Energy storage multi-dimensional value decoupling attribution method based on causal diagram attention network

Abstract

The application discloses a kind of energy storage multi-dimensional value decoupling attribution methods based on causal diagram attention network, comprising: collecting power system multi-dimensional time series operation data, constructing physical topology model, configuring initial electrical characteristic vector for each node;From the above data, extract operating variables and build causal dependence relationship, superimpose physical topology model and causal dependence relationship into causal diagram model and output graph structure information;Based on graph structure information and initial electrical characteristic vector, calculate attention weight using graph attention network;Based on operating variables and attention weight, deduce multi-dimensional physical operating state quantity under the fact scenario with energy storage intervention and the counterfactual scenario without intervention, and calculate individual intervention effect matrix, extract multi-dimensional physical value decoupling integral quantity and construct comprehensive economic allocation parameter model, output settlement and deployment instruction.The application realizes the accurate separation and fair attribution of energy storage multi-dimensional value in complex power grid environment.

Description

Technical Field

[0001] This invention relates to the fields of power system operation optimization and power market technology, and in particular to a multi-dimensional value decoupling attribution method for energy storage based on causal graph attention network. Background Technology

[0002] With the advancement of new power system construction, energy storage systems have become a core regulatory resource for mitigating power source-load fluctuations and enhancing the physical resilience of the power grid. In actual power grid operation, energy storage devices often participate in multiple application scenarios simultaneously, including peak shaving, frequency regulation, and voltage support. Their charging and discharging actions trigger local node power flow changes, which propagate globally through the physical network. This spatiotemporal interweaving of multidimensional actions results in complex nonlinear responses in the power grid, making it difficult to accurately decouple and fairly attribute the multidimensional value of energy storage.

[0003] In existing energy storage value assessment technologies, attribution analysis often relies on correlation-based time series fitting models or marginal electricity price decomposition methods. These models offer a certain level of fitting accuracy under stable grid operating conditions. However, when faced with sudden environmental stresses such as rapid changes in renewable energy output or unplanned unit outages, the grid's physical state tends to exhibit nonlinear, limit-breaking characteristics. Existing assessment models, on the one hand, lack the "counterfactual" reasoning logic found in causal inference frameworks, and on the other hand, lack topological graph features reflecting the underlying network physical propagation characteristics (such as power flow tracking and electrical distance sensitivity) as input. This makes it difficult to accurately isolate background state drift caused by environmental disturbances, resulting in significant quantification errors in the actual peak-shaving capacity and effective voltage support contributions of energy storage.

[0004] In fitting low-level control commands, existing methods often employ fixed piecewise numerical functions (such as if-else hard thresholds) for intervention command fitting. However, under grid transient conditions, equipment responses must strictly meet the requirement of first-order derivative continuity. Conventional non-smooth hard threshold mapping can trigger secondary physical shocks such as power flow distortion caused by algorithmic defects. These shocks are easily misjudged as negative effects of energy storage actions, leading to cumulative drift in the joint attribution results across multiple scenarios.

[0005] Furthermore, most existing assessment architectures treat peak shaving, frequency regulation, and voltage support as independent modules for numerical calculation, failing to incorporate the complex spatial topology interactions of the power grid (especially reactive power support attenuation under different electrical distances) into the value assessment equation in real time. This static parameter setting method ignores the power flow coupling relationship between multi-dimensional physical indicators, which can easily lead to parameter mismatch in the observation model under complex network panoramic conditions, further amplifying the overall attribution error of the system under a multi-market parallel settlement architecture.

[0006] Therefore, existing technologies urgently need a multi-dimensional value joint decoupling method for energy storage that can integrate multi-dimensional spatial topological features, be based on counterfactual deduction logic, and take into account smooth physical control characteristics, so as to improve the attribution reliability and objectivity of energy storage participation in multi-market settlement in complex grid environments. Summary of the Invention

[0007] To address the aforementioned shortcomings of existing technologies, this invention provides a decoupling and attribution method for the multidimensional value of energy storage based on causal graph attention networks. By constructing a causal graph model that integrates physical topology and causal dependencies, extracting spatial correlation features using graph attention networks, and inferring the physical state differences with / without energy storage intervention based on a counterfactual reasoning framework, this invention solves the technical problem that the contribution of multidimensional value of energy storage in complex power grid environments cannot be accurately decoupled and fairly attributed due to environmental confusion and interference, discontinuous control commands, and missing spatial coupling.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0009] A multidimensional value decoupling attribution method for energy storage based on causal graph attention networks includes the following steps:

[0010] S1. Collect multi-dimensional time-series operation data of the power system, construct a physical topology model with the power system bus as nodes and transmission lines as edges, and configure initial electrical feature vectors for each node;

[0011] S2. Extract multiple runtime variables from the multidimensional time-series runtime data, construct causal dependencies between the runtime variables, superimpose the physical topology model and the causal dependencies into a causal graph model, and output the graph structure information of the causal graph model;

[0012] S3. Input the graph structure information and the initial electrical feature vector into the graph attention network, calculate the attention coefficient of the physical space influence of the energy storage node on the grid node, and obtain the attention weight after normalization;

[0013] S4. Using the operating variables corresponding to the causal dependency relationship and the attention weight as input, the multi-dimensional physical operating state quantities under the factual scenario with energy storage intervention and the counterfactual scenario without energy storage intervention are deduced through the counterfactual control deduction model, and the factual physical state quantities and counterfactual physical state quantities are output. The counterfactual control deduction model includes a peak-shaving deduction model based on the Softplus smoothing activation function, a voltage support deduction model based on closed-loop soft lift control, and a double exponential frequency modulation deduction model that takes into account the rotor inertia and damping physical characteristics.

