Multi-energy storage state sensing network optimization method based on intelligent algorithm

Abstract

The invention discloses a multi-energy storage state sensing network optimization method based on an intelligent algorithm, and relates to the technical field of intelligent power grids, and the method comprises the steps: collecting the real-time operation data of a multi-energy storage unit, and obtaining the topological structure data of the multi-energy storage unit; constructing an energy flow diagram based on the real-time operation data and the topological structure data; extracting state feature vectors from the energy flow graph and the attribute matrix thereof through a self-supervised learning model; based on the state feature vector and the energy flow graph, topological structure data of the multi-energy energy storage unit is optimized through a graph neural network; based on the optimized topological feature matrix, generating a scheduling strategy of the multi-energy storage unit through reinforcement learning; by constructing the energy flow graph and the attribute matrix thereof, a complex system is simplified into a graph structure which is easy to process, the calculation efficiency is improved, and the transparency of the system is enhanced.

Description

Technical Field

[0001] The present invention relates to the field of smart grid technology, and in particular to a multi-energy storage state perception network optimization method based on an intelligent algorithm. Background Art

[0002] In the energy field, with the rapid development of renewable energy and the advancement of smart grid construction, energy storage systems have been widely used as an important means to smooth fluctuations and improve power supply reliability and power quality. Multi-Energy Storage Systems (MESS) can integrate various types of energy storage units, such as batteries, supercapacitors, flywheels, etc., and achieve complementarity and coordination between different energy storage forms through reasonable scheduling to meet the diverse needs of the power system.

[0003] In recent years, with the continuous advancement of artificial intelligence technology, especially the advantages of machine learning algorithms in data processing and pattern recognition, intelligent algorithms have been increasingly used in energy storage status perception and network optimization. Although the existing related technologies have achieved the monitoring and control of energy storage systems to a certain extent, in the process of network optimization, due to the inability to fully utilize global information, some local optimal solutions rather than global optimal solutions may be caused. Summary of the invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a multi-energy storage state-aware network optimization method based on an intelligent algorithm to solve the problem of lack of global perspective in topology optimization.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a multi-energy storage state perception network optimization method based on an intelligent algorithm, which comprises:

[0008] Collect real-time operation data of multi-energy storage units and obtain topological structure data of multi-energy storage units;

[0009] constructing an energy flow diagram based on the real-time operation data and the topological structure data;

[0010] Extract state feature vectors from energy flow diagrams and their attribute matrices through a self-supervised learning model;

[0011] Based on the state feature vector and energy flow diagram, the topological structure data of the multi-energy storage unit is optimized through the graph neural network;

[0012] Based on the optimized topology, the scheduling strategy of multi-energy storage units is generated through reinforcement learning.

[0013] As a preferred solution of the multi-energy storage state perception network optimization method based on intelligent algorithm described in the present invention, wherein: the real-time operation data of the multi-energy storage unit is collected, and the topological structure data of the multi-energy storage unit is obtained. The specific steps are as follows:

[0014] Real-time collection of operating data of each energy storage device in the multi-energy storage unit;

[0015] Synchronize the time axis of sensors and edge nodes that collect operation data, and set the frequency of operation data collection;

[0016] The operating data of each energy storage device is summarized in real time through edge nodes;

[0017] Through the outlier detection method of the 3σ criterion, the normal range of each operating data is set, and the values ​​outside the range are eliminated;

[0018] Fill in the missing running data in a short period of time by linear interpolation;

[0019] All collected operation data are normalized using the Min-Max normalization method;

[0020] The topology data is obtained through the control center of the multi-energy storage unit.

[0021] As a preferred solution of the multi-energy storage state perception network optimization method based on intelligent algorithm described in the present invention, wherein: based on the real-time operation data and the topological structure data, an energy flow diagram is constructed, and the specific steps are as follows:

[0022] Each energy storage device is abstracted as a node of the energy flow graph, the operation data of the energy storage device is used as the attribute of the node, and the attributes of all nodes are integrated into a node attribute matrix;

[0023] The transmission paths in the topological structure data are used as the edges of the energy flow graph, and the connection relationship and transmission efficiency between energy storage devices are used as the attributes of the edges, which are integrated into an edge attribute matrix.

