Power distribution area fluctuating load early warning method and system thereof

Abstract

The application discloses a power distribution area fluctuation load early warning method and system, and relates to the technical field of intelligent monitoring and control of power distribution networks. The method comprises the following steps: constructing a digital twin model synchronized with the running state, jointly analyzing the digital twin data set in time domain and frequency domain, extracting power amplitude change rate and phase jump index, and generating a load fluctuation vector; determining an adaptive threshold based on the statistical characteristics of the load fluctuation vector and the historical load distribution, and generating a primary early warning signal when the fluctuation exceeds the limit; identifying the dominant load source and the unbalanced category by using a distributed collaborative analysis algorithm; correlating and analyzing the dominant load source identification result and the primary early warning signal to generate comprehensive early warning information; and outputting the comprehensive early warning information to the power distribution area monitoring platform to realize dynamic visualization and intelligent control. The method can realize real-time identification and active early warning of the fluctuation load of the power distribution area, and improve the safety and intelligent level of system operation.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring and control technology for power distribution networks, and in particular to a method and system for early warning of fluctuating loads in distribution substations. Background Technology

[0002] Distribution transformer substations are crucial nodes connecting the distribution network and end users, and their operational status directly impacts the reliability and power quality of the regional power supply. With the large-scale integration of distributed power sources, electric vehicles, and controllable loads, the load in distribution transformer substations exhibits strong randomness and suddenness, with the amplitude and frequency of load fluctuations continuously increasing. This leads to increasingly prominent problems such as three-phase imbalance, voltage exceeding limits, and equipment thermal overload.

[0003] In existing technologies, load analysis of transformer substations mostly relies on static statistical models or periodic prediction algorithms. These methods assume that load changes are stable and are difficult to reflect instantaneous fluctuation characteristics. Some studies have introduced frequency domain analysis or machine learning methods to identify abnormal fluctuations, but most algorithms are based only on single-node data and ignore the electrical coupling relationship between nodes, making it impossible to accurately locate the source of fluctuations and make a global judgment.

[0004] In terms of anomaly identification and early warning, traditional solutions mostly adopt fixed threshold discrimination mechanisms, which cannot adaptively adjust according to the operating status, resulting in high false alarm rates and poor timeliness. At the same time, existing systems mostly remain at the data monitoring level, lacking the ability to link with operation control and scheduling systems, and thus failing to form an effective early warning-response closed loop.

[0005] Therefore, current distribution transformer substations still have shortcomings in real-time identification of load fluctuations, dynamic adjustment of thresholds, location of dominant load sources, and linkage display of early warning information. There is an urgent need for a technical solution that can integrate multi-source data, reflect topological characteristics, and have dynamic analysis capabilities to achieve accurate monitoring and intelligent early warning of fluctuating loads in distribution transformer substations. Summary of the Invention

[0006] To address the above problems, this invention proposes a method and system for early warning of fluctuating loads in distribution radio areas.

[0007] The present invention achieves the above objectives through the following technical solutions:

[0008] A method for early warning of fluctuating load in a distribution radio area, the method comprising the following steps:

[0009] S1: Obtain the voltage and current signals of each phase line deployed in the distribution substation area, and combine the topology and operating parameters of the distribution substation area to construct a digital twin model that is synchronized with the distribution substation area in real time. Map the voltage and current signals as input parameters to the digital twin model to generate a digital twin dataset that represents the real-time operating status of the distribution substation area.

[0010] S2: Perform multi-domain feature processing on the digital twin dataset, and use the time domain and frequency domain joint analysis method to extract dynamic feature indicators that characterize the power amplitude change and phase change. Calculate the load fluctuation vector describing the degree of load fluctuation based on the dynamic feature indicators, which is used to reflect the load change trend of the distribution substation in different time periods.

[0011] S3: Based on the statistical change characteristics of the load fluctuation vector and the historical load distribution, determine an adaptive threshold for judging abnormal fluctuations. When the load fluctuation vector is detected to exceed the adaptive threshold, generate a primary fluctuation warning signal.

[0012] S4: When the primary fluctuation warning signal is detected, the dominant load source identification process is initiated: the load fluctuation vectors corresponding to multiple monitoring nodes in the distribution area are input into the digital twin model, the dominant load source causing the load fluctuation is identified through the distributed collaborative analysis algorithm, and the node location and imbalance category corresponding to the dominant load source are determined.

[0013] S5: The results of the identification of the dominant load source are correlated with the primary fluctuation warning signal to generate comprehensive warning information including the fluctuation intensity level, dominant load source information and imbalance category. The comprehensive warning information is then output to the distribution area monitoring platform or dispatch terminal to prompt the operators to intervene or adjust.

[0014] As a preferred embodiment of the present invention, the digital twin model is composed of a graph structure consisting of multiple virtual nodes and virtual line units. The virtual nodes correspond to physical devices in the distribution substation, and the virtual line units are used to characterize the electrical coupling relationship between each virtual node.

[0015] When performing real-time parameter synchronization on the digital twin dataset, an online consistency correction mechanism based on the voltage sensitivity matrix is ​​used for parameter updates. The voltage sensitivity matrix is ​​obtained by linearizing the topology and line impedance parameters of the distribution substation and is used to characterize the response relationship of node voltage to changes in injected power.

[0016] The parameter update rule is as follows:

[0017] ;

[0018] In the formula, , These are the parameter vectors of the digital twin before and after the update, respectively; This represents the measured voltage change. The increment for correcting the parameter vector of the digital twin; This is the step size coefficient; This is the voltage sensitivity matrix; It is a weighted matrix; The sparse regularization coefficient;

[0019] This ensures that the digital twin model remains consistent with the operating status of the distribution radio area in real time.

[0020] As a preferred embodiment of the present invention, step S2 specifically includes:

[0021] The isolation forest algorithm is applied to the power sequence in the digital twin dataset to identify and remove point outliers. The positions of the outliers are filled with the median of the power sequence to obtain a clean power sequence in the time domain.

[0022] A fast Fourier transform is performed on the time-domain clean power sequence. After conversion to the frequency domain, a low-pass filter is applied to filter out high-frequency components. Then, an inverse fast Fourier transform is performed to reconstruct the frequency-domain smooth power sequence.

[0023] Based on the topology of the distribution radio station, a graph Laplacian matrix is ​​constructed. The frequency-domain smoothed power sequence of each node in the digital twin dataset is used as a graph signal. Graph frequency-domain filtering is performed to smooth spatial noise and obtain a topologically consistent power sequence.

[0024] Based on the topologically consistent power sequence, a Gaussian process regression model with time as input and power as output is constructed. The sliding window maximum likelihood estimation method is used to update the hyperparameters of the Gaussian process regression model online to correct parameter drift and output the updated prediction confidence interval.

[0025] Based on the updated Gaussian process regression model output, the upper bound of the confidence interval of the amplitude change rate of the topologically consistent power sequence within the adaptive sliding time window is calculated as the robust amplitude change rate, and the phase jump index between voltage and current signals is calculated simultaneously.

[0026] Based on the period of the dominant frequency component in the topologically consistent power sequence, the length of the adaptive sliding time window is dynamically adjusted.

[0027] The robust amplitude change rate is combined with the phase jump index to form a load fluctuation vector.

[0028] In a preferred embodiment of the present invention, step S3, determining the adaptive threshold for judging fluctuation anomalies, includes the following method:

[0029] Extract historical load fluctuation vectors within a preset time range from the historical load database, assign spatial weights related to node electrical adjacency and line impedance distance to the historical load fluctuation vectors according to the topology of the distribution substation, and calculate the weighted empirical distribution function.

[0030] A two-dimensional feature plane is constructed with the robust amplitude change rate in the historical load fluctuation vector as the horizontal axis and the phase jump index as the vertical axis. On the two-dimensional feature plane, contour lines covering a preset confidence level are calculated based on a weighted empirical distribution function. The preset confidence level is determined based on the statistical distribution of the robust amplitude change rate output by the Gaussian process regression model.