[0014] S5. Based on the stated factual physical state quantities and counterfactual physical state quantities, calculate the individual intervention effect matrix, extract the multidimensional physical value decoupling integral, construct a comprehensive economic allocation parameter model, and output economic settlement and allocation instructions.

[0015] As a preferred embodiment, in step S1, the physical topology model is represented as follows:

[0016] ;

[0017] In the formula: Represents the physical topology model; Represents the set of bus nodes in a power system; This represents the set of transmission line edges that connect adjacent nodes.

[0018] As a preferred embodiment, in step S2, the causal dependencies among the various operating variables include: using system background state characteristics as confounding variables, using the active and reactive power output sequences of the energy storage system as physical intervention variables, and using the system's physical operational safety state indicators as outcome variables; and forming a causal directed acyclic graph with the confounding variables, physical intervention variables, and outcome variables as nodes, and the causal relationships between variables characterized by the following nonlinear physical coupling function as directed edges:

[0019] ;

[0020] ;

[0021] In the formula: express The outcome variable at time; express It contains confounding variables related to fluctuations in net load demand at all times; express Physical intervention variables at any given time; and These represent the nonlinear physical coupling functions for the constraint outcome variables and the constraint physical intervention variables, respectively; and These represent the independent random observation error terms in the corresponding generation process;

[0022] The graph structure information of the causal graph model includes the corresponding physical adjacency matrix generated by the physical topology model and the causal directed acyclic graph.

[0023] As a preferred option, the specific processing steps of step S3 include:

[0024] S31. Input the initial electrical feature vectors of each node into the graph attention network, and perform linear feature transformation through a learnable shared weight matrix to obtain the feature representation vectors of each node in the hidden layer:

[0025] ;

[0026] In the formula: Indicates the first Nodes in the hidden layer The feature representation vector; Represents a node exist The initial electrical feature vector at time t; Indicates the first The shared weight transformation matrix of the layer;

[0027] S32, Computing Energy Storage Nodes For power grid nodes The physical space influences the attention coefficient, which is then normalized using the Softmax function to obtain the normalized attention weights.

[0028] ;

[0029] In the formula: Indicates energy storage node For power grid nodes Normalized attention weights; This represents an exponential function with the natural constant e as its base. This represents the modified linear unit activation function with leakage; This represents the transpose of the weight vector of a single-layer feedforward neural network. Operators for concatenating feature vectors; The nodes obtained from the physical adjacency matrix represent The set of first-order topological physical neighbor nodes; Indicates the index number for summation traversal; , , Representing the power grid nodes Energy storage nodes Neighbor nodes The feature representation vector; This represents the shared weight transformation matrix.

[0030] As a preferred embodiment, in step S4, the peak-shaving and peak-shaving deduction model is used to calculate the energy storage peak-shaving active power output component in the actual physical state quantity and the peak-shaving active power output component in the inverse physical state quantity; the processing procedure is expressed as follows:

[0031] ;

[0032] In the formula: express The active power output of energy storage for peak shaving in real-world scenarios is 0, while the active power output for peak shaving in counterfactual scenarios is 0. This represents the logarithmic operator with base e. This represents an exponential function with the natural constant e as its base. The smoothing factor represents the control curve; express It contains confounding variables related to fluctuations in net load demand at all times; This represents the physical start threshold parameter for peak shaving issued by the underlying scheduling system.

[0033] As a preferred embodiment, in step S4, the physical voltage component in the factual physical state quantity and the natural drop voltage component in the counterfactual physical state quantity are calculated using the voltage support deduction model; the processing procedure is expressed as follows:

[0034] ;

[0035] ;

[0036] In the formula: express Voltage reactive power compensation output under real-time scenarios; This represents the voltage control gain smoothing constant; This represents the set target baseline value for the physical safety operation of the system; express Natural voltage drop components in counterfactual scenarios at all times; express The physical voltage component after being safely lifted in a real-world scenario; This represents the normalized attention weights.

[0037] As a preferred embodiment, in step S4, the frequency offset component after frequency modulation response in the factual physical state quantity and the frequency natural drop component in the inverse factual physical state quantity are calculated using the double exponential frequency modulation derivation model; the processing procedure is expressed as follows:

[0038] ;

[0039] In the formula: express Frequency offset component after frequency modulation response in a real-time scenario; express The frequency of counterfactual scenarios naturally decreases in weight. The equivalent physical duration is calculated from the occurrence of the disturbance; A and B represent the physical gain coefficients that reflect the depth of energy storage droop control and inertia response, respectively. , , These represent the equivalent damping time constants in the corresponding dynamic processes.

[0040] As a preferred embodiment, in step S5, the individual intervention effect matrix is ​​represented as follows:

[0041] ;

[0042] In the formula: express Time-based individual intervention effect matrix; This represents a causal inference physical interference quantifier; Represents a factual physical state quantity; Represents counterfactual physical state quantities; Indicates physical intervention variables; This indicates that the forced energy storage operates according to actual instructions; 0 indicates that the energy storage is forced to remain inactive; 0 indicates a counterfactual zero vector representing that the energy storage has not intervened. The counterfactual zero vector refers to the baseline state where the energy storage system is completely inactive and the output is zero.

