A method and system for predicting load power of an energy storage power station

Abstract

The present application relates to the technical field of intelligent control of power energy storage system, and particularly relates to a load power prediction method and system for energy storage power station. The present application solves the technical problems of the prior art, such as the poor load power prediction effect caused by the ignorance of the electrical topology connection and functional coupling relationship between each energy storage unit in the energy storage power station, and the limited feature expression capability.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for power storage systems, and more specifically, to a method and system for predicting the load power of an energy storage power station. Background Technology

[0002] Electrochemical energy storage power stations play a crucial role in power system frequency regulation and peak shaving, renewable energy consumption, and demand-side response. Managing the power output of energy storage power stations requires forecasting the load power of each energy storage unit to formulate grid dispatch plans, optimize power dispatch strategies, and achieve more balanced management of the energy storage batteries.

[0003] In existing technologies, statistical learning-based prediction methods are commonly used to forecast the load power of energy storage power stations. These methods typically extract statistical patterns from historical load data for prediction. However, they are essentially linear or shallow nonlinear models with limited feature representation capabilities, resulting in generally low prediction accuracy. On one hand, existing methods often treat the load power prediction of each energy storage unit as an independent task, either modeling each unit separately or simply concatenating the power data of all units before inputting it into the model for prediction. This approach ignores the correlation between energy storage units within the power station caused by electrical topology connections and functional coupling. In actual operation, different energy storage units within the same power station are coupled in electrical topology, and their operational functions are constrained by a unified energy management strategy, resulting in a clear spatial dependence between their load power. Ignoring this spatial correlation further limits prediction accuracy.

[0004] Therefore, existing related technologies have technical problems such as ignoring the electrical topology connection and functional coupling relationship between the energy storage units inside the energy storage power station when predicting load power, and limited feature expression ability leading to poor load power prediction results. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for predicting the load power of an energy storage power station, which solves the technical problems in the prior art, such as ignoring the electrical topology connection and functional coupling relationship between the energy storage units inside the energy storage power station when predicting load power, and the limited feature expression capability leading to poor load power prediction effect.

[0006] To solve the above-mentioned technical problems, the solution adopted in this application is as follows:

[0007] A method for predicting the load power of an energy storage power station includes:

[0008] S1: Obtain the operation data of each energy storage node in the energy storage power station, construct the original feature vector based on the operation data, calculate the SOH derivative physical quantity based on the battery health status, and concatenate the SOH derivative physical quantity and the original feature vector to obtain the enhanced feature vector;

[0009] S2: Construct the adjacency matrix of the physical layer graph structure and the adjacency matrix of the functional layer graph structure of the energy storage node;

[0010] S3: Construct a mask matrix based on the grid connection switching status of each energy storage node, and use the mask matrix to mask the adjacency matrix of the physical layer graph structure and the adjacency matrix of the functional layer graph structure to obtain the physical layer mask adjacency matrix and the functional layer mask adjacency matrix.

[0011] S4: Construct the initial feature matrix, and define the graph convolution operator using the physical layer mask adjacency matrix and the functional layer mask adjacency matrix;

[0012] S5: Calculate the current maximum available discharge power and maximum available charging power of each energy storage node;

[0013] S6: Construct a total loss function, train a graph neural network model using the constructed total loss function, and perform graph convolution processing through the defined graph convolution operator to obtain the predicted load power values ​​of each energy storage node.

[0014] Preferably, the specific implementation method of S1 includes:

[0015] S1.1: Obtain the operating data of each energy storage node, and construct the original feature vector of the operating data of each energy storage node according to the energy storage node number and sampling time;

[0016] S1.2: Map the battery health state to the SOH derivative physical quantity;

[0017] S1.3: The battery health status and its SOH derivative physical quantities are concatenated to the original feature vector to form an enhanced feature vector.

[0018] Preferably, the specific implementation method of S2 includes:

[0019] S2.1: Based on the electrical topology parameters of the energy storage power station, the electrical distance between any two energy storage nodes is calculated using the equivalent impedance magnitude between energy storage nodes and the system reference impedance.

[0020] S2.2: Map electrical distances to an adjacency matrix of the physical layer graph structure using indicator functions;

[0021] S2.3: Calculate the SOH state similarity between energy storage nodes based on the battery health status of each energy storage node;

[0022] S2.4: Calculate the load behavior correlation between energy storage nodes based on the historical charging and discharging power of each energy storage node;

[0023] S2.5: The adjacency matrix of the functional layer graph structure is constructed by weighted fusion of the similarity between the SOH state and the correlation between load behavior.

[0024] Preferably, the specific implementation method of S3 includes:

[0025] S3.1: Obtain the grid connection status indicator variables of each energy storage node, and construct a mask matrix using the grid connection status indicator variables;

[0026] S3.2: Multiply the adjacency matrix of the physical layer graph structure by the mask matrix on the left and right respectively to obtain the physical layer mask adjacency matrix;

[0027] S3.3: Multiply the adjacency matrix of the functional layer graph structure by the mask matrix on the left and right respectively to obtain the functional layer mask adjacency matrix.

