared difference between the test value and the true value using the mean square error, as shown in equation (A13):

[0066]

[0067] In equation (A13), P i,t and ...

[0068] Considering the different data types of output power for operational condition prediction and energy management, the prediction losses for photovoltaic systems, energy storage systems, and electric locomotives are all calculated using cross-entropy loss to determine the difference between the predicted probability distribution and the actual distribution. sys Describe it as shown in equation (A14):

[0069]

[0070] In equation (A14), B is the batch size. This is a unique thermal encoding for real-world operating conditions, where sys represents the photovoltaic system, energy storage system, and electric locomotive, and M... sys,b,i To predict probabilities.

[0071] Furthermore, the online learning strategy uses incremental updates, allowing the model to gradually incorporate new data without losing existing data. This enhances the model's ability to handle anomalies and improves the energy management accuracy of the graph neural network model. The graph neural network model uses data collected at time t to manage energy flow at time t+1. After energy management, the new energy output power, electric locomotive power, and energy storage charging / discharging power at time t+1 are respectively P nes,t+1 P el,t+1 and P ess,t+1 To quantify the accuracy of energy management, the power data of electric locomotives collected at time t+1 is used to obtain the optimal management result through optimization model (Equation A1), and the mean square error loss function is designed as shown in Equation (A15).

[0072]

[0073] In equation (A15), P′ nes,t+1 For optimal output power of new energy, P′ el,t+1 P′ represents the actual power of the electric locomotive. ess,t+1 The optimal charging and discharging power for energy storage.

[0074] Furthermore, a threshold θ is set, if the loss function... If the result exceeds the set threshold θ, the adjacency matrix weights are updated, the feature information is added to the feature buffer, the model is trained according to equation (A16), and the trained graph neural network model replaces the original graph neural network model.

[0075]

[0076] In equation (A16), μ∈(0,1) is the update rate, used to control the balance between the original weights and the new weights, w new For the new transfer weights, F new Let A(i,j) be the original adjacency matrix, A'(i,j) be the updated adjacency matrix, F be the original feature matrix, and F' be the updated feature matrix.

[0077] Compared with the prior art, the present invention has the following technical effects:

[0078] (1) By dividing the operating conditions of the main operating entities in the collaborative energy supply system, an optimization model is used to construct a network diagram of different operating conditions of different operating entities, and then a graph neural network topology structure of interdependence of operating entities is constructed, revealing the characteristics of operating condition transfer within and between each operating entity.

[0079] (2) Based on the graph neural network topology, spatiotemporal feature information was extracted by using graph convolutional networks and multi-head attention mechanisms, and a multi-level feature extraction network structure was established to achieve real-time, efficient and advanced energy control.

[0080] (3) In order to enhance the applicability and robustness of the graph neural network model, the present invention sets up an online learning strategy. By collecting actual running data in real time, dynamically updating the adjacency matrix and adding characteristic information, the graph neural network model can be autonomously updated online, ensuring that the energy control results are real-time and optimal.

[0081] Attached Figure

[0082] Figure 1 This is a structural diagram of the "grid-source-storage-vehicle" coordinated energy supply system of the present invention;

[0083] Figure 2 This is a schematic diagram of a three-layer graph convolutional layer structure for a graph neural network model.

[0084] Figure 3 This is a flowchart of the energy management process of the present invention;

[0085] Figure 4 This is a diagram illustrating the online learning strategy of this invention.

[0086] Figure 5 This is a graph showing the actual power data of an electric locomotive at a certain traction substation.

[0087] Figure 6 This is a typical daily power curve for a photovoltaic power generation system;

[0088] Figure 7 This is a comparison chart of electric locomotive power predictions according to the present invention;

[0089] Figure 8 This is a power output diagram of the photovoltaic system of the present invention;

[0090] Figure 9 This is a power output diagram of the energy storage system of the present invention. Detailed Implementation

[0091] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the invention, any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art to all other embodiments obtained without creative effort should be included within the protection scope of the present invention.