[0043] As a preferred embodiment, in step S5, the multidimensional physical value decoupling integral includes the peak-shaving integrated physical contribution, the voltage support integrated physical contribution, and the frequency modulation auxiliary integrated physical contribution; the calculation formula is expressed as:

[0044] Peak-shaving integrated physical contribution:

[0045] ;

[0046] In the formula: Indicates the overall physical contribution of peak shaving; Indicates the total physical time period of the analysis; This represents the intervention effect value of the peak-to-valley difference in the corresponding cross-section load; This represents the discrete time step or the integral / differential time element;

[0047] Overall physical contribution of voltage support:

[0048] ;

[0049] In the formula: This indicates the overall physical contribution of voltage support; This represents the set of physical nodes of the partition of controllable interest within the system. This indicates the intervention value for improving node voltage deviation;

[0050] Frequency modulation-assisted overall physical contribution:

[0051] ;

[0052] In the formula: Indicates the overall physical contribution of frequency modulation assistance; This represents a definite integral operator over a specified time interval; Indicates the start time of the frequency disturbance; Indicates the duration of the recovery process; This indicates the intervention value for improving frequency transient deviation.

[0053] As a preferred embodiment, in step S5, the comprehensive economic allocation parameter model is expressed as follows:

[0054] ;

[0055] In the formula: Indicates the parameters of comprehensive economic allocation; This indicates the unit settlement rate for peak-shaving auxiliary physical services; The parameter representing the conversion factor for reactive voltage support; The physical settlement price representing the FM mileage;

[0056] Based on comprehensive economic allocation parameters The quantitative value is used to issue economic settlement and allocation instructions to the underlying trading system and scheduling system.

[0057] Compared with the prior art, the present invention has the following technical effects:

[0058] 1. This invention generates a physical adjacency matrix by constructing a physical topology model, which limits the range of physical neighbors for graph attention calculation and avoids redundancy and distortion caused by blind calculation of the entire graph. At the same time, it constructs a causal directed acyclic graph with confounding variables, physical intervention variables and outcome variables as nodes, which provides variable definition and intervention operation specifications for counterfactual inference, ensuring that the counterfactual scenario has clear causal semantics and computability.

[0059] 2. This invention uses Softplus smoothing soft threshold to replace the traditional segmented hard threshold, ensuring that the first derivative of peak-shaving output is absolutely continuous throughout the entire time domain. The peak-shaving and peak-shaving simulation model built based on the Softplus smoothing activation function ensures the continuous differentiability of energy storage charging and discharging power throughout the entire time domain while meeting basic peak-shaving requirements. This algorithm avoids secondary physical impacts on power grid flow caused by command step jumps from the control source, improving the stability of equipment grid-connected operation.

[0060] 3. The present invention is a voltage support deduction model based on closed-loop soft lift control with spatial weighted superposition based on attention weight. It integrates reactive power compensation with closed-loop feedback of voltage deviation at each node, and realizes spatial differential allocation of compensation through attention weight. This enables weak-end nodes far from the energy storage access point to obtain support strength appropriate to their electrical distance, and maintains the voltage of each node within a safe range throughout the process.

[0061] 4. The dual-exponential frequency modulation extrapolation model constructed in this invention integrates the gain coefficient reflecting the depth of energy storage droop and inertia response, as well as the equivalent time constant considering rotor inertia, to achieve seamless connection between the instantaneous frequency trajectory and steady-state waveform during disturbances. It accurately characterizes the shift of the minimum point of transient frequency and the recovery slope, overcoming the defect of estimation drift caused by static extrapolation at the moment of fault. This model, together with the causal graph, constitutes a complete counterfactual reasoning framework, providing a physically credible comparison benchmark for frequency modulation contribution stripping.

[0062] 5. This invention uses causal interference quantifiers to subtract factual and counterfactual physical states dimension by dimension, eliminating environmental background drift interference to obtain pure individual intervention effects. It then integrates the effects of each dimension over the entire analysis cycle to obtain physically quantifiable values ​​such as peak-shaving power, voltage support, and frequency regulation mileage, which are ultimately mapped to a multi-market price system to generate comprehensive economic allocation parameters. This process brings together the outputs of all the aforementioned steps, completing a transparent, traceable, and quantitative closed loop from "physical improvement" to "economic settlement," providing an objective, fair, and physically based evidentiary foundation for energy storage to participate in multi-market parallel settlement. Attached Figure Description

[0063] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0064] Figure 1 This is a flowchart of a multi-dimensional value decoupling attribution method for energy storage based on causal graph attention network disclosed in this invention;

[0065] Figure 2 This is a counterfactual comparison chart of a real net load peak shaving scenario for an independent system operator under Softplus soft threshold intervention, as shown in the example.

[0066] Figure 3 This is a comparison diagram of counterfactual inferences of a real frequency accident scenario of a country's national power grid based on the double exponential transient equation in the example;

[0067] Figure 4 This is a simulation diagram of the counterfactual safety support of the grid node voltage at the grid connection point based on closed-loop smooth control in the embodiment.

[0068] Figure 5 This is a bar chart verifying the comprehensive economic settlement voucher after precise decoupling and attribution of multidimensional values ​​in the embodiment. Detailed Implementation

[0069] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0070] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 a part of the embodiments of the present invention, not all of them. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to represent selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0071] Example:

[0072] To address the problems of poor anti-interference capability of existing evaluation models, discontinuous underlying control commands, and lack of multi-dimensional spatiotemporal coupling analysis in existing technologies, this invention proposes a multi-dimensional value decoupling and attribution method for energy storage based on causal graph attention networks. This method first constructs a causal graph model that integrates physical topology and causal dependencies. Second, it utilizes graph attention networks to calculate the attention weights of the spatial impact of energy storage nodes on grid nodes under physical topology constraints, achieving accurate quantification of power flow propagation effects. Third, it constructs a counterfactual reasoning control framework using the Softplus smoothing activation function, closed-loop soft lift, and a double-exponential frequency modulation dynamics model to deduce the factual and counterfactual physical states with / without energy storage intervention. Finally, it compares the two states dimension-by-dimensionally using causal interference quantifiers to extract the pure individual intervention effects and map them to economic configuration parameters. This invention aims to effectively eliminate the interference of random background fluctuations in the grid environment. Under the condition of adhering to the continuity constraints of underlying physical operation, it accurately extracts and quantifies the physical intervention contributions of energy storage devices in multiple time and spatial dimensions, thereby outputting comprehensive economic configuration parameters that can be directly used for parallel settlement and cost allocation in multiple electricity markets.

