Load situation space-time coupling perception and dynamic correction method based on user side behavior map

Abstract

The invention discloses a load situation space-time coupling perception and dynamic correction method based on a user side behavior map. The method comprises the following steps: dynamically constructing the user side behavior map; modeling space-time coupling characteristics; load situation awareness; dynamic error correction; and verifying and optimizing the effect. According to the method, the capabilities of accurately predicting the load change of the user side and capturing transient events are greatly improved, and meanwhile, the defects of a traditional method in the aspect of understanding the collaborative relationship between user behaviors and equipment are effectively overcome.

Description

Technical Field

[0001] The present invention relates to the field of electric power, and particularly to a method for spatio-temporal coupling perception and dynamic correction of load situation based on a user-side behavior map. Background Art

[0002] In modern power systems, accurately predicting the load situation and promptly responding to dynamic changes on the user side are crucial for improving the operating efficiency and stability of the power grid. However, due to insufficient sampling rates and the neglect of the collaborative relationship between devices in existing technologies, it is difficult to effectively capture transient events and device interaction characteristics, resulting in limited accuracy of load prediction. Especially when facing transient currents less than 20 ms generated by, for example, the start and stop of air-conditioning compressors, traditional methods often cannot provide sufficient sampling density to record the complete waveforms of these key events, thus affecting the understanding and analysis of user behavior patterns. In addition, traditional load prediction models mostly focus on time series analysis or static topological structures and lack the ability to consider spatio-temporal coupling characteristics simultaneously, which further limits their application effects in complex power environments. Summary of the Invention

[0003] The purpose of the present invention is to provide a method for spatio-temporal coupling perception and dynamic correction of load situation based on a user-side behavior map, which not only greatly improves the accurate prediction of user-side load changes and the capture ability of transient events, but also effectively solves the deficiencies of traditional methods in understanding the collaborative relationship between users and devices.

[0004] The technical solution of the present invention to solve the above technical problems is as follows:

[0005] A method for spatio-temporal coupling perception and dynamic correction of load situation based on a user-side behavior map includes the following steps:

[0006] S1, constructing a dynamic map of user-side behavior;

[0007] S2, based on the constructed dynamic map of user-side behavior, designing a two-channel spatio-temporal graph convolutional network to quantify the correlation intensity between user behavior and load characteristics and establish a spatio-temporal coupling correlation model;

[0008] S3, based on the spatio-temporal coupling correlation model, fusing multi-source data to quantitatively evaluate the load situation;

[0009] S4, performing dynamic error correction on the outputs of the spatio-temporal coupling correlation model and the quantitative evaluation of the load situation.

[0010] Optionally, in the step S1, a dynamic map of user-side behavior is constructed by high-frequency data acquisition and device fingerprint extraction;

[0011] In the high-frequency data acquisition, voltage / current waveforms and device status data are collected, and the minimum sampling interval Δt=0.1ms is set;

[0012] In the device fingerprint extraction stage, an improved time-varying window S transform is used:

[0013]

[0014] τ is the time shift parameter, which indicates the time point of the current analysis; f is the frequency parameter, which is used to specify the frequency component of the analysis; t is the time variable; x(t) is the original time domain signal; e -j2πft is a complex exponential function used for frequency decomposition; It is a time-varying Gaussian window function, whose bandwidth is adaptively adjusted with the frequency f;

[0015] Based on the data and device fingerprints obtained from high-frequency data collection and device fingerprint extraction, the probability of collaborative power consumption between devices is quantified by defining the device association matrix A.

[0016]

[0017] a ij is the correlation between device i and device j, ranging from [0,1]; T is the total observation time window length; δ(i t =1∩j t =1) is an indicator function, which is 1 when devices i and j are started at the same time at time t, otherwise it is 0; N i ,N j is the number of startups of devices i and j within the observation window.

