Intelligent management method and system of networking energy storage system based on energy prediction

Abstract

The invention relates to the technical field of electric power control, and particularly discloses an intelligent management method and system of a network construction energy storage system based on energy prediction, and the method comprises the steps: constructing a power grid topological graph embedded with virtual life nodes; extracting spatial features and a global life state from the power grid topological graph, and extracting time features from historical load data and historical frequency modulation data; inputting the spatial feature, the global life state and the time feature into a preset energy prediction model to obtain an energy prediction value, the energy prediction value including an output prediction value and a load prediction value; inputting the energy prediction value into a strategy decision model to obtain a target control strategy; and regulating and controlling the network construction energy storage system according to the target control strategy. The method solves the problem that in a traditional method, life perception isolation cannot be deeply coupled with energy prediction, and the frequency modulation precision of a power grid cannot be remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of power control technology, and in particular to an intelligent management method and system for a grid-connected energy storage system based on energy prediction. Background Art

[0002] As the scale of renewable energy grid connection continues to expand, grid frequency stability faces severe challenges. In traditional grid regulation technology, grid-connected energy storage systems are mainly used as voltage sources. However, this technology cannot meet the increasing demand for grid regulation. For this reason, technicians have proposed using grid-connected energy storage systems as a core means to smooth power fluctuations and provide fast frequency response.

[0003] The clustered operation of a grid-type energy storage system requires the coordination of power balance and life balance among multiple nodes. In existing control schemes, single nodes are often used as units for global coordinated control and independent optimization. However, in actual application, technicians have found that the existing technology has at least the following technical problems:

[0004] When a node operates at limited power due to life loss, adjacent nodes may be overloaded due to a sudden increase in load, causing regional stability risks. The existing control method of grid-connected energy storage systems mainly adopts passive control methods, that is, after the stability risk occurs, abnormal node maintenance or global adjustment is performed according to the actual usage conditions, resulting in fluctuations in the response to the power grid, greatly reducing the quality of grid access regulation and control. At the same time, it may pose a risk of overcharging or over-discharging the battery units in the energy storage system, reducing the service life of the energy storage system.

[0005] Therefore, the control strategies of existing grid-type energy storage systems have low timeliness and accuracy, high long-term operating costs, and cannot meet the actual needs of enterprises. Summary of the invention

[0006] In order to overcome the above technical problems existing in the prior art, an embodiment of the present invention provides an intelligent management method and system for a grid-connected energy storage system based on energy prediction.

[0007] In a first aspect, the present invention provides an intelligent management method for a grid-connected energy storage system based on energy prediction, the method comprising: constructing a grid topology map embedded with virtual life nodes; extracting spatial features and global life states from the grid topology map, and extracting time features from historical load data and historical frequency modulation data; inputting the spatial features, the global life states and the time features into a preset energy prediction model to obtain energy prediction values, the energy prediction values ​​including output prediction values ​​and load prediction values; inputting the output prediction values ​​and the load prediction values ​​into a strategy decision model to obtain a target control strategy; and regulating the grid-connected energy storage system according to the target control strategy.

[0008] Preferably, a power grid topology map embedded with virtual life nodes is constructed, including: determining the power grid topology nodes based on power grid equipment resource data; constructing the power grid topology map based on the physical connection relationship between the power grid topology nodes; determining the life loss rate and life loss gradient of each power grid topology node; constructing virtual life nodes based on the life loss rate and the life loss gradient; embedding the virtual life nodes into the power grid topology map to generate an embedded power grid topology map.

[0009] Preferably, the life loss rate of each power grid topology node is determined, including: obtaining operating data of each power grid topology node; constructing a life loss model based on the operating data; performing life change analysis on the corresponding power grid topology node based on the life loss model to generate life change data; generating a life loss rate based on the life change data; determining the life loss gradient of each power grid topology node, including: fitting the life loss rate of each power grid topology node based on a linear regression algorithm to generate fitting data; generating a life loss gradient based on the fitting data.

[0010] Preferably, the method further comprises: extracting bus nodes from the power grid topology nodes; binding and connecting each bus node with the virtual life node; and determining the edge weight between each bus node and the virtual life node based on the life sensitivity coefficient, wherein the edge weight is characterized as: ,in, It is represented as the connection weight between the virtual life node and the bus node. is the life sensitivity coefficient, is the global life loss rate.