[0073] Specifically, the energy storage multi-dimensional value decoupling attribution method based on causal graph attention network proposed in this invention, in a practical power system operation control and evaluation embodiment, uses a real publicly available dataset with a total duration of 120 minutes to perform underlying hardware parameter deduction. The system performs factual and counterfactual causal deductions for three different control scenarios: peak shaving, frequency regulation, and voltage support, and outputs corresponding state response curves and economic attribution evidence.

[0074] The specific method in this embodiment is as follows: Figure 1 As shown, it includes the following steps:

[0075] S1. Collect multi-dimensional time-series operation data of the power system, construct a physical topology model with the power system bus as nodes and transmission lines as edges, and configure initial electrical feature vectors for each node;

[0076] In this embodiment, the physical topology model is represented as follows:

[0077] ;

[0078] In the formula: Represents the physical topology model; Represents the set of bus nodes in a power system; This represents the set of transmission line edges that connect adjacent nodes.

[0079] definition For power system bus nodes exist The initial electrical characteristic vector at time t is composed of the historical time-series voltage, active power, and reactive power of each node, where... Indicates the physical node number in the network topology. It represents the discrete time step.

[0080] S2. Extract multiple runtime variables from the multidimensional time-series runtime data, construct causal dependencies between the runtime variables, superimpose the physical topology model and the causal dependencies into a causal graph model, and output the graph structure information of the causal graph model;

[0081] In this embodiment, the causal dependency is constructed by the following steps:

[0082] Within the potential outcome framework, the system background state characteristics are defined as confounding variables, the active and reactive power processing sequences of the energy storage system are physical intervention variables, and the system physical operation safety state indicators are outcome variables. A low-level data generation mechanism model is also established.

[0083] ;

[0084] ;

[0085] In the formula: express The outcome variable at time; express It contains confounding variables related to fluctuations in net load demand at all times; express Physical intervention variables at any given time; and These represent the nonlinear physical coupling functions for the constraint outcome variables and the constraint physical intervention variables, respectively; and These represent the independent random observation error terms in the corresponding generation process;

[0086] The graph structure information of the causal graph model includes the corresponding physical adjacency matrix and causal directed acyclic graph generated by the physical topology model; the causal directed acyclic graph uses the confusion variable, physical intervention variable, and outcome variable as nodes, and the nonlinear physical coupling function as the nodes. and The causal relationships between the variables are represented by directed edges.

[0087] This embodiment generates a physical adjacency matrix by constructing a physical topology model, which limits the range of physical neighbors for graph attention calculation and avoids redundancy and distortion caused by blind calculation of the entire graph. At the same time, it constructs a causal directed acyclic graph with confounding variables, physical intervention variables and outcome variables as nodes, which provides variable definition and intervention operation specifications for counterfactual inference, ensuring that the counterfactual scenario has clear causal semantics and computability.

[0088] S3. Input the graph structure information and the initial electrical feature vector into the graph attention network, calculate the attention coefficient of the physical spatial influence of the energy storage node on the grid node, and obtain the attention weight after normalization; in specific implementation, this includes:

[0089] S31. Input the initial electrical feature vectors of each node into the graph attention network, and perform linear feature transformation through a learnable shared weight matrix to obtain the feature representation vectors of each node in the hidden layer:

[0090] ;

[0091] In the formula: Indicates the first Nodes in the hidden layer The feature representation vector; Represents a node exist The initial electrical feature vector at time t; Indicates the first The shared weight transformation matrix of the layer;

[0092] S32, Computing Energy Storage Nodes For power grid nodes The physical space influences the attention coefficient, which is then normalized using the Softmax function to obtain the normalized attention weights.

[0093] ;

[0094] In the formula: Indicates energy storage node For power grid nodes Normalized attention weights; This represents an exponential function with the natural constant e as its base. This represents the modified linear unit activation function with leakage; This represents the transpose of the weight vector of a single-layer feedforward neural network. Operators for concatenating feature vectors; The nodes obtained from the physical adjacency matrix represent... The set of first-order topological physical neighbor nodes; Indicates the index number for summation traversal; , , Representing power grid nodes Energy storage nodes Neighbor nodes The feature representation vector; This represents the shared weight transformation matrix.

[0095] In this embodiment, the physical adjacency matrix and the initial electrical feature vector are input into the graph attention network, so that the attention weights are calculated under topological constraints, accurately quantifying the spatial attenuation intensity of the power flow caused by energy storage actions. Compared with the fully connected attention mechanism, this method retains physical reliability and reduces computational complexity. Moreover, the attention weights can be directly used for spatial weighted superposition in subsequent closed-loop control, realizing end-to-end spatial perception transmission from feature extraction to control execution.

[0096] S4. Using the operating variables corresponding to the causal dependency relationship and the attention weight as input, the multi-dimensional physical operating state quantities under the factual scenario with energy storage intervention and the counterfactual scenario without energy storage intervention are deduced through the counterfactual control deduction model, and the factual physical state quantities and counterfactual physical state quantities are output. The counterfactual control deduction model includes a peak-shaving deduction model based on the Softplus smoothing activation function, a voltage support deduction model based on closed-loop soft lift control, and a double exponential frequency modulation deduction model that takes into account the rotor inertia and damping physical characteristics.