[0018] Optionally, the step S2 is specifically:

[0019] (1) The spatial channel uses an improved graph attention network to dynamically learn the energy interaction weights between devices:

[0020]

[0021] is the feature vector of node i in the l-th layer network; N(i) is the set of neighbor nodes of device i; α ij is the attention weight of device i to device j; W (l) is the learnable weight matrix of layer l; σ is the activation function;

[0022] Attention weight α ij Calculated by the following mechanism:

[0023]

[0024] a is the parameter vector of the attention mechanism; LeakeyReLU is the leaky rectified linear unit;

[0025] (2) The time channel uses causal dilated convolution:

[0026]

[0027] y t is the output at time t; x t-d·k is the value of the input sequence at time t - d·k; w k is the weight of the k-th convolutional kernel; d is the dilation factor, which grows exponentially with the network layer;

[0028] (3) The behavior-load coupling modeling quantifies the correlation strength between user behavior and the load curve by defining the coupling degree index C bl (t):

[0029]

[0030] F b (t) is the behavior feature vector, including the device usage frequency and duration; F l (t) is the load feature vector, including active power and harmonic distortion rate.

[0031] Optionally, the step S3 is specifically:

[0032] (1) Multi-source data fusion constructs the user intention inference model P(y|X):

[0033]

[0034] y is the predicted event; X is the multi-source input data set; K is the number of data sampling points within the time window; x k is the electrical feature vector at the k-th time point; e k the environmental feature vector at the k-th time point; f elec (·) is the electrical feature encoder; f env (·) environmental feature encoder; GNN(·) is the graph neural network for fusing electrical and environmental features;

[0035] Fuse the two types of features through the graph neural network to identify the implicit user behavior logic;

[0036] (2) The situation quantification and evaluation defines the load situation index (LSI):

[0037]

[0038] P i (t) is the actual power of device i at time t; P ratedis the rated power of device i; Δf i (t) = |f i (t) - 50| is the frequency deviation; f base is the reference frequency; ω1, ω2 are weight coefficients.

[0039] Preferably, the weight coefficients are set as ω1 = 0.6 and ω2 = 0.4.

[0040] Optionally, in step S4, an error propagation model and an adversarial correction network are established;

[0041] Error propagation modeling describes the spatio-temporal evolution of errors based on the diffusion equation:

[0042]

[0043] E(x, t) is the error field intensity at position x and time t; D is the diffusion coefficient, reflecting the connectivity of the power grid topology; λ is the attenuation factor, simulating the natural dissipation rate of errors; S(x, t) is the error source term;

[0044] The adversarial correction network embeds Kirchhoff's law constraints in the generative adversarial network:

[0045]

[0046] G(z) is the load curve; D(·) is the probability value output by the discriminator; Y bus is the admittance matrix of the power grid nodes; I inj is the node injection current vector; β is the weight of the physical constraint term, balancing data-driven and physical consistency.

[0047] Optionally, it further includes an effect verification and optimization step, adopting an online monitoring index error suppression rate and an adaptive tuning mechanism;

[0048] The error suppression rate is defined as:

[0049]

[0050] P pred (t) is the predicted load value at time t; P true (t) is the actual load value at time t; T is the length of the evaluation time period;

[0051] Parameter adaptive tuning adopts a Bayesian optimization framework:

[0052]

[0053] θ is the set of hyperparameters to be optimized; θ *is the optimal hyperparameter configuration obtained through optimization; E[f(θ)|D] is the expected value of the objective function f(θ) under the condition of the existing observed data D; f(θ) is the objective function, representing the model performance index corresponding to the hyperparameter configuration θ; D is the existing observed data set, including historical hyperparameter configurations and their corresponding performance evaluation results.

[0054] The present invention has the following beneficial effects:

[0055] The present invention proposes a method for spatio-temporal coupling perception and dynamic correction of load situation based on the user-side behavior graph, which significantly improves the accuracy and reliability of load forecasting in the power system. By adopting high-frequency data acquisition, transient events that are difficult to record by traditional methods can be captured, ensuring the complete recording of the waveforms of key events. The introduction of the improved time-varying window S transform and device fingerprint extraction technology solves the problem of the fixed resolution of the traditional short-time Fourier transform and realizes the accurate identification of device interaction characteristics. In addition, the constructed behavior graph not only quantifies the probability of collaborative power consumption between devices, but also deepens the understanding of the interaction between devices and its impact on the power grid. The design of the dual-channel spatio-temporal graph convolutional network, especially the application of the improved graph attention network and causal dilated convolution, enables the model to capture the complex characteristics of the load simultaneously in the spatial and temporal dimensions, effectively solving the limitations of traditional load forecasting models and graph convolutional networks. The introduction of the multi-source data fusion and situation quantification evaluation mechanism further enhances the in-depth understanding of the load state. Combining the error propagation model with the adversarial correction network, closed-loop optimization is achieved, and error propagation is suppressed. Finally, through the joint application of the online monitoring index error suppression rate and the adaptive tuning mechanism, the present invention not only improves the model performance at the initial deployment, but also supports self-learning and optimization during long-term operation, demonstrating important progress and broad application potential in the field of smart grid technology.