[0024] Each node is connected by edges to form an energy flow graph of the multi-energy storage unit.

[0025] As a preferred solution of the multi-energy storage state perception network optimization method based on intelligent algorithm described in the present invention, wherein: the state feature vector is extracted from the energy flow diagram and its attribute matrix through the self-supervised learning model, and the specific steps are as follows:

[0026] Based on the node attribute matrix of historical operation data, the operation data of energy storage equipment at k consecutive time steps are sampled;

[0027] Based on the time step setting, calculate the rate of change and acceleration of the operating data of the energy storage device;

[0028] The operation data, change rate and change acceleration of the energy storage device at consecutive time steps are summarized into a time series feature matrix;

[0029] Extract the interaction features between nodes through the graph neural network encoder in the self-supervised learning model;

[0030] The interaction features between nodes in each layer are iteratively calculated through a multi-layer graph neural network to obtain the node interaction feature matrix;

[0031] Based on adding position encoding to the node interaction feature matrix, the time series decoder is used to predict the node attributes of the future time step and output the predicted time series feature matrix;

[0032] Use mean square error as the loss function to calculate the difference between the predicted time series feature matrix and the true time series feature matrix;

[0033] Use random masking to set part of the time series feature matrix to missing values;

[0034] Use the time series decoder to reconstruct the time series feature matrix containing missing values ​​to obtain a complete time series feature matrix;

[0035] The mean square error is used as the loss function for the mask position to calculate the difference between the reconstructed time series feature matrix and the original time series feature matrix;

[0036] Combine the prediction and reconstruction loss functions of the time series feature matrix to obtain the total loss function of the self-supervised learning model;

[0037] Update the parameters of the self-supervised learning model through the back-propagation algorithm;

[0038] Iteratively train the time series feature matrices of all time windows until the maximum number of iterations is reached;

[0039] The real-time time series feature matrix is ​​input into the self-supervised learning model for prediction and reconstruction, and the running feature vector matrix is ​​output.

[0040] As a preferred solution of the multi-energy storage state perception network optimization method based on intelligent algorithm described in the present invention, wherein: based on the state feature vector and energy flow diagram, the topological structure data of the multi-energy storage unit is optimized through the graph neural network, and the specific steps are as follows:

[0041] Based on the energy flow graph, the network density and maximum flow path of the topological structure of the multi-energy storage unit are calculated as the global characteristics of the topological structure;

[0042] The running feature vector matrix, the edge attribute matrix and the global features of the topological structure are concatenated to obtain a topological feature matrix;

[0043] Based on the topological feature matrix, the generator in the generative adversarial network is input, the topological feature matrix is ​​optimized, and the topological candidate feature matrix is ​​output;

[0044] In the discriminator, the graph isomorphism network is used to distinguish the authenticity of the topological candidate feature matrix;

[0045] Through the interaction between the generator and the discriminator, a topological candidate feature matrix whose authenticity cannot be distinguished by the discriminator is finally output.

[0046] As a preferred solution of the multi-energy storage state perception network optimization method based on intelligent algorithm described in the present invention, wherein: the generator in the adversarial network based on the topological feature matrix is ​​optimized to output the topological candidate feature matrix. The specific steps are as follows:

[0047] Through the multi-layer perceptron of the generator, nonlinear feature mapping is performed on the topological feature matrix, high-dimensional features are extracted layer by layer, and the topological candidate matrix is ​​output;

[0048] Symmetrically process the generated topology candidate matrix.

[0049] As a preferred solution of the multi-energy storage state perception network optimization method based on intelligent algorithm described in the present invention, wherein: in the discriminator, the authenticity of the topological candidate feature matrix is ​​distinguished by using a graph isomorphism network, and the specific steps are as follows:

[0050] The authenticity is distinguished by calculating the Euclidean distance and cosine similarity between the topological candidate feature matrix and the topological feature matrix;

[0051] According to the calculation results of Euclidean distance and cosine similarity, the Euclidean distance threshold and cosine similarity threshold are set by using statistical analysis method;

[0052] When the calculated Euclidean distance is less than the Euclidean distance threshold and the cosine similarity is greater than the cosine similarity threshold, the discriminator determines that the topological candidate feature matrix is ​​a topological feature matrix;

[0053] When the calculation result shows that the Euclidean distance is greater than the Euclidean distance threshold, and the cosine similarity is less than the cosine similarity threshold, the discriminator distinguishes the topological candidate feature matrix from the non-topological feature matrix, and the generator continues to regenerate it.