[0097] In this embodiment, to fully cover the value contribution of energy storage systems in the three main application scenarios of peak shaving, voltage support, and frequency regulation assistance in the electricity market, the steps of this embodiment are to construct peak shaving and peak reduction simulation models, voltage support simulation models, and dual-exponential frequency regulation simulation models, respectively, dividing the multidimensional physical operating state variables into three independent dimensions for simulation. The actual physical state variables include the active power output component of energy storage peak shaving. Physical voltage components Frequency offset component The counterfactual physical state quantities include the peak-shaving active power output component (which is always 0 in the counterfactual scenario) and the natural voltage drop component. Frequency naturally drops component .

[0098] (1) Peak shaving scenario simulation

[0099] In peak-shaving scenarios, energy storage mainly works by absorbing or releasing energy. Gong Gong This method aims to smooth out the peak-valley difference in the net load of the power grid. This embodiment employs a peak-shaving and peak-reduction simulation model based on the Softplus smoothing activation function to eliminate the discontinuity in the output curve caused by traditional segmented hard threshold control. Specifically, the peak-shaving and peak-reduction simulation model is used to calculate the energy storage peak-shaving active power output component in the actual physical state variables and the peak-shaving active power output component in the inverse physical state variables. The processing procedure is as follows:

[0100] ;

[0101] In the formula: express The active power output of energy storage for peak shaving in real-world scenarios is 0, while the active power output for peak shaving in counterfactual scenarios is 0. This represents the logarithmic operator with base e. This represents an exponential function with the natural constant e as its base. The smoothing factor representing the control curve; express It contains confounding variables related to fluctuations in net load demand at all times; This represents the physical start threshold parameter for peak shaving issued by the underlying scheduling system.

[0102] like Figure 2 As shown, this embodiment imports historical operating data from an independent system operator exhibiting a sharp increase in net load in the evening. The red dashed line represents the counterfactual natural evolution trajectory (i.e., the original net load) under conditions without energy storage intervention; the blue solid line represents the actual operating trajectory after energy storage intervention. From t=0 to t=50 min and after t=115 min, the grid load did not reach the trigger condition, and the blue and red lines completely overlapped. When the net load exceeded the set threshold, thanks to the Softplus smoothing soft threshold equation constructed in this invention, the actual intervention state (blue solid line) smoothly transitioned and the peak value was reduced. The light green shaded area in the figure represents the decoupled extracted Individual Intervention Effect (ITE-peak). The slope of the tangent at the beginning and end of the curve remains consistent with the slope at the closure point, without any first-order derivative jumps or bends, verifying the effectiveness of this invention in avoiding secondary physical shocks to the power grid.

[0103] This embodiment uses Softplus smoothing soft threshold to replace the traditional segmented hard threshold, ensuring that the first derivative of peak-shaving output is absolutely continuous throughout the entire time domain. The peak-shaving and peak-shaving simulation model built based on the Softplus smoothing activation function ensures the continuous differentiability of energy storage charging and discharging power throughout the entire time domain while meeting basic peak-shaving requirements. This algorithm avoids secondary physical impacts on power grid flow caused by command step jumps from the control source, improving the stability of equipment grid-connected operation.

[0104] (2) Voltage support scenario simulation

[0105] In voltage support scenarios, energy storage maintains the grid connection point voltage above the safe operating threshold by injecting or absorbing reactive power into the grid. This embodiment employs a voltage support deduction model based on closed-loop soft-lift control to ensure that the grid voltage does not exceed the limit throughout the entire time scale. Specifically, using the voltage support deduction model, the physical voltage component in the factual physical state quantities and the natural voltage drop component in the counterfactual physical state quantities are calculated. The processing procedure is as follows:

[0106] ;

[0107] ;

[0108] In the formula: express Voltage reactive power compensation output under real-time scenarios; This represents the voltage control gain smoothing constant; This represents the set target baseline value for the physical safety operation of the system; express The natural voltage drop component in a counterfactual scenario without energy storage intervention; express The physical voltage component after being safely lifted when energy storage intervention occurs in a real-world scenario; This represents the normalized attention weights.

[0109] This embodiment is a voltage support simulation model based on closed-loop soft lift control, which uses spatial weighted superposition based on attention weights. It integrates reactive power compensation with closed-loop feedback of voltage deviation at each node, and achieves spatially differentiated allocation of compensation through attention weights. This allows weak-end nodes far from the energy storage access point to receive support appropriate to their electrical distance, and maintains the voltage of each node within a safe range throughout the process.

[0110] (3) FM scenario simulation

[0111] In frequency regulation scenarios, energy storage suppresses transient frequency drops caused by sudden generator tripping or load disturbances through rapid active power response. Its response behavior is constrained by rotor inertia and damping physical characteristics. This embodiment employs a double-exponential frequency regulation derivation model that considers rotor inertia and damping physical characteristics to achieve a seamless approximation of the frequency drop slope and the steady-state minimum point at the moment of disturbance. Specifically, using the double-exponential frequency regulation derivation model, the frequency offset component after frequency regulation response in the actual physical state quantity and the natural frequency drop component in the counterfactual physical state quantity are calculated.

[0112] The processing procedure is as follows:

[0113] ;

[0114] In the formula: express Frequency offset component after frequency modulation response in a real-time scenario; express The frequency of counterfactual scenarios naturally decreases in weight. The equivalent physical duration is calculated from the occurrence of the disturbance; A and B represent the physical gain coefficients that reflect the depth of energy storage droop control and inertia response, respectively. , , These represent the equivalent damping time constants in the corresponding dynamic processes.