[0056] The present invention can start from the behavior of the user side, realize the accurate perception and dynamic correction of the load situation by constructing a dynamic behavior graph based on high-frequency data acquisition and device fingerprint extraction, and combining advanced spatio-temporal coupling feature modeling technology, so as to meet the growing needs of the smart grid. It greatly improves the accuracy and reliability of load forecasting, and at the same time enhances the understanding and response ability to the user-side behavior. Brief Description of the Drawings

[0057] Figure 1 is the flowchart of the method of the present invention. Detailed Embodiments

[0058] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only used to explain the present invention and are not intended to limit the scope of the present invention.

[0059] Embodiment

[0060] The present invention provides a method for spatio-temporal coupling perception and dynamic correction of load situation based on user-side behavior graph in the field of electric power, with reference to Figure 1 As shown, the method includes:

[0061] S1: Dynamically construct the user-side behavior graph;

[0062] Due to insufficient sampling rate and ignoring the device collaboration relationship in the prior art, it is difficult to capture transient events and device interaction characteristics. Therefore, in this step, a dynamically updated behavior graph is constructed through high-frequency data acquisition and device fingerprint extraction.

[0063] (1) In high-frequency data acquisition, voltage / current waveforms and device status data are collected through smart meters with a sampling rate ≥ 10 kHz. Since in actual operations, transient currents with a duration less than 20 ms are generated during the start / stop events of air conditioner compressors, if traditional sampling methods are used, some transient feature information will be lost. Therefore, the present invention sets the minimum sampling interval Δt = 0.1 ms to ensure complete recording of key event waveforms such as device start / stop and short circuits.

[0064] (2) In the device fingerprint extraction stage, to address the defect of the fixed time-frequency resolution of the traditional short-time Fourier transform (STFT), an improved time-varying window S transform is designed:

[0065]

[0066] τ is the time shift parameter, representing the current analysis time point; f is the frequency parameter, used to specify the frequency component for analysis; t is the time variable; x(t) is the original time-domain signal; e -j2πft is the complex exponential function, used for frequency decomposition; is the time-varying Gaussian window function, whose bandwidth is adaptively adjusted with the frequency f.

[0067] (3) Construct the behavior graph. Based on the data and device fingerprints obtained in the previous two steps, the collaborative power consumption probability between devices is quantified by defining the device correlation matrix A, which helps to deeply understand how devices interact with each other and their impact on the power grid.

[0068]

[0069] a ij is the correlation degree between device i and device j, with a range of [0, 1]; T is the total observation time window length; δ(i t = 1 ∩ j t = 1) is the indicator function, which is 1 when devices i and j start simultaneously at time t, otherwise 0; N i , N j are the start times of devices i and j within the observation window.

[0070] S2: Spatiotemporal Coupling Feature Modeling;

[0071] Traditional load forecasting models only deal with time series, while graph convolutional networks are mostly used for static topology analysis. Neither of them can capture spatiotemporal coupling characteristics simultaneously. In this step, a dual-channel spatiotemporal graph convolutional network is designed to achieve feature decoupling and fusion.

[0072] (1) The spatial channel adopts an improved graph attention network (GATv2) to dynamically learn the energy interaction weights between devices:

[0073]

[0074] is the feature vector of node (device) i in the l-th layer network; N(i) is the set of neighbor nodes of device i (defined by the behavior graph); α ij is the attention weight of device i to device j; W (l) is the learnable weight matrix in the l-th layer; σ is the activation function.

[0075] The attention weight α ij is calculated through the following mechanism:

[0076]

[0077] a is the parameter vector of the attention mechanism; LeakyReLU is the leaky rectified linear unit.

[0078] Compared with the traditional GAT, GATv2 can identify asymmetric device relationships by introducing a dynamic attention mechanism.

[0079] (2) The time channel adopts causal dilated convolution (Dilated TCN) to solve the problem that traditional RNN-based models are difficult to capture long-period patterns:

[0080]

[0081] y t is the output at time t; x t-d·k is the value of the input sequence at time t - d·k; w k is the weight of the k-th convolutional kernel; d is the dilation factor, which grows exponentially with the network level.

