A power distribution network power distribution state monitoring method and system

Abstract

The application discloses a power distribution network power distribution state monitoring method and system, distributed node data of a power distribution network is collected; time domain and frequency domain characteristics of the distributed node data are extracted, a power flow increment trajectory characteristic vector is obtained, including time domain characteristics, frequency domain characteristics and power flow direction characteristics; the power flow increment trajectory characteristic vector is input into a convolutional neural network model based on double channels and decision fusion, a disturbance classification label is output, and an abnormal monitoring node is marked; a power distribution network topology graph is constructed, neighborhood analysis is performed on the abnormal monitoring node, and a neighborhood node set is obtained; similarity of the power flow increment trajectory characteristic vector of the abnormal monitoring node and the power flow increment trajectory characteristic vector of each distributed node in the neighborhood node set is calculated, and a disturbance propagation path is obtained; disturbance propagation path data is collected, feature weights of the disturbance propagation path data are calculated, the disturbance propagation path data and the feature weights are combined, power distribution network operation risk indexes are calculated, and risk levels are divided.

Description

A method and system for monitoring the distribution status of a power distribution network Technical Field

[0001] This invention belongs to the field of power distribution network safety technology, and specifically relates to a method and system for monitoring the power distribution status of a power distribution network. Background Technology

[0002] With the increasing penetration of distributed generation in distribution networks and the growing complexity of load characteristics, monitoring the operational status of distribution networks faces significant challenges. The intermittent and random nature of distributed generation easily triggers disturbances such as power flow fluctuations and voltage exceedances, severely impacting power supply reliability and power quality. Therefore, timely and accurate identification and classification of disturbance types in the distribution network, and the establishment of reliable disturbance classification labels, have become crucial foundations for achieving intelligent monitoring and proactive prevention. Disturbance classification labels can quantify the operational status of nodes into categories such as "normal," "abnormal," or "potential risk," providing direct evidence for subsequent risk assessment and decision-making.

[0003] Traditional methods combine multi-dimensional features (such as time-domain and frequency-domain features) extracted from distributed nodes into a comprehensive feature vector, which is then fed as a single input into a convolutional neural network for feature learning and classification, and outputs the corresponding disturbance classification label to determine the state of the distribution network.

[0004] However, traditional methods simply merge features of different natures at the input layer, ignoring the differences between time-domain and frequency-domain features as different information modes. They lack separate analysis of the differences between these two modes, resulting in insufficient reliability of the disturbance classification results output by the model when dealing with power distribution network disturbance scenarios, leading to inaccurate state monitoring of the power distribution network. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method and system for monitoring the power distribution status of a power distribution network.

[0006] The technical solution adopted by the present invention is: a method for monitoring the distribution status of a power distribution network, comprising the following steps:

[0007] Step 1: Collect data from distributed nodes in the power distribution network to obtain distributed node data;

[0008] Step 2: Extract the time-domain and frequency-domain features of the distributed node data to obtain the power flow incremental trajectory feature vector, including time-domain features, frequency-domain features, and power flow direction features;

[0009] Step 3: Obtain the one-hot encoded vector and environmental feature vector corresponding to each distributed node. Input the one-hot encoded vector, environmental feature vector and power flow incremental trajectory feature vector into the convolutional neural network model based on dual-channel and decision fusion, and output the disturbance classification label and mark the abnormal monitoring node.

[0010] Step 4: Construct a distribution network topology based on the distributed nodes of the distribution network; perform neighborhood analysis on the abnormal monitoring nodes based on the distribution network topology to obtain a set of neighboring nodes; calculate the similarity between the power flow increment trajectory feature vector of the abnormal monitoring node and the power flow increment trajectory feature vector of each distributed node in the set of neighboring nodes to obtain the disturbance propagation path.

[0011] Step 5: Collect all distributed node data of the disturbance propagation path to obtain disturbance propagation path data, calculate the feature weights of the disturbance propagation path data, and combine the disturbance propagation path data and feature weights to calculate the distribution network operation risk index and classify the risk level, thereby realizing the status monitoring of the distribution network.