[0115] like Figure 3 As shown, this embodiment imports generator disconnection accident data from a national power grid of a certain country, including actual SCADA measurement noise floor. A sudden generator disconnection fault occurs at t=60min, and the red dashed line (natural frequency drop without energy storage) shows a severe precipitous frequency drop. The blue solid line (with energy storage intervention) and the red dashed line physically diverge strictly from the same steady-state point at the moment of the accident. Based on the derivation results of the double-exponential rotor dynamics equations, the two-stage action characteristics of energy storage are accurately separated and displayed: the first stage mitigates the initial slope of the frequency drop (RoCoF) through virtual inertia response; the second stage effectively raises the frequency extreme low point (Nadir) and the final steady-state recovery level through droop control. The light purple shaded area in the figure represents the decoupled actual frequency regulation auxiliary contribution (ITE-freq).

[0116] The dual-exponential frequency modulation extrapolation model constructed in this embodiment integrates the gain coefficient reflecting the depth of energy storage droop and inertia response, as well as the equivalent time constant taking into account the rotor inertia, to achieve seamless connection between the instantaneous frequency trajectory and steady-state waveform during disturbances. It accurately depicts the shift of the transient frequency minimum point and the recovery slope, overcoming the defect of estimation drift caused by static extrapolation at the moment of fault. This model, together with the causal graph, constitutes a complete counterfactual reasoning framework, providing a physically credible comparison benchmark for frequency modulation contribution stripping.

[0117] S5. Based on the stated factual physical state quantities and counterfactual physical state quantities, calculate the individual intervention effect matrix, extract the multidimensional physical value decoupling integral, construct a comprehensive economic allocation parameter model, and output economic settlement and allocation instructions.

[0118] In this embodiment, based on the factual physical state quantity output in step S4 and counterfactual physical state quantities These represent the physical operating states of the system at the same moment under two scenarios: with and without energy storage intervention. This step compares the differences between the two scenarios across various dimensions to isolate the independent intervention contribution of the energy storage devices and quantify it into economic parameters that can participate in market settlement. Specifically, this includes:

[0119] (1) Calculation of individual intervention effect matrix

[0120] The Individual Treatment Effect (ITE) measures the state difference between scenarios with and without intervention on the same physical dimension. In this embodiment, the factual and counterfactual physical state quantities output by S4 are aligned dimensionally and then subtracted dimension-by-dimensionally to construct the ITE matrix. The calculation model is as follows:

[0121] ;

[0122] In the formula: express Time-based individual intervention effect matrix; This represents a causal inference physical interference quantifier; Represents a factual physical state quantity; Represents counterfactual physical state quantities; Indicates physical intervention variables; This indicates that the forced energy storage operates according to actual instructions; 0 indicates that the energy storage is forced to remain inactive; 0 indicates a counterfactual zero vector representing that the energy storage has not intervened. The counterfactual zero vector refers to the baseline state in which the energy storage system is completely inactive and the output is zero. It is used as a reference for comparing and calculating the actual intervention effect.

[0123] (2) Extraction of multidimensional physical value decoupling integral

[0124] The individual intervention effect matrix provides the instantaneous physical improvement at each moment, while energy storage devices participating in electricity market settlement typically focus on the cumulative physical contribution over the entire service period. Therefore, this embodiment... Each element in the process is subjected to time integration and spatial aggregation to extract the multidimensional physical value decoupling integral.

[0125] The multidimensional physical value decoupling integral includes the peak-shaving integrated physical contribution, the voltage support integrated physical contribution, and the frequency modulation auxiliary integrated physical contribution; the calculation formula is expressed as:

[0126] Peak-shaving integrated physical contribution:

[0127] ;

[0128] In the formula: Indicates the overall physical contribution of peak shaving; Indicates the total physical time period of the analysis; This represents the intervention effect value of the peak-to-valley difference in the corresponding cross-section load; This represents the discrete time step or the integral / differential time element;

[0129] Overall physical contribution of voltage support:

[0130] ;

[0131] In the formula: This indicates the overall physical contribution of voltage support; This represents the set of physical nodes of the partition of controllable interest within the system. This indicates the intervention value for improving node voltage deviation;

[0132] Frequency modulation-assisted overall physical contribution:

[0133] ;

[0134] In the formula: Indicates the overall physical contribution of frequency modulation assistance; This represents a definite integral operator over a specified time interval; Indicates the start time of the frequency disturbance; Indicates the duration of the recovery process; This indicates the intervention value for improving frequency transient deviation.

[0135] (3) Construction of a comprehensive economic allocation parameter model

[0136] The aforementioned multidimensional physical value decoupling integrals reflect the physical contributions of energy storage in peak shaving, voltage support, and frequency regulation assistance, respectively. To map these physical contributions to economic value that can participate in market settlement, this embodiment constructs a comprehensive economic configuration parameter model based on electricity market price signals for various application scenarios:

[0137] ;

[0138] In the formula: Indicates the parameters of comprehensive economic allocation; This indicates the unit settlement rate for peak-shaving auxiliary physical services; The parameter representing the conversion factor for reactive voltage support; This indicates the physical settlement price for FM mileage.

[0139] (4) Output of economic settlement and allocation instructions

[0140] Based on comprehensive economic allocation parameters The quantitative value is used to issue economic settlement and allocation instructions to the underlying trading system and scheduling system.

[0141] Economic settlement instructions: Decomposed into three independent revenue components , , The corresponding accounts for parallel settlement in multiple markets are entered separately, serving as quantitative evidence of the revenue obtained by the energy storage device in the three markets of peak shaving, voltage support, and frequency regulation auxiliary.

[0142] Dispatch instructions: According to Based on the numerical value and composition ratio, the scheduling system dynamically optimizes the next stage of operation strategy for energy storage equipment: when the revenue share of peak shaving dimension is significantly high, the charging and discharging depth should be appropriately increased to obtain more peak shaving revenue; when the revenue share of frequency regulation dimension is high, more active power reserve capacity should be reserved first to respond to sudden frequency events; when the revenue share of voltage support dimension is low, reactive power reserve capacity can be appropriately reduced to reduce equipment losses, thereby achieving economic operation and optimal threshold allocation of energy storage equipment in multiple market environments.