[0092] Energy management methods for grid-source-storage-vehicle collaborative energy supply systems based on GNN, such as Figure 1 As shown, the energy management method is used in the "grid-source-storage-vehicle" coordinated functional system of electrified railways. This coordinated functional system includes a traction power supply system, an external power system, a single-phase step-down transformer, a power fusion device, an LCL filter, secondary equipment, an energy management system, an energy storage system, and a new energy system. The traction power supply system adopts a single-phase AC power frequency power supply system. The power fusion device is installed inside the traction substation and connects to the left and right power supply arms of the traction substation. The energy storage system and the new energy system are connected to the DC bus of the power fusion device after passing through DC / DC converters. The energy management system completes energy exchange with the DC / DC converter by controlling the power fusion device.

[0093] The energy management method is based on a graph neural network model, utilizing graph convolutional networks and attention mechanisms to achieve real-time proactive energy control, and employing an online learning strategy to enable online self-updating of the graph neural network model; the graph neural network model is as follows: Figure 2 As shown, the specific steps for constructing a graph neural network model include:

[0094] S111: Load power data from the collaborative energy supply system over several time periods, and classify the operating conditions of various operating entities such as new energy, energy storage, and electric locomotives from a data-driven perspective;

[0095] S112: Based on the results of the division of the main operating conditions, different operating conditions of the main operating conditions are regarded as nodes, and the transfer relationship between different operating conditions is regarded as edges. A weighted and directed operating condition transfer network graph composed of nodes and edges is constructed.

[0096] S113: Construct an optimization model to optimize the working condition transfer relationship between the operating entities, obtain the working condition transfer network diagram of the operating entities, and output the theoretically optimal energy control result obtained through the optimization model;

[0097] The optimization model uses maxf ％ To describe, as shown in equation (S1):

[0098]

[0099] In formula (S1), enes％ For the energy utilization efficiency of new energy systems, To improve the absorption efficiency of regenerative braking energy in energy storage systems, Energy feedback efficiency of energy storage systems;

[0100] The energy utilization efficiency e of the new energy system nes％ The calculation is as shown in equation (S2):

[0101]

[0102] In formula (S2), Let be the power output of the renewable energy system to the external power grid at time t. Let be the power output from the new energy system to the energy storage system at time t. Let be the power output of the new energy system to the electric locomotive at time t. Let t be the maximum power that the renewable energy system can output at time t, and m be the length of the historical power data sequence.

[0103] The energy storage system's regenerative braking energy absorption efficiency The calculation is as shown in equation (S3):

[0104]

[0105] In formula (S3), The charging power of the energy storage system at time t. Let t be the regenerative braking power of the electric locomotive;

[0106] Energy feedback efficiency of the energy storage system The calculation is as shown in equation (S4):

[0107]

[0108] In formula (S4), Let t be the discharge power of the energy storage system at time t.

[0109] S114: Construct a graph neural network model based on the operating condition transfer network diagram of the main operating entity.

[0110] like Figure 3 As shown, the specific steps for achieving real-time advance energy management using graph convolutional networks and attention mechanisms include:

[0111] S121: Based on the graph neural network model, load the adjacency matrix and feature matrix;

[0112] S122: By stacking the first, second, and third graph convolutional layers, deep features of nodes are extracted step by step to capture local dependencies between the running entities; local spatial feature information is extracted through the first graph convolutional layer, and the extracted feature matrix is ​​output; the specific method is: aggregating the spatial features of each node at time t to generate a preliminary representation at time t+1, and then using a learnable weight matrix W (l) A linear transformation is performed, and a nonlinearity is introduced by a nonlinear activation function σ to achieve aggregate spatial feature extraction of the graph convolutional layer, outputting the first graph convolutional layer feature matrix H. (l) As shown in equation (S5):

[0113]

[0114] In equation (S5), σ is the nonlinear activation function ReLU. Let A be the angle matrix after adding the self-loop, and let A be the characteristic matrix. D is the feature matrix of the previous layer. l-1 For feature dimension, D is a learnable weight matrix. l It is the output feature dimension of the current layer, where the initial input feature H (0) =X∈R N×F F = 2.

[0115] S123: To extract temporal feature information at a deeper level, local spatial feature information is first extracted through a first-layer graph convolutional network, and a feature matrix is ​​output. Then, the attention mechanism is based on H... (1) By using high-dimensional feature weighting to compensate, the attention weights between all node pairs are calculated to obtain the global interaction information H. (2) Extract temporal feature information.

