Abnormal state evaluation device for wide-area measurement data dynamic threshold optimization

Abstract

The invention relates to the technical field of electric power system wide area measurement and operation state evaluation, in particular to an abnormal state evaluation device for wide area measurement data dynamic threshold optimization, which comprises a data acquisition and preprocessing module, a feature engineering module, a dynamic threshold optimization module, an abnormal evaluation module and a decision output module, the feature engineering module is connected with the data acquisition and preprocessing module, the dynamic threshold optimization module is connected with the feature engineering module, and the anomaly evaluation module is connected with the dynamic threshold optimization module and the data acquisition and preprocessing module. The technical scheme of'full-process intelligentization and full-scene adaptation 'is provided for power grid abnormity monitoring and dispatching from the six dimensions of'data reliability, feature accuracy, threshold adaptability, evaluation objectivity, decision landing and operation and maintenance economy', the safety, stability and operation and maintenance efficiency of power grid operation are remarkably improved, and the method has extremely high engineering application value.

Description

Technical Field

[0001] This invention relates to the field of wide-area measurement and operation status assessment technology for power systems, specifically to an abnormal state assessment device for dynamic threshold optimization of wide-area measurement data. Background Technology

[0002] Against the backdrop of the large-scale and increasingly complex development of power systems, Wide Area Measurement Systems (WAMS) have become a core means of real-time power grid monitoring due to the high sampling rate (typically 50Hz) of Pphasor Measurement Units (PMUs). However, existing anomaly assessment technologies suffer from the following key problems: Poor threshold adaptability: Traditional assessment methods use fixed thresholds (such as a fixed percentage based on rated values), which cannot match dynamic operating conditions such as power grid load fluctuations, weather changes, and equipment aging, easily leading to false alarms or missed alarms; Insufficient utilization of multi-source data: Relying solely on PMU electrical measurement data, without effectively integrating auxiliary data such as meteorological (temperature, wind speed), equipment status (transformer oil temperature), and power grid topology, and anomalies caused by sensor failures are difficult to self-diagnose, affecting the assessment basis; Low accuracy of feature selection: Measurement data contains multiple components such as trends, seasons, periods, and randomness, and existing technologies do not specifically select key features strongly correlated with anomalies, and random noise interference leads to poor robustness in anomaly identification. Summary of the Invention

[0003] The purpose of this invention is to provide an abnormal state assessment device for dynamic threshold optimization of wide-area measurement data, so as to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] An anomaly assessment device for dynamic threshold optimization of wide-area measurement data includes:

[0006] The data acquisition and preprocessing module is used to acquire real-time measurement data and multi-source auxiliary data from the wide-area measurement system, perform data cleaning, spatiotemporal correlation processing and sensor fault self-diagnosis, and remove abnormal measurement values.

[0007] The feature engineering module, connected to the data acquisition and preprocessing module, is used to perform time-series decomposition and feature filtering on the preprocessed data.

[0008] The dynamic threshold optimization module, connected to the feature engineering module, includes a reinforcement learning unit and a threshold matrix generation unit. The reinforcement learning unit adopts a hierarchical decoupled training strategy, with the upper layer being the global operating condition layer and the lower layer being the local equipment layer. The policy gradient flow responds to real-time operating condition changes, while the value gradient flow optimizes long-term stability. The threshold matrix generation unit constructs a dynamic threshold matrix based on the power grid topology correlation, and the matrix element values ​​are obtained through the output transformation of the reinforcement learning unit.

[0009] The anomaly assessment module is connected to the dynamic threshold optimization module and the data acquisition and preprocessing module, respectively, and is used to detect and classify abnormal states.

[0010] The decision output module, connected to the anomaly assessment module, is used to generate anomaly status alarm information and decision suggestions.

[0011] Preferably, the real-time measurement data includes voltage amplitude, current value, and frequency collected by the PMU; the multi-source auxiliary data includes meteorological data, equipment status information, and power grid topology data.

[0012] The specific steps of the data acquisition and preprocessing module include:

[0013] Multi-source data acquisition: Electrical parameters of power grid nodes are acquired at a frequency of 50Hz through the PMU data acquisition subunit to form a real-time measurement dataset; meteorological data, equipment status data and power grid topology data are acquired through the multi-source data interface subunit. The meteorological data includes temperature and wind speed, the equipment status data includes node transformer oil temperature, and the power grid topology data represents the connection relationship between nodes in matrix form.