[0031] The contour lines are extended outward to form a safety margin, creating a dynamic decision boundary, which is then used as an adaptive threshold.

[0032] In a preferred embodiment of the present invention, step S3, which involves generating a primary fluctuation warning signal when the load fluctuation vector is detected to exceed the adaptive threshold, specifically includes:

[0033] Based on the predicted confidence interval width output by the Gaussian process regression model, the uncertainty level of the robust amplitude change rate is calculated. When the uncertainty level is higher than the preset threshold, the safety margin of the dynamic decision boundary is increased by the first proportional coefficient. When the uncertainty level is lower than the preset threshold, the safety margin of the dynamic decision boundary is decreased by the second proportional coefficient to achieve real-time correction of the adaptive threshold.

[0034] A warning frequency feedback mechanism is established to count the number of primary fluctuation warning signals generated within a unit time window. When the number of primary fluctuation warning signals exceeds the frequency threshold, the safety margin is increased by a preset step size; when the number of primary fluctuation warning signals is lower than the frequency threshold, the safety margin is decreased by a preset step size.

[0035] The updated dynamic decision boundary is used as an adaptive threshold and compared with the real-time load fluctuation vector. When the proportion of the load fluctuation vector exceeding the corrected adaptive threshold within a continuous time window reaches the set judgment threshold, a primary fluctuation warning signal is generated, and the corresponding time index and node number in the digital twin model are recorded.

[0036] As a preferred embodiment of the present invention, the dominant load source identification process includes node dynamic feature extraction and collaborative clustering steps:

[0037] Based on the load fluctuation vector of each node in the digital twin model, a multi-scale node dynamic feature vector is extracted that integrates long-term dynamic characteristics and short-term fluctuation characteristics. The long-term dynamic characteristics are obtained by fitting the parameters of the Ornstein-Uhlenbeck stochastic process, and the short-term fluctuation characteristics include instantaneous gradient and variance.

[0038] On the graph structure of the digital twin model, the electrical distance is calculated based on the updated digital twin parameter vector, and distributed collaborative clustering is performed on the dynamic feature vector of the multi-scale nodes using the electrical distance as the weight.

[0039] The node communities generated by clustering are matched with the historical load feature database, and the communities that successfully match are marked as candidate dominant load sources.

[0040] As a preferred embodiment of the present invention, determining the node location and imbalance category corresponding to the dominant load source includes:

[0041] Based on the node electrical connection relationship in the digital twin model, a wave impact propagation model is constructed from the candidate dominant load source to the root node of the distribution substation, and the attenuation coefficient of wave energy on the propagation path is calculated.

[0042] Based on the fluctuation impact propagation model and attenuation coefficient, the contribution weight of the candidate dominant load source to the three-phase imbalance of the distribution transformer area is calculated. The candidate dominant load source with the largest contribution weight is determined as the dominant load source, and the node position and imbalance category corresponding to the dominant load source are recorded.

[0043] As a preferred embodiment of the present invention, step S5, which involves generating comprehensive early warning information including fluctuation intensity level, dominant load source information, and imbalance category, specifically includes:

[0044] Based on the time index of the primary fluctuation early warning signal record and the node number in the digital twin model, load fluctuation vectors corresponding to the time period and node range are extracted from the digital twin dataset, and a sliding time window is constructed on the time axis.

[0045] Within each sliding time window, the primary fluctuation warning signals of the dominant load source nodes are statistically aggregated, and the average power change amplitude of each dominant load source node is calculated. Frequency of warning signals per unit time and volatility variance Construct a fluctuation intensity scoring function for the dominant load source node: ;

[0046] In the formula, As the dominant load source node The fluctuation intensity score; , , These are adaptive weighting coefficients;

[0047] Based on the continuous values ​​output by the fluctuation intensity scoring function, the continuous scoring values ​​are mapped to discrete fluctuation intensity level intervals using a fuzzy inference mapping model to generate multi-level fluctuation intensity levels.

[0048] The multi-level fluctuation intensity levels are fused with the node location, type, and imbalance category of the dominant load source to form a structured comprehensive early warning information.

[0049] As a preferred embodiment of the present invention, the method for outputting the comprehensive early warning information to the distribution area monitoring platform or dispatch terminal includes:

[0050] Convert the structured comprehensive early warning information data frames into a preset data exchange format;

[0051] Based on the fluctuation intensity level and dominant load source type in the comprehensive early warning information data frame, the rule matching engine retrieves the control strategy template corresponding to the fluctuation intensity level from the control strategy knowledge base to generate a set of candidate control suggestions.

[0052] Through the distribution area communication link and the preset communication protocol interface, the converted comprehensive early warning information data frame and the candidate control suggestion set are simultaneously sent to the distribution area monitoring platform or dispatch terminal.

[0053] In the human-machine interface of the distribution area monitoring platform or dispatch terminal, the comprehensive early warning information is displayed in a visual manner according to the alarm level display rules;

[0054] In the visualization interface, based on the topology of the digital twin model, the location of the dominant load source node is marked, and the fluctuation causal path from the dominant load source to the root node of the distribution area is rendered, realizing dynamic topology display and synchronous update of alarm status.

[0055] A distribution area load fluctuation early warning system, the system comprising:

[0056] The data acquisition module is used to acquire voltage and current signals of each phase line in the distribution transformer area, and to receive the topology and operating parameters of the distribution transformer area.

[0057] The digital twin modeling module is used to construct a digital twin model that is synchronized with the distribution substation in real time based on the topology and operating parameters of the distribution substation, and to map the voltage and current signals as input parameters to the digital twin model to generate a digital twin dataset that represents the real-time operating status of the distribution substation.

[0058] The multi-domain feature processing module is used to perform joint time-domain and frequency-domain analysis on the digital twin dataset, extract dynamic feature indicators that characterize power amplitude and phase changes, and calculate a load fluctuation vector that describes the degree of load fluctuation, which is used to reflect the load change trend of the distribution substation in different time periods.

[0059] An adaptive threshold calculation module is used to determine an adaptive threshold for judging abnormal fluctuations based on the statistical change characteristics of the load fluctuation vector and the historical load distribution, and to generate a primary fluctuation warning signal when the load fluctuation vector is detected to exceed the adaptive threshold.

[0060] The dominant load source identification module is used to input the load fluctuation vectors corresponding to multiple monitoring nodes in the distribution area into the digital twin model when the primary fluctuation warning signal is detected, and to identify the dominant load source causing the load fluctuation through a distributed collaborative analysis algorithm, and determine the node location and imbalance category corresponding to the dominant load source.

[0061] The comprehensive early warning generation module is used to perform correlation analysis between the identification results of the dominant load source and the primary fluctuation early warning signal to generate comprehensive early warning information including fluctuation intensity level, dominant load source information and imbalance category;

[0062] The information interaction and visualization module is used to output the comprehensive early warning information to the distribution area monitoring platform or dispatch terminal in a preset data exchange format, and to dynamically visualize it in the topology interface of the digital twin model to prompt operators to intervene or adjust.

[0063] The beneficial effects of this invention are as follows: By constructing a digital twin model synchronized with the operating status of the distribution substation, multi-source data fusion modeling and real-time simulation are achieved, improving the accuracy and response speed of fluctuation monitoring; through multi-dimensional feature extraction using joint time and frequency domain analysis, power change characteristics can be comprehensively depicted, enabling accurate capture of load fluctuation trends and disturbance characteristics; by adopting an adaptive threshold determination mechanism based on historical distribution, the warning threshold can be dynamically adjusted according to the operating status, significantly reducing false alarms and missed alarms; by introducing a distributed collaborative analysis algorithm to identify the dominant load source, the precise location of the fluctuation source and the determination of the imbalance category are achieved, improving the pertinence and interpretability of fluctuation warnings; by generating comprehensive warning information including fluctuation intensity level, dominant load source information, and imbalance category through comprehensive correlation analysis, and displaying it in conjunction with the monitoring platform, the automatic generation and intelligent control closed loop of warning information are realized. Attached Figure Description

[0064] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0065] Figure 1 This is a flowchart of the method of the present invention;

[0066] Figure 2 This is a schematic diagram of system modules in an embodiment of the present invention. Detailed Implementation

[0067] 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 some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0068] like Figure 1 As shown, this is an embodiment of the present invention, which provides a method for early warning of fluctuating loads in a distribution radio area, including the following steps:

[0069] S1: Obtain the voltage and current signals of each phase line deployed in the distribution substation area, and combine the topology and operating parameters of the distribution substation area to construct a digital twin model that is synchronized with the distribution substation area in real time. Map the voltage and current signals as input parameters to the digital twin model to generate a digital twin dataset that represents the real-time operating status of the distribution substation area.