[0028] Preferably, the specific implementation method of S4 includes:

[0029] S4.1: Stack the enhanced feature vectors of each energy storage node row by row to form an initial feature matrix;

[0030] S4.2: Normalize the physical layer mask adjacency matrix and the functional layer mask adjacency matrix to obtain the normalized physical layer mask adjacency matrix and functional layer mask adjacency matrix.

[0031] S4.3: A graph convolution operator is defined based on the normalized physical layer mask adjacency matrix, the functional layer mask adjacency matrix, and the initial feature matrix. This graph convolution operator uses the normalized physical layer mask adjacency matrix and the functional layer mask adjacency matrix as the graph structure input and the initial feature matrix as the node feature input.

[0032] Preferably, in step S4.2, before the normalization process, the following processing method is further included:

[0033] An identity matrix is ​​superimposed on the diagonal of the physical layer mask adjacency matrix and the functional layer mask adjacency matrix, respectively.

[0034] Preferably, the specific implementation method of S5 includes:

[0035] S5.1: Establish the mapping relationship between battery health status and battery internal resistance, and calculate the battery internal resistance;

[0036] S5.2: Calculate the current maximum available discharge power of the energy storage node based on the battery internal resistance;

[0037] S5.3: Calculate the current maximum available charging power of the energy storage node based on the battery internal resistance.

[0038] Preferably, the specific implementation method of S6 includes:

[0039] S6.1: Construct a total loss function that includes a prediction error loss term, a SOH physical constraint penalty term, and a weight decay regularization term;

[0040] S6.2: Training a graphical neural network model based on the constructed total loss function;

[0041] S6.3: Input the enhanced feature matrix into the trained graph neural network model. The trained graph neural network model loads the trained weight parameters and performs graph convolution layer by layer according to the defined graph convolution operator to output the predicted load power of each energy storage node at each sampling time.

[0042] Preferably, the specific implementation method of S6.2 includes:

[0043] S6.2.1: Input the enhanced feature matrix into the graph neural network model, perform graph convolution layer by layer forward according to the graph convolution operator defined in S4, and output the predicted load power of each energy storage node at each sampling time;

[0044] S6.2.2: Based on the total loss function, calculate the total loss value under the current graph neural network model parameters using the maximum available discharge power and the maximum available charging power;

[0045] S6.2.3: Using the chain rule, calculate the gradient of the total loss function with respect to each weight parameter to be trained;

[0046] S6.2.4: Update the parameters of each weight to be trained by updating the gradient along the gradient descent direction according to the preset step size;

[0047] S6.2.5: Repeat the above steps until the total loss value no longer decreases or the maximum number of training rounds is reached, and save the trained weight parameters.

[0048] A load power prediction system for an energy storage power station, applicable to the aforementioned load power prediction method for an energy storage power station, includes:

[0049] The data acquisition and feature enhancement module is used to acquire the operating data of each energy storage node in the energy storage power station, construct the original feature vector based on the operating data, calculate the SOH derivative physical quantity based on the battery health status, and concatenate the SOH derivative physical quantity and the original feature vector to obtain the enhanced feature vector;

[0050] The graph structure construction module, connected to the data acquisition and feature enhancement module, is used to construct the adjacency matrix of the physical layer graph structure and the adjacency matrix of the functional layer graph structure of the energy storage node.

[0051] The mask processing module, connected to the graph structure construction module, is used to construct a mask matrix based on the grid-connected switching state of each energy storage node. The mask matrix is ​​used to perform mask processing on the adjacency matrix of the physical layer graph structure and the adjacency matrix of the functional layer graph structure to obtain the physical layer mask adjacency matrix and the functional layer mask adjacency matrix.

[0052] The graph convolution operator definition module is connected to the mask processing module. It is used to construct the initial feature matrix and define the graph convolution operator using the physical layer mask adjacency matrix and the functional layer mask adjacency matrix.

[0053] The constraint calculation module, connected to the graph convolution operator definition module, is used to calculate the current maximum available discharge power and maximum available charging power of each energy storage node.

[0054] The model training output module is connected to the constraint calculation module to construct the total loss function. The constructed total loss function is used to train the graph neural network model, and graph convolution is performed through the defined graph convolution operator to obtain the load power prediction value of each energy storage node.