[0014] Data cleaning: Missing data is filled using linear interpolation. The formula for handling missing values ​​is:

[0015] ,

[0016] in, To be filled Time Node The measured value; , For missing moments The moment before The next moment The known measured values; For calculation arrive The rate of change per unit time at any given moment; for arrive The time interval between moments;

[0017] Outlier handling: Extreme noise is removed using the 3σ rule. If it is, then it is marked as an outlier. and replace with .

[0018] Preferably, the sensor fault self-diagnosis of the data acquisition and preprocessing module includes the following steps:

[0019] Adjacent node data consistency verification: compute node With neighboring nodes Measurement deviation: ,

[0020] In the formula, for Time Node The measured value, for Time Node The measured value;

[0021] like Then, further combining historical data , Identify the faulty sensor; if and Then remove and replace with ,

[0022] In the formula, , For missing moments The moment before The next moment The known measured values, This is the standard deviation of the measurement.

[0023] Preferably, the feature engineering module uses the STL time series decomposition method to decompose the measurement data into trend components, seasonal components, periodic components and random components, and removes the random components; it calculates the contribution of trend components, seasonal components and periodic components through an improved Q learning algorithm, and dynamically selects key features.

[0024] The improved Q-learning algorithm of the feature engineering module specifically includes:

[0025] The state space is defined as trend component, seasonal component, and periodic component, and the action space retains the corresponding feature components.

[0026] Define state space Action space ;

[0027] improve Value update formula:

[0028] ,

[0029] in: For state Next action Value estimation; To preserve the original Historical information weight of the value The learning rate; For tradition The temporal difference update term of the learning process. For the reward function; The learning rate; Discount factor; For subsequent status The largest below value.

[0030] Preferably, the reinforcement learning unit of the dynamic threshold optimization module specifically includes:

[0031] upper layer:

[0032] state ,

[0033] in, Total power of the power grid; Action ;

[0034] The reward function formula is:

[0035] ,

[0036] in, For real-time measurements, The current threshold, To provide real-time response weights, The deviation between real-time measurement and threshold;

[0037] Long-term stable portion:

[0038] ,

[0039] in, For a period of time Total deviation within, For long-term stable weights;

[0040] Lower layer:

[0041] state ,action .

[0042] Preferably, the policy gradient flow of the reinforcement learning unit adopts the REINFORCE algorithm, and the gradient formula is:

[0043] ,

[0044] in, For policy gradient flow, For policy functions; The logarithmic gradient of the policy is used to adjust the policy parameters. ; It is the state-action value function.

[0045] Preferably, the long-term stability of the value gradient flow optimization is achieved by updating the state value function using TD error, including:

[0046] ,

[0047] in, The state value function; For instant rewards; For the next state Value estimation; Discount factor; The learning rate; For target value.

[0048] Preferably, the generation of the dynamic threshold matrix specifically includes:

[0049] Based on topology data Build threshold matrix ,element Represents a node and Voltage threshold during connection;

[0050] Matrix element mapping formula:

[0051]

[0052] in, Matrix element mapping, For nodes Rated voltage, proportionality coefficient It is the state-action value function.

[0053] Preferably, the anomaly assessment module constructs an element-level comparison mechanism between the real-time operating condition matrix and the dynamic threshold matrix, calculates the deviation matrix, calculates the anomaly probability by combining the KNN density peak detection algorithm, and outputs the anomaly severity level based on the anomaly probability.

[0054] Compared with the prior art, the beneficial effects of the present invention are:

[0055] This invention uses PMU acquisition and multi-source data fusion to unify meteorological and equipment status data to a 50Hz time step, and establishes a node spatial association mapping table based on the power grid topology to achieve time synchronization and spatial association of multi-source data fusion, ensuring data dimensional integrity and spatiotemporal consistency, and providing an unbiased basis for subsequent anomaly identification.

[0056] This invention uses STL decomposition to accurately separate four types of components: long-term load variation, 24-hour cycle, equipment temperature rise cycle, and sensor interference. After removing random components, the signal-to-noise ratio of abnormal characteristic signals is improved, preventing noise from masking the true anomaly.