[0143] like Figure 5 As shown, this embodiment will Figures 2 to 4 The multidimensional physical value decoupling integral obtained by decoupling ( , , The system is comprehensively configured through a market mapping equation. The bar chart visually outputs the set of decision-making execution variables: peak-shaving comprehensive value of 10709.0 yuan, voltage support value of 15210.5 yuan, and frequency regulation auxiliary value of 11812.5 yuan. The three-dimensional value magnitudes are balanced and independent, effectively eliminating double-counting caused by overlapping environmental variables. The underlying electricity market trading system or grid settlement center can use this objective data as evidence to execute subsequent market quotation optimization and cost allocation instructions.

[0144] In summary, the energy storage multidimensional value decoupling attribution method based on causal graph attention networks proposed in this embodiment overcomes the limitations of traditional evaluation methods based on static correlation fitting and hard threshold numerical calculation, and establishes a comprehensive attribution architecture that deeply integrates the underlying physical constraints of the power system and the causal evolution logic of data. At the theoretical decoupling level, this method effectively eliminates confounding variables such as random fluctuations in the environmental background by constructing a power grid physical topology model and counterfactual causal inference logic, blocking the transmission of spurious correlations under strong coupling of multidimensional physical states, and achieving precise separation of energy storage intervention effects in complex spatiotemporal dimensions.

[0145] Specifically, at the physical control level, this method abandons non-smooth step numerical mapping and simple translational deduction, and introduces a peak-shaving and peak-shaving deduction model based on Softplus, a voltage support deduction model based on closed-loop soft support, and a dual-exponential frequency regulation deduction model that takes into account rotor inertia response. This implementation architecture ensures the absolute continuity of the first derivative of the control commands of the underlying equipment, avoids secondary physical shocks to the power grid induced by command mutations, and ensures that the system's operating state under extreme transient conditions is strictly controlled above the physical safety baseline, thus achieving a substantial leap from "theoretical numerical calculation" to "high-fidelity physical safety control." At the engineering application and market economy level, the technical solution of this embodiment provides an objective, transparent, and quantifiable execution certificate for energy storage assets to participate in the new power system multi-market joint clearing mechanism. The system dynamically maps multi-dimensional physical indicators such as precisely decoupled peak-shaving transfer power, voltage support efficiency, and effective frequency regulation mileage to the ancillary service market settlement model according to their physical contribution, generating mutually independent and non-overlapping comprehensive economic configuration parameters. This parameter can serve as a highly credible data basis and can be directly applied to the underlying power market trading system and dispatch center. It effectively resolves the current industry pain points such as the intertwining of multi-dimensional values ​​of energy storage and the ambiguity in cost sharing, and improves the comprehensive dispatch efficiency and resource allocation rationality of energy storage assets.

[0146] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit the technical solutions. Those skilled in the art should understand that any modifications or equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the present invention should be covered within the scope of the claims of the present invention.

Claims

1. A multi-dimensional value decoupling attribution method for energy storage based on causal graph attention networks, characterized in that, Includes the following steps: S1. Collect multi-dimensional time-series operation data of the power system, construct a physical topology model with the power system bus as nodes and transmission lines as edges, and configure initial electrical feature vectors for each node; S2. Extract multiple runtime variables from the multidimensional time-series runtime data and construct the causal dependencies between the runtime variables; The physical topology model and the causal dependencies are superimposed to form a causal graph model, and the graph structure information of the causal graph model is output. S3. Input the graph structure information and the initial electrical feature vector into the graph attention network, calculate the attention coefficient of the physical space influence of the energy storage node on the grid node, and obtain the attention weight after normalization; S4. Using the operating variables corresponding to the causal dependency relationship and the attention weight as input, the multi-dimensional physical operating state quantities under the factual scenario with energy storage intervention and the counterfactual scenario without energy storage intervention are deduced through the counterfactual control deduction model, and the factual physical state quantities and counterfactual physical state quantities are output. The counterfactual control deduction model includes a peak-shaving deduction model based on the Softplus smoothing activation function, a voltage support deduction model based on closed-loop soft lift control, and a double exponential frequency modulation deduction model that takes into account the rotor inertia and damping physical characteristics. S5. Based on the stated factual physical state quantities and counterfactual physical state quantities, calculate the individual intervention effect matrix, extract the multidimensional physical value decoupling integral, construct a comprehensive economic allocation parameter model, and output economic settlement and allocation instructions.

2. The energy storage multidimensional value decoupling attribution method based on causal graph attention network according to claim 1, characterized in that, In step S1, the physical topology model is represented as follows: ； In the formula: Represents the physical topology model; Represents the set of bus nodes in a power system; This represents the set of transmission line edges that connect adjacent nodes.

3. The energy storage multidimensional value decoupling attribution method based on causal graph attention network according to claim 2, characterized in that, In step S2, the causal dependencies between the various operating variables include: using the system background state characteristics as confounding variables, the active and reactive power output sequences of the energy storage system as physical intervention variables, and the system physical operation safety state indicators as outcome variables; and a causal directed acyclic graph consisting of the confounding variables, physical intervention variables, and outcome variables as nodes, and the causal relationships between variables characterized by the following nonlinear physical coupling function as directed edges: ； ； In the formula: express The outcome variable at time; express It contains confounding variables related to fluctuations in net load demand at all times; express Physical intervention variables at any given time; and These represent the nonlinear physical coupling functions for the constraint outcome variables and the constraint physical intervention variables, respectively; and These represent the independent random observation error terms in the corresponding generation process; The graph structure information of the causal graph model includes the corresponding physical adjacency matrix generated by the physical topology model and the causal directed acyclic graph.