[0055] The technical solution of this application has at least the following advantages and beneficial effects:

[0056] The energy storage power station load power prediction method and system provided by this invention maps battery health status to three derivative physical quantities—SOH degradation rate, charge / discharge efficiency, and available capacity—as input features by constructing an enhanced feature matrix. Based on this, a physical layer graph structure and a functional layer graph structure are constructed to extract electrical coupling and functional coupling relationships between energy storage nodes, enabling the graph convolution model to simultaneously aggregate spatial neighborhood information from both electrical topology and functional coupling dimensions, thus compensating for the information loss problem of single topology modeling. Furthermore, a mask matrix is ​​constructed using the grid-connected switch status to mask the physical layer adjacency matrix and the functional layer adjacency matrix, allowing the load power prediction to adapt to the start-stop state switching of energy storage nodes, avoiding the need for specific targeting. Each topology scenario is trained or fine-tuned separately. Based on this, the present invention adopts a dual-channel graph convolution processing method to extract the information of adjacent energy storage nodes in the physical layer and the information of adjacent energy storage nodes in the functional layer, and then convolutionally fuse them so that the output features of each energy storage node simultaneously include physical electrical topology features and functional layer features. Finally, the maximum available charge and discharge power of each energy storage node is calculated by the mapping relationship between battery health status and battery internal resistance, and a total loss function is constructed to constrain the prediction of load power. This solves the technical problems in the prior art that ignore the electrical topology connection and functional coupling relationship between the energy storage units inside the energy storage power station when predicting load power, and the limited feature expression ability leads to poor load power prediction effect. Attached Figure Description

[0057] Figure 1 This is a flowchart of the energy storage power station load power prediction method of the present invention. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] This invention discloses a method for predicting the load power of an energy storage power station, comprising the following steps:

[0060] S1: Obtain the operation data of each energy storage node in the energy storage power station, construct the original feature vector based on the operation data, calculate the SOH derivative physical quantity based on the battery health status, and concatenate the SOH derivative physical quantity and the original feature vector to obtain the enhanced feature vector;

[0061] S2: Construct the adjacency matrix of the physical layer graph structure and the adjacency matrix of the functional layer graph structure of the energy storage node;

[0062] S3: Construct a mask matrix based on the grid connection switching status of each energy storage node, and use the mask matrix to mask the adjacency matrix of the physical layer graph structure and the adjacency matrix of the functional layer graph structure to obtain the physical layer mask adjacency matrix and the functional layer mask adjacency matrix.

[0063] S4: Construct the initial feature matrix, and define the graph convolution operator using the physical layer mask adjacency matrix and the functional layer mask adjacency matrix;

[0064] S5: Calculate the current maximum available discharge power and maximum available charging power of each energy storage node;

[0065] S6: Construct a total loss function, train a graph neural network model using the constructed total loss function, and perform graph convolution processing through the defined graph convolution operator to obtain the predicted load power values ​​of each energy storage node.

[0066] In this embodiment, in S1, the operating data of each energy storage node in the energy storage power station is acquired, an original feature vector is constructed based on the operating data, the SOH derivative physical quantity is calculated based on the battery health state, and the SOH derivative physical quantity and the original feature vector are concatenated to obtain an enhanced feature vector. The specific implementation method includes the following steps:

[0067] S1.1: Obtain the operating data of each energy storage node, and construct the original feature vector of the operating data of each energy storage node according to the energy storage node number and sampling time;

[0068] Specifically, the operational data includes historical charge and discharge power. Battery state of charge Battery health status Terminal voltage Charging and discharging current Battery temperature This includes the electrical topology parameters of the energy storage power station and the grid connection status. Indicates the energy storage node number. Indicates the sampling time;

[0069] For any energy storage node At sampling time Take its current sampling time and before The historical charge / discharge power, battery state of charge, terminal voltage, charge / discharge current, and battery temperature at each sampling time constitute the original feature vector, namely:

[0070] ;

[0071] in, The preset historical sequence length is used to provide time-series context information for load forecasting.

[0072] As an example, in this embodiment The value is 4, meaning that the original feature vector of each energy storage node at any sampling time contains 5 types of operational data from 5 sampling times.

[0073] S1.2: Map the battery health state to the SOH derivative physical quantity;

[0074] Specifically, the physical properties of SOH derivatives include the SOH degradation rate. Charge and discharge efficiency Available capacity ;

[0075] Among them, the degradation rate of SOH The result is obtained using the following formula:

[0076] ;

[0077] in, The preset time window length;

[0078] As an example, The preferred time is 1 hour;

[0079] The state of harmonics (SOH) degradation rate reflects how fast the batteries in an energy storage node age. The higher the degradation rate, the faster the power output capability of the energy storage node degrades.

[0080] Charge and discharge efficiency The result is obtained using the following formula:

[0081] ;

[0082] in, This refers to the initial charge and discharge power of the battery. This is the efficiency degradation coefficient, which is determined by the factory parameters of the energy storage battery.

[0083] Battery health status The lower the efficiency, the lower the charge / discharge efficiency. The lower the value, the greater the actual heat generation and energy loss of the energy storage battery under the same power command.

[0084] Available capacity The result is obtained using the following formula:

[0085] ;

[0086] in, Rated capacity of the energy storage battery, battery health status The lower the value, the smaller the usable capacity of the energy storage battery, and the shorter the time the energy storage battery can continuously output power.

[0087] S1.3: The battery health status and its SOH derivative physical properties are concatenated to the original feature vector to form an enhanced feature vector;

[0088] Specifically, the splicing process is as follows:

[0089] ;

[0090] in, This is a vector concatenation operation; The original feature vector, To enhance the feature vector.

