Method and system for predicting residual electric quantity of energy storage system based on deep learning
By constructing a topological relationship diagram and integrating multi-dimensional features and temperature information, the problem of inaccurate estimation of available capacity of lithium batteries is solved, and more accurate prediction of residual power of energy storage systems is achieved, improving user experience.
Patent Information
- Application Number
- CN202410306477.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-03-18
- Publication Date
- 2025-07-25
AI Technical Summary
In the prior art, the accuracy of the available capacity estimation of lithium batteries is difficult to meet user needs, resulting in large errors in battery SOC estimation and affecting user experience.
Based on deep learning methods, by constructing a topological relationship diagram, integrating multi-dimensional features and temperature information, training a local capacity prediction model, and improving the accuracy of residual power prediction of energy storage systems.
It improves the accuracy of the residual power forecast of the energy storage system, alleviates users' mileage anxiety, and improves user experience.
Smart Images
Figure CN120370186A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the technical field of batteries, and particularly to a method and system for predicting the remaining power of an energy storage system based on deep learning. Background Art
[0002] The SOC (State-of-Charge) of a battery refers to the state of the remaining power of the battery. Currently, battery technology has developed rapidly, but in the existing technology, the accuracy of estimating the available capacity of lithium batteries is difficult to meet the user's usage requirements, and accurately estimating the SOC of the battery in actual applications is what users expect. A very important reason is to let users "have a clear idea". Taking electric vehicles as an example, if according to past experience, users know that the driving range is about 500 km when the battery is fully charged, and now there is still 50 km to the nearest charging pile, then users know that when the remaining power reaches 10% - 15%, they have to go charging, or for safety, they go charging when the SOC is 20%.
[0003] If the SOC estimation error is greater than 20%, then when the vehicle shows that the remaining power is 20%, it may stop working at any time and break down on the roadside. Generally speaking, when the power of an electric vehicle is lower than 30%, the driver will have range anxiety, worrying that the power is not enough to drive to the charging station. When the power is between 20% - 30%, the vehicle should be prepared to be charged. Obviously, inaccurate estimation of the available capacity of lithium batteries and large errors will directly result in poor user experience.
[0004] Of course, the above example is only about the experience problems brought by the application of lithium batteries in vehicles. In other scenarios where lithium batteries are applied, inaccurate estimation of the available capacity of lithium batteries will also cause troubles to users from different perspectives. Summary of the Invention
[0005] The embodiments of this application provide a method and system for predicting the remaining power of an energy storage system based on deep learning, which are used to fuse multi-dimensional features, introduce temperature as a reference, and mainly consider the main factors affecting the remaining power of the energy storage system, so as to improve the accuracy of predicting the remaining power of the energy storage system.
[0006] The embodiments of this application propose a method for predicting the remaining power of an energy storage system based on deep learning, which is applied to an energy storage system including multiple energy storage subsystems, and any energy storage subsystem includes multiple energy storage units. The remaining power prediction method includes the following steps:
[0007] For any energy storage subsystem, obtain the layout information of each energy storage unit in the energy storage subsystem, as well as historical charge and discharge data, temperature information, and corresponding historical energy storage parameters;
[0008] Based on the acquired layout information, a topological relationship graph is constructed for each energy storage unit of any energy storage subsystem. In the topological relationship graph, the distance between nodes is used to describe the proximity position relationship of the energy storage units, and the series and parallel relationships of the energy storage units are described by connecting lines;
[0009] The historical charge-discharge data and historical energy storage parameters of each energy storage unit of the acquired any energy storage subsystem are screened, and the charge-discharge data and energy storage parameters of each energy storage unit in the same time period are assigned to the corresponding nodes of the topological relationship graph; and, a temperature background is assigned to the topological relationship graph based on the acquired temperature information;
[0010] Based on the formed topological relationship graphs, a training data set is constructed, and a local capacity prediction model is trained based on the training data set;
[0011] The charge-discharge data, temperature information, and energy storage parameters of each energy storage unit of the newly collected energy storage subsystem are processed and input into the trained local capacity prediction model;
[0012] Based on the obtained local capacity, the remaining capacity of any energy storage subsystem is determined.