[0070] In one embodiment, by constructing a virtual network in the digital space corresponding to the physical distribution area and correcting its model parameters in real time, the digital twin is made to maintain dynamic consistency with the physical distribution area, thereby providing accurate input data for load fluctuation early warning.

[0071] First, voltage and current acquisition devices are deployed at each key monitoring point in the distribution transformer area (such as low-voltage outgoing lines, branch nodes, and main load terminals). These devices upload real-time signals to a digital twin platform via a communication network. On this platform, a topology structure consisting of virtual nodes and virtual line units is established. Each virtual node corresponds to an actual device in the physical transformer area (e.g., a transformer, feeder connection, or user load terminal). Each virtual line unit describes the electrical connection characteristics between two virtual nodes, with parameters including line impedance, admittance, and line power loss coefficient. This structure allows for the simulation of the actual physical processes of voltage, current, and power flow in the transformer area within a virtual space.

[0072] During operation, the operating conditions of the physical distribution area will constantly change, such as load fluctuations, temperature changes, and changes in connected equipment, all of which will cause deviations in line parameters. In order to ensure that the digital twin model continuously reflects the real state, this embodiment adopts an online consistency correction mechanism based on the voltage sensitivity matrix.

[0073] Voltage sensitivity matrix This describes the linear response of each virtual node's voltage change to injected power change. For example, when the power of a node increases, the proportion of voltage drop represents the node's sensitivity to its own power. This matrix can be obtained through linearized modeling of the distribution substation's topology and line impedance parameters.

[0074] Digital twin parameter vector This includes a set of parameters for each virtual circuit element in the digital twin model, such as line impedance and node admittance. These parameters are updated over time to reflect the dynamic changes in the physical system. (Correction increment) This represents the correction amount of the model parameters at the current moment. It is a variable in the optimization process, and its solution objective is to minimize the deviation between the virtual node voltage change calculated by the model and the measured voltage change.

[0075] In each update cycle, the system collects the real-time voltage changes of each monitoring node. The voltage change is then compared with that predicted by the digital twin model. Based on this difference, the system performs the following optimization calculations:

[0076] ;

[0077] In the formula, , These are the parameter vectors of the digital twin before and after the update, respectively; This is the measured voltage change. The correction increment for the digital twin parameter vector is calculated by measuring the voltage change predicted by the virtual model. Compared with the measured voltage change The difference between them represents the degree to which the model deviates from the actual working conditions; is the step size coefficient, used to control the update step size and prevent over-correction from causing model oscillation; is the voltage sensitivity matrix. The weighted matrix is ​​used to calculate the weighted matrix. Different weights are assigned to the errors at different nodes, and the critical nodes that have a greater impact on voltage stability are corrected first. Here are the sparse regularization coefficients and the sparse regularization term. This is used to limit the magnitude of parameter adjustments, ensuring that only a few parameters related to actual fluctuations are updated, thus avoiding over-adjustment that could cause system instability.

[0078] Updated digital twin parameter vector The virtual node voltage is written into the digital twin model and recalculated. It is then compared with the measured voltage. If the error converges to within a preset threshold, the model is considered to be consistent with the physical transformer area.

[0079] The mathematical calculations involved in this algorithm can all be implemented using existing open-source optimization tools (such as PyTorch, TensorFlow, or CVXPy), and the model parameters and voltage sampling data can be obtained from common intelligent acquisition terminals and transformer monitoring devices. Weighting matrix Step size coefficient and sparse regularization coefficient The value of can be flexibly set according to the size of the distribution area, the number of nodes, and the data update cycle to ensure the stability and computational efficiency of the algorithm in real-time operation. Among them, the diagonal elements of the weighting matrix... It can be determined based on node voltage sensitivity or node importance. For example, when determined based on voltage sensitivity, it can be taken as... , These are the diagonal elements of the voltage sensitivity matrix. To prevent division by zero, when the value is determined based on node importance, a constant can be taken as... , The calculation is normalized based on the node's voltage level, load capacity, and fluctuation frequency, with a value ranging from 0.2 to 1.0. The step size factor can be a fixed value or an adaptively decreasing form. For example, with a fixed step size... , The value range is 0.05 to 0.3; for adaptive step size, it can be taken as... , This is the initial step size (e.g., 0.2). The attenuation coefficient is (0.01~0.05). To update the cycle, in actual distribution substations, the step size can be adjusted according to the data update period (e.g., every 10 to 30 seconds) to ensure stable convergence of the model under high-frequency sampling conditions. The sparsity regularization coefficient can be taken as a value between 0.001 and 0.1, and can be dynamically adjusted according to the intensity of load fluctuations. , This is the initial value (e.g., 0.01). The adjustment coefficient is (0.1 to 0.5). This represents the standard deviation of load fluctuation within the current period; when system load fluctuations are severe, the sparse regularization coefficient automatically increases to enhance the robustness of the model; when the system is stable, Reduce the value to improve the calibration sensitivity.

[0080] Using the methods described above, the digital twin model can complete a global parameter correction within seconds to tens of seconds. In situations such as sudden load changes in the distribution substation, the connection of distributed power sources, or the charging of electric vehicle groups, model deviations can be identified and quickly corrected in real time. The corrected model provides high-precision input data for subsequent load fluctuation early warnings, enabling the predictive model to be trained and inferred based on real-world operating conditions, significantly improving the accuracy of early warnings.

[0081] S2: Perform multi-domain feature processing on the digital twin dataset, and use the joint analysis method of time domain and frequency domain to extract dynamic feature indicators that characterize the changes in power amplitude and phase. Based on the dynamic feature indicators, calculate the load fluctuation vector that describes the degree of load fluctuation, which is used to reflect the load change trend of the distribution substation in different time periods.

[0082] In one embodiment, step S2 involves: removing outliers using the isolated forest algorithm → smoothing the time signal using Fourier transform and low-pass filtering → enhancing spatial consistency through graph signal filtering → achieving dynamic prediction and confidence interval calculation using Gaussian process regression → extracting the rate of change of amplitude and phase jump to form the load fluctuation vector. The entire implementation process is as follows:

[0083] S21: Abnormal Sample Identification and Temporal Clean Power Sequence Generation

[0084] Real-time load signals in distribution substations often contain isolated anomalies caused by sensor drift, communication jitter, or temporary interference. To prevent these anomalies from affecting the accuracy of subsequent fluctuation detection, this embodiment first selects the power time series of each node in the digital twin dataset and applies the Isolation Forest algorithm for anomaly detection.

[0085] The principle of isolated forests is to generate multiple decision trees by randomly partitioning the feature space. Abnormal samples have shorter paths within the trees and are easily isolated. The algorithm outputs an anomaly score for each sample point. If the score is higher than a preset threshold (e.g., 0.55), it is considered an anomalous sample. The contamination rate parameter is typically set between 0.01 and 0.05 to adapt to different data quality scenarios. For the locations of samples determined to be anomalous, the median value of the power sequence within the time window is used to replace it, thereby generating a continuous, abrupt, clean power sequence in the time domain. Median padding effectively avoids error amplification caused by mean shift, allowing the reconstructed power sequence to maintain its fluctuation trend while suppressing transient interference.