[0054] As a preferred solution of the multi-energy storage state perception network optimization method based on intelligent algorithm described in the present invention, wherein: based on the optimized topological feature matrix, the scheduling strategy of the multi-energy storage unit is generated by reinforcement learning, and the specific steps are as follows:

[0055] After the generator of the generative adversarial network is optimized, a topological posterior feature matrix is ​​obtained that the discriminator cannot distinguish the authenticity;

[0056] The topological complementary characteristic matrix is ​​used as the state space vector of the multi-energy storage unit;

[0057] For each energy storage device of the multi-energy storage unit, the output power adjustment, energy transmission path selection and load distribution ratio are used as action space vectors;

[0058] The reward function is set with the optimization objectives of maximizing energy transmission efficiency and load rate;

[0059] Build policy networks and value networks based on deep deterministic policy gradient algorithms;

[0060] Input the state space vector into the policy network, generate a new action space vector, and calculate the immediate reward function;

[0061] Evaluate the value of the new action space vector and state space vector by calculating the target value function on the value network;

[0062] Merge the state space vector, action space vector, new action space vector and immediate reward function into a strategy vector and put it into the experience replay pool;

[0063] Extract the policy vector from the experience replay pool, calculate the target value, and update the value network by minimizing the value network loss function;

[0064] Update the policy network by maximizing the policy network loss function;

[0065] According to the real-time state space vector of the multi-energy storage unit, the action space vector is output through the deep deterministic policy gradient algorithm to obtain the scheduling strategy of the multi-energy storage unit;

[0066] According to the soft update rules, it continuously interacts with the environment and updates the parameters of the policy network and value network.

[0067] In a second aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the multi-energy storage state-aware network optimization method based on intelligent algorithm as described in the first aspect of the present invention is implemented.

[0068] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the multi-energy storage state-aware network optimization method based on intelligent algorithm as described in the first aspect of the present invention.

[0069] The beneficial effects of the present invention are as follows: through precise data collection and preprocessing, the time consistency, integrity and reliability of the operating data are ensured, providing a solid foundation for subsequent analysis, constructing an energy flow diagram and its attribute matrix, simplifying the complex system into an easy-to-process graph structure, improving computing efficiency and enhancing the transparency of the system, using a self-supervised learning model to extract state feature vectors, capturing the dynamic behavior patterns of the equipment, revealing implicit relationships, and compensating for missing information through data reconstruction, thereby improving the intelligence level of state perception, combining state characteristics and energy flow diagrams, using graph neural networks and generative adversarial networks to optimize the topology structure, generating a more reasonable and efficient network layout, and significantly improving the energy efficiency and load rate of the system, and generating a dynamic scheduling strategy based on the optimized topology structure through reinforcement learning, so that the system can be flexibly adjusted according to the real-time state to maximize energy storage efficiency and economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0071] Figure 1 This is a flow chart of the multi-energy storage state perception network optimization method based on intelligent algorithm in Example 1.

[0072] Figure 2 This is a flow chart for optimizing the topological feature matrix in Example 1. DETAILED DESCRIPTION

[0073] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0074] Example 1, reference Figure 1 and Figure 2 , which is the first embodiment of the present invention, and provides a multi-energy storage state perception network optimization method based on an intelligent algorithm, comprising the following steps:

[0075] S1. Collect real-time operation data of the multi-energy storage unit and obtain topological structure data of the multi-energy storage unit.

[0076] Real-time collection of operating data (voltage, current, power output, SOC, SOH and temperature, etc.) of each energy storage device (such as battery pack, hydrogen energy storage device, thermal energy storage unit, etc.) in the multi-energy energy storage unit;

[0077] Synchronize the time axis of sensors and edge nodes that collect operation data to ensure that the collected multi-source data can be aligned on the time axis to avoid data conflicts, and set the frequency of operation data collection;

[0078] The operating data of each energy storage device is summarized in real time through edge nodes;

[0079] Through the outlier detection method of the 3σ criterion, the normal range of each operating data is set, and the values ​​outside the range are eliminated;

[0080] Fill in the missing running data in a short period of time by linear interpolation;

[0081] In order to eliminate the influence of each parameter dimension, the Min-Max normalization method is used to normalize all the collected operating data;

[0082] The topology data (connection relationship between energy storage devices, transmission path, transmission efficiency and load conditions of energy storage devices) is obtained through the control center of the multi-energy storage unit.