[0082] (3) Behavior-load coupling modeling quantifies the correlation strength between user behavior and load curves by defining the coupling degree index C bl (t):

[0083]

[0084] F b(t) is the behavior feature vector, including device usage frequency, duration, etc.; F l (t) is the load feature vector, including active power, harmonic distortion rate, etc.

[0085] Behavior-load coupling modeling aims to quantify the association strength between user behavior and the power load curve by defining a coupling degree index, so as to effectively combine the analysis results in two dimensions of space and time with the actual usage behavior of users. This model utilizes the interaction information between different devices obtained from spatial channel analysis, such as which devices are often turned on simultaneously and their mutual dependencies, etc. At the same time, combined with the long-term trends and periodic data provided by the time channel, it deeply understands the specific impact ways of these behavior patterns on the power load. Finally, through a detailed quantitative analysis of user behavior characteristics (including but not limited to device usage frequency, duration, etc.) and load characteristics (such as active power, harmonic distortion rate, etc.), a clear connection between the two is established.

[0086] S3: Load situation awareness;

[0087] Existing technologies mostly rely on pure electrical measurement data and lack comprehensive consideration of user intentions and environmental factors. This step realizes in-depth understanding of the load state through multi-source data fusion and situation quantification evaluation.

[0088] (1) Multi-source data fusion constructs a user intention inference model P(y|X):

[0089]

[0090] y is the predicted event; X is the multi-source input data set; K is the number of data sampling points within the time window; x k is the electrical feature vector at the k-th time point; e k is the environmental feature vector at the k-th time point; f elec (·) is the electrical feature encoder; f env (·) is the environmental feature encoder; GNN(·) is the graph neural network for fusing electrical and environmental features.

[0091] By fusing the two types of features through the graph neural network (GNN), the implicit user behavior logic can be identified.

[0092] (2) Situation quantification evaluation defines the load situation index (LSI):

[0093]

[0094] P i (t) is the actual power of device i at time t; P rated is the rated power of device i; Δf i (t) = |fi (t) - 50| is the frequency deviation; f base is the reference frequency; ω1, ω2 are weight coefficients, and the weights are set as ω1 = 0.6 and ω2 = 0.4, reflecting the operation and maintenance strategy that power over - limit takes precedence over frequency deviation.

[0095] The weights are set as ω1 = 0.6 and ω2 = 0.4 because in actual operation and maintenance, power over - limit (P i / P rated > 1) directly threatens the safety of equipment and needs to be given priority attention; while the frequency deviation (Δf i / f base ) reflects the stability of the power grid and needs to be considered but is not the primary indicator. When LSI(t)>0.8, an alarm is triggered to guide the intervention of operation and maintenance personnel.

[0096] The multi - source data fusion to construct the user intention inference model provides the necessary input for the situation quantitative assessment. Through the comprehensive analysis of electrical and environmental characteristics, more accurate user behavior information can be obtained, which is one of the bases for calculating LSI. Based on the results of the user intention inference model, it can help determine which areas or devices may face higher risks (such as power over - limit), thus guiding the setting of LSI and the triggering conditions of the alarm mechanism.

[0097] S4: Dynamic error correction;

[0098] The traditional threshold alarm mechanism can only correct anomalies at a single point and cannot suppress error propagation. The present invention realizes closed - loop optimization by establishing an error propagation model and an adversarial correction network.

[0099] The error propagation modeling describes the spatio - temporal evolution of errors based on the diffusion equation:

[0100]

[0101] E(x, t) is the error field strength at position x and time t; D is the diffusion coefficient, reflecting the connectivity of the power grid topology; λ is the attenuation factor, simulating the natural dissipation rate of errors; S(x, t) is the error source term (such as local errors caused by equipment failures).

[0102] The adversarial correction network embeds Kirchhoff's law constraints in the generative adversarial network (GAN):

[0103]

[0104] G(z) is the load curve; D(·) is the probability value output by the discriminator; Y bus is the admittance matrix of the power grid nodes; I inj is the node injection current vector; β is the weight of the physical constraint term, balancing data - driven and physical consistency.