[0011] Preferably, the preset energy prediction model includes a feature processing network and a prediction network; the feature processing network processes the spatial features, the global life state and the time features based on gated attention to obtain fused features; and the prediction network outputs predicted values ​​based on the edge weights and the fused features.

[0012] Preferably, the spatial features, the global life state and the temporal features are processed based on gated attention to obtain fused features, including: performing feature splicing on the spatial features, the global life state and the temporal features to obtain spliced ​​features; determining initial weights according to the spliced ​​features; dynamically adjusting the initial weights according to the global life state to obtain fused weights; and performing feature fusion on the spatial features and the temporal features according to the fusion weights to obtain fused features.

[0013] Preferably, the strategy decision model is a strategy decision model based on a proximal strategy optimization algorithm, and the reward function of the strategy decision model is: ,in, The grid frequency deviation, represents the life loss rate, Represents the difference between the predicted energy value and the actual value. represents a control value, which is used to indicate a target control strategy, , , , are all constants.

[0014] Preferably, the strategy decision model is also preceded by a multi-rolling optimization model, and the multi-rolling optimization model includes an objective function, and the objective function is: ,in, Indicates The grid frequency deviation at the moment, Indicates The life loss rate at each moment, Indicates The operating cost of the moment, represents the life protection weight, represents the economic weight, and the objective function satisfies the life loss constraint, which is: .

[0015] In a second aspect, an embodiment of the present invention provides an intelligent management system for a grid-connected energy storage system based on energy prediction, the system comprising: a graph construction module, for constructing a grid topology graph embedded with virtual life nodes; a feature extraction module, for extracting spatial features and global life states from the grid topology graph, and extracting time features from historical load data and historical frequency modulation data; an energy prediction module, for inputting the spatial features, the global life states and the time features into an energy prediction model to obtain energy prediction values, the energy prediction values ​​including output prediction values ​​and load prediction values; a strategy generation module, for inputting the energy prediction values ​​into a strategy decision model to obtain a target control strategy; and a regulation module, for regulating the grid-connected energy storage system according to the target control strategy.

[0016] The present invention has at least the following technical effects:

[0017] The present invention solves the problem that the isolated lifespan perception in the traditional method cannot be deeply coupled with the energy prediction, and significantly improves the frequency regulation accuracy and system reliability of the power grid.

[0018] Other features and advantages of the embodiments of the present invention will be described in detail in the subsequent detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The accompanying drawings are used to provide further understanding of the embodiments of the present invention and constitute a part of the specification. Together with the following specific implementation methods, they are used to explain the embodiments of the present invention, but do not constitute a limitation to the embodiments of the present invention.

[0020] Figure 1 The embodiment of the present invention provides a flowchart of an intelligent management method of a grid-connected energy storage system based on energy prediction;

[0021] Figure 2 A schematic diagram of the structure of an intelligent management system for a grid-connected energy storage system based on energy prediction provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0022] The specific implementation of the embodiment of the present invention is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation described here is only used to illustrate and explain the embodiment of the present invention, and is not used to limit the embodiment of the present invention.

[0023] The terms "system" and "network" in the embodiments of the present invention can be used interchangeably. "Multiple" means two or more than two. In view of this, "multiple" can also be understood as "at least two" in the embodiments of the present invention. "And / or" describes the association relationship of associated objects, indicating that three relationships can exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / ", unless otherwise specified, generally indicates that the previous and subsequent associated objects are in an "or" relationship. In addition, it should be understood that in the description of the embodiments of the present invention, the words "first", "second", etc. are only used to distinguish the purpose of description, and cannot be understood as indicating or implying relative importance, nor can they be understood as indicating or implying order.

[0024] Please refer to Figure 1 The embodiment of the present invention provides an intelligent management method of a grid-connected energy storage system based on energy prediction, the method comprising:

[0025] Step 1: construct a power grid topology map embedded with virtual life nodes.

[0026] In one possible implementation, in order to solve the technical problems in the prior art of strong volatility of the power grid and high randomness of power compensation demand, and to reduce the impact of power grid fluctuations on the grid-connected energy storage system, an intelligent learning model (such as a conventional deep learning model such as a convolutional neural network model) is used to predict the energy of the power grid, especially for a power grid with a photovoltaic system connected, to predict the energy it receives and consumes, so as to determine the control strategy of the grid-connected energy storage system for the power grid.