[0083] S2. Constructing an energy flow diagram based on the real-time operation data and the topological structure data.

[0084] Each energy storage device is abstracted as a node of the energy flow graph, the operation data of the energy storage device is used as the attribute of the node, and the attributes of all nodes are integrated into a node attribute matrix;

[0085] The transmission paths in the topological structure data are used as the edges of the energy flow graph, and the connection relationship and transmission efficiency between energy storage devices are used as the attributes of the edges, which are integrated into an edge attribute matrix.

[0086] Each node is connected by edges to form an energy flow graph of the multi-energy storage unit;

[0087] Specifically, each row in the node attribute matrix in the energy flow diagram corresponds to the attribute of the node, and the columns represent different operating data.

[0088] The edge attribute matrix is ​​an adjacency matrix, which represents the connection relationship and energy transmission efficiency between energy storage devices. The element value in the edge attribute matrix represents the energy transmission efficiency between two nodes. When the element value in the edge attribute matrix is ​​equal to 0, it means that there is no direct connection between the two nodes.

[0089] S3. Extract state feature vectors from the energy flow diagram and its attribute matrix through a self-supervised learning model. Based on the node attribute matrix of historical operation data, sample the operation data of energy storage devices for k consecutive time steps;

[0090] Based on the setting of the time step, the change rate and change acceleration of the operating data of the energy storage device in the data are calculated;

[0091] The energy storage equipment operation data, change rate and change acceleration of continuous time steps are summarized into a time series feature matrix

[0092] The interaction features between nodes are extracted through the graph neural network encoder in the self-supervised learning model. The expression is as follows:

[0093]

[0094] in, is the node interaction feature matrix of the l+1th layer at time t, σ is the activation function (such as ReLU), A t The edge attribute matrix at time t, F t is the time series feature matrix at time window t, W (l) is the weight matrix of the lth layer, b (l) is the bias term of the lth layer, t is the time point, and l is the layer index of the graph neural network feature extraction layer;

[0095] Through the multi-layer graph neural network, the interaction features between nodes in each layer are iteratively calculated to obtain the node interaction feature matrix Among them, L is the number of layers of the multi-layer graph neural network, which represents the running state of the node in the time window t and its edge attributes with other nodes;

[0096] Use a time series decoder (such as GRU or Transformer) to predict node attributes for future time steps and output a predicted time series feature matrix;

[0097] Specifically, since the Transformer decoder itself does not have a time series mechanism, position encoding is added to the time step of the interaction feature matrix of each node to provide time step information;

[0098] Add the position code to the node’s interaction feature matrix to get the node feature matrix, which is used as the final input of the decoder.

[0099] Map the node feature matrix to the query, key, and value matrices respectively, and calculate the weights between different nodes and time steps through multi-head attention;

[0100] The node features of each time step are transformed through the fully connected feedforward network of the graph neural network, and the predicted time series feature matrix is ​​output;

[0101] Use mean square error as the loss function to calculate the difference between the predicted time series feature matrix and the true time series feature matrix;

[0102] Furthermore, a random masking method is used to set part of the time series feature matrix to missing values ​​(i.e., the value of the node attribute matrix is ​​0).

[0103] Extract the interaction features between nodes through the graph neural network encoder in the self-supervised learning model;

[0104] Use the time series decoder to reconstruct the time series feature matrix containing missing values ​​to obtain a complete time series feature matrix;

[0105] The mean square error is used as the loss function for the mask position to calculate the difference between the reconstructed time series feature matrix and the original time series feature matrix;

[0106] Combining the prediction and reconstruction loss functions of the time series feature matrix, we get the total loss function of the self-supervised learning model, which is expressed as follows:

[0107] K = 0.5K1 + 0.5K2;

[0108] Among them, K is the total loss function of the self-supervised learning model, K1 is the time series data prediction loss function, and K2 is the time series data reconstruction loss function;

[0109] Update the parameters of the self-supervised learning model through a back-propagation algorithm (such as the Adam optimizer);

[0110] Iteratively train the time series feature matrices of all time windows until the maximum number of iterations (e.g., 100 rounds) is reached;

[0111] After training is completed, the real-time time series feature matrix is ​​input into the self-supervised learning model for prediction and reconstruction, and the operating feature vector matrix is ​​output. The rows of the feature vector matrix represent the operating feature vectors of the energy storage device.