[0012] Preferably, in step 1, the three-phase active power, three-phase reactive power, effective voltage value, effective current value, and power flow direction of the distributed nodes are synchronously collected and recorded at a sampling interval of not less than 100 milliseconds.

[0013] Preferably, in step 2, the power flow incremental trajectory feature vector FX = [ , , , , , , , ];

[0014] Among them, the fluctuation amplitude of the Xth feature of the tidal current increment trajectory feature vector. , Represents the feature vector of power flow increment trajectory The maximum value of the Xth feature. Represents the trajectory of power flow increment The minimum value of the Xth feature; the measured value of each distributed node i at each sampling time t within the current time window win. and its benchmark value Perform a subtraction operation to generate the power flow increment trajectory for that monitoring node. ;in, This represents the power flow increment trajectory of the Xth feature of the data from the i-th distributed node. Let X represent the measured value of the Xth feature of the i-th distributed node data at time t. (t) represents the baseline value of the Xth feature of the i-th distributed node data at time t;

[0015] Average rate of change of each feature of the power flow increment trajectory ;in, Let N be the average rate of change of the Xth feature of the power flow increment trajectory, N be the total number of sampling points within the time window win, and k be the index of the sampling point. Indicates the k-th sampling time The increment value of the Xth feature in the +1 power flow increment trajectory. Indicates the k-th sampling time The increment value of the Xth feature in the power flow increment trajectory. The sampling time interval;

[0016] The absolute and integral quantities of the Xth feature of the power flow increment trajectory ;

[0017] Frequency domain features are extracted from the tidal current increment trajectory. First, a discrete Fourier transform is performed on the tidal current increment trajectory to obtain the spectrum X(f); based on the spectrum X(f), the dominant oscillation frequency is calculated. and the proportion of dominant frequency energy Feature extraction is performed on the tidal current direction state sequence in the tidal current increment trajectory to obtain the reversal frequency. Where M is the number of direction flips. The length of the time window; extract the set of stable state duration values ​​for all power flow directions remaining unchanged within the time window win. , ,..., } Calculate the average of the durations in the set of duration values. The shortest duration in the set of duration values .

[0018] Preferably, the convolutional neural network model based on dual-channel and decision fusion includes a time-domain path and a frequency-domain path;

[0019] Both the time-domain path and the frequency-domain path include a first convolutional layer, a first max-pooling layer, a second convolutional layer, a second max-pooling layer, a third convolutional layer, a flattening layer, and a fully connected layer, all connected in sequence. The first convolutional layer is followed by a batch normalization layer and a ReLU activation function; the second convolutional layer is followed by a batch normalization layer and a ReLU activation function; the third convolutional layer is followed by a batch normalization layer and a ReLU activation function; the fully connected layer includes a first fully connected layer and a second fully connected layer, and the output layer uses the Softmax activation function to output a probability distribution vector.

[0020] The input to the time-domain path is a time-domain hybrid vector of dimension K1, and the input to the frequency-domain path is a frequency-domain hybrid vector of dimension K2. Finally, the perturbation classification label is output based on the convolutional neural network model of dual-channel and decision fusion, where K1 and K2 are preset values.

[0021] Preferably, in step 3, the one-hot encoded vector and environmental feature vector corresponding to each distributed node are obtained; then, the temporal features of the power flow incremental trajectory features and the power flow direction features are combined into a temporal feature sub-vector, and the frequency domain features of the power flow incremental trajectory features are called frequency domain feature sub-vectors; next, the standardized temporal feature sub-vectors, one-hot encoded vectors, and environmental feature vectors are combined into a temporal hybrid vector and input into the temporal path of the convolutional neural network model based on dual-channel and decision fusion; the standardized frequency domain feature sub-vectors, one-hot encoded vectors, and environmental feature vectors are combined into a frequency domain hybrid vector and input into the frequency domain path of the convolutional neural network model based on dual-channel and decision fusion; the perturbation classification label is output, and the distributed nodes with the perturbation classification label as abnormal are marked as abnormal monitoring nodes;

[0022] Based on the weighted probability voting principle, the final perturbation classification label is calculated. ;in, Indicates the disturbance classification label. This represents the probability that a sample output from the time-domain path belongs to class c. This represents the probability that a sample output from the frequency domain path belongs to class c. Represents the time-domain path weights. This represents the frequency domain path weight.