[0091] In this step, the battery health status is expanded from a single numerical feature to a multi-dimensional set of physical features, including SOH degradation rate, charge / discharge efficiency, and available capacity. This approach is taken because the impact of battery health status on energy storage node load prediction cannot be reflected by a single physical feature alone, but rather by the combined limitation of the energy storage battery's power output capability by three physical features: internal resistance, efficiency, and capacity.

[0092] In this embodiment, in S2, the adjacency matrix of the physical layer graph structure and the adjacency matrix of the functional layer graph structure of the energy storage node are constructed. The specific implementation method includes:

[0093] S2.1: Based on the electrical topology parameters of the energy storage power station, the electrical distance between any two energy storage nodes is calculated using the equivalent impedance magnitude between energy storage nodes and the system reference impedance.

[0094] Specifically, the electrical distance between any two energy storage nodes is calculated using the following formula:

[0095] ;

[0096] in, For energy storage nodes To energy storage nodes In the The equivalent impedance magnitude along the transmission path, The system reference impedance, For energy storage nodes To energy storage nodes The smaller the electrical distance between the two energy storage nodes, the tighter the electrical coupling between them, and the greater the mutual influence of their power.

[0097] It should be noted that the electrical topology of an energy storage power station is a circuit network consisting of energy storage nodes (such as battery clusters) and branches (such as transformers). Current needs to pass through several branches to be transmitted from one energy storage node to another. These branches constitute a transmission path. There are multiple transmission paths between two energy storage nodes. The lines on each transmission path have impedance. The magnitude of the impedance is the equivalent impedance magnitude, which reflects the degree to which the path impedes the current.

[0098] In power system analysis, to eliminate the order-of-magnitude differences caused by different voltage levels, the actual impedance value is usually divided by a reference value for per-unit processing. This reference value is the system reference impedance, which is calculated from the power system reference capacity and reference voltage.

[0099] S2.2: Map electrical distances to an adjacency matrix of the physical layer graph structure using indicator functions;

[0100] Specifically, the mapping process described above is as follows:

[0101] ;

[0102] in, The adjacency matrix of the physical layer graph structure; The electrical distance attenuation coefficient, The maximum effective electrical distance threshold, This is an indicator function. The indicator function takes the value 1 when its condition is true, and takes the value 0 otherwise.

[0103] and The values ​​are all determined by the actual electrical topology of the energy storage power station, and this invention does not limit their specific values.

[0104] Energy storage nodes that are electrically closer have higher initial association weights in the physical layer graph structure, and when the electrical distance exceeds a threshold... When there is no connection between two energy storage nodes, the edge weight is 0. In a power system, the degree of mutual influence between two nodes is mainly determined by the electrical distance. Therefore, in this step, the adjacency matrix of the physical layer graph structure is initialized with the electrical distance.

[0105] It should be noted that graph neural networks organize the information transmission relationships between nodes through a graph structure. A graph structure consists of vertices and edges. In this invention, each energy storage node is a vertex; an edge represents an information transmission relationship between two vertices, and the edge weight represents the strength of this information transmission relationship. All edge weights are arranged in vertex order to form a square matrix called the adjacency matrix. In this matrix, the... Line number The elements of the column are the vertices. With vertex Edge weights between them.

[0106] S2.3: Calculate the SOH state similarity between energy storage nodes based on the battery health status of each energy storage node;

[0107] Specifically, the formula for calculating the similarity of SOH states is as follows:

[0108] ;

[0109] in, This is the SOH similarity bandwidth parameter. For energy storage nodes and energy storage nodes Similarity of SOH states between them.

[0110] It should be noted that the SOH similarity bandwidth parameter is used to control the rate at which similarity decays with differences in battery health. The higher the value, the more gradual the similarity decay caused by differences in the same battery health status; conversely, the lower the value, the steeper the similarity decay.

[0111] State of Health (SOH) similarity measures the degree of similarity between any two energy storage nodes in terms of battery health state, with a value range of [value missing]. The closer the value is to 1, the closer the battery health status of the two energy storage nodes is.

[0112] S2.4: Calculate the load behavior correlation between energy storage nodes based on the historical charging and discharging power of each energy storage node;

[0113] Specifically, the formula for calculating the correlation between load behavior is as follows:

[0114] ;

[0115] in, The historical window length is expressed in the number of sampling points.

[0116] As an example, This represents the 12 most recent sampling points.

[0117] For energy storage nodes Average power within a historical window.

[0118] It should be noted that load behavior correlation is used to measure the similarity between any two energy storage nodes in their historical charge and discharge power fluctuations, and the value range is [value range missing]. The closer the value is to 1, the more consistent the load power change trends of the two energy storage nodes are; conversely, the values ​​are opposite.

[0119] S2.5: The adjacency matrix of the functional layer graph structure is constructed by weighted fusion of the similarity between SOH states and the correlation between load behavior;

[0120] Specifically, the adjacency matrix of the functional layer graph structure is constructed as follows:

[0121] ;

[0122] in, , To integrate the weighting coefficients, and .

