A power distribution network multi-time scale state estimation method based on PMU

Abstract

This invention belongs to the field of distribution network state estimation technology, specifically disclosing a multi-timescale state estimation method for distribution networks based on a Power Management Unit (PMU). The method includes: integrating high-frequency PMU data and low-frequency SCADA data; extracting multi-timescale features reflecting the dynamics and steady state of the power grid after quality assessment and preprocessing; identifying the current operating conditions of the power grid through real-time analysis of these multi-scale features; and adaptively generating a data fusion and state estimation strategy containing dynamic weighting coefficients and model selection instructions using a dynamic collaborative decision-making mechanism. This strategy guides the system to perform weighted fusion of multi-timescale features and match them with corresponding state estimation models for solution, adapting to different operating conditions such as changes in power grid topology or data gaps. This invention constitutes a closed-loop adaptive state estimation framework from data processing to model solving, thereby improving the risk resistance and emergency response capabilities of distribution network operation.

Description

Technical Field

[0001] This invention belongs to the field of distribution network state estimation technology, and relates to a multi-timescale state estimation method for distribution networks based on PMU. Background Technology

[0002] As a crucial link in the power system connecting users and transmitting electrical energy, the stable and efficient operation of the distribution network is of great significance to the national economy. In modern distribution network operation and management, real-time and accurate monitoring of the network's operating status is fundamental to achieving advanced application functions, such as fault diagnosis, optimized dispatching, voltage and reactive power control, and fault self-healing. State estimation technology is the core means of obtaining the operating status of the distribution network. It processes measurement data distributed throughout the network, eliminating the influence of measurement errors and poor data, thereby calculating the current voltage amplitude and phase angle of all nodes in the network. With the deployment of advanced measurement technologies such as synchronous phasor measurement units (PMUs), the amount and types of data available to the distribution network have increased significantly, making more accurate state estimation possible.

[0003] In the existing technology, there are also some solutions involving distribution network state estimation. For example, Chinese patent publication number CN117691584A discloses a dynamic state estimation method for distribution network intervals based on hybrid measurement fusion. It is mainly aimed at distribution networks with a high proportion of new energy sources. It integrates multiple data sources such as SCADA, micro PMU and pseudo measurement, and dynamically adjusts the weights according to the uncertainty and correlation of the measurement data. Its ultimate goal is to solve for the upper and lower boundaries of the state quantity, that is, to output an interval rather than a single precise value, so as to quantify the impact of uncertainty caused by new energy power generation.

[0004] However, after in-depth analysis, the existing technical solutions have the following significant technical defects and fail to solve the core challenges faced by the distribution network in actual operation: 1. The core algorithm framework of the existing methods is usually single and fixed. Although the weights can be adjusted, this static framework is difficult to cope with complex and ever-changing operating conditions such as distribution network line faults and topology reconfiguration. Once the actual operating conditions do not match the model's preset scenario, the estimation accuracy and convergence will drop significantly, and true operating condition adaptation cannot be achieved.

[0005] 2. Existing methods focus more on the inherent uncertainties of data sources, but lack sophisticated online assessment and proactive utilization mechanisms for data quality issues that arise in real time and change dynamically during the measurement process. When the quality of some data drops sharply, the system cannot proactively detect and adjust its fusion strategy or estimation model, resulting in low-quality data passively participating in the calculation, polluting the final result, and causing a sharp deterioration in system performance.

[0006] 3. In existing technologies, data preprocessing, feature extraction, working condition identification, and state estimation are often relatively independent processes, and there is a lack of collaborative optimization mechanisms between modules. Once an error or deviation occurs in a certain process, it will be directly passed down and affect the accuracy and robustness of the final estimation result, making it impossible to achieve overall system-level adaptability and anti-interference capabilities. Summary of the Invention

[0007] In order to overcome the above-mentioned defects of the prior art and to achieve the above objectives, the present invention proposes the following technical solution: a multi-time-scale state estimation method for distribution networks based on PMU, comprising: step 1, acquiring the high-frequency data stream of the synchronous phasor measurement unit (PMU) and the low-frequency data stream of the monitoring and data acquisition system (SCADA) in the distribution network, and generating multi-source measurement data.

[0008] Step 2: Perform quality assessment and preprocessing on the multi-source measurement data to generate standardized measurement data and corresponding measurement quality indicators.

[0009] Step 3: Extract multi-timescale features that characterize the power grid's operating status from standardized measurement data and generate a multi-timescale feature set.

[0010] Step 4: Based on multi-timescale feature sets and measurement quality indicators, identify the power grid operating conditions and generate labels for the current power grid operating conditions.

[0011] Step 5: Input the current power grid operating condition label and multi-time scale feature set into the dynamic estimation strategy generator to generate an adaptive fusion and estimation strategy.

[0012] Step 6: Based on the adaptive fusion and estimation strategy, perform data fusion on the feature sets of multiple time scales to generate fused feature vectors.

[0013] Step 7: Based on the adaptive fusion and estimation strategy, select a corresponding state estimation model and calculate the fusion feature vector to generate the distribution network state estimation result.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The present invention introduces data quality assessment and multi-time scale feature extraction, and uses a dynamic triangular collaborative decision-making mechanism to intelligently match the preferred fusion strategy and state estimation model, so as to ensure that the system can adaptively adjust under complex working conditions such as changes in power grid topology, partial data loss or high noise interference, avoid error amplification caused by fixed strategies, and thus continuously output high-precision power grid state results.

[0015] (2) By real-time monitoring and analysis of the quality and characteristics of multi-source data and multiple time scales, this invention can identify the specific operating conditions of the current power grid. This intelligent condition perception capability enables the state estimation strategy to be dynamically optimized in a targeted manner, allowing the entire system to shift from passive response to active adaptation, thereby reducing reliance on human experience and improving the automation level of distribution network operation and maintenance.

[0016] (3) By accurately identifying the power grid operating conditions and quickly generating adaptive strategies, this invention can provide more rapid and accurate comprehensive information on the current power grid status when facing sudden events or abnormal situations. This provides a reliable basis for dispatchers to make rapid decisions and take emergency measures, thereby improving the risk resistance and emergency response capabilities of the distribution network. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying 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.

[0018] Figure 1 This is a schematic diagram of the implementation steps of the method of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 As shown, the present invention proposes a multi-time-scale state estimation method for distribution networks based on PMU, including: Step 1, acquiring the high-frequency data stream of the synchronous phasor measurement unit (PMU) and the low-frequency data stream of the monitoring and data acquisition system (SCADA) in the distribution network to generate multi-source measurement data.

[0021] In a preferred embodiment, the engineering objective of this step is to establish two parallel real-time data acquisition channels, which are used to receive measurement information from the PMU and SCADA systems respectively, and to initially integrate the two data streams of different rates and formats into a unified raw dataset to be processed through standardized data interface protocols and data structures.

[0022] First, the system's data acquisition module configures and receives the high-frequency data stream from the Synchronous Phasor Measurement Unit (PMU). In engineering terms, this means configuring a client conforming to the IEEE C37.118 standard to connect to the Phasor Data Concentrator (PDC) in the distribution network. The PMU is a measurement device capable of precise time synchronization using GPS second pulse signals. It reports voltage phasors, current phasors, and frequencies of grid nodes at a rate of 25 to 100 frames per second. The data acquisition module initiates a connection based on the preset PDC server IP address and port number, subscribing to the data stream of the designated PMU device. The received data frames are parsed in real time, extracting the raw PMU measurements containing high-precision timestamps, phasor amplitudes, phasor phase angles, frequencies, and rates of change of frequency. These continuous data frames constitute the high-frequency data stream.