[0086] S22: Frequency Domain Smoothing and Low-Pass Reconstruction

[0087] After obtaining the clean power sequence in the time domain, a Fast Fourier Transform (FFT) is performed on the power sequence to further remove measurement noise and extract the dominant periodic components, transforming the signal from the time domain to the frequency domain. In the frequency domain, the energy distribution of each frequency component can be clearly observed. Low-frequency components typically reflect normal periodic changes in the load (e.g., power fluctuations between day and night), while high-frequency components are mostly caused by transient interference or sampling errors. Therefore, a low-pass filter is applied to retain only the low-frequency components and filter out high-frequency noise. After filtering, an Inverse Fast Fourier Transform (IFFT) is performed on the frequency domain signal to reconstruct the signal back to the time domain, thus obtaining a smoothed power sequence in the frequency domain. This sequence exhibits a more stable and continuous fluctuation pattern, making it suitable for subsequent analysis.

[0088] S23: Graph Signal Filtering and Topology Consistency Correction

[0089] To ensure spatial data consistency, the physical structure of the distribution substation is mapped to a graph model. Nodes in the graph represent measurement points within the substation (e.g., branch nodes, user access points, transformer terminals), and edges represent line connections. Based on this topology, a graph Laplacian matrix is ​​constructed, with matrix elements reflecting the electrical coupling strength between nodes. The frequency-domain smoothed power sequence of each node is treated as a graph signal, and the signal is smoothed using a graph spectroscopic filter. For example, by setting a spectral response function... This is achieved through a thermonuclear filter, where... For smoothing coefficients, These are the eigenvalues ​​of the graph Laplacian matrix. The purpose of graph frequency domain filtering is to smooth out any abrupt changes or deviations from neighboring nodes in the data at a given node, ensuring that the resulting topologically consistent power sequence maintains both temporal continuity and spatial consistency. This guarantees the electrical rationality of the overall data for the transformer substation area.

[0090] S24: Dynamic Update of Gaussian Process Regression Model

[0091] After obtaining the topologically consistent power sequence, a Gaussian process regression (GPR) model with time as input and power as output is constructed to achieve adaptive prediction of load change trends and uncertainty estimation.

[0092] Gaussian process regression (GPR) is a non-parametric statistical modeling method that can predict output values ​​at any time point based on historical observations, while simultaneously providing the prediction confidence interval. In actual operation, load distribution drifts with time and seasonal variations. To maintain consistency between the prediction model and the current state, this embodiment employs sliding window maximum likelihood estimation to update the hyperparameters of the GPR model (including kernel length scale, noise variance, etc.) online. The update rules ensure that the model can automatically correct parameter drift, allowing the prediction confidence interval to dynamically adjust with the load.

[0093] The updated model outputs two key quantities in each time window:

[0094] The mean of the predicted power (representing the expected trend);

[0095] Forecast confidence interval (representing the range of uncertainty in fluctuations).

[0096] S25: Extraction of robust amplitude change rate and phase jump index

[0097] Based on the upper bound of the confidence interval output by the Gaussian process regression model above, this embodiment defines a robust amplitude change rate, which is the relative rate of change of power over time calculated within the upper limit of the confidence interval. This indicator is more stable than the traditional difference method and can effectively suppress instantaneous spikes or abnormal fluctuations.

[0098] Simultaneously, the instantaneous phase angle is calculated from the voltage and current signals at the corresponding moment in the digital twin dataset (e.g., through arctangent operation). The phase difference between adjacent moments is the phase jump index. This index reflects the dynamic changes in power factor or electrical load type.

[0099] S26: Adaptive sliding time window adjustment

[0100] Since the dominant frequency components (such as daily load cycles and photovoltaic output cycles) of different distribution area nodes may differ, this embodiment automatically adjusts the sliding window length based on the dominant frequency of the topologically consistent power sequence. When the dominant cycle is short (e.g., photovoltaic fluctuations), the window length automatically decreases to improve response speed; when the cycle is long (e.g., residential load), the window length increases accordingly to maintain statistical stability. This ensures that robust amplitude change rate and phase indices are always calculated on an appropriate time scale.

[0101] S27: Load Fluctuation Vector Generation

[0102] robust amplitude change rate Phase jump index Combined synchronously according to time, forming a binary vector. :

[0103] ;

[0104] This vector is the load fluctuation vector, used to describe the load fluctuation characteristics of the distribution substation at the current moment. The load fluctuation vector will be used in subsequent steps for abnormal fluctuation identification, dominant load source location, and comprehensive early warning analysis.

[0105] Each sub-process of the above method can be automatically executed by the computing module in the digital twin system. Anomaly detection, FFT filtering, graph signal filtering, Gaussian process prediction, and sliding window control can all run in parallel. The entire method, while ensuring data quality, achieves accurate multi-domain extraction of load fluctuation characteristics in distribution substations, forming the basis for subsequent fluctuation identification.

[0106] S3: Based on the statistical change characteristics of the load fluctuation vector and the historical load distribution, determine an adaptive threshold for judging abnormal fluctuations. When the load fluctuation vector is detected to exceed the adaptive threshold, generate a primary fluctuation warning signal.

[0107] In one embodiment of the present invention, a threshold decision boundary that reflects the dynamic characteristics of the distribution substation load is established based on historical load fluctuation patterns and real-time prediction model output, used to determine whether load fluctuations belong to an abnormal state. Specifically, the process includes the following:

[0108] S31: Retrieve the historical load database of the distribution area for the operating records within a preset time range (e.g., the past 30 days). Each historical record contains a historical load fluctuation vector generated by a data twin model. ,in The robust amplitude change rate at a historical moment is used to reflect the relative rate of change of load power at that moment. It is a phase transition index between voltage and current signals, used to reflect changes in load phase characteristics.

[0109] To ensure that the distribution of historical samples reflects the actual electrical connection characteristics, the system assigns spatial weights to each historical sample based on the topology of the distribution substation. This weight consists of two parts:

[0110] Electrical adjacency coefficient between nodes: This describes the electrical coupling strength between nodes. The closer the nodes are and the tighter the connection, the higher the weight.

[0111] Line impedance distance factor: used to describe the effect of electrical distance between nodes on power propagation. The greater the impedance, the lower the weight.

[0112] The system obtains the weighting coefficients for each sample through normalization, and then constructs a weighted empirical distribution function (WEDF) based on these weighted samples to characterize the overall statistical features of load fluctuations in the distribution substation. The implementation can be carried out using the following steps:

[0113] Perform cumulative distribution statistics on all historical sample points in weighted order;

[0114] Calculate and normalize the weighted frequency for each dimension (rate of change of amplitude, phase jump);

[0115] Forming a two-dimensional weighted empirical distribution function ,in For robust amplitude change rate, This is a phase jump indicator.

[0116] Through this weighted modeling, the system can reflect the relative contribution of different nodes to global fluctuations, improving the adaptability of threshold decision-making to spatial distribution differences.

[0117] S32: Map all historical samples to a two-dimensional coordinate plane with "robust amplitude change rate" as the horizontal axis and "phase jump index" as the vertical axis. The coordinates of each sample point represent its load fluctuation state within a specific time period. Based on the weighted empirical distribution function... Calculate the statistical coverage area at different confidence levels (e.g., 90%, 95%, 99%). The confidence level is determined by the robust magnitude rate of change prediction distribution output by the Gaussian process regression (GPR) model. The GPR model provides the mean and variance of the predicted values ​​at each time point; a larger variance indicates higher prediction uncertainty, and the corresponding confidence level should be lower.

[0118] Based on this statistical characteristic, the confidence level is adaptively selected (e.g., 90% when the prediction uncertainty is high and 99% when it is low). Then, contour lines corresponding to the confidence level are drawn on the two-dimensional feature plane to represent the statistical distribution boundary of most normal load fluctuation samples.

[0119] S33: To prevent the boundary from overly adhering to historical data and becoming overly sensitive to new fluctuations, the system expands the contour lines outwards by a safety margin. The expansion ratio can be set according to the system's preset risk coefficient or empirical coefficient (e.g., 5%–10%). The expanded boundary forms the Dynamic Decision Boundary (DDB). The adaptive threshold is the functional relationship or numerical set corresponding to this dynamic decision boundary, used to determine whether the real-time load fluctuation vector is within the normal range.