[0112] It should be noted that by predicting the state of the energy storage equipment in the future time step, mastering the dynamic behavior pattern of the energy storage equipment, learning the evolution law of the equipment state, discovering potential abnormal trends, and reconstructing the time series feature matrix, the missing operating data is completed to ensure the accuracy of the prediction.

[0113] S4. Based on the state feature vector and energy flow diagram, the topological structure data of the multi-energy storage unit is optimized through the graph neural network.

[0114] Based on the energy flow graph, the network density and maximum flow path of the topological structure of the multi-energy storage unit are calculated as the global characteristics of the topological structure;

[0115] Specifically, based on the node attribute matrix and the edge attribute matrix, the network density of the topological structure of the multi-energy storage unit is calculated;

[0116] Use a maximum flow algorithm (such as the Ford-Folkerson algorithm) to calculate the maximum flow path capacity under the current topology;

[0117] The running feature vector matrix, the edge attribute matrix and the global features of the topological structure are concatenated to obtain a topological feature matrix;

[0118] Based on the generator in the topological feature matrix generative adversarial network, the topological feature matrix is ​​optimized and the topological candidate feature matrix is ​​output;

[0119] Through the multi-layer perceptron of the generator, nonlinear feature mapping is performed on the topological feature matrix, high-dimensional features are extracted layer by layer, and the topological candidate matrix is ​​output;

[0120] Specifically, in the first layer of perceptron, the topological feature matrix is ​​mapped to the intermediate feature matrix, and the expression is as follows:

[0121] Z1=ReLU(W1×C+b1);

[0122] Among them, Z1 is the intermediate feature matrix, ReLU represents the activation function, W1 is the weight matrix of the first layer of the multilayer perceptron, C is the topological feature matrix, and b1 is the bias term of the first layer of the multilayer perceptron;

[0123] In the second layer of perceptron, the intermediate feature matrix is ​​mapped to the topological candidate matrix, and the expression is as follows:

[0124] D=SIgmoid(W2×Z1+b2),

[0125] Where D is the topology candidate matrix, Sigmoid means limiting the output to the interval [0,1], W2 is the weight matrix of the second layer of the perceptron, and b2 is the bias term of the second layer of the multilayer perceptron;

[0126] Symmetrically process the generated topological candidate matrix;

[0127] In the discriminator, the graph isomorphism network is used to distinguish the authenticity of the topological candidate feature matrix;

[0128] The authenticity is distinguished by calculating the Euclidean distance and cosine similarity between the topological candidate feature matrix and the topological feature matrix. The expression is as follows:

[0129] E = ∥FD∥2;

[0130]

[0131] Among them, E is the Euclidean distance between the topological candidate feature matrix and the topological feature matrix, F is the topological feature matrix, and G is the cosine similarity between the topological candidate feature matrix and the topological feature matrix;

[0132] The Euclidean distance measures the distance between the topological candidate feature matrix and the topological feature matrix in high-dimensional space. The smaller the distance, the more realistic the topological candidate feature matrix is.

[0133] The cosine similarity measures the directional similarity between the topological candidate feature matrix and the topological feature matrix. The value range is [-1,1]. The closer the value is to 1, the more similar the structural distribution of the topological candidate feature matrix and the topological feature matrix is.