[0023] Simultaneously, the system's data acquisition module activates another acquisition channel to obtain low-frequency data streams from the SCADA (Supervisory and Data Acquisition) system. The SCADA system is the core of traditional power automation systems, primarily providing steady-state operation information for the power grid. This channel accesses the SCADA system's real-time database via a client configured with industrial communication protocols such as IEC 60870-5-104 or DNP3, or through an OPC interface. This operation is triggered by periodic polling, typically set to a period of 2 to 5 seconds. During each poll, the system actively requests and acquires specified telemetry and teleindication data, such as active power, reactive power, bus voltage amplitude, and the status information of circuit breakers and disconnectors. These data sets, updated at a lower frequency, constitute the low-frequency data stream.

[0024] Finally, to generate multi-source measurement data, a shared data buffer is established within the system. Data from the PMU high-frequency data stream and the SCADA low-frequency data stream are both sent to this buffer after being received and initially parsed. The system archives and organizes the data in the buffer using a fixed time base, such as 1 second. Within a time base period, all data frames reported by the PMUs, along with the data obtained from the latest SCADA poll within that period, are collectively encapsulated into a unified data structure unit. This structure unit represents the multi-source measurement data for a time slice, containing the dynamic and steady-state information of the power grid at that moment, and flows in this form to subsequent data quality assessment and preprocessing steps.

[0025] Step 2: Perform quality assessment and preprocessing on the multi-source measurement data to generate standardized measurement data and corresponding measurement quality indicators.

[0026] In a preferred embodiment, the step of performing quality assessment and preprocessing on multi-source measurement data to generate standardized measurement data and corresponding measurement quality indicators includes: time synchronization and alignment of the high-frequency data stream of the PMU and the low-frequency data stream of the SCADA.

[0027] Calculate the signal-to-noise ratio, data integrity rate, and internal consistency index of PMU data and SCADA data within a specific time window.

[0028] Based on the signal-to-noise ratio, data integrity rate, and internal consistency index, the measurement quality index of each data point is calculated through a scoring model.

[0029] Based on measurement quality indicators, the raw data is cleaned, corrected, and normalized to generate standardized measurement data.

[0030] Specifically, this method aims to perform quality assessment and preprocessing on multi-source measurement data collected by PMU and SCADA systems in the distribution network to ensure the accuracy of subsequent state estimation. Its engineering implementation process is as follows: Step 1: Time Synchronization and Alignment. The purpose of this project is to eliminate the time offset between the PMU high-frequency data stream and the SCADA low-frequency data stream, ensuring that all data is processed on a unified time base. PMU data typically includes a high-precision GPS timestamp, with a sampling frequency of, for example, 100 times / second, while SCADA data has a sampling frequency of, for example, 1 time / second, and its timestamp accuracy is relatively low. Matching is achieved through data timestamps. For example, the SCADA data timestamp can be upsampled or interpolated to the most recent PMU data timestamp, or the PMU data can be aggregated and averaged according to the SCADA data's time resolution to achieve synchronization. For instance, for one SCADA measurement per second, we will select the corresponding timestamp or the average or median of the PMU data within the most recent time window.

[0031] Step 2: Calculation of quality metrics within a specific time window. The purpose of this project is to quantitatively evaluate the reliability and availability of PMU and SCADA data within a local time period. For both PMU and SCADA data, a sliding time window is set, for example, 10 seconds or 30 seconds. Within this time window, the signal-to-noise ratio, data integrity rate, and internal consistency metrics are calculated respectively.

[0032] Signal-to-noise ratio calculation: For any measurement channel, such as PMU voltage amplitude measurement, let the average effective signal power within a time window be... and the average power of noise is Then the signal-to-noise ratio This value reflects the purity of the measurement data. The average power of the signal within a specific time window is obtained by calculating the average of the squares of the signal amplitude. The average power of the noise is obtained by calculating the variance of the residual after subtracting the smoothing trend from the signal.

[0033] Data integrity rate calculation: Within a time window, let the expected total number of data points be... The actual number of valid data points received was Then the data integrity rate This value reflects the presence of missing data.

[0034] Internal self-consistency index calculation: This index assesses the consistency of data from different sources by verifying whether they follow the basic physical laws of the power grid.

[0035] Step 3: Calculation of Measurement Quality Indicators. The purpose of this project is to synthesize the aforementioned local quality assessments and generate a unified measurement quality index for each data point. A scoring model, such as a weighted linear combination model or a fuzzy logic-based expert system, is used to calculate the signal-to-noise ratio of each data point. Data integrity rate and internal consistency indicators Input. Comprehensive measurement quality indicators for each data point. It can be calculated as ,in These are the weighting coefficients, and their sum is 1, for example... . It is and A normalization function that maps to the 0-1 interval. A higher value indicates the first The better the quality of each data point.

[0036] Step 4: Raw data cleaning, correction, and normalization to generate standardized measurement data. The purpose of this step is to optimize the raw data based on measurement quality indicators to make it suitable for subsequent state estimation algorithms. The data generated in this step also includes measurement quality indicators for each data point for use in subsequent decision-making.

[0037] First, perform data cleaning: set a rejection threshold for data points whose measured quality indicators are below this threshold, for example, by setting a preset rejection threshold. These data points are marked as unreliable and can be removed or corrected using interpolation to replace the original values. For example, if the quality index of five consecutive data points is below the removal threshold, these data segments are considered to contain a large amount of noise and are corrected.

[0038] Secondly, perform data correction: set a correction threshold that is higher than the removal threshold. For data points whose measurement quality index is between the removal threshold and the correction threshold, apply smoothing filters such as moving average filters or Kalman filters for optimization.

[0039] Finally, normalization is performed: for example, all voltage amplitude, current amplitude, active power and reactive power measurements are normalized to the 0 to 1 range of their respective nominal values ​​or maximum measured values, to ensure that measurement data of different dimensions and values ​​are compared and processed under a unified scale, and to generate standardized measurement data.

[0040] In a further preferred embodiment, the internal self-consistency index is obtained by calculating the power difference of the associated region using voltage phase information in the PMU data stream.

[0041] The power difference is compared with the power change in the same area obtained from the SCADA data stream for consistency.

[0042] Based on the results of the consistency comparison, an internal consistency index is generated quantitatively.

[0043] Specifically, the engineering objective of this step is to quantify the inherent reliability of multi-source measurement data by cross-validating the consistency of data from different measurement sources, thereby generating an internal self-consistency index. This process utilizes the fundamental physical laws of the power grid, namely the relationship between power change and phase angle change, to perform physical-level calibration between the high-frequency dynamic information of the PMU and the low-frequency steady-state information of the SCADA system.

[0044] The system first uses voltage phase information from the PMU data stream to calculate the power change within a specific associated region. In engineering terms, an associated region is defined as a portion of the power grid bounded by at least two nodes equipped with PMUs, such as a distribution feeder with PMUs at both ends. The system then locks the nodes based on a pre-configured power grid topology model. With nodes The related areas between them.