4. The energy storage multidimensional value decoupling attribution method based on causal graph attention network according to claim 3, characterized in that, The specific processing steps of step S3 include: S31. Input the initial electrical feature vectors of each node into the graph attention network, and perform linear feature transformation through a learnable shared weight matrix to obtain the feature representation vectors of each node in the hidden layer: ； In the formula: Indicates the first Nodes in the hidden layer The feature representation vector; Represents a node exist The initial electrical feature vector at time t; Indicates the first The shared weight transformation matrix of the layer; S32, Computation of Energy Storage Node For power grid nodes The physical space influences the attention coefficient, which is then normalized using the Softmax function to obtain the normalized attention weights. ； In the formula: Indicates energy storage node For power grid nodes Normalized attention weights; This represents an exponential function with the natural constant e as its base. This represents the modified linear unit activation function with leakage; This represents the transpose of the weight vector of a single-layer feedforward neural network. Operators for concatenating feature vectors; The nodes obtained from the physical adjacency matrix represent... The set of first-order topological physical neighbor nodes; Indicates the index number for summation traversal; , , Representing power grid nodes Energy storage nodes Neighbor nodes The feature representation vector; This represents the shared weight transformation matrix.

5. The energy storage multidimensional value decoupling attribution method based on causal graph attention network according to claim 1, characterized in that, In step S4, the peak shaving and peak reduction simulation model is used to calculate the energy storage peak shaving active power output component in the factual physical state quantity and the peak shaving active power output component in the antifactual physical state quantity. The processing procedure is expressed as follows: ； In the formula: express The active power output of energy storage for peak shaving in real-world scenarios is 0, while the active power output for peak shaving in counterfactual scenarios is 0. This represents the logarithmic operator with base e. This represents an exponential function with the natural constant e as its base. The smoothing factor represents the control curve; express It contains confounding variables related to fluctuations in net load demand at all times; This represents the physical start threshold parameter for peak shaving issued by the underlying scheduling system.

6. The energy storage multidimensional value decoupling attribution method based on causal graph attention network according to claim 5, characterized in that, In step S4, the physical voltage component in the factual physical state quantity and the natural drop voltage component in the counterfactual physical state quantity are calculated using the voltage support deduction model. The processing procedure is expressed as follows: ； ； In the formula: express Voltage reactive power compensation output under real-time scenarios; This represents the voltage control gain smoothing constant; This represents the set target baseline value for the physical safety operation of the system; express Natural voltage drop components in counterfactual scenarios at all times; express The physical voltage component after being safely lifted in a real-world scenario; This represents the normalized attention weights.

7. The energy storage multidimensional value decoupling attribution method based on causal graph attention network according to claim 6, characterized in that, In step S4, the frequency offset component after frequency modulation response in the factual physical state quantity and the frequency natural drop component in the counterfactual physical state quantity are calculated using the double exponential frequency modulation deduction model. The processing procedure is expressed as follows: ； In the formula: express Frequency offset component after frequency modulation response in a real-time scenario; express The frequency of counterfactual scenarios naturally decreases in weight. The equivalent physical duration is calculated from the occurrence of the disturbance; A and B represent the physical gain coefficients that reflect the depth of energy storage droop control and inertia response, respectively. , , These represent the equivalent damping time constants in the corresponding dynamic processes.

8. The energy storage multidimensional value decoupling attribution method based on causal graph attention network according to claim 1, characterized in that, In step S5, the individual intervention effect matrix is ​​represented as follows: ； In the formula: express Time-based individual intervention effect matrix; This represents a causal inference physical interference quantifier; Represents a factual physical state quantity; Represents counterfactual physical state quantities; Indicates physical intervention variables; This indicates that the forced energy storage operates according to actual instructions; 0 indicates that the energy storage is forced to remain inactive; 0 indicates a counterfactual zero vector representing that the energy storage has not intervened. The counterfactual zero vector refers to the baseline state where the energy storage system is completely inactive and the output is zero.

9. The energy storage multidimensional value decoupling attribution method based on causal graph attention network according to claim 8, characterized in that, In step S5, the multidimensional physical value decoupling integral includes the peak-shaving integrated physical contribution, the voltage support integrated physical contribution, and the frequency modulation auxiliary integrated physical contribution; the calculation formula is expressed as: Peak-shaving integrated physical contribution: ； In the formula: Indicates the overall physical contribution of peak shaving; Indicates the total physical time period of the analysis; This represents the intervention effect value of the peak-to-valley difference in the corresponding cross-section load; This represents the discrete time step or the integral / differential time element; Overall physical contribution of voltage support: ； In the formula: This indicates the overall physical contribution of voltage support; This represents the set of physical nodes of the partition of controllable interest within the system. This indicates the intervention value for improving node voltage deviation; Frequency modulation-assisted overall physical contribution: ； In the formula: Indicates the overall physical contribution of frequency modulation assistance; This represents a definite integral operator over a specified time interval; Indicates the start time of the frequency disturbance; Indicates the duration of the recovery process; This indicates the intervention value for improving frequency transient deviation.

10. The energy storage multidimensional value decoupling attribution method based on causal graph attention network according to claim 9, characterized in that, In step S5, the comprehensive economic allocation parameter model is expressed as follows: ； In the formula: Indicates the parameters of comprehensive economic allocation; This indicates the unit settlement rate for peak-shaving auxiliary physical services; The parameter representing the conversion factor for reactive voltage support; The physical settlement price representing the FM mileage; Based on comprehensive economic allocation parameters The quantitative value is used to issue economic settlement and allocation instructions to the underlying trading system and scheduling system.

Images