[0116] The specific steps of the attention mechanism include:

[0117] S1231: Through linear transformation, the feature matrix H of the first convolutional layer is transformed... (l) Projecting onto three different spaces generates corresponding query matrix Q, key matrix K, and value matrix V, as shown in equation (S6):

[0118]

[0119] In formula (S6), There are three learnable projection matrices, d k =D1 / h is the feature dimension of each attention head, and h is the number of attention heads. This is achieved by learning W... Q W K W V Three weight matrices yield three shapes. The new matrices Q, K, V.

[0120] S1232: To measure the attention level of the query matrix Q to the key matrix K, click operations are performed on both the query matrix Q and the key matrix K, and then divided by a scaling factor. Calculate the attention score and obtain the attention weight matrix F. attn As shown in equation (A7):

[0121]

[0122] In equation (S7), the attention weight matrix F attn The attention score is converted into a probability distribution, representing the degree of attention each node pays to other nodes.

[0123] S1233: Obtain the attention weight matrix F attn Next, information from other nodes needs to be fused to generate a new feature representation for each node, and the single-head attention output h needs to be calculated. i As shown in equation (S8):

[0124] h i =F attn V(S8)

[0125] Concatenate all single-head attention outputs along the feature dimension to fuse the multi-head attention, resulting in a multi-head attention concatenation matrix. Simultaneously, the multi-head attention concatenation matrix H′ is mapped to the original dimension to obtain the output feature matrix H. attn As shown in equation (S9):

[0126]

[0127] In formula (S9), To output the projection matrix, the feature matrix H output by the multi-head attention mechanism attn This is the input to the second graph convolutional layer, H (2) =H attn Substitute into equation (S5) and continue the convolution operation.

[0128] S124: Extract comprehensive spatiotemporal features through the second graph convolutional layer;

[0129] S125: Extract global features through the third graph convolutional layer;

[0130] S126: Spatial features are extracted through graph convolutional layers and temporal features are extracted through a multi-head attention mechanism. The energy control layer receives the feature information extracted by the graph convolutional layers and the attention mechanism, and outputs the final feature matrix H. (L) Predict the operating conditions of the main body at the next moment.

[0131] The specific process of using full connectivity to predict the operating conditions of the main body is shown in equation (S10):

[0132] M = softmax(W) m H (L) +b m (S10)

[0133] In formula (S10), W m b m These are all parameters of the classification layer, M∈R N×K Let K be the probability distribution of working conditions for each node, where K is the number of working condition categories and N is the total number of working conditions.

[0134] Based on the received feature matrix, the energy management layer further predicts the power of the electric locomotive, the optimal output power of the photovoltaic system and the charging and discharging power of the energy storage system at the next moment, according to the prediction results of the main operating conditions. The specific power output process is shown in equation (S11):

[0135]

[0136] In formula (S11), All are learnable parameters, P∈R N×1 This represents the output power value.

[0137] S127: If a stop command is received, the process ends; if no stop command is received, repeat S121 to S126.

[0138] Through such Figure 4 The online learning strategy shown enables online self-updating learning of the graph neural network model. This strategy uses incremental updates, allowing the model to gradually incorporate new data without losing existing data, enhancing its ability to handle anomalies and improving the energy management accuracy of the graph neural network model. The graph neural network model uses data collected at time t to manage energy flow at time t+1. Specific steps include:

[0139] S131: Load the energy control results obtained through the graph neural network model, and the theoretically optimal energy control results obtained through the optimization model;

[0140] S132: To measure the error between the test value and the true value, determine the control accuracy of the graph neural network model, and improve the energy control effect of the graph neural network, it is necessary to calculate the loss function of the graph neural network. Since the graph neural network needs to predict the operating conditions of the same entity and control the energy flow of the coordinated energy supply system, in order to fully utilize the potential correlation between the operating conditions and power values ​​of different entities, a weighted combination of the operating condition prediction loss and the energy control result loss is used to describe the loss function of the graph neural network model, as shown in equation (S12):

[0141] L=λ(L NES +L ESS +L EL )+L MSE (S12)

[0142] In equation (S12), the predicted loss under the working condition is represented by λ(L). NES +L ESS +L EL ) to describe, L NES L ESS L EL The predicted losses for photovoltaic systems, energy storage systems, and electric locomotives are L, respectively. MSE λ represents the energy management result loss, and is a hyperparameter used to balance the loss from operating condition prediction and the loss from energy management results.

[0143] The energy management result loss L MSE The quantification is achieved by calculating the squared difference between the test value and the true value using the mean square error, as shown in equation (S13):

[0144]

[0145] In formula (S13), P i,t and ...