[0134] According to the calculation results of Euclidean distance and cosine similarity, the Euclidean distance threshold and cosine similarity threshold are set by using statistical analysis method;

[0135] Specifically, according to the calculation results of Euclidean distance and cosine similarity, statistical analysis is performed and distribution graphs are drawn;

[0136] According to the distribution graph, the dividing point between the topological candidate feature matrix and the topological feature matrix pair distribution is calculated using the 95% quantile method in statistics;

[0137] The average values ​​corresponding to the 95% quantile and the 5% quantile in the Euclidean distance and cosine similarity distribution graphs are used as the Euclidean distance threshold and the cosine similarity threshold;

[0138] When the calculated Euclidean distance is less than the Euclidean distance threshold and the cosine similarity is greater than the cosine similarity threshold, the discriminator determines that the topological candidate feature matrix is ​​a topological feature matrix;

[0139] When the calculation result shows that the Euclidean distance is greater than the Euclidean distance threshold, and the cosine similarity is less than the cosine similarity threshold, the discriminator distinguishes the topological candidate feature matrix from the non-topological feature matrix, and the generator continues to regenerate it;

[0140] Through the interaction between the generator and the discriminator, a topological candidate feature matrix that the discriminator cannot distinguish from the real is finally output;

[0141] Specifically, when the discriminator distinguishes that the topological candidate feature matrix is ​​not a real topological feature matrix, the generator updates the parameters of the generator according to the distinction basis of the discriminator and re-outputs a new topological candidate feature matrix;

[0142] The discriminator uses the distinction of the topological candidate feature matrix as training, constantly adapting to the changes in the real topology, and more strictly distinguishing the topological candidate feature matrix output by the generator.

[0143] S5. Based on the optimized topological feature matrix, the scheduling strategy of the multi-energy storage unit is generated through reinforcement learning.

[0144] After the generator of the generative adversarial network is optimized, a topological posterior feature matrix is ​​obtained that the discriminator cannot distinguish the authenticity;

[0145] The topological complement feature matrix is ​​used as the state space vector of the multi-energy storage unit;

[0146] For each energy storage device of the multi-energy storage unit, the output power adjustment, energy transmission path selection and load distribution ratio are used as action space vectors;

[0147] Taking maximizing energy transmission efficiency and load rate as optimization goals, the reward function is set and the expression is as follows:

[0148] R = 0.5r1 + 0.5r2;

[0149] Among them, R is the reward function value, r1 is the maximum energy transmission efficiency value, and r2 is the maximum load rate value;

[0150] Build policy networks and value networks based on deep deterministic policy gradient algorithms;

[0151] Input the state space vector into the policy network, generate a new action space vector, and calculate the immediate reward function;

[0152] On the value network, by calculating the target value function Q(S t ,A t ), evaluate the value of the new action space vector and state space vector, the expression is as follows:

[0153]

[0154] Among them, S t is the state space vector, Y t is the action space vector, represents the weighted average of rewards in all possible future paths, γ is the discount factor, and R t+k+1 is the reward function value from time t+k+1, k is the time step, t is the time point;

[0155] Merge the state space vector, action space vector, new action space vector and immediate reward function into a strategy vector and put it into the experience replay pool;

[0156] Extract the policy vector from the experience replay pool, calculate the target value, and update the value network by minimizing the value network loss function;

[0157] Update the policy network by maximizing the policy network loss function;

[0158] According to the real-time state space vector of the multi-energy storage unit, the action space vector is output through the deep deterministic policy gradient algorithm to obtain the scheduling strategy of the multi-energy storage unit;

[0159] According to the soft update rules, it continuously interacts with the environment and updates the parameters of the policy network and value network.

[0160] This embodiment also provides a computer device, which is suitable for the case of a multi-energy storage state-aware network optimization method based on an intelligent algorithm, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the multi-energy storage state-aware network optimization method based on an intelligent algorithm as proposed in the above embodiment.

[0161] The computer device may be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a key, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse, etc.

[0162] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, the multi-energy storage state-aware network optimization method based on an intelligent algorithm as proposed in the above embodiment is implemented; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (Static Random Access Memory, referred to as SRAM), electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, referred to as EEPROM), erasable programmable read-only memory (Erasable Programmable Read Only Memory, referred to as EPROM), programmable read-only memory (Programmable Red-Only Memory, referred to as PROM), read-only memory (Read-Only Memory, referred to as ROM), magnetic storage, flash memory, disk or optical disk.