[0057] This invention improves the Q-learning algorithm by introducing "feature sensitivity weights," making Q-value updates more inclined towards "features that are inherently sensitive to anomalies." It quantifies feature importance through a feature contribution formula, retaining only components with a contribution exceeding 0.3. Compared to traditional experience-based screening, this improves the correlation between key features and anomalies and reduces the interference of "redundant features" on subsequent threshold optimization.

[0058] This invention employs an upper layer (global operating condition layer) that uses the total power of the power grid as the state and dynamically adjusts the global threshold to respond to load increases or decreases; a lower layer (local equipment layer) uses equipment oil temperature and periodic components as the state and fine-tunes the threshold by adjusting the points; and then optimizes the real-time response through the strategy gradient flow (REINFORCE algorithm) + ensures long-term stability through the value gradient flow. The real-time performance and stability of the threshold adjustment are significantly better than traditional methods.

[0059] This invention constructs an N×N threshold matrix based on the power grid topology to represent the voltage threshold when node ij is connected. By mapping the reinforcement learning output with the node's rated voltage and connection relationship, the threshold of nodes with different connection relationships and different voltage levels is differentiated, and the scenario adaptability covers a variety of power grid structures.

[0060] This invention eliminates the influence of voltage levels by normalizing the deviation matrix and sets clear grading thresholds to avoid errors in subjective human judgment. It can be directly applied to regional or provincial power grids across voltage levels, significantly improving the versatility of the assessment.

[0061] This invention eliminates the need for human intervention throughout the entire process, from data preprocessing (sensor self-diagnosis), feature selection (improved Q-learning), threshold optimization (hierarchical reinforcement learning), to decision recommendations (knowledge graph); thus avoiding power grid risks caused by human error and indirectly reducing losses from maintenance accidents.

[0062] This invention achieves synchronous expansion of the dynamic threshold matrix with the number of nodes (the N×N matrix automatically adapts to newly added nodes); it improves Q-learning and hierarchical reinforcement learning to be adaptive, eliminating the need to retrain the model due to the increase in nodes; and its scalability covers "regional power grid - provincial power grid - cross-regional power grid", providing technical support for the large-scale development of power grids.

[0063] This invention not only solves the core defects of traditional anomaly assessment technology, but also provides a "full-process intelligent and full-scenario adaptable" technical solution for power grid anomaly monitoring and dispatch from six dimensions: "data reliability, feature accuracy, threshold adaptability, assessment objectivity, decision implementation, and operation and maintenance economy". It significantly improves the safety, stability and operation and maintenance efficiency of power grid operation and has extremely high engineering application value. Attached Figure Description

[0064] Figure 1 This is a schematic diagram of the structure of the present invention;

[0065] Figure 2 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0066] 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.

[0067] like Figure 1 As shown, the anomaly assessment device for dynamic threshold optimization of wide-area measurement data includes:

[0068] The data acquisition and preprocessing module is used to acquire real-time measurement data and multi-source auxiliary data from the wide-area measurement system, perform data cleaning, spatiotemporal correlation processing, and sensor fault self-diagnosis, and remove abnormal measurement values. Real-time measurement data includes voltage amplitude, current value, and frequency acquired by the PMU; multi-source auxiliary data includes meteorological data, equipment status information, and power grid topology data.

[0069] The feature engineering module, connected to the data acquisition and preprocessing module, is used to perform time-series decomposition and feature selection on the preprocessed data. The STL time series decomposition method is used to decompose the measurement data into trend components, seasonal components, periodic components and random components, and the random components are removed. The contribution of trend components, seasonal components and periodic components is calculated through an improved Q-learning algorithm, and key features are dynamically selected.

[0070] The dynamic threshold optimization module, connected to the feature engineering module, includes a reinforcement learning unit and a threshold matrix generation unit. The reinforcement learning unit adopts a hierarchical decoupled training strategy, with the upper layer being the global operating condition layer and the lower layer being the local equipment layer. The policy gradient flow responds to real-time operating condition changes, while the value gradient flow optimizes long-term stability. The threshold matrix generation unit constructs a dynamic threshold matrix based on the power grid topology correlation, and the matrix element values ​​are obtained through the output transformation of the reinforcement learning unit.