[0146] Considering the different data types of output power for operational condition prediction and energy management, the prediction losses for photovoltaic systems, energy storage systems, and electric locomotives are all calculated using cross-entropy loss to determine the difference between the predicted probability distribution and the actual distribution. sys The description is as follows, as shown in equation (S14):

[0147]

[0148] In equation (S14), B is the batch size. This is a unique thermal encoding for real-world operating conditions, where sys represents the photovoltaic system, energy storage system, and electric locomotive, and M... sys,b,i To predict probabilities.

[0149] After energy management, at time t+1, the output power of new energy sources, the power of electric locomotives, and the charging and discharging power of energy storage are respectively P nes,t+1 P el,t+1 and P ess,t+1 To quantify the accuracy of energy management, the power data of electric locomotives collected at time t+1 is used to obtain the optimal management result through model optimization (Equation S1), and the mean square error loss function is designed as shown in Equation (S15).

[0150]

[0151] In equation (S15), P′ nes,t+1 For optimal output power of new energy, P′ el,t+1 P′ represents the actual power of the electric locomotive. ess,t+1 The optimal charging and discharging power for energy storage.

[0152] S133: Compare whether the loss function result exceeds the set threshold θ. If the loss function... If the result exceeds the set threshold θ, the adjacency matrix weights are updated, the feature information is added to the feature buffer, the model is trained according to equation (S16), and the trained graph neural network model replaces the original graph neural network model.

[0153]

[0154] In equation (S16), μ∈(0,1) is the update rate, used to control the balance between the original weights and the new weights, w new For the new transfer weights, F new Let A(i,j) be the original adjacency matrix, A'(i,j) be the updated adjacency matrix, F be the original feature matrix, and F' be the updated feature matrix.

[0155] S134: If the set threshold θ is not exceeded, no processing is performed; if the set threshold is exceeded, the data group is added to the data buffer.

[0156] S135: Update the adjacency matrix, add feature information, and wait for the next training update of the graph neural network model.

[0157] against Figure 1 The illustrated "grid-source-storage-vehicle" coordinated energy supply system assumes that the energy storage system uses batteries as the energy storage medium and the new energy power generation system uses a photovoltaic system. Taking the actual power data of a certain electric locomotive at a certain traction substation on a single day as an example, such as... Figure 5 As shown, using the typical daily output power of photovoltaic systems as the output power data, the power curve for a certain day is as follows: Figure 6 As shown in Table 1, the parameters of the collaborative energy supply system are as follows.

[0158] Table 1 Parameter Table of Collaborative Energy Supply System

[0159]

[0160] Through actual experimental verification, experimental results were obtained. To better observe the predicted results, a continuous 30-minute period of power data was selected as the observation period. Figure 7It can be seen that the predicted power and the actual power are basically consistent. Furthermore, comparing the electric locomotive power prediction results of this method with those of the BP neural network reveals that the prediction results of this method are more accurate, have smaller errors, and exhibit better power tracking.

[0161] from Figure 8 It can be seen that the actual output power of the photovoltaic system changes in roughly the same way as the power of the electric locomotive, which greatly improves the local consumption capacity of the new energy system.

[0162] from Figure 9 It can be seen that the energy storage system can adaptively charge and discharge according to the changes in the vehicle's traction power, effectively playing a role in peak shaving and valley filling.

[0163] Table 2 Power Parameter Information from Numerical Simulation

[0164]

[0165] The power simulation data are shown in Table 2. The analysis results show that for the photovoltaic system, the photovoltaic energy absorption capacity of this method reaches 10556.59 kWh, and the photovoltaic system energy utilization efficiency is as high as 81.81%, which is 31.78% and 17.85% higher than the fixed threshold strategy and dynamic programming strategy, respectively, demonstrating a significant improvement in photovoltaic energy utilization. On the other hand, for the energy storage system, the regenerative braking energy absorption efficiency of this method is 38.35%, higher than the other two strategies. Meanwhile, the feedback efficiency of the energy storage system using this method is 53.82%, higher than the 31.37% of the fixed threshold strategy and the 35.10% of the dynamic programming strategy.