[0105] This design ensures that the generated corrected load curve satisfies the circuit laws, avoiding physically infeasible solutions generated by traditional data-driven methods.

[0106] In the dynamic error correction step, closed-loop optimization is achieved by establishing an error field dynamics model and a physical constraint adversarial correction mechanism. First, the error propagation modeling describes the spatio-temporal evolution process of errors in the power grid based on the diffusion equation. This model can detail the variation of the error field intensity with position and time, where the diffusion coefficient reflects the connectivity of the power grid topology, the attenuation factor simulates the rate of natural error dissipation, and the error source term identifies the origin of local errors such as equipment failures. This modeling approach provides us with a profound understanding of how errors are generated and spread within the power grid.

[0107] Meanwhile, the adversarial correction network adopts the method of generative adversarial networks (GANs) and embeds key physical constraints in its framework. This method ensures that the corrected load curve is not only closer to the true value but also strictly follows the physical rules of the power system. The adversarial correction network adjusts the load curve, evaluates the consistency between the correction result and the actual power grid behavior using the probability value output by the discriminator, and at the same time, by introducing the weight of the physical constraint term, finds the best balance between the data-driven result and physical consistency, thus avoiding generating solutions that do not conform to physical laws.

[0108] The relationship between the two is complementary and mutually reinforcing. The error propagation modeling reveals the propagation pattern and influence range of errors in the power grid, providing a theoretical basis and direction guidance for the adversarial correction network; while the adversarial correction network dynamically corrects errors based on this information combined with specific physical constraints, effectively suppressing the propagation of errors and improving the prediction accuracy. Ultimately, this strategy not only enhances the accuracy and reliability of load forecasting but also ensures that all correction measures are physically reasonable and feasible, forming an efficient and reliable closed-loop optimization system. This method significantly improves the stability and operation and maintenance efficiency of the power system, reflecting an important progress in smart grid technology.

[0109] S5: Effect verification and optimization;

[0110] To verify the effectiveness of the method and achieve parameter self-optimization, an online monitoring index error suppression rate and an adaptive tuning mechanism are designed.

[0111] The error suppression rate is defined as:

[0112]

[0113] P pred (t) is the predicted load value at time t; P true (t) is the actual load value at time t; T is the length of the evaluation time period.

[0114] The parameter adaptive tuning adopts the Bayesian optimization framework:

[0115]

[0116] θ is the set of hyperparameters to be optimized; θ * is the optimal hyperparameter configuration obtained through optimization; E[f(θ)|D] is the expected value of the objective function f(θ) under the condition of the existing observed data D; f(θ) is the objective function, representing the model performance index corresponding to the hyperparameter configuration θ; d is the existing observed data set, including historical hyperparameter configurations and their corresponding performance evaluation results.

[0117] In the effect verification and optimization stage, in order to ensure that the proposed method is both effective and can be automatically adjusted to achieve the best performance, an online monitoring index error suppression rate and an adaptive tuning mechanism are designed. These two work closely together to jointly ensure the continuous improvement and optimization of the system performance.

[0118] First of all, the error suppression rate, as a key online monitoring index, is used to quantify the prediction accuracy of the model. It is calculated by comparing the difference between the predicted load value and the actual load value at a certain moment, and gives a comprehensive evaluation within the entire evaluation time period. The error suppression rate directly reflects the ability of the model to capture the actual load changes and is an important criterion for measuring the effectiveness of the method. A lower error suppression rate means that the model has a higher prediction accuracy and can more accurately reflect the real load situation.

[0119] Secondly, in order to further improve the model performance and achieve the optimal configuration of parameters, an adaptive tuning mechanism is introduced. This mechanism is based on the Bayesian optimization framework and aims to efficiently search the hyperparameter space to find the optimal hyperparameter configuration. In this process, the objective function is defined as the model performance index corresponding to the hyperparameter configuration, and it is evaluated based on the existing observed data set. This data set contains historical hyperparameter configurations and their corresponding performance evaluation results, enabling each optimization to be based on previous experience and gradually approaching the optimal solution. In this way, the adaptive tuning mechanism can not only significantly improve the prediction accuracy of the model, but also enhance its generalization ability and stability.