[0045] Then, extract the starting time of these two nodes within a specific time window. and the end time The voltage phasor information is used to calculate the change in active power within the time window based on the AC power system power flow equations. The calculation formula is as follows: ,in, This represents the change in active power calculated based on PMU data. It is a coefficient related to line parameters, and its expression is: , and These are nodes and The voltage amplitude is obtained from PMU data and is treated as a constant or taken as the average value within a time window in this calculation; It is a connection node and The line reactance is a static parameter retrieved from the power grid model database, with typical values ​​ranging from 0.01 to 0.5 of the baseline value. and These are nodes and At any moment The voltage phase angle is reported in real time by the PMU.

[0046] Subsequently, the system calculates the power change using this PMU. The system performs a consistency comparison with the power change data for the same area obtained from the SCADA data stream of the monitoring and data acquisition system. From the time-synchronized and aligned dataset, the system extracts data from the same related area within the same time window. The active power value measured by the SCADA system and Power change measured by SCADA It can be obtained directly by subtraction, that is .

[0047] Finally, based on the consistency comparison results, the system quantifies and generates an internal consistency index using a decay function. This index reflects the degree of consistency between the two data sources. Its calculation formula is as follows: ,in, It is an internal self-consistency index with a value range of [0,1]. The closer the value is to 1, the better the self-consistency. It is a dimensionless sensitivity adjustment coefficient, which is usually tuned based on offline simulation and historical data analysis. The value range is generally between 5 and 20, and it is used to adjust the sensitivity of the index to power deviation. It is the absolute value of the difference between the power changes calculated by the two methods, representing the degree of inconsistency between the measurement data. It is the rated transmission power of the associated area, which is obtained from the equipment ledger database as a standardized benchmark value. Its purpose is to eliminate the impact of differences in line capacity and ensure the universality of the indicator. It is an exponential function with the natural constant e as its base.

[0048] Step 3: Extract multi-timescale features that characterize the power grid's operating status from standardized measurement data and generate a multi-timescale feature set.

[0049] In a preferred embodiment, the step of extracting multi-timescale features characterizing the power grid operating status from standardized measurement data and generating a multi-timescale feature set includes: filtering standardized measurement data based on measurement quality indicators, filtering out data whose measurement quality indicators are higher than a preset filtering threshold, and forming a high-confidence feature subset for feature extraction.

[0050] From the high-confidence feature subset, we extract fast-changing features that characterize the instantaneous dynamics of the power grid and slow-changing features that characterize the steady-state trend of the power grid.

[0051] The fast-changing characteristics characterizing the instantaneous dynamics of the power grid include at least one of instantaneous voltage phase angle difference, frequency change rate, or short-time power fluctuation; the slow-changing characteristics characterizing the steady-state trend of the power grid include at least one of steady-state voltage amplitude, steady-state active / reactive power flow, or equipment status information.

[0052] The extracted fast-changing features are combined with slow-changing features to generate a multi-time-scale feature set.

[0053] Specifically, this method aims to effectively extract and combine features reflecting different operating time scales of the power grid from preprocessed standardized measurement data to form a comprehensive multi-time scale feature set. The engineering implementation process is as follows: Step 1: Standardized measurement data are filtered based on measurement quality indices. Data with measurement quality indices higher than a preset screening threshold are selected to form a high-confidence feature subset for feature extraction. The purpose of this process is to ensure that the input data for subsequent feature extraction has sufficiently high reliability, thereby avoiding the adverse effects of noise or low-quality data on the state estimation results. In Step 2, each standardized measurement data point is assigned a measurement quality index. At this stage, the system sets a predefined quality screening threshold. ,For example .all Measurement data points are considered high-quality data and included in a high-confidence feature subset. If the measurement data fails to meet this threshold, it is not considered, thus ensuring that only the most reliable data participates in subsequent feature extraction. This screening mechanism is like setting an entry threshold for data quality, so that only data that passes the "qualification test" can enter the analysis process.

[0054] The second step involves extracting fast-changing features characterizing the instantaneous dynamics of the power grid and slow-changing features characterizing the steady-state trend from the high-confidence feature subset. The aim of this project is to capture two core aspects of the distribution network's operating state: instantaneous, rapid changes and stable, long-term operating trends. Instantaneous, rapid changes include fault impacts or sudden load fluctuations, while stable, long-term operating trends include daily load curves.

[0055] The rapidly changing characteristics reflect the instantaneous dynamic behavior of the power grid and are typically represented by high-frequency data from the power management unit (PMU). These characteristics usually capture phenomena such as high-frequency oscillations, fault transients, rapid voltage / current drops, or sudden changes. For example, the following rapidly changing characteristics can be extracted: Instantaneous voltage phase angle difference: the voltage phase angle difference between adjacent nodes. , They represent nodes respectively. and nodes The absolute value of the voltage phase difference; abrupt changes in this value may indicate a fault or power oscillation. For example, setting a time window of 50 milliseconds, the rate of change of the instantaneous phase angle difference is calculated every 10 milliseconds. ,in Represents the difference between two adjacent phase angles The change describes the amount of change in How much did the phase angle difference between the two nodes change in this extremely short period of time? This indicates the time interval used to calculate the rate of change of phase angle.

[0056] Frequency deviation and rate of change of frequency: the actual frequency obtained from the PMU. With rated frequency deviation and the rate of change of frequency .

[0057] Instantaneous fluctuations in voltage and current: such as the root mean square value, standard deviation, or instantaneous maximum / minimum deviation of voltage and current amplitudes within a short time window. For example, calculate the deviation of the voltage amplitude from the average value of the previous period within a 100-millisecond time window.

[0058] Short-term power fluctuations: for example, the rate of change of node injected power on a second-scale. ,in This represents the total injected active power at a certain power grid node or the active power transmitted by a certain line at the current moment. Indicates the previous moment and The power value measured at the same location. This indicates the time interval used to calculate the rate of change of power.

[0059] The slow-varying characteristics reflect the steady-state and slow dynamic processes of the power grid, typically represented by low-frequency data from SCADA combined with average or trend data from the PMU. These characteristics usually capture load flow, voltage levels, and system structure changes. For example, the following slow-varying characteristics can be extracted: Steady-state voltage level: The long-term average or trend value of node voltage amplitude. For example, averaging the PMU voltage amplitude over a 1-minute or 5-minute time window yields... .

[0060] Steady-state active / reactive power flow: the average active power on a line or transformer. and reactive power For example, a 10-second rolling average based on real-time SCADA readings or PMU data.

[0061] Equipment status information, such as the open / closed state of switches and the position of transformer taps, may only change over a relatively long timescale.

[0062] Load level or load curve: The current level of the regional total load or its degree of matching with historical load curves. For example, it can be classified according to the proportion of the current day's load to the maximum load of the past 24 hours. The current level of the regional total load is divided into peak, off-peak, and average off-peak, etc.

[0063] The third step involves combining the extracted fast-changing and slow-changing features to generate a multi-time-scale feature set. The purpose of this project is to effectively integrate the extracted information representing different dynamic characteristics of the power grid, forming a comprehensive, multi-dimensional feature vector. This feature set can capture both the transient behavior and steady-state trends of the power grid, providing complete information for subsequent state estimation. For example, for a specific time point, the instantaneous rate of change of voltage phase angle and frequency deviation extracted from PMU data are concatenated with the steady-state voltage amplitude and average active power flow extracted from SCADA data or the long-term average value of PMU data to form a high-dimensional feature vector. .in This represents a rapidly changing feature vector. This represents a slowly varying eigenvector. For example, an eigenvector could include: the maximum voltage phase angle difference within the current 10 milliseconds, the current standard deviation of frequency fluctuation, the average voltage over the past minute, the average active power over the past 5 minutes, and the current substation main switch status. This combination ensures that the overall operating status of the power grid is fully characterized in both time and space.