[0013] Optionally, screening the historical charge-discharge data of each energy storage unit of the acquired any energy storage subsystem includes:
[0014] In each time period, the historical charge-discharge data of each energy storage unit is trimmed or queried forward and backward in time, so that the charge-discharge data within any time period covers the charge data and discharge data with the same power ratio for charging and discharging any energy storage unit;
[0015] Judge the number of energy storage units in the charging or discharging state at the start time of any time period, and take the mode of the number as the alignment basis;
[0016] On the time axis of any time period, based on the determined alignment basis, align the start times of each data segment of the charge data and discharge data, and arrange the charge data and discharge data at intervals;
[0017] Assigning the energy storage parameters in the same time period to the corresponding nodes of the topological relationship graph includes:
[0018] Determine the energy storage parameters at the end time of any time period, and assign the energy storage parameters at the end time of any time period to the corresponding nodes in the topological relationship graph.
[0019] Optionally, training the local capacity prediction model based on the training data set includes:
[0020] Extract the charge-discharge characteristics of the charge-discharge data after alignment; and,
[0021] According to the temperature background, nodes with a temperature deviation greater than a preset temperature deviation threshold are taken as nodes of interest, and based on the formed topological relationship graph, topological features are extracted;
[0022] Node strengthening is performed on the extracted features to enhance the weights of the nodes of interest;
[0023] The extracted charge and discharge features, topological features, and features after weight enhancement are adaptively fused through an adaptive weighted fusion function;
[0024] The features after weight enhancement are subjected to feature extraction again, and the extracted features are added to the adaptively fused features;
[0025] The added features are input into a fully connected layer, and the output of the fully connected layer is used as the prediction result of the local capacity prediction model.
[0026] Optionally, the extraction of the charge and discharge features of the aligned charge and discharge data is implemented through a Transformer encoder;
[0027] Based on the formed topological relationship graph, the extraction of topological features is to first divide the topological relationship graph into multiple sub-topological graphs, and for any sub-topological graph, a node contribution rate convolutional network is used to extract them, where the node contribution rate convolutional network processes the input sub-topological graph in the following way:
[0028] D1 = ISRU{CNN(D, D node ).norm}
[0029] δ = top n (D T )
[0030]
[0031] Among them, ISRU represents the ISRU activation function, D and D node respectively represent the nodes and the features of the nodes in the sub-topological graph, as the input of the CNN convolutional neural network, norm is a mathematical norm, CNN represents the convolutional neural network, D1 represents the features after convolutional operation, top n represents n nodes of interest in the sub-topological graph with a temperature deviation greater than the preset temperature deviation threshold D T , δ represents the weight factor, represents feature multiplication, and D2 represents the output features of the node contribution rate convolutional network.
[0032] Optionally, the node strengthening of the extracted features to enhance the weights of the nodes of interest includes:
[0033] Neighbor nodes with temperature correlation for each node of interest will be determined;
[0034] Based on the determined neighbor nodes, the weight of the node of interest is enhanced, satisfying:
[0035] a(d m ,d k )=c T ReLU(H·[d m ,d k )
[0036]
[0037] where tanh represents the tanh activation function, ReLU represents the ReLU activation function, H represents the weight matrix, m represents the serial number of the node of interest, k represents the serial number of the neighbor node with temperature correlation, k′ represents the variable for performing summation, K represents the number of neighbor nodes with temperature correlation, d m ,d k represents the unenhanced feature, d′ m represents the feature of the enhanced node of interest m, a(d m ,d k ) represents the attention metric of the node of interest, c is a learnable hyperparameter, and T represents transpose.
[0038] Optionally, the extracted charge-discharge features, topological features, and features after weight enhancement are adaptively fused through an adaptive weighted fusion function, satisfying:
[0039] X out =[Sigmoid(r)X1+(1-Sigmoid(r))X2]
[0040] where X out represents the feature after adaptive fusion, Sigmoid(*) represents the Sigmoid function, r represents the randomly initialized weight factor, X1 represents the extracted charge-discharge features and topological features, and X2 represents the features after weight enhancement.