[0120] When a real-time load fluctuation vector is detected to fall outside this boundary in subsequent steps, it is determined to be in an "abnormal fluctuation" state, and the generation process of a primary fluctuation warning signal is triggered.

[0121] The adaptive threshold determination process can be implemented in software within the central processing unit of the distribution area digital twin system. Its core algorithm can be implemented using Python, MATLAB, or C++, and weighted statistics and contour line calculations can be performed using numerical libraries such as NumPy and SciPy. The data input consists of historical digital twin datasets and the current GPR prediction output. The data output is a two-dimensional dynamic threshold matrix or boundary equation, which can be used in real-time by the subsequent anomaly detection section.

[0122] In this embodiment, after detecting that the real-time load fluctuation vector exceeds the adaptive threshold, the threshold boundary is dynamically corrected based on the model prediction uncertainty and historical warning frequency, and a warning signal is generated when the continuity determination condition is met.

[0123] S34: Obtain the predicted confidence interval of the Gaussian process regression model output in real time. This confidence interval is characterized by the model's variance term. Based on the width of the predicted confidence interval, calculate the uncertainty level of the robust amplitude change rate, specifically by obtaining the uncertainty level parameter at the current time through normalization. Its calculation form can be expressed as: ,in The reference standard deviation represents the fluctuation level of the model under normal operating conditions.

[0124] Set two scaling factors:

[0125] First proportional coefficient Used to increase the threshold boundary under conditions of high uncertainty;

[0126] Second proportionality coefficient Used to reduce threshold boundaries under low uncertainty conditions.

[0127] when Execute when (the value is higher than the preset threshold): ;

[0128] when Execute when (below preset threshold):

[0129] Through the above real-time updates, the adaptive threshold is corrected online, ensuring a dynamic balance between the sensitivity of fluctuation detection and the stability of the model.

[0130] S35: Establish an early warning frequency feedback mechanism to count the number of primary fluctuation early warning signals generated within a unit time window (e.g., 5 minutes or 1 hour). And set the target frequency threshold. and adjust step size .like If the current threshold is deemed too low, the possibility of false alarms increases, therefore execution is performed. ;like If the threshold is too high, the risk of missed detection increases, therefore the following action is taken: This feedback adjustment process ensures that the threshold is optimized through self-learning based on system behavior over long-term operation, keeping the early warning mechanism within a reasonable sensitivity range.

[0131] S36: Use the updated dynamic decision boundary as an adaptive threshold, along with the real-time load fluctuation vector. Comparison, when detected When the current time index is outside the adaptive threshold, the corresponding node number in the digital twin model is recorded. To avoid misjudgment due to occasional noise, the system introduces a continuous time window mechanism: if the time window is of length [missing information], [missing information]. The proportion of judgments exceeding the corrected threshold within the time window. satisfy: If this is confirmed, a persistent load fluctuation anomaly is observed at the node, triggering the generation of a primary fluctuation warning signal. Among these, This represents the number of times within the window that exceed the threshold. This represents the total number of samples within the window. The threshold ratio for continuous determination (e.g., 0.6 or 0.7).

[0132] The generated primary fluctuation warning signal contains the following information fields:

[0133] Time Index ;

[0134] Corresponding node number (derived from digital twin model mapping);

[0135] Current corrected threshold parameter ;

[0136] Fluctuation intensity level (can be based on) (Hierarchical definition).

[0137] The method for generating primary load fluctuation early warning signals can be deployed in the online analysis layer of a digital twin platform, employing a parallel computing architecture to ensure real-time performance. Input data includes: dynamic decision boundary parameters, confidence intervals predicted by a Gaussian process regression model, and real-time load fluctuation vector sequences. The output is a structured primary early warning event data stream, which can be sent to a scheduling terminal or intelligent control system via a message queue (such as MQTT or Kafka) for subsequent load adjustments or manual intervention.

[0138] This embodiment achieves a double-closed-loop threshold self-tuning system by coupling model uncertainty analysis with a feedback frequency mechanism.

[0139] Short-term correction loop (based on model uncertainty): quickly responds to model prediction fluctuations, fine-tunes the safety margin in real time, and prevents threshold lag;

[0140] Medium- to long-term correction loop (based on early warning frequency): Automatically adjusts system sensitivity through statistical periodic feedback to ensure long-term stability;

[0141] Continuous judgment logic: avoids single-point false triggers and enhances the reliability of anomaly identification through time-series cumulative judgment.

[0142] S4: When a primary load fluctuation warning signal is detected, the dominant load source identification process is initiated: the load fluctuation vectors corresponding to multiple monitoring nodes in the distribution area are input into the digital twin model, and the dominant load source causing the load fluctuation is identified through a distributed collaborative analysis algorithm, and the node location and imbalance category corresponding to the dominant load source are determined.

[0143] In power distribution areas, accurately identifying specific load sources (such as a large motor or charging stations in a residential area) that cause voltage fluctuations and three-phase imbalances is crucial for precise control. Traditional methods often only determine that "an anomaly exists" but cannot quickly locate the "source of the anomaly," leading to low operation and maintenance efficiency. This embodiment details how to achieve automatic identification and location of the dominant load source through digital twin models and advanced data analysis techniques.

[0144] S41: Extracting Multi-Scale Node Dynamic Feature Vectors

[0145] Different types of loads exhibit different fluctuation characteristics. For example, motor startup shows a sudden, large gradient change (short-term characteristic), while residential electricity consumption shows a periodic, slow fluctuation (long-term characteristic). To provide a comprehensive picture, both long-term and short-term characteristics need to be extracted simultaneously.

[0146] Long-term dynamic characteristic extraction: For each node, its load fluctuation vector is treated as a time-varying signal, and an Ornstein-Uhlenbeck (OU) stochastic process is used to model this signal. This process can well describe the physical process that fluctuates around the mean and has the characteristic of regressing to the long-term mean, which is consistent with the dynamic behavior of power load. The maximum likelihood estimation method is used to fit the OU process, and three core parameters are obtained:

[0147] Mean regression rate: indicates the speed at which the load recovers from a fluctuating state to a normal level.

[0148] Long-term mean: Represents the average level of load over a long period of operation.

[0149] Volatility: Indicates the degree of fluctuation in load.

[0150] Short-term volatility characteristics extraction includes instantaneous gradient and variance:

[0151] Instantaneous change gradient: Calculate the difference between the current load value and the previous load value (or use the slope of a linear regression with a shorter time window) to capture sudden changes in load.

[0152] Variation variance: The variance of the load value is calculated within a short sliding time window (e.g., 1 minute) to measure the stability of the load during that period.

[0153] The three parameters of the above OU process, the instantaneous gradient and the variance of change, are combined with the robustness amplitude change rate and phase jump index calculated in step S2 to form a comprehensive multi-scale node dynamic feature vector. This vector completely describes the load behavior pattern of a node from multiple dimensions.

[0154] S42: Electrically Distance-Weighted Distributed Cooperative Clustering

[0155] In a power grid, nodes that are electrically close together have a greater influence on each other. This step aims to identify groups of nodes that are electrically closely connected and have similar electricity consumption behaviors. These groups are likely caused by the same dominant load source (such as a factory) or the same type of load source (such as a residential area). Using the updated digital twin parameter vector (containing accurate line impedance and other parameters) obtained through online correction of the voltage sensitivity matrix, the electrical distance between any two nodes is calculated based on the grid topology and line impedance. A common method is to calculate the magnitude of the equivalent impedance between nodes; the larger the impedance, the greater the electrical distance.

[0156] Perform distributed collaborative clustering: Treat the digital twin model of the entire distribution substation as a graph, where nodes are electrical nodes, edges are lines, and the weight of an edge is the reciprocal of its electrical distance (i.e., the closer the electrical distance, the greater the weight).

[0157] A distributed clustering algorithm, such as weighted spectral clustering or message-passing-based community detection, is employed. Each node sends its multi-scale dynamic feature vector to its electrical neighbors. The clustering algorithm compares not only the similarity of feature vectors between nodes (e.g., cosine similarity) but also uses electrical distance as a weight. This means that two nodes, even if their features are similar, will not be grouped together if they are electrically far apart; conversely, if they are electrically close and their features are similar, they are highly likely to be grouped into the same cluster.