[0123] In the actual operation of energy storage power stations, even if energy storage nodes with similar battery health states are geographically dispersed, they will exhibit coordinated charging and discharging behaviors under grid dispatch commands due to the similarity of available battery capacity and charging and discharging efficiency. This coordinated relationship driven by the degradation state cannot be obtained solely through load behavior correlation. Therefore, in this step, the SOH state similarity is calculated and weighted and fused with the load behavior correlation, so that the functional layer graph structure can simultaneously obtain the two-dimensional relationship between different energy storage nodes, namely load behavior correlation and SOH state similarity.

[0124] In this embodiment, in S3, a mask matrix is ​​constructed based on the grid-connected switching state of each energy storage node. The mask matrix is ​​then used to mask the adjacency matrix of the physical layer graph structure and the adjacency matrix of the functional layer graph structure, resulting in the physical layer mask adjacency matrix and the functional layer mask adjacency matrix. The specific implementation method includes:

[0125] S3.1: Obtain the grid connection status indicator variables of each energy storage node, and construct a mask matrix using the grid connection status indicator variables;

[0126] Specifically, grid connection status indicator variable A value of 1 indicates an energy storage node. The grid is in a connected state; the grid connection status indicator variable is... A value of 0 indicates an energy storage node. Currently offline;

[0127] The mask matrix is:

[0128] ;

[0129] in, The total number of energy storage nodes is given by the mask matrix. A diagonal matrix, whose diagonal line is the first... Each element is a node. Grid connection status indicator variable All off-diagonal elements are 0.

[0130] S3.2: Multiply the adjacency matrix of the physical layer graph structure by the mask matrix on the left and right respectively to obtain the physical layer mask adjacency matrix;

[0131] Specifically, the above masking process is performed as follows:

[0132] ;

[0133] in, This is the physical layer mask adjacency matrix.

[0134] For any two energy storage nodes and energy storage nodes The corresponding elements in the physical layer mask adjacency matrix after masking are:

[0135] .

[0136] S3.3: Multiply the adjacency matrix of the functional layer graph structure by the mask matrix on the left and right respectively to obtain the functional layer mask adjacency matrix;

[0137] Specifically, the above masking process is performed as follows:

[0138] ;

[0139] in, This is the adjacency matrix of the functional layer mask.

[0140] For any two energy storage nodes and energy storage nodes The corresponding elements in the adjacency matrix of the functional layer mask after masking are:

[0141] .

[0142] In this step, if the energy storage node or energy storage node If either node is in an off-grid state, the edge weight between the two nodes is set to 0. This isolates the off-grid energy storage node from the graph structure, and its historical topology is temporarily isolated. When the energy storage node reconnects to the grid, the topology is restored. The reason for this is that energy storage nodes need to frequently switch between on-grid and off-grid states during maintenance or fault protection. Existing load forecasting methods require retraining the load forecasting model to adapt to such topology changes. In this step, a mask matrix is ​​used to achieve adaptive graph structure changes with topology.

[0143] In this embodiment, in step S4, an initial feature matrix is ​​constructed, and a graph convolution operator is defined using the physical layer mask adjacency matrix and the functional layer mask adjacency matrix. The specific implementation method includes the following steps:

[0144] S4.1: Stack the enhanced feature vectors of each energy storage node row by row to form an initial feature matrix;

[0145] Specifically, the initial feature matrix :

[0146] ;

[0147] in, This represents the total number of energy storage nodes.

[0148] In this matrix, each column corresponds to a feature dimension, and each row corresponds to an energy storage node.

[0149] S4.2: Normalize the physical layer mask adjacency matrix and the functional layer mask adjacency matrix to obtain the normalized physical layer mask adjacency matrix and functional layer mask adjacency matrix.

[0150] S4.3: A graph convolution operator is defined based on the normalized physical layer mask adjacency matrix, the functional layer mask adjacency matrix, and the initial feature matrix. This graph convolution operator takes the normalized physical layer mask adjacency matrix and the functional layer mask adjacency matrix as the graph structure input and the initial feature matrix as the node feature input.

[0151] Specifically, the formula for the above graph convolution processing operator is:

[0152] ;

[0153] In this step, the graph convolution operator mentioned above is only used as a mathematical definition of the graph convolution operator in the graph neural network model. Its actual calculation process is performed during the graph neural network model training and inference process in step S6.

[0154] This is the normalized adjacency matrix of the physical layer mask adjacency matrix; This is the normalized adjacency matrix of the functional layer mask adjacency matrix; For activation functions;

[0155] The graph convolutional network is set to... layer, The value range is from 2 to 4. As an example, The preferred value is 2, when In this process, each energy storage node can aggregate the feature information of its second-order neighboring nodes. Excessively deep graph convolutional layers can lead to homogenization of features among energy storage nodes. Therefore, this embodiment employs a two-layer graph convolutional structure to obtain topological and functional associations between energy storage nodes while avoiding overly smooth features.

[0156] As an example, the activation function can be the ReLU function. The normalization of the physical layer mask adjacency matrix and the functional layer mask adjacency matrix can be performed by normalization methods such as symmetric normalization and random walk normalization. Those skilled in the art will understand that symmetric normalization and random walk normalization are common normalization methods for matrix normalization in graph convolutional networks, so their specific processing procedures and formulas will not be elaborated here.