[0041] Optionally, based on the obtained local capacity, determining the remaining capacity of any energy storage subsystem includes:
[0042] Based on the capacity prediction value output by the local capacity prediction model, the proportion of the number of energy storage units corresponding to each node of interest is used to determine the remaining capacity of any energy storage subsystem in segments.
[0043] An embodiment of the present application also provides a remaining power prediction system for an energy storage system based on deep learning, which includes a processor and a memory. A computer program is stored on the memory, and when the computer program is executed by the processor, the steps of the remaining power prediction method for the energy storage system based on deep learning as described above are implemented.
[0044] The prediction method in the embodiment of the present application uses multi-dimensional features for fusion, introduces temperature as a reference, and focuses on the main factors affecting the remaining power of the energy storage system, thereby improving the accuracy of predicting the remaining power of the energy storage system.
[0045] The above description is only an overview of the technical solution of the present application. In order to be able to understand the technical means of the present application more clearly, it can be implemented according to the content of the specification. And in order to make the above and other purposes, features and advantages of the present application more obvious and understandable, the following specifically illustrates the specific embodiments of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] By reading the following detailed description of the preferred embodiments, various other advantages and benefits will become clear to those of ordinary skill in the art. The drawings are only for the purpose of showing the preferred embodiments and are not considered to be a limitation of the present invention. Moreover, throughout the drawings, the same reference numerals are used to represent the same components. In the drawings:
[0047] Figure 1 It is a schematic diagram of the basic process of the remaining power prediction method for the energy storage system based on deep learning in this embodiment. DETAILED DESCRIPTION
[0048] Hereinafter, the exemplary embodiments of the present disclosure will be described in more detail with reference to the drawings. Although the exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be fully conveyed to those skilled in the art.
[0049] An embodiment of the present application provides a remaining power prediction method for an energy storage system based on deep learning, which is applied to an energy storage system including multiple energy storage subsystems, and any energy storage subsystem includes multiple energy storage units. For example, the energy storage system includes multiple energy storage subsystems managed by partitions, and any energy storage subsystem may include multiple energy storage units. For example, the energy storage unit may be a lithium battery. The remaining power prediction method in the embodiment of the present application includes the following steps:
[0050] In step S101, for any energy storage subsystem, obtain the layout information of each energy storage unit in any energy storage subsystem, as well as historical charge-discharge data, temperature information, and corresponding historical energy storage parameters. In subsequent specific implementations, the embodiments of the present application comprehensively consider the possible impacts of the layout of energy storage units, temperature data, and historical charge-discharge data on battery performance, and thus collect relevant data. The historical energy storage parameters referred to in the embodiments of the present application may be, for example, one or more parameters such as the rated capacity, internal resistance, and open-circuit voltage of the energy storage unit.
[0051] In step S102, based on the obtained layout information, construct a topological relationship graph for each energy storage unit of any energy storage subsystem, where in the topological relationship graph, the distance between nodes is used to describe the proximity relationship of the energy storage units, and the series-parallel relationship of the energy storage units is described by connecting lines. That is, by establishing the positional relationship of the layout in the form of a topological graph, one energy storage unit is used as a node, and the nodes with electrical connections are represented by connecting lines.
[0052] In step S103, screen the historical charge-discharge data and historical energy storage parameters of each energy storage unit of any energy storage subsystem obtained, and assign the charge-discharge data and energy storage parameters of each energy storage unit in the same time period to the corresponding nodes of the topological relationship graph; and, assign a temperature background to the topological relationship graph based on the obtained temperature information. Specifically, the historical charge-discharge data and historical energy storage parameters of each energy storage unit of any energy storage subsystem obtained can be screened and processed, and the historical energy storage parameters are assigned to the corresponding nodes after screening. Specific energy storage parameters, charge-discharge data, and the remaining power under the corresponding state can be obtained for different specifications of energy storage units, such as lithium batteries in a test environment. Then, add a temperature background to the nodes in the topology. For example, a temperature field distribution can be established in the topological relationship graph based on the measured temperature data of each energy storage unit, so as to add a temperature background to the topological relationship graph. In subsequent examples, features can be extracted based on the topological relationship graph.