[0158] Through multiple rounds of information exchange and aggregation between nodes, several node clusters are eventually formed.

[0159] S43: Candidate dominant load source marker

[0160] A historical load feature database is pre-established, storing multi-scale node dynamic feature vector templates for various typical load sources (such as "water pump motors," "electric arc furnaces," "residential areas," and "data centers"). The overall features of each node community identified in step S42 (such as the mean of all node feature vectors within the community) are matched with the templates in the feature database using similarity calculations (e.g., Euclidean distance or cosine similarity). If the similarity between a community's features and a template exceeds a preset matching threshold, the community is marked as a candidate dominant load source. For example, an identified community whose features highly match the "electric arc furnace" template is marked as a "candidate electric arc furnace load."

[0161] S44: Final determination of the dominant load source

[0162] Fluctuations from a load source propagate along the power grid lines, attenuating due to line impedance. Load sources closer to the point of problem manifestation (e.g., a busbar with excessive three-phase imbalance) and with lower path impedance have a greater actual impact. Based on the precise node connections and line impedance parameters in a digital twin model, an electrical path is constructed from the location of each candidate dominant load source to the root node of the entire distribution substation. A fluctuation propagation model is then established along this path. This model can be simplified to a transfer function, and by calculating line impedance and network topology, the attenuation coefficient of the fluctuation energy from the source node to the root node is determined. A larger attenuation coefficient indicates that the impact of the load source is less likely to be perceived by the system.

[0163] Based on the fluctuation impact propagation model and attenuation coefficient, the contribution weight of candidate dominant load sources to the three-phase imbalance of the distribution area is calculated. That is, for each candidate dominant load source, the formula for calculating its contribution weight can be designed as: contribution weight = (core characteristic intensity of the community, such as volatility) / (attenuation coefficient).

[0164] Compare the contribution weights of all candidate dominant load sources, determine the candidate dominant load source with the largest contribution weight as the final dominant load source, and automatically record the node location (e.g., "node 5 downstream of transformer 12") and imbalance category (e.g., "causing phase B to be overloaded").

[0165] S5: Correlate the identification results of the dominant load source with the primary fluctuation warning signal to generate comprehensive warning information including the fluctuation intensity level, dominant load source information and imbalance category, and output the comprehensive warning information to the distribution area monitoring platform or dispatch terminal to prompt the operators to intervene or adjust.

[0166] In one specific embodiment, the method for generating comprehensive early warning information including fluctuation intensity level, dominant load source information, and imbalance category includes the following steps:

[0167] S51: Data Extraction and Time Window Construction

[0168] When a primary load fluctuation warning signal is detected, the system first extracts the load fluctuation vector matching the warning signal time from the digital twin dataset based on the time index of the warning signal record and the corresponding node number in the digital twin model. To facilitate statistical calculations, a sliding time window is established on the time axis. The length of the sliding window is determined by the system's set observation period, such as 5 minutes, 10 minutes, or 30 minutes. Each window covers continuous load fluctuation data within that time period. As the window slides forward, new real-time data continuously enters, while expired data is discarded, thus achieving continuous dynamic updates.

[0169] S52: Statistical Calculation of Fluctuation Characteristics

[0170] Within each sliding time window, the primary fluctuation warning signals of the dominant load source node are statistically aggregated.

[0171] The following three features are calculated from the extracted load fluctuation vector:

[0172] (1) Amplitude of average power change : Represents the average magnitude of power change of the dominant load source node within this time window, and is calculated as follows: ,in, For nodes In the Power value at time, The average power within this window. This represents the number of sampling points within the sliding time window.

[0173] (2) Frequency of warning signals per unit time This represents the ratio of the number of times a node triggers a primary fluctuation warning signal to the length of the window within a given time window. A higher frequency indicates more frequent load fluctuations.

[0174] (3) Variance : Indicates the degree of instability of the node's load fluctuations within the window, which is defined as: The larger the variance, the more drastic the fluctuation.

[0175] The above three indicators can comprehensively reflect the fluctuation intensity, fluctuation frequency and stability characteristics of the dominant load source node.

[0176] S53: Fluctuation Intensity Scoring and Grading

[0177] To comprehensively assess the volatility of each dominant load source node, a volatility intensity scoring function is constructed in this embodiment:

[0178] ;

[0179] In the formula, As the dominant load source node The fluctuation intensity score, The higher the value, the more drastic the fluctuations at the corresponding node; , , The adaptive weighting coefficients set for the system are used to balance the importance of the three indicators;

[0180] The adaptive weighting coefficients can be determined through regression analysis of historical data, for example, by setting them empirically in the initial stage. The system can then automatically adjust based on real-time error feedback, gradually optimizing the scoring model.

[0181] Using fuzzy inference mapping model to Discretization is performed. The model pre-defines the correspondence between several scoring intervals and volatility level labels, for example:

[0182] If the range is between 0 and 0.3, the fluctuation intensity level is Level I (slight fluctuation).

[0183] If the value is between 0.3 and 0.6, the volatility level is Level II (moderate volatility).

[0184] If the value is between 0.6 and 0.8, the fluctuation intensity level is Level III (significant fluctuation).

[0185] When the range is greater than 0.8, the fluctuation intensity level is IV (severe fluctuation).

[0186] S54: Integrated Information Fusion and Output

[0187] The fuzzy inference module automatically assigns a corresponding fluctuation intensity level label to each node based on the membership function calculation results of the score value and the threshold interval. After obtaining the multi-level fluctuation intensity level, the system integrates the results with the basic information of the dominant load source node, including the node location, dominant load source type, and imbalance category.

[0188] S55: Visualizing the process

[0189] The generated comprehensive early warning information is encapsulated in a structured format. This comprehensive early warning information data includes fields such as the early warning signal time index, the dominant load source node number, the fluctuation intensity level, the dominant load source type, the imbalance category, the fluctuation causal path index, and the confidence level output by the Gaussian process regression model. To achieve cross-system data interoperability, the system converts the structured data frames into a preset data exchange format. This data exchange format can be JSON, XML, or a message format conforming to the IEC 61850 MMS standard. In cloud monitoring scenarios, JSON format is typically used for ease of parsing and secondary development; when interacting with power automation terminal equipment, the IEC 61850 standard format can be used to ensure communication compatibility with the distribution automation system.

[0190] The system is internally configured with a control strategy knowledge base, which pre-stores various executable strategy templates for different fluctuation levels and load source types. Each template is indexed by a combination of fluctuation level and load source type, forming a dual-keyword retrieval structure. For example, when the fluctuation level is medium and the dominant load source type is inductive, the corresponding strategy templates in the knowledge base may include control measures such as starting and stopping reactive power compensation equipment, adjusting transformer tap changer voltage levels, or automatic power factor correction. The system automatically reads the fluctuation level and dominant load source type fields from the comprehensive early warning information through a rule matching engine and retrieves the corresponding strategy templates from the control strategy knowledge base. When multiple templates meet the criteria, the system integrates them into a candidate control suggestion set and fuses them with the comprehensive early warning information data frame to form a complete instruction set that can be directly invoked by the dispatch terminal.

[0191] The converted comprehensive early warning information data frame and candidate control suggestion set are synchronously sent to the distribution area monitoring platform or dispatch terminal via the distribution area communication link and the preset communication protocol interface. The communication link can use industrial Ethernet, 4G or 5G industrial wireless private network, and the communication protocol can be a standardized protocol such as IEC 60870-5-104, MQTT or HTTPS. Before data transmission, the communication interface module is responsible for encoding and protocol encapsulation to ensure reliable interoperability between different platforms. When the system uses the MQTT protocol, the comprehensive early warning information message can be published to a unified topic path, such as " / district / grid / alerts / composite_warning". The monitoring platform or dispatch terminal can receive early warning data from the distribution area in real time by subscribing to this topic, achieving cloud synchronization.