[0163] In summary, the present invention ensures the time consistency, integrity and reliability of operating data through precise data collection and preprocessing, providing a solid foundation for subsequent analysis, constructing an energy flow diagram and its attribute matrix, simplifying complex systems into an easy-to-process graph structure, improving computing efficiency and enhancing system transparency, using a self-supervised learning model to extract state feature vectors, capturing the dynamic behavior patterns of equipment, revealing implicit relationships, and compensating for missing information through data reconstruction, thereby improving the intelligence level of state perception, combining state characteristics and energy flow diagrams, using graph neural networks and generative adversarial networks to optimize the topology structure, generating a more reasonable and efficient network layout, and significantly improving the energy efficiency and load rate of the system, and generating a dynamic scheduling strategy based on the optimized topology structure through reinforcement learning, so that the system can be flexibly adjusted according to the real-time state to maximize energy storage efficiency and economic benefits.

[0164] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A multi-energy storage state perception network optimization method based on intelligent algorithm, characterized by: include, Collect real-time operation data of multi-energy storage units and obtain topological structure data of multi-energy storage units; constructing an energy flow diagram based on the real-time operation data and the topological structure data; Extract state feature vectors from energy flow diagrams and their attribute matrices through a self-supervised learning model; Based on the state feature vector and energy flow diagram, the topological structure data of the multi-energy storage unit is optimized through the graph neural network; Based on the optimized topological feature matrix, the scheduling strategy of multi-energy storage units is generated through reinforcement learning.

2. The multi-energy storage state perception network optimization method based on intelligent algorithm according to claim 1 is characterized in that: The specific steps of collecting the real-time operation data of the multi-energy storage unit and obtaining the topological structure data of the multi-energy storage unit are as follows: Real-time collection of operating data of each energy storage device in the multi-energy storage unit; Synchronize the time axis of the acquisition sensor and edge node, and set the frequency of operation data acquisition; The operating data of each energy storage device is summarized in real time through edge nodes; Through the outlier detection method of the 3σ criterion, the normal range of each operating data is set, and the values ​​outside the range are eliminated; Fill in the missing running data in a short period of time by linear interpolation; The Min-Max normalization method is used to normalize all collected operating data; The topology data is obtained through the control center of the multi-energy storage unit.

3. The multi-energy storage state perception network optimization method based on intelligent algorithm as claimed in claim 2 is characterized by: The energy flow diagram is constructed based on the real-time operation data and the topological structure data. The specific steps are as follows: Each energy storage device is abstracted as a node of the energy flow graph, the operation data of the energy storage device is used as the attribute of the node, and the attributes of all nodes are integrated into a node attribute matrix; The transmission paths in the topological structure data are used as the edges of the energy flow graph, and the connection relationship and transmission efficiency between energy storage devices are used as the attributes of the edges, which are integrated into an edge attribute matrix. Each node is connected by edges to form an energy flow graph of the multi-energy storage unit.

4. The multi-energy storage state perception network optimization method based on intelligent algorithm as claimed in claim 3 is characterized by: The specific steps of extracting the state feature vector from the energy flow diagram and its attribute matrix through the self-supervised learning model are as follows: Based on the node attribute matrix of historical operation data, the operation data of energy storage equipment at k consecutive time steps are sampled; Based on the time step setting, calculate the rate of change and acceleration of the operating data of the energy storage device; The operation data, change rate and change acceleration of the energy storage device at consecutive time steps are summarized into a time series feature matrix; Extract the interaction features between nodes through the graph neural network encoder in the self-supervised learning model; The interaction features between nodes in each layer are iteratively calculated through a multi-layer graph neural network to obtain the node interaction feature matrix; Based on adding position encoding to the node interaction feature matrix, the time series decoder is used to predict the node attributes of the future time step and output the predicted time series feature matrix; Use mean square error as the loss function to calculate the difference between the predicted time series feature matrix and the true time series feature matrix; Use random masking to set part of the time series feature matrix to missing values; Use the time series decoder to reconstruct the time series feature matrix containing missing values ​​to obtain a complete time series feature matrix; The mean square error is used as the loss function for the mask position to calculate the difference between the reconstructed time series feature matrix and the original time series feature matrix; Combine the prediction and reconstruction loss functions of the time series feature matrix to obtain the total loss function of the self-supervised learning model; Update the parameters of the self-supervised learning model through the back-propagation algorithm; Iteratively train the time series feature matrices of all time windows until the maximum number of iterations is reached; The real-time time series feature matrix is ​​input into the self-supervised learning model for prediction and reconstruction, and the running feature vector matrix is ​​output.