[0027] However, in actual application, if only the additional attributes of each grid node, such as extended fields such as voltage and power, are used as prediction parameters, once a node becomes abnormal due to its service life, the grid demand will fluctuate greatly instantly, which will affect the use of the grid energy storage system. Therefore, in an embodiment of the present invention, by topologically transforming the grid nodes and further considering the virtual life node in the analysis parameters of each topological node, the life state-related parameters of different grid topological nodes can be correlated with each other. After being processed by the fully connected layer, the overall perception of the global life state can be effectively improved. The model trained by this method can comprehensively consider the life state-related parameters of the grid nodes to make more accurate energy predictions, thereby managing the grid energy storage system more intelligently and improving the system control accuracy.

[0028] In an embodiment of the present invention, a power grid topology map embedded with virtual life nodes is constructed, including: determining the power grid topology nodes based on power grid equipment resource data; constructing the power grid topology map based on the physical connection relationship between the power grid topology nodes; determining the life loss rate and life loss gradient of each power grid topology node; constructing a virtual life node based on the life loss rate and the life loss gradient; embedding the virtual life node into the power grid topology map to generate an embedded power grid topology map.

[0029] In one possible implementation, a virtual life node is embedded in the power grid topology map. The attribute information of the virtual life node includes the life loss rate and the life loss gradient. The life loss rate indicates the proportion of the life loss of the system within a certain period of time. The loss gradient refers to the rate of change of the life loss rate over time, which is used to describe the acceleration or deceleration trend of the life loss. For example, in the initial stage, the characteristic vector of each virtual life node can be initialized to [0,0]. As the system runs, these characteristic parameters are continuously updated to reflect the global life status of each topological node in the entire power grid.

[0030] In an embodiment of the present invention, the life loss rate of each power grid topology node is determined, including: obtaining operating data of each power grid topology node; constructing a life loss model based on the operating data; performing life change analysis on the corresponding power grid topology node based on the life loss model to generate life change data; and generating a life loss rate based on the life change data.

[0031] Specifically, when evaluating the life loss rate of each power grid topology node, the historical operation data of each power grid topology node can be obtained, and the historical operation data includes but is not limited to the voltage data, temperature change data, active power and reactive power data of the power grid topology node. According to the above historical operation data, a life loss model can be constructed. For example, the life loss model is based on a commonly used life prediction model (such as a Weibull model or an artificial neural network (ANN) model). Through the life loss model, the life change data of the current topology node can be calculated, and the life loss rate of the current topology node can be determined according to the change in the life change data per unit time. Since the service life change of each power grid node is low and the life deviation is often small, in order to reduce system fluctuations and improve the stability of the control system, the life loss rate of all power grid topology nodes can be further weighted averaged to obtain the life loss rate of the virtual life node.

[0032] At this time, the life loss gradient is further calculated. For example, a technician calculates and compares the life loss rate in different time periods to obtain the life loss gradient. The historical data of the life loss rate can also be fitted using methods such as linear regression to calculate the life loss gradient.

[0033] In an embodiment of the present invention, by further embedding virtual life in the grid topology node, the energy prediction accuracy of the entire grid can be further improved, especially the impact of abnormal grid nodes on the grid can be quickly perceived, and the grid-forming energy storage system can be controlled to quickly respond to abnormal nodes and adjust the control strategy in time to avoid affecting the grid-forming energy storage system, thereby improving the stability and reliability of the grid-forming energy storage system.

[0034] After constructing the power grid topology graph embedded with virtual life nodes, the bus nodes of the power grid can be further extracted, and then each bus node is connected to the virtual life node. After the connection, the edge weight between each bus node and the virtual life node is further calculated.

[0035] In an embodiment of the present invention, the method further includes: extracting bus nodes from the power grid topology nodes; binding and connecting each bus node with the virtual life node; determining the edge weight between each bus node and the virtual life node based on the life sensitivity coefficient, and the edge weight is characterized as: ,in, It is represented as the connection weight between the virtual life node and the bus node. is the life sensitivity coefficient, is the global life loss rate.