[0071] The anomaly assessment module is connected to the dynamic threshold optimization module and the data acquisition and preprocessing module, respectively, to realize the detection and classification of abnormal states; it constructs an element-level comparison mechanism between the real-time operating condition matrix and the dynamic threshold matrix, calculates the deviation matrix, calculates the anomaly probability by combining the KNN density peak detection algorithm, and outputs the anomaly severity level according to the anomaly probability;

[0072] The decision output module, connected to the anomaly assessment module, is used to generate anomaly status alarm information and decision suggestions; it implements the linkage response of the early warning strategy based on knowledge graph modeling technology.

[0073] like Figure 2 As shown, the evaluation method for an anomaly assessment device with dynamic threshold optimization for wide-area measurement data includes the following steps:

[0074] Step S1, Data Acquisition and Preprocessing: Acquire high-quality, spatiotemporally aligned multi-source data and remove sensor fault data.

[0075] Step S11, Multi-source data acquisition:

[0076] Electrical parameters of the power grid nodes are collected at a frequency of 50Hz using the PMU data acquisition unit and recorded as the real-time measurement dataset. ;

[0077] in, For time steps, Total duration Number the nodes. The total number of nodes. for Voltage amplitude at time node i;

[0078] Through the multi-source data interface unit Frequent meteorological data collection ;in, For temperature, Wind speed;

[0079] Device status data ;in, For nodes Transformer oil temperature;

[0080] Topology data ;in, Represents a node and Direct connection, This indicates no connection.

[0081] Step S12, Data Cleaning:

[0082] Missing value handling: Missing data is filled using linear interpolation. The formula for handling missing values ​​is:

[0083]

[0084] in, To be filled Time, Node The measured value;

[0085] , For missing moments The moment before The next moment The known measured values;

[0086] To calculate The rate of change of a measurement per unit time at a given moment, i.e., the average rate of change of the measured value;

[0087] for arrive The time interval between moments;

[0088] Outlier handling: The 3σ rule (based on the statistical properties of the normal distribution) is used to remove extreme noise. If it is, then it is marked as an outlier. and replace with ;

[0089] in, This represents the average value of the voltage measurement data.

[0090] This represents the standard deviation of the voltage measurement data.

[0091] Step S13, Spatiotemporal correlation processing:

[0092] Time alignment: Meteorological and equipment status data are resampled and interpolated to 50Hz to match the time step of PMU data;

[0093] Spatial Association: Based on Topological Data Establish nodes With neighboring nodes ( The associated mapping table of ).

[0094] Step S14, Sensor fault self-diagnosis:

[0095] Adjacent node data consistency verification: compute node With neighboring nodes Measurement deviation:

[0096] like (This is a consistency threshold, set based on line impedance, typically 0.05kV), which is then further combined with historical data. , Identifying faulty sensors:

[0097] like and Then remove and replace with .

[0098] Step S2, Time Series Decomposition and Key Feature Screening: Separate the multi-component features of the measurement data and screen key features that are sensitive to anomalies.

[0099] Step S21, STL timing decomposition:

[0100] Preprocessed voltage data The STL decomposition formula is as follows:

[0101]

[0102] in:

[0103] The trend component reflects long-term changes in grid load, such as intraday load growth.

[0104] Seasonal components reflect periodic fluctuations, such as the 24-hour load cycle;

[0105] This is a periodic component that reflects the inherent cycle of the equipment, such as the temperature rise cycle of a transformer.

[0106] These are random components that reflect noise, such as sensor interference.

[0107] Decomposition parameters: The time window is set to 24 hours, corresponding to the seasonal component period, and the LOESS smoothing coefficient is set to 0.75.