5. The multi-energy storage state perception network optimization method based on intelligent algorithm as claimed in claim 4 is characterized by: The topological structure data of the multi-energy storage unit is optimized by graph neural network based on the state feature vector and energy flow diagram. The specific steps are as follows: Based on the energy flow graph, the network density and maximum flow path of the topological structure of the multi-energy storage unit are calculated as the global characteristics of the topological structure; The running feature vector matrix, edge attribute matrix and global features are concatenated to obtain the topological feature matrix; Based on the generator in the topological feature matrix generative adversarial network, the topological feature matrix is ​​optimized and the topological candidate feature matrix is ​​output; In the discriminator, the graph isomorphism network is used to distinguish the authenticity of the topological candidate feature matrix; Through the interaction between the generator and the discriminator, a topological candidate feature matrix whose authenticity cannot be distinguished by the discriminator is finally output.

6. The multi-energy storage state perception network optimization method based on intelligent algorithm as claimed in claim 5 is characterized by: The generator in the adversarial network based on the topological feature matrix is ​​used to optimize the topological feature matrix and output the topological candidate feature matrix. The specific steps are as follows: Through the multi-layer perceptron of the generator, nonlinear feature mapping is performed on the topological feature matrix, high-dimensional features are extracted layer by layer, and the topological candidate matrix is ​​output; Symmetrically process the generated topology candidate matrix.

7. The multi-energy storage state perception network optimization method based on intelligent algorithm according to claim 6 is characterized in that: In the discriminator, a graph isomorphism network is used to distinguish the authenticity of the topological candidate feature matrix. The specific steps are as follows: The authenticity is distinguished by calculating the Euclidean distance and cosine similarity between the topological candidate feature matrix and the topological feature matrix; According to the calculation results of Euclidean distance and cosine similarity, the Euclidean distance threshold and cosine similarity threshold are set by using statistical analysis method; When the calculated Euclidean distance is less than the Euclidean distance threshold and the cosine similarity is greater than the cosine similarity threshold, the discriminator determines that the topological candidate feature matrix is ​​a topological feature matrix; When the calculation result shows that the Euclidean distance is greater than the Euclidean distance threshold, and the cosine similarity is less than the cosine similarity threshold, the discriminator distinguishes the topological candidate feature matrix from the non-topological feature matrix, and the generator continues to regenerate it.

8. The multi-energy storage state perception network optimization method based on intelligent algorithm as claimed in claim 7 is characterized by: The scheduling strategy of the multi-energy storage unit is generated by reinforcement learning based on the optimized topological feature matrix. The specific steps are as follows: After the generator of the generative adversarial network is optimized, a topological posterior feature matrix is ​​obtained that the discriminator cannot distinguish the authenticity; The topological complementary characteristic matrix is ​​used as the state space vector of the multi-energy storage unit; For each energy storage device of the multi-energy storage unit, the output power adjustment, energy transmission path selection and load distribution ratio are used as action space vectors; The reward function is set with the optimization objectives of maximizing energy transmission efficiency and load rate; Build policy networks and value networks based on deep deterministic policy gradient algorithms; Input the state space vector into the policy network, generate a new action space vector, and calculate the immediate reward function; Evaluate the value of the new action space vector and state space vector by calculating the target value function on the value network; Merge the state space vector, action space vector, new action space vector and immediate reward function into a strategy vector and put it into the experience replay pool; Extract the policy vector from the experience replay pool, calculate the target value, and update the value network by minimizing the value network loss function; Update the policy network by maximizing the policy network loss function; According to the real-time state space vector of the multi-energy storage unit, the action space vector is output through the deep deterministic policy gradient algorithm to obtain the scheduling strategy of the multi-energy storage unit; According to the soft update rules, it continuously interacts with the environment and updates the parameters of the policy network and value network.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the multi-energy storage state perception network optimization method based on intelligent algorithm according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the multi-energy storage state perception network optimization method based on intelligent algorithm according to any one of claims 1 to 8 are implemented.

Images