[0036] In a preferred embodiment, in order to reduce the amount of preliminary calculations and improve calculation efficiency, the technician can set the edge weight between the bus node and the virtual life node to a fixed value based on work experience or actual conditions. In an embodiment of the present invention, the fixed value is preferably 0.5. However, in actual application, the above-mentioned setting as a fixed value will cause the model to have a global perception deviation of the life of the power grid node. Therefore, further, in the process of model prediction, a life sensitivity coefficient is set, and the edge weight between each bus node and the virtual life node is dynamically determined based on the life sensitivity coefficient, thereby effectively solving the problem of global perception deviation caused by fixed weights.

[0037] In an embodiment of the present invention, by selecting the bus as the core node in the power grid topology, compared with device-level nodes such as inverters and transformers, the bus node has a standardized and high-precision data acquisition system, which can synchronously obtain key parameters such as voltage and power with millisecond-level accuracy, avoiding the data integration problem caused by heterogeneous protocols of devices. By utilizing the hub status and global characterization capabilities of the bus in the power system, the topological modeling at the bus level greatly reduces the computational complexity, and improves the training efficiency of the graph neural network by more than 10 times, meeting the real-time control requirements. In addition, the hard connection design of the bus node and the virtual life node realizes the deep coupling of the life state and electrical characteristics: the virtual node generates a global health vector by aggregating the life data of the bus-related equipment, and reversely embeds the bus feature space, so that the control strategy can dynamically perceive the limitations of the power regulation capability of the equipment associated with the bus node. This design not only retains the physical authenticity of the topological structure, but also expands the life perception dimension through virtual nodes, avoiding the problems of local information overload and global state fragmentation in the inverter-level topology.

[0038] Step 2: extracting spatial features and global life state from the power grid topology map, and extracting time features from historical load data and historical frequency modulation data.

[0039] Specifically, the spatial features are extracted using the graph convolution network GCN, and the temporal features are extracted using the dual-channel LSTM. The features of the busbar node and the virtual life node are processed independently in the graph convolution network GCN to output the spatial features and the global life state respectively. The graph convolution network GCN can normalize the features of the neighbor nodes and use the degree to adjust the weights to prevent the deviation caused by the different node degrees. The graph convolution network includes 3 layers of GCN, each layer has an output dimension of 64, the activation function is ReLU, and the neighborhood aggregation adopts weighted average. The graph convolution network GCN performs feature processing as a conventional technology, which will not be repeated in this embodiment.

[0040] Specifically, the dual-channel LSTM includes two LSTM networks, one of which includes two layers of LSTM for processing historical load data, and one of which includes one layer of LSTM for processing historical frequency modulation data. Feature processing is a conventional technology and is not described in detail in this embodiment.

[0041] For example, the historical 72-hour load data (time step of 1 hour) is input into two LSTM networks to output the load baseline for the next 24 hours. The historical frequency regulation instruction data is input into one layer of LSTM (time step of 5 minutes to output the frequency regulation demand for the next hour). The load baseline is downsampled to a 1-hour step, and the frequency regulation demand is interpolated to a 1-hour step, and they are merged into a unified time series feature.

[0042] After obtaining the above features, they are input into the preset prediction model for further prediction processing.

[0043] Step 3: Input the spatial characteristics, the global life state and the time characteristics into a preset energy prediction model to obtain an energy prediction value, wherein the energy prediction value includes an output prediction value and a load prediction value.

[0044] In an embodiment of the present invention, the preset energy prediction model includes a feature processing network and a prediction network. For example, the preset energy prediction model is generated in advance by a technician based on deep neural network DNN model training; the spatial features, the global life state and the time features are processed by the feature processing network based on gated attention to obtain fused features; and the prediction network outputs a predicted value based on the edge weights and the fused features.

[0045] Specifically, the prediction network can adopt a multi-layer cascaded fully connected structure to map high-dimensional fusion features to the output and load prediction space through nonlinear transformation. The network input layer receives the fusion feature vector output by the feature processing network, which encodes the spatiotemporal correlation and equipment life status information. The first-layer fully connected network performs a global linear combination of the fusion features through a weight matrix composed of edge weights, and then introduces an activation function with adaptive saturation characteristics to suppress the interference of outliers while retaining key pattern features. The second-layer fully connected network further extracts high-order interactive features, and adopts a sparse weight distribution strategy to reduce the impact of redundant parameters on the prediction results. Finally, the output layer generates the new energy output and load prediction values ​​respectively through a dual-channel independent mapping mechanism to ensure that the prediction process of the two shares the underlying feature representation and avoids target conflicts. During the training process, the network dynamically adjusts the weights of each layer through back propagation, and combines the early stopping mechanism and the gradient clipping strategy to balance the fitting ability and generalization performance, so that the prediction stability can be maintained when the equipment health changes dynamically. Through hierarchical abstraction and feature decoupling, this structure realizes the accurate solution of multi-objective prediction in complex power grid environments, and effectively overcomes the problem of prediction deviation accumulation caused by feature coupling in traditional single network models.