[0053] In step S104, based on the formed topological relationship graphs, construct a training data set, and train a local capacity prediction model based on the training data set. Specifically, the training data set can be constructed in chronological order, and the local capacity prediction model can be trained. In subsequent examples of the present application, segmentation is performed based on the topological relationship graph, and then the segmented images of the topological relationship graph are trained.
[0054] In step S105, the charge-discharge data, temperature information, and energy storage parameters of each energy storage unit of the newly collected energy storage subsystem are processed and input into the trained local capacity prediction model. In the embodiments of the present application, considering the charge-discharge data, temperature information, and battery parameters of the energy storage unit, feature fusion is performed through the local capacity prediction model, so as to predict the remaining battery capacity of, for example, some interested energy storage units by using the local capacity prediction model. By introducing the charge-discharge data and temperature information for processing and then fusing them into features, the main factors affecting the remaining power of the energy storage system are mainly considered, and the accuracy of predicting the remaining power of the energy storage system is improved.
[0055] In step S106, based on the obtained local capacity, the remaining capacity of any energy storage subsystem is determined. In some embodiments, determining the remaining capacity of any energy storage subsystem based on the obtained local capacity may include: based on the capacity prediction value output by the local capacity prediction model, the proportion of the number of energy storage units corresponding to each interested node is segmented to determine the remaining capacity of any energy storage subsystem.
[0056] Based on the capacity prediction value output by the local capacity prediction model for the interested nodes, and the proportion of the interested nodes in all the energy storage units in the energy storage subsystem. For example, if only a small part of the interested nodes, such as less than 5%, then the remaining capacity of a subsystem that conforms to the actual working conditions of the energy storage unit is determined based on the capacity prediction value of the non-interested nodes. For example, if the proportion exceeds 30%, then the average value of the capacity prediction values of the interested nodes can be used as the remaining capacity that conforms to the current working conditions of the energy storage system.
[0057] In the embodiments of the present application, the interested nodes may be, for example, nodes with a relatively large deviation in working temperature from the normal working temperature range. The energy storage units corresponding to such nodes correspond to the environmental impact factors of the operation of the energy storage system. For example, the energy storage unit is in a low-temperature environment, or some work in a high-temperature environment. In the embodiments of the present application, taking the capacity of the energy storage unit of the interested node as the basis for calculating the system capacity can further improve the authenticity and accuracy of the capacity prediction and improve the user experience.
[0058] In some embodiments, screening the historical charge-discharge data of each energy storage unit of the obtained any energy storage subsystem includes:
[0059] In each time period, the historical charge-discharge data of each energy storage unit is trimmed or queried forward and backward in time, so that the charge-discharge data within any time period covers the charging data and discharging data with the same power proportion for charging and discharging any energy storage unit.
[0060] Judge the number of energy storage units in the charging or discharging state at the start time of any time period, and take the mode of the number as the alignment basis.
[0061] On the time axis of any time period, based on the determined alignment basis, align the start times of the data segments of the charging data and the discharging data, and arrange the charging data and the discharging data at intervals. In some specific examples, the charge-discharge data can be the corresponding charge-discharge curves. Through alignment and interval arrangement, features such as the curve slope and duration of the charge-discharge curves with the same charge-discharge power ratio can be highlighted in subsequent feature extraction, improving the representative value of the extracted features.
[0062] In some embodiments, assigning the energy storage parameters in the same time period to the corresponding nodes of the topological relationship graph includes:
[0063] Determine the energy storage parameters at the end time of the any time period, and assign the energy storage parameters at the end time of the any time period to the corresponding nodes in the topological relationship graph. Specifically, in the embodiments of the present application, the energy storage parameters at the end time of the same time period are considered to characterize the energy storage situation at the end time of each time period.