[0064] Step 4: Based on multi-timescale feature sets and measurement quality indicators, identify the power grid operating conditions and generate labels for the current power grid operating conditions.

[0065] In a preferred embodiment, the step of identifying power grid operating conditions based on multi-timescale feature sets and measurement quality indicators, and generating current power grid operating condition labels, includes: inputting the multi-timescale feature sets extracted in real time into a pre-trained operating condition identification model.

[0066] The operating condition identification model makes predictions based on the feature-operating condition mapping relationship established by historical data, and generates prediction results.

[0067] The confidence level of the prediction result is evaluated by combining the measurement quality indicators. When the confidence level is higher than the preset confidence level threshold, the prediction result is output as the label of the current power grid operating condition.

[0068] Specifically, the purpose of this project is to identify the specific operating state of the current distribution network based on the extracted multi-timescale feature set and associated measurement quality indicators, and to generate a corresponding operating condition label. This step is crucial for the subsequent dynamic selection of the optimal state estimation strategy. A typical operating condition identification process includes, but is not limited to, the following general engineering steps: Step 1: Establish a historical operating condition database and feature-operating condition mapping relationship. The purpose of this project is to provide a reference benchmark and discrimination rules for operating condition identification. In the offline phase, a large amount of historical distribution network operation data is collected, including various types of multi-timescale features and their corresponding actual operating conditions. These various types of multi-timescale features include normal operation, peak load, off-peak load, renewable energy grid connection, distributed power source fluctuations, line faults, transformer faults, short circuits, line breaks, and distributed power source disconnection. Using data mining and machine learning techniques, such as cluster analysis and classifier training, a mapping model between the feature space and known operating condition labels is established. For example, clustering algorithms such as K-means can be used to divide historical data into different clusters based on feature similarity, with each cluster representing an operating condition. Alternatively, a supervised learning model can be used, taking multi-timescale features as input and outputting corresponding operating condition labels. During training, measurement quality indicators are comprehensively considered to screen and weight training samples; for example, higher weights are given to high-quality data.

[0069] Step 2: Real-time Feature Extraction and Data Preprocessing. The purpose of this step is to prepare real-time data that meets the input requirements of the established working condition recognition model. After the real-time multi-timescale feature set is generated, it undergoes necessary standardization or normalization to ensure consistency with the feature data format used during training. Simultaneously, this feature set also carries corresponding measurement quality indicators. During preprocessing, if the proportion of the original measurements upon which a feature depends that have quality indicators below a preset threshold exceeds a specified limit, the feature is marked as low confidence, prompting the model to reduce its weight or disregard it in decision-making.

[0070] Step 3: Apply the operating condition identification model for prediction. The purpose of this project is to determine the current operating condition category of the power grid based on real-time input using a trained model. The pre-processed real-time multi-timescale feature set is used as input and substituted into the pre-trained operating condition identification model. The model will output a classification result, i.e., a predicted label for the current power grid operating condition. For example, the model may identify specific operating condition labels such as "normal operation - medium load," "branch short-circuit fault," and "excessive output of distributed power sources." Some models may also output the probability distribution of each operating condition, for example, a 90% probability of "normal operation - medium load" and a 10% probability of "minor anomaly."

[0071] Step 4: Confidence Assessment and Output of Operating Condition Labels. This step aims to further evaluate the operating condition labels provided by the model to ensure their reliability. If the operating condition identification model provides a probability distribution, the label with the highest probability is selected as the initial result. The confidence of the operating condition label is further evaluated by combining the measurement quality index. Only when both the model's output probability and the comprehensive quality index of the feature set meet preset threshold conditions—for example, output probability > 0.95 and comprehensive quality index > 0.98—is the operating condition label rated as high confidence. Subsequently, a predefined function maps the model's output probability and the measurement quality index of the input features to a unified confidence score. Finally, the system outputs the current power grid operating condition label and its corresponding confidence score.

[0072] Step 5: Input the current power grid operating condition label and multi-time scale feature set into the dynamic estimation strategy generator to generate an adaptive fusion and estimation strategy.

[0073] In a preferred embodiment, the step of inputting the current power grid operating condition label and multi-timescale feature set into a dynamic estimation strategy generator to generate an adaptive fusion and estimation strategy includes: inputting the multi-timescale feature set into an identifier for identifying operating conditions to generate preliminary operating condition hypotheses.

[0074] The initial operating condition assumptions are input into a generator for generating a strategy, which generates at least one candidate fusion and estimation strategy.

[0075] At least one candidate fusion and estimation strategy is simulated using digital mirroring, and the simulation results are compared and verified with the data anchors obtained through screening to generate strategy verification feedback.

[0076] The policy verification feedback is sent to both the recognizer and the generator simultaneously to perform collaborative optimization and iteration of the initial working condition assumptions and candidate fusion and estimation strategies.

[0077] After the iteration meets the preset convergence condition, the optimized operating condition label is output as the current power grid operating condition label, and the corresponding optimization strategy is output as the adaptive fusion and estimation strategy. The adaptive fusion and estimation strategy includes dynamic weight coefficients for guiding data fusion and state estimation model selection instructions for specifying the calculation model.

[0078] Specifically, this method aims to adaptively generate a fusion and state estimation strategy based on the current power grid operating conditions and multi-timescale feature sets through a dynamic estimation strategy generator. This generation process is implemented through a triangular collaborative decision-making mechanism to ensure the robustness and optimality of the strategy. The engineering implementation process is as follows: Step 1: Input the multi-timescale feature set into an identifier for identifying operating conditions to generate preliminary operating condition hypotheses. The purpose of this project is to make a preliminary judgment on the current operating state of the power grid. The multi-timescale feature set generated in Step 3, which includes both rapidly changing and slowly changing features, is then input into the pre-trained operating condition identifier. In the middle. This identifier. Based on patterns learned from historical data, the system outputs one or more preliminary operating condition hypotheses. For example, the identifier might output labels such as "normal high load" or "slight frequency fluctuations." This hypothesis is a preliminary classification and may contain some ambiguity or uncertainty. For instance, the identifier might assign a certain operating condition as the primary probability while providing several other alternative operating conditions, each with its own probability score.

[0079] Step 2: Input the initial operating condition assumptions into a generator to generate at least one candidate fusion and estimation strategy. The goal of this project is to formulate potential data fusion and state estimation schemes suitable for the current operating condition based on the initial assessment. Generator Receive from the recognizer The initial operating assumptions are based on its internally stored policy knowledge base and heuristic rules. Propose at least one candidate fusion and estimation strategy , Each strategy package is numbered, and each strategy includes a specific fusion algorithm, the selected state estimation model, and parameter configurations. For example, for the "slight frequency fluctuation" condition, the generator might suggest a fusion and estimation strategy that emphasizes high-frequency PMU data and uses a Kalman filter.