[0120] The relationship between the two is reflected in that the error suppression rate provides real-time feedback to help identify performance bottlenecks under the current model configuration; while the adaptive tuning mechanism dynamically adjusts model parameters based on this feedback with the aim of minimizing the error suppression rate, thereby continuously improving the overall performance of the model. This combination not only ensures the effectiveness of the model at the initial deployment but also supports self-learning and optimization during long-term operation, ensuring that the system can adapt to changing actual conditions and maintain a high level of prediction accuracy and stability. In this way, through the synergistic effect of the error suppression rate and the adaptive tuning mechanism, our solution achieves comprehensive optimization from theory to practice, demonstrating strong application potential and adaptability.

[0121] In step S1 of the present invention, device fingerprints are extracted through high-frequency data collection and an improved time-varying window S transform to construct a dynamic user behavior map. This map quantifies the probability of collaborative power consumption between devices through a device correlation matrix, providing device interaction relationships and transient event characteristics for the spatial channel in step S2.

[0122] Based on the dynamic map in step S1, step S2 designs a dual-channel spatio-temporal graph convolutional network to quantify the correlation strength between user behavior and load characteristics and establish a spatio-temporal coupling correlation model. Step S2 provides a model basis for the situation awareness in step S3 and the error correction in step S4.

[0123] On the basis of the correlation model in step S2, step S3 further integrates environmental factors and user intentions, constructs a multi-source data fusion model through a graph neural network, and defines a load situation index. The model in step S2 provides the correlation rules between behavior and load, while step S3 transforms it into an actionable situation assessment (such as LSI warning) through multi-source data fusion. Step S3 converts the theoretical model in step S2 into dynamic prediction and warning capabilities in actual applications by supplementing environmental and user intention data.

[0124] Step S4 dynamically corrects the theoretical outputs in step S2 (such as correlation strength, device weights, periodic patterns) and the situation awareness outputs in step S3 through an error propagation model and an adversarial correction network. Step S4 directly depends on the error identification in the theoretical output of the spatio-temporal coupling model in step S2 and the LSI warning in step S3. The correction results are fed back to steps S2 and S3 to form a closed-loop optimization, improving the model robustness and prediction accuracy.

[0125] Step S5 provides a basis for verification and optimization of the entire process of steps S1 - S4, ensuring that the system can continuously adapt to conditions such as user behavior and environmental changes in a dynamic environment; the optimization results are fed back to each step to form a full-link iterative improvement mechanism.

[0126] An embodiment of the present invention also provides a storage medium. The storage medium stores a computer program, and when the computer program is executed by a processor, some or all of the steps in the embodiment of the method for spatio-temporal coupling perception and dynamic correction of load situation based on the user-side behavior map provided by the present invention are implemented. The storage medium may be a magnetic disk, an optical disk, a read-only memory (ROM), a random access memory (RAM), or the like.

[0127] Those skilled in the art can clearly understand that the technologies in the embodiments of the present invention can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solutions in the embodiments of the present invention, in essence, or the part that contributes to the prior art can be embodied in the form of a software product. The computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments of the present invention.

[0128] The above are only the preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A load situation spatio-temporal coupling perception and dynamic correction method based on a user-side behavior map, characterized in that, It includes the following steps: S1. Construct a dynamic behavior graph of the user side; S2. Based on the constructed dynamic behavior graph of the user side, design a dual-channel spatio-temporal graph convolutional network, quantify the correlation strength between user behavior and load characteristics, and establish a spatio-temporal coupling correlation model; S3. Based on the spatio-temporal coupling correlation model, fuse multi-source data and quantitatively evaluate the load situation; S4. Dynamically correct the errors in the outputs of the spatio-temporal coupling correlation model and the quantitative evaluation of the load situation.

2. The load situation spatio-temporal coupling perception and dynamic correction method based on the user-side behavior atlas according to claim 1, characterized in that In step S1, a dynamic behavior graph of the user side is constructed through high-frequency data collection and device fingerprint extraction; In the high-frequency data collection, voltage / current waveforms and device status data are collected, and the minimum sampling interval Δt = 0.1 ms is set; In the device fingerprint extraction stage, an improved time-varying window S transform is adopted; τ is the time translation parameter, representing the time point of the current analysis; f is the frequency parameter, used to specify the frequency components of the analysis; t is the time variable; x(t) is the original time-domain signal; e -j2πft is the complex exponential function, used for frequency decomposition; is the time-varying Gaussian window function, whose bandwidth is adaptively adjusted with the frequency f; Based on the data and device fingerprints obtained from high-frequency data collection and device fingerprint extraction, the collaborative power consumption probability between devices is quantified by defining a device correlation matrix A; a ij is the correlation degree between device i and device j, with a range of [0, 1]; T is the length of the total observation time window; δ(i t = 1 ∩ j t = 1) is an indicator function, which is 1 when devices i and j start simultaneously at time t, and 0 otherwise; N i , N j is the number of startups of devices i and j within the observation window.