[0064] In some embodiments, training the local capacity prediction model based on the training data set includes:
[0065] Extract the charge-discharge features of the charge-discharge data after alignment; and,
[0066] According to the temperature background, use the nodes with a temperature deviation greater than the preset temperature deviation threshold as the nodes of interest, and based on the formed topological relationship graph, extract topological features. Specifically, the formed topological relationship graph can be divided into multiple sub-topological graphs according to a specified specification. In the case of giving the temperature background, referring to the normal charge-discharge operating temperature range, the nodes corresponding to the energy storage units outside this temperature range are used as the nodes of interest, and the topological features (topological image features) of the sub-topological graphs are proposed.
[0067] Perform node strengthening on the extracted features to enhance the weights of the nodes of interest.
[0068] Perform adaptive fusion on the extracted charge-discharge features, topological features, and the features after weight enhancement through an adaptive weighted fusion function.
[0069] Perform feature extraction on the features after weight enhancement again, and add the extracted features to the adaptively fused features. Specifically, the network structure used for performing feature extraction on the features after weight enhancement again can be the same as that used for the first extraction of charge-discharge features and the extraction of topological features.
[0070] Input the added features into a fully connected layer, and use the output of the fully connected layer as the prediction result of the local capacity prediction model.
[0071] In some embodiments, the charge and discharge characteristics of the charge and discharge data after extraction and alignment are implemented by a Transformer encoder.
[0072] Based on the formed topological relationship graph, to extract the topological features, the topological relationship graph is first segmented into multiple sub-topological graphs, and for any sub-topological graph, the node contribution rate convolutional network is used to extract the features, where the node contribution rate convolutional network processes the input sub-topological graph in the following manner:
[0073] D1 = ISRU{CNN(D, D node ).norm}
[0074] δ = top n (D T )
[0075]
[0076] where ISRU represents the ISRU activation function, D and D node respectively represent the nodes and the features of the nodes in the sub-topological graph, as the input of the CNN convolutional neural network, norm is the mathematical norm, CNN represents the convolutional neural network, D1 represents the features after the convolutional operation, top n represents n interesting nodes in the sub-topological relationship graph where the reserved temperature deviation is greater than the preset temperature deviation threshold D T , δ represents the weight factor, represents the multiplication of features, and D2 represents the output features of the node contribution rate convolutional network.
[0077] In some embodiments, to perform node strengthening on the extracted features to enhance the weights of the interesting nodes includes:
[0078] Determine the neighbor nodes that have temperature correlation with each interesting node. In some examples, for instance, after a certain node changes to an interesting node, if based on the temperature background, the neighbor node also changes to an interesting node, then it is judged to have temperature correlation. A temperature correlation coefficient can be added to the neighbor nodes of the interesting node to represent the strength of the correlation between the nodes.
[0079] Based on the determined neighbor nodes, enhance the weights of the interesting nodes, satisfying:
[0080] a(d m , d k ) = c T ReLU(H·[d m , d k )
[0081]
[0082] Among them, tanh represents the tanh activation function, ReLU represents the ReLU activation function, H represents the weight matrix, m represents the serial number of the node of interest, k represents the serial number of the temperature-dependent neighbor nodes, k′ represents the variable for performing summation, K represents the number of temperature-dependent neighbor nodes, d m , d k represents the unenhanced feature, d′ m represents the enhanced feature of the node of interest m, a(d m , d k ) represents the attention metric of the node of interest, c is a learnable hyperparameter, and T represents the transpose.
[0083] In some embodiments, the extracted charge-discharge features, topological features, and features after weight enhancement are adaptively fused through an adaptive weighted fusion function to satisfy:
[0084] X out =[Sigmoid(r)X1+(1 - Sigmoid(r))X2]
[0085] Among them, X out represents the feature after adaptive fusion, Sigmoid(*) represents the Sigmoid function, r represents the randomly initialized weight factor, X1 represents the extracted charge-discharge features and topological features, and X2 represents the features after weight enhancement.