Claims

1. An energy management method for a "grid-source-storage-vehicle" collaborative energy supply system based on GNN, characterized in that, The energy management method is based on a graph neural network model, which uses graph convolutional networks and attention mechanisms to achieve real-time proactive energy control, and uses an online learning strategy to achieve online self-updating learning of the graph neural network model. The energy management method is used in the "network-source-storage-vehicle" coordinated functional system of electrified railways, and completes energy exchange with DC / DC converters by controlling the power fusion device. The specific steps for constructing the graph neural network model include: S111: Load power data from several time periods of the collaborative energy supply system, and divide the operating conditions of each operating entity from a data-driven perspective. The operating entities include at least new energy sources, energy storage, and electric locomotives. S112: Based on the results of the division of the main operating conditions, different operating conditions of the main operating conditions are regarded as nodes, and the transfer relationship between different operating conditions is regarded as edges. A weighted and directed operating condition transfer network graph composed of nodes and edges is constructed. S113: Construct an optimization model to optimize the working condition transfer relationship between the operating entities, obtain the working condition transfer network diagram of the operating entities, and output the theoretically optimal energy control result obtained through the optimization model; S114: Construct a graph neural network model based on the operating condition transfer network diagram of the main operating entity; The specific steps for achieving real-time advance energy management using graph convolutional networks and attention mechanisms include: S121: Based on the graph neural network model, load the adjacency matrix and feature matrix; S122: Extract local spatial feature information through the first graph convolutional layer and output the extracted feature matrix; S123: The attention mechanism compensates for the high-dimensional feature weighting by calculating the attention weights between all node pairs to obtain global interaction information and extract temporal feature information; S124: Extract comprehensive spatiotemporal features through the second graph convolutional layer; S125: Extract global features through the third graph convolutional layer; S126: The energy management layer predicts the operating conditions of each operating entity at the next moment based on the feature matrix output by the last graph convolutional layer of the graph convolutional network and attention mechanism, further predicts the power of the electric locomotive, and outputs the energy management results after the graph neural network model. S127: If a stop command is received, the process ends; if no stop command is received, repeat S121 to S126. The specific steps for implementing online self-updating learning of graph neural network models using online learning strategies include: S131: Load the energy control results obtained through the graph neural network model, and the theoretically optimal energy control results obtained through the optimization model; S132: After calculating the mean squared error loss function, determine the control accuracy of the graph neural network model; S133: Compare whether the result of the loss function exceeds the set threshold; S134: If the set threshold is not exceeded, no processing is performed; if the set threshold is exceeded, the data group is added to the data buffer. S135: Update the adjacency matrix, add feature information, and wait for the next training update of the graph neural network model.

2. The energy management method for a "grid-source-storage-vehicle" collaborative energy supply system based on GNN as described in claim 1, characterized in that, The optimization model uses maxf ％ To describe it, as in equation (1): In equation (1), e nes％ For the energy utilization efficiency of new energy systems, To improve the absorption efficiency of regenerative braking energy in energy storage systems, Energy feedback efficiency of energy storage systems; The energy utilization efficiency e of the new energy system nes％ The calculation is as shown in equation (2): In equation (2), Let be the power output of the renewable energy system to the external power grid at time t. Let be the power output from the new energy system to the energy storage system at time t. Let be the power output of the new energy system to the electric locomotive at time t. Let t be the maximum power that the renewable energy system can output at time t, and m be the length of the historical power data sequence. The energy storage system's regenerative braking energy absorption efficiency The calculation is as shown in equation (3): In equation (3), The charging power of the energy storage system at time t. Let t be the regenerative braking power of the electric locomotive; Energy feedback efficiency of the energy storage system The calculation is as shown in equation (4): In equation (4), Let t be the discharge power of the energy storage system at time t.

3. The energy management method for a "grid-source-storage-vehicle" collaborative energy supply system based on GNN as described in claim 2, characterized in that, By stacking the first, second, and third graph convolutional layers, deep features of nodes are extracted step by step to capture local dependencies between the running entities; The specific method for extracting aggregated spatial features using the first graph convolutional layer to obtain the first layer's spatial feature information and output feature matrix is ​​as follows: Aggregate the spatial features of each node at time t to generate a preliminary representation at time t+1, and then use a learnable weight matrix W. (l) A linear transformation is performed, and a nonlinearity is introduced by a nonlinear activation function σ to achieve aggregate spatial feature extraction of the graph convolutional layer, outputting the first graph convolutional layer feature matrix H. (l) .