[0046] The embodiment of the present invention splits the preset energy prediction model into a two-layer structure of a feature processing network and a prediction network, realizing the decoupled optimization of feature fusion and mapping decision. The feature processing network focuses on the deep interaction and dynamic calibration of multimodal features, effectively overcoming the problem of feature drowning and over-smoothing caused by static fusion in traditional methods. The prediction network performs nonlinear mapping based on high-purity fusion features, focusing on establishing an accurate association from feature space to output-load prediction values, avoiding the interference of redundant noise on the final prediction. This staged processing mechanism makes the model both flexible and stable: the feature processing layer can independently adapt to changes in the grid topology, while the prediction layer maintains robust modeling of the core output law. The two achieve global optimization through gradient collaborative updating, which not only improves the prediction robustness in extreme scenarios, but also ensures the strategy consistency of life-sensitive control, and ultimately achieves long-term balanced optimization of frequency regulation performance and system health.

[0047] Traditional prediction models usually adopt a staged modeling approach, processing temporal and spatial features independently and fusing them through concatenation or simple splicing. Such models rely on a static weight allocation mechanism and cannot dynamically adjust the fusion weights, resulting in a significant decrease in model prediction accuracy.

[0048] In an embodiment of the present invention, the spatial features, the global life state and the time features are processed based on gated attention to obtain fused features, including: performing feature splicing on the spatial features, the global life state and the time features to obtain spliced ​​features; determining initial weights according to the spliced ​​features; dynamically adjusting the initial weights according to the life loss rate to obtain fused weights; and performing feature fusion on the spatial features and the time features according to the fusion weights to obtain fused features.

[0049] Specifically, the spatial features, time features and global life state are spliced ​​along the feature dimension. Since the dimension of the global life state is low, the global life state needs to be expanded during splicing, and then splicing is performed. After the splicing is completed, based on the splicing features, the corresponding fusion weights are generated using the fully connected layer, the fusion weights are normalized using the Sigmoid activation function, and then the normalized fusion weights are adjusted according to the dynamic weight adjustment rules that consider the life loss rate.

[0050] More specifically, the dynamic adjustment rule is: when the life loss rate is greater than a set value, the fusion weight increases by a first reference value; when the life loss gradient is greater than zero, the fusion weight decreases by a second reference value.

[0051] In the embodiment of the present invention, the first reference value is preferably 0.2, and the second reference value is preferably 0.1.

[0052] It should be noted that, when fusing features, the embodiment of the present invention only fuses the time features and the space features, and does not directly fuse the global life state with the time features and the space features, thereby avoiding repeated introduction of noise into the life state.

[0053] The embodiment of the present invention is based on the gated attention mechanism and introduces the global life state to dynamically calibrate the spatiotemporal feature fusion weights: first, the spatial features, temporal features and the extended life state are spliced ​​to generate the initial weights; then the weights are dynamically adjusted according to the life loss rate and the health gradient to achieve adaptive allocation of feature contributions according to working conditions.

[0054] After the processing prediction value and the compliance prediction value are generated, they are further input into the strategy decision model. However, in actual application, traditional strategy decision models mostly use a single-objective reward function with fixed weights, such as optimizing only frequency deviation or economy, without considering the dynamic coupling of life loss and prediction error, resulting in reduced effectiveness of the control strategy and failure to meet actual needs.

[0055] Step 4: input the energy prediction value into the strategy decision model to obtain the target control strategy.

[0056] In an embodiment of the present invention, the policy decision model is a policy decision model based on a proximal policy optimization algorithm, and the reward function of the policy decision model is:

[0057]

[0058] in, The grid frequency deviation, represents the life loss rate, Represents the difference between the predicted energy value and the actual value. represents a control value, which is used to indicate a target control strategy, , , , are all constants.