[0086] In the embodiments of the present application, by designing a Transformer encoder + a node contribution rate convolutional network based on CNN for feature extraction at different levels, and by enhancing the temperature-dependent neighbor nodes of each node of interest, the processed features can contain both deep feature information and important shallow information that is beneficial to improving the recognition effect. And in the fusion branch, by fusing the extracted charge-discharge features, topological features, and features after weight enhancement through a designed adaptive weighted fusion function, the recognition effect of unknown working conditions under actual operating conditions can be further improved through self-adaptive fusion, the compatibility of the algorithm with unknown working conditions can be improved, and the forgetting problem of the model can be alleviated. Through the remaining power prediction method of the present application, the authenticity and accuracy of the remaining capacity prediction of the energy storage unit can be further improved, user anxiety can be alleviated, and user experience can be improved.
[0087] The embodiments of the present application also propose a remaining power prediction system for an energy storage system based on deep learning, including a processor and a memory. A computer program is stored on the memory, and when the computer program is executed by the processor, the steps of the aforementioned remaining power prediction method for an energy storage system based on deep learning are implemented.
[0088] Moreover, although exemplary embodiments have been described herein, the scope includes any and all embodiments based on the present disclosure that have equivalent elements, modifications, omissions, combinations (e.g., schemes that cross various embodiments), adaptations, or alterations. It is not limited to the examples described in this specification or during the implementation of this application, and the examples will be construed as non-exclusive.
[0089] The above description is intended to be illustrative and not restrictive. For example, the above examples (or one or more of them) can be used in combination with each other. For example, those of ordinary skill in the art can use other embodiments when reading the above description.
[0090] The above embodiments are only exemplary embodiments of the present disclosure. Those skilled in the art can make various modifications or equivalent replacements to the present invention within the spirit and scope of the present disclosure, and such modifications or equivalent replacements should also be regarded as falling within the scope of protection of the present invention.
Claims
1. A method for predicting the remaining power of an energy storage system based on deep learning, characterized in that, Applied to an energy storage system including multiple energy storage subsystems, and any one of the energy storage subsystems includes multiple energy storage units, the remaining power prediction method includes the following steps: For any one of the energy storage subsystems, obtain the layout information of each energy storage unit in the energy storage subsystem, as well as historical charge and discharge data, temperature information, and corresponding historical energy storage parameters; Based on the obtained layout information, construct a topological relationship graph for each energy storage unit of any one of the energy storage subsystems. In the topological relationship graph, the distance between nodes is used to describe the proximity relationship of the energy storage units, and the series and parallel relationships of the energy storage units are described by connecting lines; Screen the historical charge and discharge data and historical energy storage parameters of each energy storage unit of any one of the obtained energy storage subsystems, and assign the charge and discharge data and energy storage parameters of each energy storage unit in the same time period to the corresponding nodes of the topological relationship graph; and, assign a temperature background to the topological relationship graph based on the obtained temperature information; Based on the formed topological relationship graphs, construct a training data set, and train a local capacity prediction model based on the training data set; Process the newly collected charge and discharge data, temperature information, and energy storage parameters of each energy storage unit of the energy storage subsystem, and input them into the trained local capacity prediction model; Based on the obtained local capacity, determine the remaining capacity of any one of the energy storage subsystems.
2. The method for predicting the remaining power of the energy storage system based on deep learning according to claim 1, wherein The screening of the historical charge and discharge data of each energy storage unit of any one of the obtained energy storage subsystems includes: In each time period, clip or perform forward-backward time query on the historical charge and discharge data of each energy storage unit, so that the charge and discharge data in any time period cover the charge data and discharge data with the same power ratio for charging and discharging any one of the energy storage units; Judge the number of energy storage units in the charging or discharging state at the start time of any one of the time periods, and take the mode of the number as the alignment basis; On the time axis of any one of the time periods, based on the determined alignment basis, align the start times of each data segment of the charge data and discharge data, and arrange the charge data and discharge data at intervals; Assigning the energy storage parameters in the same time period to the corresponding nodes of the topological relationship graph includes: Determine the energy storage parameters at the end time of any one of the time periods, and assign the energy storage parameters at the end time of any one of the time periods to the corresponding nodes in the topological relationship graph.