[0108] Step S22, Improve the calculation of Q-learning feature contribution:

[0109] Define state space (Remove random components) Action space ( Indicates "reserve the first" (each characteristic component)

[0110] improve Value update formula (introducing feature sensitivity weights) , For components The correlation coefficient with historical outlier data is an improvement on traditional Q-learning, used to dynamically evaluate "feature components". The value of anomaly detection:

[0111]

[0112] in:

[0113] For state (characteristic components) ) Execute the action Value estimation (retaining this characteristic component);

[0114] To preserve the original Historical information weight of the value (Learning rate, set to 0.1) Controls the degree of influence of new information;

[0115] For tradition The learned "temporal difference update term" includes:

[0116] : Reward function (if components) It can identify historical anomalies. (Positive reward), otherwise (Negative rewards) guide the algorithm to focus on "features that are sensitive to anomalies";

[0117] The learning rate;

[0118] As a discount factor, it emphasizes "long-term returns" and allows the algorithm to consider the cumulative value of subsequent states;

[0119] For subsequent status The largest below Value represents "the optimal value that can be obtained in the future";

[0120] For feature sensitivity weights, feature components The correlation coefficient with historical outlier data reflects the "inherent sensitivity of this feature to outliers"; [The text then abruptly shifts to a seemingly unrelated topic:] ...introducing... back, Value updates will be more inclined to "features that are inherently related to anomalies", thus improving the targeting of feature selection.

[0121] Formula for calculating feature contribution:

[0122] Used to quantize "characteristic components" "At the node" The relative importance of anomaly identification;

[0123] Among them, molecules The "value estimate" of this feature component (obtained by improved Q-learning; the larger the value, the more effectively the feature can identify anomalies);

[0124] denominator : The "sum of value estimates" for all candidate feature components (such as trend, seasonal, and cyclical components);

[0125] final This is the "normalized contribution" of the feature component; a larger value indicates that the feature contributes more to the node. The more critical the anomaly detection, the better.

[0126] Key feature selection logic:

[0127] like ( ), empirical threshold), then retain the component. As a node Key features (such as) and (This is usually preserved). For example, the trend component of power grid measurement data. (Reflecting long-term load changes), seasonal component (Reflecting periodic fluctuations such as 24-hour intervals), it is usually more closely associated with anomalies and its contribution is likely to exceed the threshold. Therefore, it will be selected as a key feature for subsequent dynamic threshold optimization and anomaly assessment.

[0128] Step S3, Dynamic Threshold Optimization: Based on real-time operating conditions, dynamically adjust the thresholds and construct a threshold matrix associated with the power grid topology.

[0129] Step S31, Layered Reinforcement Learning Training:

[0130] Upper layer (global operating condition layer):

[0131] state ( (Total power of the power grid);

[0132] action (Global threshold adjustment, range ±0.02kV);

[0133] Reward function (balancing real-time response and long-term stability):

[0134]

[0135] in, These are real-time measurement values; This is the current threshold. Weights for real-time response; It is the "deviation between real-time measurement and threshold". After taking the negative sign, the smaller the deviation, the greater the reward for this part.

[0136] Long-term stable portion:

[0137]

[0138] in, For a period of time The total deviation within the range is averaged and then passed through an exponential function. "Compression range; the smaller the total deviation, the closer the exponential result is to 1, and the greater the reward for that part;"

[0139] To ensure the long-term stability of the power grid, the threshold is adjusted frequently due to short-term fluctuations.

[0140] Lower layer (local device layer):

[0141] state (Equipment oil temperature and cycle component);

[0142] action (node and (Associated local threshold adjustment amount).

[0143] Policy gradient flow (optimizing real-time response):

[0144] The REINFORCE algorithm is used to optimize "real-time action selection," and its gradient formula is:

[0145]

[0146] in, The policy function represents "in state Take action below The probability (e.g., the magnitude of adjusting the local threshold);

[0147] The logarithmic gradient of the policy is used to adjust the policy parameters. This increases the probability of "better actions" (actions that bring high value) being selected;

[0148] The state-action value function measures "in state..." Next action "Total future revenue";

[0149] The expected value (average) of the "state-action sample" is updated through gradient. The strategy will gradually tend to select the "more valuable action", so as to achieve real-time and accurate adjustment of local thresholds (such as quickly adapting the threshold when the equipment oil temperature changes suddenly).

[0150] Value gradient flow (optimizing long-term stability):

[0151] The TD error update state value function is used to optimize the "long-term stable value estimation", including:

[0152]

[0153] in, The state value function measures "being in a state". Long-term cumulative benefits over time (such as the stable operating value of equipment at a certain oil temperature and cycle component); For immediate rewards (e.g., if the measurement deviation decreases after a local threshold adjustment, the reward is positive). For the next state Value estimation; This is a discount factor (emphasizing the weight of "future returns," such as setting it to 0.9 to indicate a greater emphasis on long-term stability). The learning rate (controls the step size for updating value); The target value is the sum of the immediate reward of the current action and the discounted value of the future state.