[0080] The internally stored strategy knowledge base pre-stores a large number of "operating condition-strategy" mapping pairs. This means that for each typical power grid operating condition, such as "normal peak load" or "line short-circuit fault," one or more corresponding optimal fusion and estimation strategies are pre-set. Heuristic rules, on the other hand, are a set of logical judgment processes followed by the strategy generator when making decisions. Specifically, based on the real-time identified power grid operating condition labels and additional information such as measurement quality, the most suitable strategy instructions are intelligently matched, filtered, or adjusted in the knowledge base. For example, a heuristic rule might stipulate that when the operating condition is identified as "short-circuit fault," the strategy marked as "high robustness" in the knowledge base should be selected first. If the data integrity rate is also detected to be below a certain threshold, the specific parameters in the strategy regarding reducing the weight of the corresponding data source are further invoked, thereby dynamically generating the final adaptive strategy.

[0081] The third step involves simulating at least one candidate fusion and estimation strategy using a digital mirror, and comparing the simulation results with the selected data anchors to generate strategy verification feedback. The purpose of this project is to evaluate the effectiveness and accuracy of each candidate strategy through virtual simulation, avoiding the potential risks of directly trying them in the actual power grid.

[0082] A digital mirror is a precise mathematical model of a power grid, simulating its operational behavior and measurement processes. In the digital mirror, virtual calculations are performed on the current multi-timescale feature set according to the method specified by each candidate fusion and estimation strategy, yielding corresponding virtual state estimates. Simultaneously, data points with quality scores exceeding a preset threshold are selected from measurement quality indicators and used as data anchors. These data anchors are considered highly reliable real measurements.

[0083] The system calculates the degree of agreement between the virtual measurements derived from the virtual state estimation results and the data anchor points. For example, the agreement can be quantified by calculating the root mean square error (RMSE) of the two. Simultaneously, the physical plausibility of the virtual state estimation results is evaluated, such as whether the voltage amplitude is within a reasonable range and whether the power flow conforms to Kirchhoff's laws. Combining the agreement and physical plausibility, a simulation score is generated for each candidate strategy. Based on these simulation scores, the system generates strategy verification feedback, including which strategy performs best and the advantages and disadvantages of each strategy.

[0084] Step 4: Policy verification feedback is simultaneously sent to both the recognizer and the generator to collaboratively optimize and iterate the initial operating condition assumptions and candidate fusion and estimation strategies. The purpose of this project is to establish a closed-loop optimization mechanism, using continuous feedback to improve the accuracy of operating condition identification and the quality of strategy generation. Policy verification feedback will be simultaneously sent to the recognizer. and generator Recognizer It will use this feedback information to adjust its internal operating condition classification logic, enabling it to more accurately identify the current operating condition in subsequent recognitions, such as updating the parameters of its classification model or adjusting the weights of features. Generator Based on the feedback, the system will optimize its strategy generation rules to generate more effective candidate strategies that are better suited to the current working conditions. This iterative process will continue until the preset convergence condition is met.

[0085] Step 5: After the iterations satisfy the preset convergence condition, output the optimized operating condition label as the current power grid operating condition label, and output the corresponding optimized strategy as the adaptive fusion and estimation strategy. The purpose of this project is to obtain a final reliable operating condition judgment and the optimal execution strategy. The convergence condition can be defined as... In each iteration, the change in the simulation score of the optimal strategy is less than a preset small threshold. Or, it may reach the preset maximum number of iterations, for example... , The maximum number of iterations is 20. Once the convergence condition is met, the system outputs the optimal operating condition label determined during the iteration process as the current power grid operating condition label, and outputs its corresponding validated and optimized strategy as the adaptive fusion and estimation strategy. This final strategy will be used to guide the subsequent data fusion and state estimation processes.

[0086] In a further preferred embodiment, the step of simulating at least one candidate fusion and estimation strategy using digital mirroring, and comparing and verifying the simulation results with the data anchors obtained through screening to generate strategy verification feedback, includes: selecting data points whose quality scores exceed a preset score threshold from the measurement quality indicators as data anchors.

[0087] In the digital mirror, virtual computation is performed on the feature set at multiple time scales according to the method specified by each candidate fusion and estimation strategy to obtain the corresponding virtual state estimation results.

[0088] The digital mirror refers to a simulation model that includes the distribution network topology and line parameters.

[0089] The system calculates the degree of agreement between the virtual measurements derived from the virtual state estimation results and the data anchors, evaluates the physical rationality of the virtual state estimation results, and generates a simulation score for each candidate fusion and estimation strategy.

[0090] Based on simulation scoring, strategy verification feedback is generated.

[0091] Specifically, this method aims to elaborate on how to simulate and run candidate fusion and estimation strategies using digital mirroring, and then compare and verify the simulation results with data anchors to generate strategy verification feedback. This step is the core of the triangular collaborative decision-making mechanism, used to evaluate the merits of strategies in a virtual environment. Its engineering implementation process is as follows: Step 1: Select data points from the measurement quality indicators that exceed a preset scoring threshold as data anchors. The purpose of this process is to identify and select the most reliable actual measurement data as a "true" reference benchmark for evaluating simulation results. In Step 2, each measurement data point is assigned a measurement quality indicator. The system sets a predefined scoring threshold. ,For example or All data points whose measured quality indicators are greater than or equal to the scoring threshold will be filtered out and marked as data anchors. These anchor points may include voltage and current phasors measured by the PMU, as well as power and topology status measured by SCADA. These data are considered to represent the true operating state of the current power grid due to their high signal-to-noise ratio, high integrity, and good internal consistency.

[0092] The second step: In the digital mirror, virtual calculations are performed on the multi-timescale feature set according to the method specified by each candidate fusion and estimation strategy to obtain the corresponding virtual state estimation results. The purpose of this project is to use a precise mathematical model of the distribution network to simulate the execution of various candidate strategies to obtain the expected grid state after strategy execution. The digital mirror is a high-precision mathematical model of the power grid, containing all necessary static and dynamic information such as the distribution network topology, line parameters, transformer parameters, generator and load models. For each candidate fusion and estimation strategy, it specifies how to fuse multi-timescale features and which state estimation model to select, using the current multi-timescale feature set as input. In the digital mirror, data fusion is performed on the current multi-timescale feature set according to the fusion algorithm specified by each candidate fusion and estimation strategy to generate a fused feature vector. Subsequently, according to the state estimation model selected by each candidate fusion and estimation strategy, the fused feature vector is input into the model, outputting the virtual state estimation results of all node voltage phasors of the power grid. This result represents the best estimate of the power grid operating state under the assumption that the strategy is actually applied.

[0093] The third step involves calculating the degree of agreement between the virtual measurements derived from the virtual state estimation results and the data anchor points, evaluating the physical rationality of the virtual state estimation results, and generating a simulation score for each candidate fusion and estimation strategy. The aim of this project is to quantitatively evaluate the accuracy of each strategy's estimation results and its physical feasibility.

[0094] First, based on the virtual state estimation results, a set of data anchor points is calculated using the distribution network's measurement equations, such as converting voltage amplitude and phase angle into active power, reactive power, and current. Virtual measurement values ​​corresponding to location and type Then, the virtual measurement value is calculated. With data anchors The degree of agreement between them can be quantified by calculating the error norm between them, such as the root mean square error (RMSE). ,in It refers to the number of data anchors. Used to uniquely identify the candidate virtual state scheme being evaluated.