3. The method for predicting the remaining power of the energy storage system based on deep learning according to claim 2, wherein Training the local capacity prediction model based on the training data set includes: Extract the charge and discharge characteristics of the charge and discharge data after alignment; and, According to the temperature background, use the nodes with temperature deviation greater than the preset temperature deviation threshold as the nodes of interest, and based on the formed topological relationship graph, extract topological characteristics;; Perform node enhancement on the extracted characteristics to enhance the weights of the nodes of interest; Perform adaptive fusion on the extracted charge and discharge characteristics, topological characteristics, and the characteristics after weight enhancement through an adaptive weighted fusion function; Perform feature extraction on the characteristics after weight enhancement again, and add the extracted characteristics to the adaptively fused characteristics; Input the added characteristics into the fully connected layer, and use the output of the fully connected layer as the prediction result of the local capacity prediction model.
4. The method for predicting the remaining power of the energy storage system based on deep learning according to claim 3, wherein The charge-discharge characteristics of the charge-discharge data after extraction and alignment are realized by a Transformer encoder; Based on the formed topological relationship graph, to extract topological features, the topological relationship graph is first divided into multiple sub-topological graphs, and any sub-topological graph is extracted by using a node contribution rate convolutional network, where the node contribution rate convolutional network processes the input sub-topological graph in the following way: D1 = ISRU{CNN(D, D node ).norm} δ = top n (D T ) Among them, ISRU represents the ISRU activation function, D, D node respectively represent the nodes and their features in the sub-topological graph, which are used as the input of the CNN (Convolutional Neural Network). norm is a mathematical norm, CNN represents the Convolutional Neural Network, D1 represents the features after convolutional operation, top n represents n interesting nodes in the sub-topological graph where the retained temperature deviation is greater than the preset temperature deviation threshold D T , δ represents the weight factor, represents feature multiplication, and D2 represents the output features of the node contribution rate convolutional network.
5. The method for predicting the remaining power of the energy storage system based on deep learning according to claim 4, wherein, Node enhancement is performed on the extracted features to enhance the weights of the nodes of interest, including: Determine the neighbor nodes that have temperature correlation with each node of interest; Based on the determined neighbor nodes, enhance the weights of the nodes of interest, satisfying: a(d m ,d k ) = c T ReLU(H·[d m ,d k ) Among them, tanh represents the tanh activation function, ReLU represents the ReLU activation function, H represents the weight matrix, m represents the serial number of the node of interest, k represents the serial number of the neighbor nodes with temperature correlation, k' represents the variable for performing the summation, K represents the number of neighbor nodes with temperature correlation, and d m , d k represents the unenhanced feature, and d' m represents the feature of the enhanced node of interest m, and a(d m , d k ) represents the attention measure of the node of interest, c is a learnable hyperparameter, and T represents the transpose.
6. The method for predicting the remaining power of the energy storage system based on deep learning according to claim 5, wherein The extracted charge-discharge characteristics, topological features, and the features with enhanced weights are adaptively fused through an adaptive weighted fusion function, satisfying: X out = {Sigmoid(r)X1+(1 - Sigmoid(r))X2] Among them, X out represents the feature after adaptive fusion, Sigmoid(*) represents the Sigmoid function, r represents the randomly initialized weight factor, X1 represents the extracted charge-discharge feature and topological feature, and X2 represents the feature after weight enhancement.
7. The method for predicting the remaining power of the energy storage system based on deep learning according to claim 1, wherein Based on the obtained local capacity, determine the remaining capacity of any energy storage subsystem, including: Based on the capacity prediction value output by the local capacity prediction model, for the proportion of the number of energy storage units corresponding to each node of interest, determine the remaining capacity of any energy storage subsystem in segments.
8. A remaining power prediction system for an energy storage system based on deep learning, characterized in that, It includes a processor and a memory, and a computer program is stored on the memory. When the computer program is executed by the processor, the steps of the method for predicting the remaining power of an energy storage system based on deep learning according to any one of claims 1 to 7 are implemented.