[0154] Step S32, Dynamic threshold matrix generation:

[0155] Based on topology data Build threshold matrix ,element Represents "node" and "Voltage threshold during connection";

[0156] Matrix element mapping formula (converting reinforcement learning output Q-value to threshold):

[0157]

[0158] in, For nodes Rated voltage (e.g., 220kV). (Proportional coefficient, controls the adjustment range of the threshold).

[0159] Step S4, Anomaly Assessment: Quantify the probability of anomalies and classify them.

[0160] Step S41: Compare the real-time operating condition matrix with the threshold matrix:

[0161] Constructing a real-time operating condition matrix (element Only retain The elements are set to 0, and the rest are set to 0.

[0162] Calculate the element-order deviation matrix : ( Deviation normalization to eliminate voltage level differences).

[0163] Step S42, KNN density peak anomaly detection:

[0164] Calculate local density (The degree of clustering of deviations around node i):

[0165]

[0166] in, For the cutoff distance (take the deviation matrix) maximum value ), (Smoothing coefficient);

[0167] Calculate distance (node Minimum deviation to higher density nodes):

[0168] (like To be the global maximum, then ;

[0169] Anomaly Probability Calculation (Normalization) ):

[0170]

[0171] Step S43, Anomaly Classification:

[0172] Based on anomaly probability Divided into 4 levels:

[0173] normal: ;

[0174] Mild abnormality: ;

[0175] Moderate abnormality: ;

[0176] Severe abnormality: .

[0177] Step S5, Decision Output: Provide dispatchers with intuitive anomaly information and actionable handling suggestions.

[0178] Step S51, alarm information generation:

[0179] The content includes: abnormal node number, abnormal parameters (e.g., “node 3 voltage abnormal”), abnormal probability (e.g., “P=0.65”), severity (e.g., “moderate abnormal”), and abnormal time (accurate to milliseconds);

[0180] Output format: Audible and visual alarm (buzzer + red indicator light), LCD screen display (deviation matrix) Heat map), transmitted remotely to the dispatch center (IEC61850 protocol).

[0181] Step S52: Generation of decision suggestions based on knowledge graphs:

[0182] Knowledge graph triple example: (nodes) Moderate voltage anomaly, switch to node Reactive power compensation device), (transformer) (Slight oil temperature abnormality, start the cooling fan).

[0183] Strategy linkage logic: When "moderate voltage anomaly at node 3" is detected, the knowledge graph is traversed to find the associated strategy and the output is "It is recommended to put the reactive power compensation device at node 3#1 into operation, and the voltage is expected to recover to 220kV±2%".

[0184] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An anomaly assessment device for dynamic threshold optimization of wide-area measurement data, characterized in that, include: The data acquisition and preprocessing module is used to acquire real-time measurement data and multi-source auxiliary data from the wide-area measurement system, perform data cleaning, spatiotemporal correlation processing and sensor fault self-diagnosis, and remove abnormal measurement values. The feature engineering module, connected to the data acquisition and preprocessing module, is used to perform time-series decomposition and feature filtering on the preprocessed data. The dynamic threshold optimization module, connected to the feature engineering module, includes a reinforcement learning unit and a threshold matrix generation unit; The reinforcement learning unit adopts a hierarchical decoupled training strategy, with the upper layer being the global operating condition layer and the lower layer being the local equipment layer. The policy gradient flow responds to real-time operating condition changes, while the value gradient flow optimizes long-term stability. The threshold matrix generation unit constructs a dynamic threshold matrix based on the power grid topology correlation, and the matrix element values ​​are obtained through the output transformation of the reinforcement learning unit. The anomaly assessment module is connected to the dynamic threshold optimization module and the data acquisition and preprocessing module, respectively, and is used to detect and classify abnormal states. The decision output module, connected to the anomaly assessment module, is used to generate anomaly status alarm information and decision suggestions.