[0192] After receiving comprehensive early warning information, the monitoring platform or dispatch terminal visualizes the information through the human-machine interface module according to preset alarm level display rules. The system automatically assigns colors and display priorities based on the fluctuation intensity level; for example, level I corresponds to green, level II to yellow, level III to orange, and level IV to red, to achieve visual differentiation of different alarm levels. The interface can provide both table view and topology view simultaneously. In the table view, the system lists the early warning records, load types, imbalance categories, and corresponding control suggestions for each dominant load source node in chronological order. In the topology view, based on a digital twin model, the spatial location and fluctuation status of the dominant load sources are marked in real time.

[0193] Dynamic rendering is performed on the digital twin topology interface to visualize and annotate the dominant load source nodes and graphically represent the causal path of fluctuations. The digital twin model is based on the actual electrical topology of the transformer substation and uses a 3D graphics rendering engine (such as WebGL, Cesium, or Three.js) to display nodes, lines, and fluctuation paths in real time. When new comprehensive early warning information arrives, the topology interface automatically refreshes and updates the node status. The system uses color gradients and dynamic line animations to represent the attenuation direction and intensity changes of fluctuation energy along the propagation path, thus forming a three-dimensional visual diagram of fluctuation propagation. Users can click on the dominant load source node in the interface to view detailed information about that node, including fluctuation intensity level, load source type, imbalance category, and candidate control suggestion set. The system also supports historical playback functionality, allowing users to track the evolution trajectory of fluctuation events within a specified time period.

[0194] In this embodiment, the data interface and protocol definitions are based on existing communication standards and can be directly implemented in power distribution monitoring systems or dispatch automation platforms. The rule matching engine uses key-value mapping and priority sorting algorithms, which have low computational complexity and can complete policy matching within millisecond response time. The visualization rendering module adopts an event-driven mechanism, which can respond to alarm data updates in real time and keep the interface state synchronized, thereby ensuring the continuity and real-time performance of the system.

[0195] like Figure 2 As shown, another embodiment of the present invention provides a distribution area fluctuating load early warning system, comprising:

[0196] The data acquisition module is used to acquire voltage and current signals of each phase line in the distribution transformer area, and to receive the topology and operating parameters of the distribution transformer area.

[0197] The digital twin modeling module is used to construct a digital twin model that is synchronized with the distribution substation in real time based on the topology and operating parameters of the distribution substation, and to map voltage and current signals as input parameters to the digital twin model to generate a digital twin dataset representing the real-time operating status of the distribution substation.

[0198] The multi-domain feature processing module is used to perform joint time-domain and frequency-domain analysis on the digital twin dataset, extract dynamic feature indicators that characterize power amplitude and phase changes, and calculate the load fluctuation vector that describes the degree of load fluctuation, which is used to reflect the load change trend of the distribution substation in different time periods.

[0199] The adaptive threshold calculation module is used to determine an adaptive threshold for judging abnormal fluctuations based on the statistical change characteristics of the load fluctuation vector and the historical load distribution, and to generate a primary fluctuation warning signal when the load fluctuation vector exceeds the adaptive threshold.

[0200] The dominant load source identification module is used to input the load fluctuation vectors corresponding to multiple monitoring nodes in the distribution area into the digital twin model when a primary fluctuation warning signal is detected. The module then uses a distributed collaborative analysis algorithm to identify the dominant load source causing the load fluctuation and determine the node location and imbalance category corresponding to the dominant load source.

[0201] The comprehensive early warning generation module is used to perform correlation analysis between the identification results of the dominant load source and the primary fluctuation early warning signal to generate comprehensive early warning information that includes the fluctuation intensity level, dominant load source information and imbalance category.

[0202] The information interaction and visualization module is used to output comprehensive early warning information to the distribution area monitoring platform or dispatch terminal in a preset data exchange format, and to dynamically visualize it in the topology interface of the digital twin model to prompt operators to intervene or adjust.

[0203] In summary, this invention establishes a systematic and innovative mechanism in areas such as data fusion, dynamic feature analysis, adaptive threshold judgment, dominant load source identification, and integrated early warning linkage. By introducing digital twin technology and multi-domain data modeling, it achieves real-time simulation and intelligent early warning of the operating status of distribution substations. This method has a clear structure and is feasible to implement, significantly improving the detection accuracy and operational safety of fluctuating loads in distribution substations. It has broad engineering application prospects and can be widely applied to the real-time monitoring and proactive control of smart distribution networks, regional energy internet, and distributed energy systems.

[0204] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in this application, and these should all be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A power distribution district fluctuating load early warning method, characterized in that, The method comprises the following steps: S1: Obtain the voltage and current signals arranged on each phase line of the power distribution area, and combine the topological structure and operating parameters of the power distribution area to construct a digital twin model synchronized with the power distribution area in real time, map the voltage and current signals as input parameters to the digital twin model, and generate a digital twin data set representing the real-time operating state of the power distribution area; The digital twin model is composed of a graph structure of multiple virtual nodes and virtual line units, wherein the virtual nodes correspond to physical devices in the power distribution area, and the virtual line units represent the electrical coupling relationship between the virtual nodes; When performing real-time parameter synchronization on the digital twin data set, an online consistency correction mechanism based on a voltage sensitivity matrix is used for parameter updating, wherein the voltage sensitivity matrix is obtained by linear modeling of the topological structure and line impedance parameters of the power distribution area, and is used to describe the response relationship of node voltage to injected power changes; The parameter updating rule is: ； In the formula, , are the digital twin parameter vectors before and after updating, respectively; is the measured voltage variation; is the correction increment of the digital twin parameter vector; is the step coefficient; is the voltage sensitivity matrix; is the weighting matrix; is the sparse regularization coefficient; to ensure that the digital twin model and the operating state of the power distribution area are kept in real-time consistency; S2: Perform multi-domain feature processing on the digital twin data set, extract dynamic feature indicators representing power amplitude changes and phase changes using a time and frequency domain joint analysis method, calculate a load fluctuation vector describing the load fluctuation degree according to the dynamic feature indicators, and use it to reflect the load change trend of the power distribution area at different time periods; S3: Based on the statistical change characteristics of the load fluctuation vector and the historical load distribution, determine an adaptive threshold for judging fluctuation abnormalities, and generate a preliminary fluctuation warning signal when the load fluctuation vector is detected to exceed the adaptive threshold; In step S3, the method for determining the adaptive threshold for judging fluctuation abnormalities comprises: extracting historical load fluctuation vectors within a preset time range from a historical load database, assigning spatial weights related to node electrical adjacency and line impedance distance to the historical load fluctuation vectors according to the topological structure of the power distribution area, and calculating a weighted empirical distribution function; construct a two-dimensional feature plane with the robust amplitude change rate in the historical load fluctuation vector as the horizontal axis and the phase jump indicator as the vertical axis, calculate the contour line covering the preset confidence level on the two-dimensional feature plane based on the weighted empirical distribution function; the preset confidence level is determined according to the statistical distribution of the robust amplitude change rate output by the Gaussian process regression model; extend the contour line outward by a safety margin to form a dynamic decision boundary, and use the dynamic decision boundary as the adaptive threshold; S4: When the preliminary fluctuation warning signal is detected, start the dominant load source identification process: input the load fluctuation vectors corresponding to multiple monitoring nodes in the power distribution area into the digital twin model, identify the dominant load source causing the load fluctuation through a distributed collaborative analysis algorithm, and determine the node position and unbalanced category corresponding to the dominant load source; S5: Correlate the dominant load source identification result with the primary fluctuation early warning signal, generate comprehensive early warning information containing fluctuation intensity level, dominant load source information and imbalance category, and output the comprehensive early warning information to the distribution area monitoring platform or dispatching terminal to prompt the operation personnel to intervene or adjust.