[0157] , , These are the training weight parameter matrices for the physical layer, the functional layer, and the self-connection layer, respectively.

[0158] A linear transformation representing the characteristics of the energy storage node itself, where For the first The self-connected weight parameter matrix of the layer to be trained For the first The dimension of the layer input features, For the first The dimension of the layer output features.

[0159] Furthermore, prior to normalization, the following processing methods are also included:

[0160] An identity matrix is ​​superimposed on the diagonal of both the physical layer mask adjacency matrix and the functional layer mask adjacency matrix.

[0161] The dimensions and size of the identity matrix are consistent with those of the physical layer mask adjacency matrix and the functional layer mask adjacency matrix. In this step, the operation of superimposing the identity matrix enables each energy storage node to simultaneously aggregate its own feature information and the feature information of its neighboring nodes when performing graph convolution processing.

[0162] In this step, each convolutional layer simultaneously aggregates information from adjacent energy storage nodes from both the physical and functional layers, and fuses it with the features of the energy storage node itself. The convolutional result of that layer is then output through an activation function. This approach is because the mutual influence between energy storage nodes includes both physical topology factors such as electrical distance and functional factors such as SOH similarity and load similarity. By extracting these two types of influence information in parallel through two-layer convolution, the load prediction results are more comprehensive and accurate.

[0163] In this embodiment, in S5, the current maximum available discharge power and maximum available charging power of each energy storage node are calculated. The specific implementation method includes:

[0164] S5.1: Establish the mapping relationship between battery health status and battery internal resistance, and calculate the battery internal resistance;

[0165] Specifically, battery health status With battery internal resistance The parsing mapping relationship is as follows:

[0166] ;

[0167] in, The initial internal resistance of the battery. and The fitting parameters are related to the battery type and are determined by offline aging calibration experiments for this battery model.

[0168] S5.2: Calculate the current maximum available discharge power of the energy storage node based on the battery internal resistance;

[0169] Specifically, no. Each energy storage node at the sampling time Maximum available discharge power for:

[0170] ;

[0171] in, It is an open-circuit voltage function. This is the battery discharge cutoff voltage.

[0172] S5.3: Calculate the current maximum available charging power of the energy storage node based on the battery internal resistance;

[0173] Specifically, no. Each energy storage node at the sampling time Maximum available charging power for:

[0174] ;

[0175] in, The battery charging cutoff voltage.

[0176] In this embodiment, in step S6, a total loss function is constructed, a graph neural network model is trained using the constructed total loss function, and graph convolution processing is performed using a defined graph convolution operator to obtain the predicted load power value of each energy storage node. The specific implementation method includes:

[0177] S6.1: Construct a total loss function that includes a prediction error loss term, a SOH physical constraint penalty term, and a weight decay regularization term;

[0178] Specifically, the total loss function Represented as:

[0179] ;

[0180] in, , These are the weighting coefficients for each loss term;

[0181] This is the weight decay regularization term. The set of all weight parameters to be trained, i.e., the matrix of all weight parameters in the graph convolutional model that need to be adjusted through training. , , The set of all parameters in the middle;

[0182] The definitions of each loss term in the total loss function are as follows:

[0183] Prediction error loss term The mean square error between predicted power and actual power:

[0184] ;

[0185] in, For the first One energy storage node in Power prediction at any given time; For the first One energy storage node in The actual power value at that moment, To predict the length of the time window.

[0186] SOH physical constraint penalty for:

[0187] ;

[0188] The SOH physical constraint penalty term indicates that when the predicted discharge power exceeds the maximum available discharge power under the current battery health state of the energy storage node, or when the absolute value of the predicted charging power exceeds the maximum available charging power, a proportional penalty is applied to the excess portion.

[0189] S6.2: Training a graphical neural network model based on the constructed total loss function;

[0190] S6.3: Input the enhanced feature matrix into the trained graph neural network model. The trained graph neural network model loads the trained weight parameters and performs graph convolution layer by layer according to the defined graph convolution operator to output the predicted load power of each energy storage node at each sampling time.

[0191] Furthermore, in S6.2, the specific implementation method for training the graphical neural network model based on the constructed total loss function includes:

[0192] S6.2.1: Input the enhanced feature matrix into the graph neural network model, perform graph convolution layer by layer forward according to the graph convolution operator defined in S4, and output the predicted load power of each energy storage node at each sampling time;

[0193] In this step, the obtained load power prediction value is the unoptimized raw output calculated based on the current model parameters.

[0194] S6.2.2: Based on the total loss function, calculate the total loss value under the current graph neural network model parameters using the maximum available discharge power and the maximum available charging power;

[0195] S6.2.3: Using the chain rule, calculate the gradient of the total loss function with respect to each weight parameter to be trained;

[0196] S6.2.4: Update the parameters of each weight to be trained by updating the gradient along the gradient descent direction according to the preset step size;

[0197] S6.2.5: Repeat the above steps until the total loss value no longer decreases or the maximum number of training rounds is reached, and save the trained weight parameters;

[0198] The final saved weight parameters are the optimal parameters after training.