2. The anomaly assessment device for dynamic threshold optimization of wide-area measurement data according to claim 1, characterized in that, The real-time measurement data includes voltage amplitude, current value, and frequency collected by the PMU; The multi-source auxiliary data includes meteorological data, equipment status information, and power grid topology data; The specific steps of the data acquisition and preprocessing module include: Multi-source data acquisition: Electrical parameters of power grid nodes are acquired at a frequency of 50Hz through the PMU data acquisition subunit to form a real-time measurement dataset; meteorological data, equipment status data and power grid topology data are acquired through the multi-source data interface subunit. The meteorological data includes temperature and wind speed, the equipment status data includes node transformer oil temperature, and the power grid topology data represents the connection relationship between nodes in matrix form. Data cleaning: Missing data is filled using linear interpolation. The formula for handling missing values ​​is: ， in, To be filled Time Node The measured value; , For missing moments The moment before The next moment The known measured values; For calculation arrive The rate of change per unit time at any given moment; for arrive The time interval between moments; Outlier handling: Extreme noise is removed using the 3σ rule. If it is, then it is marked as an outlier. and replace with .

3. The anomaly assessment device for dynamic threshold optimization of wide-area measurement data according to claim 2, characterized in that, The sensor fault self-diagnosis of the data acquisition and preprocessing module includes the following steps: Adjacent node data consistency verification: compute node With neighboring nodes Measurement deviation: , In the formula, for Time Node The measured value, for Time Node The measured value; like Then, further combining historical data , Identify the faulty sensor; if and Then remove and replace with , In the formula, , For missing moments The moment before The next moment The known measured values, This is the standard deviation of the measurement.

4. The anomaly assessment device for dynamic threshold optimization of wide-area measurement data according to claim 3, characterized in that, The feature engineering module uses the STL time series decomposition method to decompose the measurement data into trend components, seasonal components, periodic components and random components, and removes the random components. The contribution of trend component, seasonal component, and periodic component is calculated by using an improved Q-learning algorithm, and key features are dynamically selected. The improved Q-learning algorithm of the feature engineering module specifically includes: The state space is defined as trend component, seasonal component, and periodic component, and the action space retains the corresponding feature components. Define the state space Action space ; improve Value update formula: ， in: For state Next action Value estimation; To preserve the original Historical information weight of the value The learning rate; For tradition The temporal difference update term of the learning process. For the reward function; The learning rate; Discount factor; For subsequent status The largest below value.

5. The anomaly assessment device for dynamic threshold optimization of wide-area measurement data according to claim 4, characterized in that, The reinforcement learning unit of the dynamic threshold optimization module specifically includes: upper layer: state , in, Total power of the power grid; Action ; The reward function formula is: ， in, For real-time measurements, The current threshold, To provide real-time response weights, The deviation between real-time measurement and threshold; Long-term stable portion: ， in, For a period of time Total deviation within, For long-term stable weights; Lower layer: state ,action .

6. The anomaly assessment device for dynamic threshold optimization of wide-area measurement data according to claim 5, characterized in that, The policy gradient flow of the reinforcement learning unit adopts the REINFORCE algorithm, and the gradient formula is: ， in, For policy gradient flow, For policy functions; The logarithmic gradient of the policy is used to adjust the policy parameters. ; It is the state-action value function.

7. The anomaly assessment device for dynamic threshold optimization of wide-area measurement data according to claim 6, characterized in that, The long-term stability of the value gradient flow optimization is achieved by updating the state value function using TD error, including: ， in, The state value function; For instant rewards; For the next state Value estimation; Discount factor; The learning rate; For target value.

8. The anomaly assessment device for dynamic threshold optimization of wide-area measurement data according to claim 7, characterized in that, The generation of the dynamic threshold matrix specifically includes: Based on topology data Build threshold matrix ,element Represents a node and Voltage threshold during connection; Matrix element mapping formula: , in, Matrix element mapping, For nodes Rated voltage, proportionality coefficient It is the state-action value function.

9. The anomaly assessment device for dynamic threshold optimization of wide-area measurement data according to claim 8, characterized in that, The anomaly assessment module constructs an element-level comparison mechanism between the real-time operating condition matrix and the dynamic threshold matrix, calculates the deviation matrix, calculates the anomaly probability by combining the KNN density peak detection algorithm, and outputs the anomaly severity level based on the anomaly probability.

Images