2. The power distribution area fluctuating load early warning method according to claim 1, characterized in that, Step S2 specifically comprises: Applying the isolation forest algorithm to the power sequence in the digital twin dataset to identify and remove point abnormal samples, filling the abnormal point position with the median of the power sequence to obtain a time-domain clean power sequence; Performing fast Fourier transform on the time-domain clean power sequence, filtering high-frequency components after conversion to the frequency domain by applying a low-pass filter, and then reconstructing by inverse fast Fourier transform to obtain a frequency-domain smooth power sequence; Based on the topology structure of the distribution area, a graph Laplacian matrix is constructed, and the frequency-domain smooth power sequence of each node in the digital twin dataset is taken as a graph signal to perform graph frequency domain filtering to smooth spatial noise and obtain a topologically consistent power sequence; Based on the topologically consistent power sequence, a Gaussian process regression model is constructed with time as input and power as output, and a sliding window maximum likelihood estimation method is used to update the hyperparameters of the Gaussian process regression model to correct parameter drift and output the updated prediction confidence interval; According to the output of the updated Gaussian process regression model, the upper bound of the amplitude change rate confidence interval of the topologically consistent power sequence in the adaptive sliding time window is calculated as the robust amplitude change rate, and the phase jump index between the voltage and current signals is also calculated; According to the period of the dominant frequency component in the topologically consistent power sequence, the length of the adaptive sliding time window is dynamically adjusted; The robust amplitude change rate and the phase jump index are combined into a load fluctuation vector.

3. The power distribution area fluctuating load early warning method of claim 1, wherein, In step S3, when the load fluctuation vector exceeds the adaptive threshold, a primary fluctuation early warning signal is generated, specifically including: According to the prediction confidence interval width output by the Gaussian process regression model, the uncertainty level of the robust amplitude change rate is calculated, and when the uncertainty level is higher than the preset threshold, the safety margin of the dynamic decision boundary is increased by a first proportion coefficient, and when the uncertainty level is lower than the preset threshold, the safety margin of the dynamic decision boundary is reduced by a second proportion coefficient, to realize real-time correction of the adaptive threshold; An early warning frequency feedback mechanism is established, and the number of primary fluctuation early warning signals generated in a unit time window is counted, and when the number of primary fluctuation early warning signals exceeds the frequency threshold, the safety margin is increased by a preset step; when the number of primary fluctuation early warning signals is lower than the frequency threshold, the safety margin is decreased by a preset step; The updated dynamic decision boundary is taken as the adaptive threshold, and compared with the real-time load fluctuation vector, and when the proportion of the load fluctuation vector exceeding the corrected adaptive threshold in the continuous time window reaches a set determination threshold, a primary fluctuation early warning signal is generated, and the corresponding time index and node number in the digital twin model are recorded.

4. The power distribution area fluctuating load early warning method of claim 3, wherein, The dominant load source identification process includes node dynamic feature extraction and collaborative clustering steps: Based on the load fluctuation vector of each node in the digital twin model, a multi-scale node dynamic feature vector fusing long-term dynamic characteristics and short-term fluctuation characteristics is extracted; the long-term dynamic characteristics are obtained by fitting Ornstein-Uhlenbeck random process parameters, and the short-term fluctuation characteristics include instantaneous change gradient and change variance; On the graph structure of the digital twin model, the electrical distance is calculated based on the updated digital twin parameter vector, and the multi-scale node dynamic feature vector is executed distributed collaborative clustering with the electrical distance as the weight; The node community generated by clustering is matched with the historical load feature library, and the community that matches successfully is marked as a candidate dominant load source.

5. The power distribution area fluctuating load early warning method of claim 4, wherein, The determination of the node position and the unbalanced category corresponding to the dominant load source includes: Based on the node electrical connection relationship in the digital twin model, a fluctuation influence propagation model from the candidate dominant load source to the distribution substation root node is constructed, and the attenuation coefficient of the fluctuation energy on the propagation path is calculated; Based on the fluctuation influence propagation model and the attenuation coefficient, the contribution weight of the candidate dominant load source to the distribution substation three-phase unbalance degree is calculated, the candidate dominant load source with the maximum contribution weight is determined as the dominant load source, and the node position and the unbalanced category corresponding to the dominant load source are recorded.

6. The power distribution area fluctuating load early warning method of claim 5, wherein, In step S5, the generation of comprehensive early warning information containing fluctuation intensity level, dominant load source information and unbalanced category includes: Based on the time index recorded by the primary fluctuation early warning signal and the node number in the digital twin model, the load fluctuation vector corresponding to the time period and node range is extracted from the digital twin data set, and a sliding time window is constructed on the time axis; In each sliding time window, the preliminary fluctuation early warning signals of the dominant load source nodes are statistically aggregated, and the average power variation amplitude of each dominant load source node is calculated , the early warning signal frequency per unit time , and the fluctuation variance , the fluctuation intensity scoring function of the dominant load source nodes is constructed: ; In the formula, the wave fluctuation intensity score value of the dominant load source node; , , the adaptive weighting coefficient;​ According to the continuous value output by the fluctuation intensity scoring function, the continuous scoring value is mapped to the discrete fluctuation intensity level interval by using a fuzzy reasoning mapping model, and a multi-level fluctuation intensity level is generated; The multi-level fluctuation intensity level is data fused with the node position of the dominant load source, the dominant load source type and the unbalanced category to form structured comprehensive early warning information.

7. The power distribution area fluctuating load early warning method of claim 6, wherein, The comprehensive early warning information is output to the distribution substation monitoring platform or the dispatching terminal, and the method includes: The structured comprehensive early warning information data frame is converted into a preset data exchange format; Based on the fluctuation intensity level and the dominant load source type in the comprehensive early warning information data frame, the control strategy template corresponding to the fluctuation intensity level is retrieved in the control strategy knowledge base through a rule matching engine, and a candidate control suggestion set is generated; Through the distribution substation communication link and the preset communication protocol interface, the converted comprehensive early warning information data frame and the candidate control suggestion set are synchronously sent to the distribution substation monitoring platform or the dispatching terminal; In the man-machine interaction interface of the distribution substation monitoring platform or the dispatching terminal, the comprehensive early warning information is visually displayed according to the alarm level display rule; In the visual interface, the dominant load source node position is labeled based on the topology structure of the digital twin model, and the fluctuation causal path from the dominant load source to the distribution substation root node is rendered, realizing dynamic topology display and alarm state synchronous update.

8. A power distribution area fluctuating load early warning system applied to the power distribution area fluctuating load early warning method of any one of claims 1-7, characterized in that, The system includes: The data acquisition module is configured to acquire voltage and current signals of each phase line in the power distribution area, and receive a topology structure and operating parameters of the power distribution area. The digital twin modeling module is configured to construct a digital twin model that is real-time synchronized with the power distribution area based on the topology structure and operating parameters of the power distribution area, and map the voltage and current signals as input parameters to the digital twin model to generate a digital twin data set representing a real-time operating state of the power distribution area. The multi-domain feature processing module is configured to perform time-domain and frequency-domain joint analysis on the digital twin data set, extract dynamic feature indexes representing power amplitude changes and phase changes, and calculate a load fluctuation vector describing a load fluctuation degree, for reflecting a load change trend of the power distribution area at different time periods. The adaptive threshold calculation module is configured to determine an adaptive threshold for judging fluctuation abnormalities based on statistical change characteristics of the load fluctuation vector and historical load distribution, and generate a primary fluctuation early warning signal when the load fluctuation vector exceeds the adaptive threshold. The dominant load source identification module is configured to input load fluctuation vectors corresponding to a plurality of monitoring nodes in the power distribution area to the digital twin model when the primary fluctuation early warning signal is detected, identify a dominant load source causing the load fluctuation through a distributed collaborative analysis algorithm, and determine a node position and an imbalance category corresponding to the dominant load source. The comprehensive early warning generation module is configured to perform correlation analysis on the dominant load source identification result and the primary fluctuation early warning signal to generate comprehensive early warning information including a fluctuation intensity level, dominant load source information, and an imbalance category. The information interaction and visualization module is configured to output the comprehensive early warning information to a power distribution area monitoring platform or a dispatching terminal in a preset data exchange format, and dynamically visualize the comprehensive early warning information in a topology structure interface of the digital twin model to prompt an operator to intervene or adjust.

Images