[0199] Specifically, hyperparameters need to be pre-set before training a graph neural network model. As an example, the Adam optimizer can be used as the gradient descent optimization algorithm, with the following hyperparameters set: learning rate of 0.001, first-order moment decay coefficient of 0.9, second-order moment decay coefficient of 0.999, and maximum training epochs of 500.

[0200] It should be noted that graph neural network models are a deep learning method for graph-structured data. They learn the embedded representations of nodes by transmitting, transforming, and aggregating feature information between nodes in a graph. The graph neural network model in this invention is implemented using a graph convolutional network (GCN), which propagates information on the graph through graph convolution operations. Graph convolutional networks (GCN) are existing technology in this field, and this invention does not improve the structure of existing graph neural network models; therefore, they will not be described in detail here.

[0201] Another aspect of the present invention discloses a load forecasting system for an energy storage power station, applicable to the aforementioned load power forecasting method for an energy storage power station, comprising:

[0202] The data acquisition and feature enhancement module is used to acquire the operating data of each energy storage node in the energy storage power station, construct the original feature vector based on the operating data, calculate the SOH derivative physical quantity based on the battery health status, and concatenate the SOH derivative physical quantity and the original feature vector to obtain the enhanced feature vector;

[0203] The graph structure construction module, connected to the data acquisition and feature enhancement module, is used to construct the adjacency matrix of the physical layer graph structure and the adjacency matrix of the functional layer graph structure of the energy storage node.

[0204] The mask processing module, connected to the graph structure construction module, is used to construct a mask matrix based on the grid-connected switching state of each energy storage node. The mask matrix is ​​used to perform mask processing on the adjacency matrix of the physical layer graph structure and the adjacency matrix of the functional layer graph structure to obtain the physical layer mask adjacency matrix and the functional layer mask adjacency matrix.

[0205] The graph convolution operator definition module is connected to the mask processing module. It is used to construct the initial feature matrix and define the graph convolution operator using the physical layer mask adjacency matrix and the functional layer mask adjacency matrix.

[0206] The constraint calculation module, connected to the graph convolution operator definition module, is used to calculate the current maximum available discharge power and maximum available charging power of each energy storage node.

[0207] The model training output module is connected to the constraint calculation module to construct the total loss function. The constructed total loss function is used to train the graph neural network model, and graph convolution is performed through the defined graph convolution operator to obtain the load power prediction value of each energy storage node.

[0208] The various embodiments of the present invention have now been described in detail. To avoid obscuring the concept of the invention, some details known in the art have not been described. Those skilled in the art will fully understand how to implement the technical solutions of this invention based on the above description, and the scope of the invention is defined by the appended claims.

Claims

1. A method for predicting the load power of an energy storage power station, characterized in that, include: S1: Obtain the operation data of each energy storage node in the energy storage power station, construct the original feature vector based on the operation data, calculate the SOH derivative physical quantity based on the battery health status, and concatenate the SOH derivative physical quantity and the original feature vector to obtain the enhanced feature vector; S2: Construct the adjacency matrix of the physical layer graph structure and the adjacency matrix of the functional layer graph structure of the energy storage node; S3: Construct a mask matrix based on the grid connection switching status of each energy storage node, and use the mask matrix to mask the adjacency matrix of the physical layer graph structure and the adjacency matrix of the functional layer graph structure to obtain the physical layer mask adjacency matrix and the functional layer mask adjacency matrix. S4: Construct the initial feature matrix, and define the graph convolution operator using the physical layer mask adjacency matrix and the functional layer mask adjacency matrix; S5: Calculate the current maximum available discharge power and maximum available charging power of each energy storage node; S6: Construct a total loss function, train a graph neural network model using the constructed total loss function, and perform graph convolution processing through the defined graph convolution operator to obtain the predicted load power values ​​of each energy storage node.

2. The method for predicting the load power of an energy storage power station according to claim 1, characterized in that, The specific implementation method of S1 includes: S1.1: Obtain the operating data of each energy storage node, and construct the original feature vector of the operating data of each energy storage node according to the energy storage node number and sampling time; S1.2: Map the battery health state to the SOH derivative physical quantity; S1.3: The battery health status and its SOH derivative physical quantities are concatenated to the original feature vector to form an enhanced feature vector.

3. The method for predicting the load power of an energy storage power station according to claim 2, characterized in that, The specific implementation method of S2 includes: S2.1: Based on the electrical topology parameters of the energy storage power station, the electrical distance between any two energy storage nodes is calculated using the equivalent impedance magnitude between energy storage nodes and the system reference impedance. S2.2: Map electrical distances to an adjacency matrix of the physical layer graph structure using indicator functions; S2.3: Calculate the SOH state similarity between energy storage nodes based on the battery health status of each energy storage node; S2.4: Calculate the load behavior correlation between energy storage nodes based on the historical charging and discharging power of each energy storage node; S2.5: The adjacency matrix of the functional layer graph structure is constructed by weighted fusion of the similarity between the SOH state and the correlation between load behavior.