[0095] Secondly, the physical rationality of the virtual state estimation results is evaluated. This includes checking whether the estimated node voltage amplitudes are within the preset safe operating range; for example, whether all node voltage amplitudes meet the reference range. Whether the phase angle difference meets the line's transmission capacity limits, and whether the active and reactive power flow satisfies Kirchhoff's laws, etc., are all physical rationality indicators. This can be obtained by scoring the degree of violation of each physical constraint, for example... ,in This is a dimensionless adjustment coefficient used to adjust the sensitivity of the index to power deviation, with a preferred value range of 5 to 20, and 10 in this embodiment. This represents a comprehensive constraint violation value, such as the violation value for a single voltage limit. ,here It is the ratio of the actual voltage value to its selected reference value. It is only applied when the voltage exceeds... The violation degree is greater than 0 only when the range is specified. Similarly, the violation degree of line overload and other physical constraints are calculated to obtain a comprehensive constraint violation degree value. The closer this value is to 1, the better the physical rationality; the closer it is to 0, the more serious the violation of physical constraints.

[0096] Finally, the fit and physical rationality indicators Combined, a simulation score is generated for each candidate fusion and estimation strategy. A simple way to combine them is... ,in , These are weighting coefficients, representing the importance of the fit and physical plausibility indicators in the final score. They are manually set by the algorithm designer based on prior knowledge and experiments. For example, , . This is the preset maximum permissible error. Based on experience with the power system, an upper limit of error that is acceptable in engineering is set. For example, for power error, it may be set to 1% of the total system load. The smaller the error, the higher the rationality and the higher the simulation score.

[0097] Step 4: Generate policy validation feedback based on simulation scores. The purpose of this step is to summarize the evaluation results of each policy and provide them in a structured form to the upper-level decision-making mechanism for collaborative optimization. By comparing the simulation scores of all candidate policies, the system can determine the best-performing policy and the relative merits of each policy. The policy validation feedback includes the simulation scores of all candidate policies, their advantages and disadvantages, and suggestions for improvement. For example, the feedback might indicate that one policy performs well in fast transient estimation but has slightly poor physical plausibility, while another policy provides stable estimation results but lacks sensitivity to instantaneous changes. This feedback information will be sent to the recognizer and generator to guide subsequent collaborative optimization iterations.

[0098] Step 6: Based on the adaptive fusion and estimation strategy, perform data fusion on the feature sets of multiple time scales to generate fused feature vectors.

[0099] In a preferred embodiment, the step of performing data fusion on a multi-timescale feature set according to an adaptive fusion and estimation strategy to generate a fused feature vector includes: parsing the adaptive fusion and estimation strategy to obtain dynamic weight coefficients for features at different time scales in the multi-timescale feature set.

[0100] Based on the dynamic weighting coefficient, the fast-changing features and the slow-changing features are weighted and fused.

[0101] The weighted features are then reduced in dimensionality and concatenated to generate a fused feature vector.

[0102] Specifically, this method aims to intelligently fuse multi-timescale feature sets based on an adaptive fusion and estimation strategy generated by a triangular collaborative decision-making mechanism, thereby generating a fused feature vector that comprehensively and accurately reflects the current power grid operating status. Its engineering implementation process is as follows: Step 1: Analyze the adaptive fusion and estimation strategy to obtain the dynamic weight coefficients for features at different time scales within a multi-time-scale feature set. The purpose of this step is to understand the specific instructions of the current strategy, especially how to adjust the importance of different types of features during the data fusion stage. In Step 6, the triangular collaborative decision-making mechanism outputs an optimized adaptive fusion and estimation strategy. This strategy is a structured data packet that explicitly contains the configuration information for the fusion module. (System parsing) Extract a dynamic set of weight coefficients specifically designed to guide feature fusion across multiple time scales. , This represents the dynamic weighting coefficient assigned to each rapidly changing feature. This represents the dynamic weighting coefficients assigned to each slowly changing feature. These weighting coefficients are dynamically generated based on the current grid operating condition label and historical learning results, reflecting the relative importance of different fast and slow changing features to state estimation under the current operating condition. For example, under transient operating conditions, the weight assigned to fast changing features may be higher, such as 0.7, while under steady-state operating conditions, the weight assigned to slow changing features may be higher, such as 0.8.

[0103] Step 2: Based on the dynamic weighting coefficients, perform weighted fusion calculations on fast-changing and slow-changing features. The purpose of this project is to integrate features of different time scales and importance through a weighting mechanism, ensuring that their contribution to the final fused feature vector matches their actual value. Each sub-feature in the multi-time-scale feature set generated in Step 4 is multiplied by its corresponding dynamic weighting coefficient.

[0104] For rapid change characteristics The weighted fast change feature is represented as .

[0105] For slow-varying characteristics The weighted slow-varying features are represented as .

[0106] The third step involves dimensionality reduction and concatenation of the weighted features to generate a fused feature vector. The purpose of this step is to further optimize the fused feature representation, eliminate potential redundant information, and provide a concise and comprehensive input to the final state estimation model through a unified output format.

[0107] First, the weighted fast change features and slow variation characteristics Dimensionality reduction can be performed independently or in combination. For example, techniques such as principal component analysis (PCA) or autoencoders can be used to eliminate redundancy among features and extract the most representative information. Let the dimensionality-reduced fast-changing features be... The slow-varying characteristics after dimensionality reduction are The dimensionality reduction dimension is usually set based on experience or obtained through cross-validation; for example, reducing the original 100-dimensional features to 30 dimensions.

[0108] Finally, the fast-changing features after dimensionality reduction are... and slow variation characteristics The features are concatenated according to a predefined method. The most common concatenation method is to simply link them end-to-end to form a unified fused feature vector. For example, if It is a 15-dimensional vector. If it is a 15-dimensional vector, then It will be a 30-dimensional vector. This fused feature vector It condenses key operational information of the power grid at different time scales and has been optimally weighted and optimized according to the current operating conditions, so it can be directly used as input for the state estimation model.

[0109] Step 7: Based on the adaptive fusion and estimation strategy, select a corresponding state estimation model and calculate the fusion feature vector to generate the distribution network state estimation result.

[0110] In a preferred embodiment, the step of selecting a corresponding state estimation model and calculating the fusion feature vector to generate a distribution network state estimation result according to the adaptive fusion and estimation strategy includes: parsing the adaptive fusion and estimation strategy and selecting a state estimation model that matches the current power grid operating condition label from a model library containing multiple models.

[0111] The fused feature vector is input into the selected state estimation model for computation.

[0112] The solution results of the output model are used as the state estimation results of the distribution network.

[0113] Specifically, this embodiment details the engineering steps involved in selecting the corresponding state estimation model and calculating the fused feature vector based on an adaptive fusion and estimation strategy, ultimately generating the distribution network state estimation result. The engineering objective is to intelligently match the most suitable estimation model according to the specific operating conditions of the current power grid, thereby maximizing the effectiveness of the fused feature vector and ensuring high-precision and robust distribution network state estimation under various complex scenarios. The engineering implementation process is as follows: The first step is model selection. The system receives the adaptive fusion and estimation strategy output from the previous step. This strategy explicitly specifies the type of state estimation model and parameter configuration to be used under the current operating condition. The model management module in the system parses this strategy and selects a corresponding model instance from a pre-set model library. This model library is a collection containing various state estimation algorithms and their parameter templates, including at least the following types of models: 1. Models suitable for topology changes: For example, weighted least squares based on the latest topology information can quickly reflect the switching of circuit breakers and switches, and update the mathematical model describing the connection relationship of power grid lines in real time. Such models will be preferred when the "Current Power Grid Operating Condition Label" indicates "Branch Switching" or "Network Reconfiguration".