3. A method for spatio-temporal coupling perception and dynamic correction of load situation based on user-side behavior map according to claim 1, characterized in that Step S2 is specifically as follows: (1) The spatial channel adopts an improved graph attention network to dynamically learn the energy interaction weights between devices; is the feature vector of node i in the l-th layer network; N(i) is the set of neighbor nodes of device i; α ij is the attention weight of device i to device j; W (l) is the learnable weight matrix of the l-th layer; σ is the activation function; Attention weight α ij Calculated by the following mechanism: a is the parameter vector of the attention mechanism; LeakyReLU is the leaky rectified linear unit; (2) The time channel adopts causal dilated convolution; y t is the output at time t; x t-d·k is the value of the input sequence at time t - d·k; w k is the k-th convolutional kernel weight; d is the dilation factor, which grows exponentially with the network layer; (3) Behavior-load coupling modeling quantifies the correlation strength between user behavior and the load curve by defining the coupling degree index C bl (t): F b (t) is a behavior feature vector, including device usage frequency and duration; F l (t) is a load feature vector, including active power and harmonic distortion rate.

4. A spatio-temporal coupling perception and dynamic correction method for load situation based on user-side behavior map according to claim 1, characterized in that Step S3 is specifically as follows: (1) Fuse multi-source data to construct a user intention inference model P(y|X); y is the predicted event; X is the multi-source input data set; K is the number of data sampling points within the time window; x k is the electrical feature vector at the k-th time point; e k is the environmental feature vector at the k-th time point; f elec (·) is the electrical feature encoder; f env (·) Environmental feature encoder; GNN(·) is a graph neural network used to fuse electrical and environmental characteristics; Fuse the two types of characteristics through a graph neural network to identify the implicit user behavior logic; (2) The situation quantitative evaluation defines a load situation index (LSI); P i P(t) is the actual power of device i at time t; rated is the rated power of device i; Δf i (t) = |f i (t) - 50| is the frequency deviation; f base is the reference frequency; ω1 and ω2 are weighting factors.

5. A method for spatio-temporal coupling perception and dynamic correction of load situation based on user-side behavior map according to claim 4, characterized in that The weight coefficients are set as ω1 = 0.6 and ω2 = 0.

4.

6. A method for spatio-temporal coupling perception and dynamic correction of load situation based on user-side behavior map according to claim 1, characterized in that, In step S4, an error propagation model and an adversarial correction network are established; The error propagation modeling describes the spatio-temporal evolution of errors based on the diffusion equation; E(x,t) is the error field strength at position x and time t; D is the diffusion coefficient, reflecting the connectivity of the power grid topology; λ is the attenuation factor, simulating the natural dissipation rate of errors; S(x,t) is the error source term; The adversarial correction network embeds Kirchhoff's law constraints in the generative adversarial network; G(z) is the load curve; D(·) is the probability value output by the discriminator; Y bus is the admittance matrix of the power grid nodes; I inj is the vector of node injection currents; β is the weight of the physical constraint term, which balances data-driven and physical consistency.

7. A load situation spatio-temporal coupling perception and dynamic correction method based on the user-side behavior map according to claim 1, characterized in that It also includes an effect verification and optimization step, adopting an online monitoring index error suppression rate and an adaptive tuning mechanism; The error suppression rate is defined as: P pred P(t) is the predicted load value at time t; true P(t) is the actual load value at time t; T is the length of the evaluation time period; The parameter adaptive tuning adopts a Bayesian optimization framework. θ is the set of hyperparameters to be optimized; θ * is the optimal hyperparameter configuration obtained through optimization; E[f(θ)|D] is the expected value of the objective function f(θ) under the condition of the existing observed data D; f(θ) is the objective function, representing the model performance index corresponding to the hyperparameter configuration θ; D is the existing observed data set, including historical hyperparameter configurations and their corresponding performance evaluation results.

Images