[0059] The grid frequency deviation is the load forecast value minus the output forecast value minus the control value, and the control value is the regulated power.

[0060] In an embodiment of the present invention, the strategy decision model is also preceded by a multi-rolling optimization model, and the multi-rolling optimization model constructs an objective function with the goal of minimizing the life loss rate, the grid frequency and the operating cost;

[0061] The objective function is:

[0062]

[0063] in, Indicates The grid frequency deviation at the moment, Indicates The life loss rate at each moment, Indicates The operating cost of the moment, represents the life protection weight, represents the economic weight, and the objective function satisfies the life loss constraint, which is: .

[0064] Based on the current state vector, the multi-step rolling optimization calculates the decision variables required to achieve goals such as minimizing the life loss rate in the future. These optimized decision variables serve as a reference and provide direction for the training and adjustment of the policy network. During the learning process, the policy network will adjust the system decision in the direction of the optimal decision determined by the multi-step rolling optimization to generate a better policy. The policy output by the policy network will affect the actual operating state of the system, and then affect the state vector at the next moment. The multi-step rolling optimization will re-optimize the calculation according to the updated state vector and adjust the decision variables. This cycle continues. The dynamic adjustment of the policy network provides real-time changing system state information for the multi-step rolling optimization, so that the multi-step rolling optimization can continuously optimize the decision according to the actual situation.

[0065] Step 5: regulating the grid-connected energy storage system according to the target control strategy.

[0066] The embodiment of the present invention achieves a multi-objective dynamic balance of life protection, frequency stability and economy through deep coupling of the strategy decision model and multi-rolling optimization, solves the core problems of target conflict, response lag and prediction fragmentation in traditional methods, and provides a highly robust intelligent management solution for the operation of the grid-connected energy storage system.

[0067] The present invention solves the problem that the traditional method of isolated life perception cannot be deeply coupled with energy prediction, and significantly improves the frequency regulation accuracy of the power grid and the reliability of the equipment throughout its life cycle.

[0068] like Figure 2 As shown, based on the same inventive concept, an embodiment of the present invention further provides an intelligent management system for a grid-connected energy storage system based on energy prediction, the system comprising:

[0069] A graph construction module, used to construct a power grid topology graph embedded with virtual life nodes;

[0070] A feature extraction module, used to extract spatial features and global life state from the power grid topology map, and to extract time features from historical load data and historical frequency modulation data;

[0071] An energy prediction module, used for inputting the spatial characteristics, the global life state and the time characteristics into an energy prediction model to obtain energy prediction values, wherein the energy prediction values ​​include output prediction values ​​and load prediction values;

[0072] A strategy generation module, used for inputting the output prediction value and the load prediction value into a decision model to obtain a target control strategy;

[0073] A control module is used to control the grid-connected energy storage system according to the target control strategy.

[0074] It should be understood that the intelligent management method of a grid-connected energy storage system based on energy prediction provided in an embodiment of the present invention and the intelligent management system of a grid-connected energy storage system based on energy prediction provided in the above embodiment are based on the same inventive concept. For more specific working principles of each module in the embodiment of the present invention, please refer to the above embodiment and will not be repeated in this embodiment.

[0075] The optional implementation modes of the embodiments of the present invention are described in detail above in conjunction with the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details in the above implementation modes. Within the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical scheme of the embodiments of the present invention, and these simple modifications all belong to the protection scope of the embodiments of the present invention.

[0076] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the embodiments of the present invention will not further describe various possible combinations.

[0077] Those skilled in the art can understand that all or part of the steps in the above-mentioned embodiment system can be completed by instructing the relevant hardware through a program, and the program is stored in a storage medium, including several instructions to enable a single-chip microcomputer, a chip or a processor to execute all or part of the steps of the system described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0078] In addition, various implementation modes of the embodiments of the present invention may be arbitrarily combined, and as long as they do not violate the concept of the embodiments of the present invention, they should also be regarded as the contents disclosed by the embodiments of the present invention.