[0114] 2. Robust models suitable for missing data: These include advanced algorithms that can automatically reduce the impact of outliers. These models maintain good estimation performance even with significant data loss or bad data and are less susceptible to individual outliers. Such models are selected when the "Current Power Grid Operating Condition Label" indicates "Missing Measurements" or "Partial Communication Failure."

[0115] 3. Filtering models suitable for high-noise scenarios: such as Extended Kalman Filter (EKF) or Unscented Kalman Filter (UKF). These models can combine system dynamic characteristics to smooth and predict measurement data containing Gaussian or non-Gaussian noise, improving the accuracy of estimation. Such models are selected when the "Current Power Grid Operating Condition Label" indicates "High Noise Interference" or "Severe Measurement Fluctuations".

[0116] In addition, the model library may contain other models, such as estimation models for specific load characteristics or distributed power integration.

[0117] After selecting the appropriate state estimation model, the system uses the "fused feature vector" generated in the previous step as input to drive the model's calculations. The model's calculation process involves using its internal mathematical algorithm, along with the input feature vector and the current power grid topology information, to iteratively solve for an optimal state variable. For common state estimation models, such as weighted least squares, the core is to solve the following optimization problem: ;

[0118] in, The state vector to be estimated is usually defined as a vector consisting of the voltage magnitude and phase angle of all nodes. It is a measurement vector transformed or mapped from the fused feature vector, i.e., the model input. It is a comprehensive set of measurement values ​​containing information from multiple time scales, obtained after feature fusion processing. This represents the measurement residual vector, where each element represents the difference or error between the measured value calculated by the first-order theory and the actual measured value. This indicates that the vector is transposed. The measurement function describes the state vector. The nonlinear relationship between the measured value and the power system is an inherent mathematical model uniquely determined by the physical laws of the power system and the network topology. It is a weighted matrix, composed of the weights of each measurement value. The model uses a numerical iterative algorithm to find the state vector that minimizes the objective function. .

[0119] Finally, the model outputs the solved state vector as the distribution network state estimation result. This result typically includes at least the voltage magnitude and phase of each node in the distribution network. Based on these fundamental state variables, the system can further calculate derived results such as active power and reactive power of each branch, providing real-time and accurate data support for power grid monitoring, optimized scheduling, and fault analysis.

[0120] In a further preferred embodiment, the method further includes a backtracking learning step, comprising: associating and storing the current power grid operating condition label, adaptive fusion and estimation strategy, fusion feature vector and distribution network state estimation result to generate historical scenario strategy pairs.

[0121] By utilizing historical scenario strategies to update the parameters in a feature selection rule base and a triangular collaborative decision-making mechanism, the subsequent feature selection and collaborative optimization iteration process can be optimized.

[0122] Specifically, this method aims to introduce a backtracking learning step. Its engineering objective is to continuously accumulate and analyze historical operational data and its corresponding strategy selection and estimation results, thereby continuously optimizing the parameters of the feature selection rules and the triangular collaborative decision-making mechanism, ultimately improving the adaptability and accuracy of the entire state estimation system. This parameter update process can be implemented based on a reinforcement learning model, whose core elements are defined as follows: a state space, consisting of the current condition label and feature vectors; an action space, a set of selectable fusion and estimation strategies; and a reward function, set according to the residual size of the state estimation results, with smaller residuals resulting in higher rewards. Through this model learning, the system can autonomously optimize which strategy to choose in different states to obtain the highest reward, and use the learned parameters to update the rule base and decision-making mechanism. The specific engineering implementation process is as follows: Step 1: Associate and store the current power grid operating condition label, adaptive fusion and estimation strategy, fusion feature vector, and distribution network state estimation results to generate historical scenario strategy pairs. The purpose of this project is to establish a traceable historical knowledge base that records the decisions made and results generated by the system under specific operating conditions. Once a complete state estimation process is completed, the system will obtain the current power grid operating condition label. Adaptive fusion and estimation strategies generated for this working condition fusion feature vectors used for performing state estimation and the final distribution network state estimation results This information, along with corresponding timestamps, system configuration versions, and other metadata, will be linked as a complete "historical scenario-policy pair" and stored in a persistent database. For example, the storage structure may include: record ID, timestamp, operating condition label, policy ID, policy parameters, hash values ​​or key statistics of feature vectors, a summary of the estimation results (e.g., voltage amplitude statistics, power balance deviation), and an indicator indicating whether the estimation was successful or failed. This knowledge base will grow over time, providing a rich dataset for subsequent learning and optimization.

[0123] The second step involves updating the parameters in a feature selection rule base and a triangular collaborative decision-making mechanism using historical scenario strategies to optimize subsequent feature selection and collaborative optimization iterations. The aim of this project is to automatically improve the system's core algorithms and decision-making logic by learning from historical data.

[0124] Update the feature selection rule base: The feature selection rule base determines which fast-changing and slow-changing features are extracted from the standardized measurement data in step 4. By analyzing the performance of different feature vectors in historical scenario strategy pairs under different operating conditions, especially when the calculation residual of the state estimation results exceeds a preset threshold, the system can analyze in reverse which feature selections or insufficient importance caused the problem. For example, if a specific type of voltage transient feature is always strongly correlated with the estimation result under a certain type of fault condition, then the feature selection rules can be adjusted to increase the priority of this type of feature or increase its weight. Specifically, the update method can be to adjust the parameters of the feature importance evaluation model or update the rule-based selection threshold. This process can employ machine learning methods, such as reinforcement learning, to reward accurate estimation results and penalize inaccurate estimations.

[0125] Update the parameters in the triangular collaborative decision-making mechanism: The triangular collaborative decision-making mechanism includes a working condition identifier, a policy generator, and a comparison and verification logic for digital mirrors. Through backtracking learning, the parameters of these components can be optimized in a targeted manner.

[0126] Parameter updates for the identifier: Adjust the operating condition identifier based on the accuracy recognition rate of operating condition tags in historical data. The model parameters, such as the classifier weights and decision boundaries, are adjusted to make it more accurate in subsequent working condition identification.

[0127] Generator parameter updates: if certain strategies If the policy generator frequently produces unsatisfactory estimation results, then... This will reduce the probability of generating such strategies or modify their generation rules. Conversely, strategies that consistently perform well will have their generation priority increased.

[0128] Parameter updates for the digital mirror verification mechanism: This includes adjusting the scoring thresholds for data anchors. For example, if it is found Being too conservative results in sparse anchor points, which can be appropriately reduced; or the simulation score can be adjusted. Medium consistency and physical rationality indicators Weighting coefficients , This makes it more reflective of the preferences for accuracy and robustness in practical applications. For example, if physical plausibility becomes more important in practice, it could be increased... Furthermore, the convergence criteria can be adaptively adjusted based on historical performance to balance efficiency and accuracy. This update process is typically performed in an offline, periodic training task to avoid excessive consumption of real-time computing resources and to ensure the stability of the updates.