4. The method for predicting the load power of an energy storage power station according to claim 3, characterized in that, The specific implementation method of S3 includes: S3.1: Obtain the grid connection status indicator variables of each energy storage node, and construct a mask matrix using the grid connection status indicator variables; S3.2: Multiply the adjacency matrix of the physical layer graph structure by the mask matrix on the left and right respectively to obtain the physical layer mask adjacency matrix; S3.3: Multiply the adjacency matrix of the functional layer graph structure by the mask matrix on the left and right respectively to obtain the functional layer mask adjacency matrix.

5. The method for predicting the load power of an energy storage power station according to claim 4, characterized in that, The specific implementation method of S4 includes: S4.1: Stack the enhanced feature vectors of each energy storage node row by row to form an initial feature matrix; S4.2: Normalize the physical layer mask adjacency matrix and the functional layer mask adjacency matrix to obtain the normalized physical layer mask adjacency matrix and functional layer mask adjacency matrix. S4.3: A graph convolution operator is defined based on the normalized physical layer mask adjacency matrix, the functional layer mask adjacency matrix, and the initial feature matrix. This graph convolution operator uses the normalized physical layer mask adjacency matrix and the functional layer mask adjacency matrix as the graph structure input and the initial feature matrix as the node feature input.

6. The method for predicting the load power of an energy storage power station according to claim 5, characterized in that, In step S4.2, before the normalization process, the following processing method is also included: An identity matrix is ​​superimposed on the diagonal of the physical layer mask adjacency matrix and the functional layer mask adjacency matrix, respectively.

7. The method for predicting the load power of an energy storage power station according to claim 6, characterized in that, The specific implementation method of S5 includes: S5.1: Establish the mapping relationship between battery health status and battery internal resistance, and calculate the battery internal resistance; S5.2: Calculate the current maximum available discharge power of the energy storage node based on the battery internal resistance; S5.3: Calculate the current maximum available charging power of the energy storage node based on the battery internal resistance.

8. The method for predicting the load power of an energy storage power station according to claim 7, characterized in that, The specific implementation method of S6 includes: S6.1: Construct a total loss function that includes a prediction error loss term, a SOH physical constraint penalty term, and a weight decay regularization term; S6.2: Training a graphical neural network model based on the constructed total loss function; S6.3: Input the enhanced feature matrix into the trained graph neural network model. The trained graph neural network model loads the trained weight parameters and performs graph convolution layer by layer according to the defined graph convolution operator to output the predicted load power of each energy storage node at each sampling time.

9. The method for predicting the load power of an energy storage power station according to claim 8, characterized in that, The specific implementation method of S6.2 includes: S6.2.1: Input the enhanced feature matrix into the graph neural network model, perform graph convolution layer by layer forward according to the graph convolution operator defined in S4, and output the predicted load power of each energy storage node at each sampling time; S6.2.2: Based on the total loss function, calculate the total loss value under the current graph neural network model parameters using the maximum available discharge power and the maximum available charging power; S6.2.3: Using the chain rule, calculate the gradient of the total loss function with respect to each weight parameter to be trained; S6.2.4: Update the parameters of each weight to be trained by updating the gradient along the gradient descent direction according to the preset step size; S6.2.5: Repeat the above steps until the total loss value no longer decreases or the maximum number of training rounds is reached, and save the trained weight parameters.

10. A load power prediction system for an energy storage power station, applicable to the load power prediction method for an energy storage power station as described in any one of claims 1-9, characterized in that, include: The data acquisition and feature enhancement module is used to acquire the operating data of each energy storage node in the energy storage power station, construct the original feature vector based on the operating data, calculate the SOH derivative physical quantity based on the battery health status, and concatenate the SOH derivative physical quantity and the original feature vector to obtain the enhanced feature vector; The graph structure construction module, connected to the data acquisition and feature enhancement module, is used to construct the adjacency matrix of the physical layer graph structure and the adjacency matrix of the functional layer graph structure of the energy storage node. The mask processing module, connected to the graph structure construction module, is used to construct a mask matrix based on the grid-connected switching state of each energy storage node. The mask matrix is ​​used to perform mask processing on the adjacency matrix of the physical layer graph structure and the adjacency matrix of the functional layer graph structure to obtain the physical layer mask adjacency matrix and the functional layer mask adjacency matrix. The graph convolution operator definition module is connected to the mask processing module. It is used to construct the initial feature matrix and define the graph convolution operator using the physical layer mask adjacency matrix and the functional layer mask adjacency matrix. The constraint calculation module, connected to the graph convolution operator definition module, is used to calculate the current maximum available discharge power and maximum available charging power of each energy storage node. The model training output module is connected to the constraint calculation module to construct the total loss function. The constructed total loss function is used to train the graph neural network model, and graph convolution is performed through the defined graph convolution operator to obtain the load power prediction value of each energy storage node.

Images