4. The energy management method for a "grid-source-storage-vehicle" coordinated energy supply system based on GNN as described in claim 3, characterized in that, The specific calculation process of the attention mechanism includes: S41: Through linear transformation, the feature matrix H of the first convolutional layer is transformed... (l) Projecting the matrix onto three different spaces generates the corresponding query matrix Q, key matrix K, and value matrix V, respectively. S42: Perform click operations on the query matrix Q and the key matrix K, divide by the scaling factor to calculate the attention score, and obtain the attention weight matrix; S43: Integrate information from other nodes to generate a new feature representation for each node, calculate the single-head attention output, concatenate all single-head attention outputs along the feature dimension, fuse multi-head attention to obtain a multi-head attention concatenation matrix, and simultaneously map the multi-head attention concatenation matrix back to the original dimension to obtain the output feature matrix as the input to the second graph convolutional layer.

5. The energy management method for a "grid-source-storage-vehicle" coordinated energy supply system based on GNN as described in claim 4, characterized in that, The energy control layer receives feature information extracted by the graph convolutional layer and attention mechanism, and outputs the final feature matrix H. (L) The specific process of predicting the operating condition of the main body at the next moment using a fully connected method is shown in equation (5): M=softmax(W m H (L) +b m ) (5) In equation (5), W m b m These are all parameters of the classification layer, M∈R N×K Let K be the probability distribution of working conditions for each node, where K is the number of working condition categories and N is the total number of working conditions.

6. The energy management method for a "grid-source-storage-vehicle" coordinated energy supply system based on GNN as described in claim 5, characterized in that, Based on the operating condition prediction results of the main body, the energy management layer further predicts the power of the electric locomotive, the optimal output power of the photovoltaic system and the charging and discharging power of the energy storage system at the next moment, and the specific power output process is as shown in equation (6): In equation (6), All are learnable parameters, P∈R N×1 This represents the output power value.

7. The energy management method for a "grid-source-storage-vehicle" collaborative energy supply system based on GNN as described in claim 6, characterized in that, The loss function of the graphical neural network model is described by a weighted combination of the loss from operating condition prediction and the loss from energy management results, as shown in equation (7): L=λ(L NES +L ESS +L EL )+L MSE (7) In equation (7), the predicted loss under the working condition is represented by λ(L). NES +L ESS +L EL To describe, L NES L ESS L EL The predicted losses for photovoltaic systems, energy storage systems, and electric locomotives are L, respectively. MSE λ represents the energy management result loss, and is a hyperparameter used to balance the loss from operating condition prediction and the loss from energy management results. The energy management result loss L MSE The quantification is achieved by calculating the squared difference between the test value and the true value using the mean square error, as shown in equation (8): In equation (8), P i,t and These are the test power and the actual power of node i at time t, respectively, and M is the training sample set; The predicted losses for photovoltaic systems, energy storage systems, and electric locomotives are all calculated using cross-entropy loss to determine the difference between the predicted probability distribution and the actual distribution. sys Describe it as shown in equation (9): In equation (9), B is the batch size. This is a unique thermal encoding for real-world operating conditions, where sys represents the photovoltaic system, energy storage system, and electric locomotive, and M... sys,b,i To predict probabilities.

8. The energy management method for a "grid-source-storage-vehicle" coordinated energy supply system based on GNN as described in claim 7, characterized in that, The energy control results obtained through the graph neural network model and the theoretically optimal energy control results obtained through the optimization model are compared, as shown in equation (10), to design the mean squared error loss function. In equation (10), P′ nes,t+1 For optimal output power of new energy, P′ el,t+1 P′ represents the actual power of the electric locomotive. ess,t+1 The optimal charging and discharging power for energy storage.

9. The energy management method for a "grid-source-storage-vehicle" coordinated energy supply system based on GNN as described in claim 8, characterized in that, Set a threshold θ, if the loss function If the result exceeds the set threshold θ, the adjacency matrix weights are updated, the feature information is added to the feature buffer, the model is trained according to equation (11), and the trained graph neural network model replaces the original graph neural network model: In equation (11), μ∈(0,1) is the update rate, used to control the balance between the original weights and the new weights, w new For the new transfer weights, F new Let A(i,j) be the original adjacency matrix, A'(i,j) be the updated adjacency matrix, F be the original feature matrix, and F' be the updated feature matrix.

Images