[0129] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, and all such modifications and additions should fall within the protection scope of the present invention.

Claims

1. A multi-time-scale state estimation method for distribution networks based on PMU, characterized in that, include: Step 1: Acquire the high-frequency data stream from the synchronous phasor measurement unit (PMU) and the low-frequency data stream from the monitoring and data acquisition system (SCADA) in the distribution network to generate multi-source measurement data; Step 2: Perform quality assessment and preprocessing on multi-source measurement data to generate standardized measurement data and corresponding measurement quality indicators; Step 3: Extract multi-timescale features characterizing the power grid's operating status from standardized measurement data to generate a multi-timescale feature set; Step 4: Based on multi-timescale feature sets and measurement quality indicators, identify the power grid operating conditions and generate labels for the current power grid operating conditions; Step 5: Input the current power grid operating condition label and multi-time scale feature set into the dynamic estimation strategy generator to generate an adaptive fusion and estimation strategy; Step 6: Based on the adaptive fusion and estimation strategy, perform data fusion on the feature sets of multiple time scales to generate fused feature vectors; Step 7: Based on the adaptive fusion and estimation strategy, select a corresponding state estimation model and calculate the fusion feature vector to generate the distribution network state estimation result.

2. The method for multi-time-scale state estimation of distribution networks based on PMU according to claim 1, characterized in that, The process of quality assessment and preprocessing of multi-source measurement data to generate standardized measurement data and corresponding measurement quality indicators includes: Time synchronization and alignment are performed between the high-frequency data stream of the PMU and the low-frequency data stream of the SCADA system. Calculate the signal-to-noise ratio, data integrity rate, and internal consistency index of PMU data and SCADA data within a specific time window; Based on the signal-to-noise ratio, data integrity rate, and internal consistency index, the measurement quality index of each data point is calculated through a scoring model. Based on measurement quality indicators, the raw data is cleaned, corrected, and normalized to generate standardized measurement data.

3. The method for multi-time-scale state estimation of distribution networks based on PMU according to claim 2, characterized in that, The internal consistency index is obtained through the following method: Calculate the power difference in the associated region using the voltage phase information in the PMU data stream; Compare the power difference with the power change in the same area obtained from the SCADA data stream for consistency. Based on the results of the consistency comparison, an internal consistency index is generated quantitatively.

4. The method for multi-time-scale state estimation of distribution networks based on PMU according to claim 1, characterized in that, The process of extracting multi-timescale features characterizing the power grid's operating state from standardized measurement data to generate a multi-timescale feature set includes: Standardized measurement data are filtered based on measurement quality indicators, and data with measurement quality indicators higher than the preset screening threshold are selected to form a high-confidence feature subset for feature extraction. From the high-confidence feature subset, we extract fast-changing features that characterize the instantaneous dynamics of the power grid and slow-changing features that characterize the steady-state trend of the power grid. The fast-changing characteristics characterizing the instantaneous dynamics of the power grid include at least one of instantaneous voltage phase angle difference, frequency change rate, or short-time power fluctuation; the slow-changing characteristics characterizing the steady-state trend of the power grid include at least one of steady-state voltage amplitude, steady-state active / reactive power flow, or equipment status information. The extracted fast-changing features are combined with slow-changing features to generate a multi-time-scale feature set.

5. The method for multi-time-scale state estimation of distribution networks based on PMU according to claim 1, characterized in that, The process of identifying power grid operating conditions based on multi-timescale feature sets and measurement quality indicators, and generating current power grid operating condition labels, includes: The multi-timescale feature set extracted in real time is input into a pre-trained working condition recognition model; The operating condition identification model makes predictions based on the feature-operating condition mapping relationship established by historical data, and generates prediction results. The confidence level of the prediction result is evaluated by combining the measurement quality indicators. When the confidence level is higher than the preset confidence level threshold, the prediction result is output as the label of the current power grid operating condition.

6. The method for multi-time-scale state estimation of distribution networks based on PMU according to claim 1, characterized in that, The step of inputting the current power grid operating condition label and multi-timescale feature set into the dynamic estimation strategy generator to generate an adaptive fusion and estimation strategy includes: The multi-timescale feature set is input into an identifier for identifying operating conditions to generate preliminary operating condition hypotheses; The initial operating condition assumptions are input into a generator for generating a strategy, which generates at least one candidate fusion and estimation strategy; At least one candidate fusion and estimation strategy is simulated using digital mirroring, and the simulation results are compared and verified with the data anchors obtained through screening to generate strategy verification feedback. The policy verification feedback is sent to both the recognizer and the generator to perform collaborative optimization and iteration of the preliminary working condition assumptions, candidate fusion and estimation strategies. After the iteration meets the preset convergence condition, the optimized operating condition label is output as the current power grid operating condition label, and the corresponding optimization strategy is output as the adaptive fusion and estimation strategy. The adaptive fusion and estimation strategy includes dynamic weight coefficients for guiding data fusion and state estimation model selection instructions for specifying the calculation model.

7. The method for multi-time-scale state estimation of distribution networks based on PMU according to claim 6, characterized in that, The process involves simulating at least one candidate fusion and estimation strategy using digital mirroring, and comparing the simulation results with the data anchors obtained through screening to generate strategy verification feedback, including: Data points whose quality scores exceed a preset score threshold are selected from the measured quality indicators and used as data anchors; In the digital mirror, based on the method specified by each candidate fusion and estimation strategy, virtual calculations are performed on the feature set at multiple time scales to obtain the corresponding virtual state estimation results; The digital mirror refers to a simulation model that includes the distribution network topology and line parameters; Calculate the degree of agreement between the virtual measurements derived from the virtual state estimation results and the data anchor points, evaluate the physical rationality of the virtual state estimation results, and generate a simulation score for each candidate fusion and estimation strategy. Based on simulation scoring, strategy verification feedback is generated.

8. The method for multi-time-scale state estimation of distribution networks based on PMU according to claim 4, characterized in that, The step of performing data fusion on multi-timescale feature sets according to an adaptive fusion and estimation strategy to generate a fused feature vector includes: We analyze the adaptive fusion and estimation strategy to obtain dynamic weight coefficients for features at different time scales in a multi-time scale feature set. Based on the dynamic weighting coefficients, a weighted fusion calculation is performed on the fast-changing features and the slow-changing features; The weighted features are then reduced in dimensionality and concatenated to generate a fused feature vector.

9. The method for multi-time-scale state estimation of distribution networks based on PMU according to claim 1, characterized in that, The step of selecting a corresponding state estimation model and calculating the fused feature vector according to the adaptive fusion and estimation strategy to generate the distribution network state estimation result includes: The adaptive fusion and estimation strategy is analyzed, and a state estimation model that matches the current power grid operating condition label is selected from a model library containing multiple models. The fused feature vector is input into the selected state estimation model for calculation; The solution results of the output model are used as the state estimation results of the distribution network.

10. The method for multi-time-scale state estimation of distribution networks based on PMU according to claim 1, characterized in that, The method also includes a backtracking learning step, comprising: The current power grid operating condition label, adaptive fusion and estimation strategy, fusion feature vector and distribution network state estimation results are associated and stored to generate historical scenario strategy pairs; By utilizing historical scenario strategies to update the parameters in a feature selection rule base and a triangular collaborative decision-making mechanism, the subsequent feature selection and collaborative optimization iteration process can be optimized.

Images