Claims

1. An intelligent management method for a grid-connected energy storage system based on energy prediction, characterized in that: The method comprises: Constructing a power grid topology diagram embedded with virtual life nodes; Extracting spatial features and global life state from the power grid topology map, and extracting time features from historical load data and historical frequency modulation data; Inputting the spatial characteristics, the global life state and the time characteristics into a preset energy prediction model to obtain an energy prediction value, wherein the energy prediction value includes an output prediction value and a load prediction value; Inputting the output prediction value and the load prediction value into a strategy decision model to obtain a target control strategy; The grid-connected energy storage system is regulated according to the target control strategy.

2. According to claim 1, an intelligent management method for a grid-connected energy storage system based on energy prediction is characterized in that: Construct a power grid topology diagram with embedded virtual life nodes, including: Determine the grid topology nodes based on grid equipment resource data; Constructing a power grid topology map according to the physical connection relationship between the power grid topology nodes; Determine the life loss rate and life loss gradient of each grid topology node; Constructing a virtual life node based on the life loss rate and the life loss gradient; The virtual life node is embedded in the power grid topology map to generate an embedded power grid topology map.

3. According to claim 1, an intelligent management method for a grid-connected energy storage system based on energy prediction is characterized in that: Determine the life loss rate of each grid topology node, including: Obtain the operating data of each grid topology node; constructing a life loss model based on the operating data; Based on the life loss model, a life change analysis is performed on the corresponding power grid topology node to generate life change data; generating a life loss rate based on the life change data; Determine the life loss gradient of each grid topology node, including: Fitting the life loss rate of each grid topology node based on a linear regression algorithm to generate fitting data; A lifetime loss gradient is generated based on the fitted data.

4. The intelligent management method of a grid-connected energy storage system based on energy prediction according to claim 1 is characterized in that: The method further comprises: Extracting busbar nodes from the power grid topology nodes; Binding and connecting each bus node to the virtual life node; The edge weight between each bus node and the virtual life node is determined based on the life sensitivity coefficient, and the edge weight is characterized as: ， in, It is represented as the connection weight between the virtual life node and the bus node. is the life sensitivity coefficient, is the global life loss rate.

5. The intelligent management method of a grid-connected energy storage system based on energy prediction according to claim 4 is characterized in that: The preset energy prediction model includes a feature processing network and a prediction network; The spatial feature, the global life state and the temporal feature are processed by the feature processing network based on gated attention to obtain a fusion feature; Outputting a prediction value based on the edge weight and the fused feature through the prediction network.

6. The intelligent management method of a grid-connected energy storage system based on energy prediction according to claim 5 is characterized in that: The spatial feature, the global life state and the temporal feature are processed based on gated attention to obtain a fusion feature, including: Performing feature splicing on the spatial feature, the global life state and the time feature to obtain a spliced ​​feature; Determining an initial weight according to the splicing feature; Dynamically adjusting the initial weight according to the global life state to obtain a fusion weight; The spatial feature and the temporal feature are fused according to the fusion weight to obtain a fusion feature.

7. The intelligent management method of a grid-connected energy storage system based on energy prediction according to claim 1 is characterized in that: The strategy decision model is a strategy decision model based on the proximal strategy optimization algorithm, and the reward function of the strategy decision model is: ， in, The grid frequency deviation, represents the life loss rate, Represents the difference between the predicted energy value and the actual value. represents a control value, which is used to indicate a target control strategy, , , , are all constants.

8. The intelligent management method of a grid-connected energy storage system based on energy prediction according to claim 7 is characterized in that: The strategy decision model is also preceded by a multi-rolling optimization model, and the multi-rolling optimization model includes an objective function, which is: ， in, Indicates The grid frequency deviation at the moment, Indicates The life loss rate at each moment, Indicates The operating cost of the moment, represents the life protection weight, represents the economic weight, and the objective function satisfies the life loss constraint, which is: .

9. An intelligent management system for a grid-connected energy storage system based on energy prediction, characterized in that: The system comprises: A graph construction module, used to construct a power grid topology graph embedded with virtual life nodes; A feature extraction module, used to extract spatial features and global life state from the power grid topology map, and to extract time features from historical load data and historical frequency modulation data; An energy prediction module, used for inputting the spatial characteristics, the global life state and the time characteristics into an energy prediction model to obtain an energy prediction value, wherein the energy prediction value includes an output prediction value and a load prediction value; A strategy generation module, used for inputting the energy prediction value into a strategy decision model to obtain a target control strategy; A control module is used to control the grid-connected energy storage system according to the target control strategy.

Images