[0023] Preferably, in step 4, complete network topology data of the distribution network is obtained, including the location information and connection relationship of all distributed nodes and designated key electrical nodes. Based on this, a topology graph G=(V, BB) of the distribution network is constructed, where vertex set V represents all distributed nodes and key electrical nodes, and edge set BB represents the electrical connection relationship between nodes.

[0024] Centered on the identified abnormal monitoring node, a neighborhood analysis is performed along the topology based on its feeder location: searching upstream, downstream, and through bypass feeders connected by closed tie switches; the physical neighborhood is defined using topological distance as a metric, and all monitoring nodes within a preset topological distance are defined as the direct neighbors of the abnormal node, ultimately yielding a set of neighborhood nodes.

[0025] Preferably, in step 4, the similarity between the power flow incremental trajectory feature vector of the anomaly monitoring node and the power flow incremental trajectory feature vector of each distributed node in the neighborhood node set is calculated using a dynamic time warping algorithm based on the shape of local extrema. The specific implementation process is as follows:

[0026] Using the anomaly monitoring node as the initial anomaly node, and taking its initial anomaly window as the starting point, continuously extract feature data from the subsequent WIN windows to form a sequence Source; After standardization, the calculation The L2 norm of the vector under each time window yields the initial scalar sequence of outlier nodes over WIN time windows. =[ This process extracts local maxima and local minima from the initial scalar sequence of outlier nodes; finally, it outputs the index sequence corresponding to the extreme points in the initial scalar sequence of outlier nodes. ;

[0027] For each extreme point index ,by Centered on the initial abnormal node scalar sequence Extract a subsequence of length b. , Use the HOG-1D function to process subsequences. Finally, the shape descriptor vector MS = [h1, h2, ..., hb, ..., hB], where hb represents the normalized value of the b-th interval, then the shape descriptor vector for each extreme point is... The final output is a comprehensive shape descriptor for the anomaly monitoring points. =[ , ,..., ,…, Similarly, calculate the comprehensive shape descriptor for each distributed node in the neighborhood node set. ;

[0028] Calculate the similarity between the anomaly monitoring node and each distributed node in the neighboring node set:

[0029] ;

[0030] in, This represents the similarity between the initial abnormal node and the i-th distributed node in the set of neighboring nodes. This is the composite shape descriptor for the initial abnormal node. It is the comprehensive shape descriptor of the i-th distributed node in the neighborhood node set.

[0031] Preferably, in step 5, the collection After standardizing the historical perturbation propagation path data using Z-Score, the entropy values ​​of the features in the perturbation propagation path data are determined based on the entropy values:

[0032] ;

[0033] in, It is the entropy value of the r-th feature in the historical disturbance propagation path data. It is the data sample size of historical disturbance propagation path data. It refers to the first The proportion of each sample data in the r-th feature column;

[0034] Calculate the feature weights of the disturbance propagation path data:

[0035] ;

[0036] in, R is the feature weight of the r-th feature in the historical disturbance propagation path data, and R is the total number of features in the historical disturbance propagation path data.

[0037] Preferably, in step 5, after standardizing the disturbance propagation path data, the disturbance propagation path data and feature weights are combined. The following operational risk indicators for the power distribution network were calculated:

[0038] ;

[0039] in, I represents the risk indicator for power distribution network operation. R' represents the r-th feature value of the disturbance propagation path data; R' is the total number of features in the disturbance propagation path data.

[0040] The technical solution adopted by the system of the present invention is: a power distribution network status monitoring system, comprising:

[0041] One or more processors;

[0042] A storage device is provided for storing one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the power distribution status monitoring method for the power distribution network.

[0043] Compared with the prior art, the beneficial effects of the present invention include:

[0044] (1) The convolutional neural network model adopted in this invention is the core technical support for power distribution network status monitoring. It can deeply mine the complex patterns in the characteristics of power flow increment trajectory, learn the deep correlation information related to time domain, frequency domain and power flow direction, effectively distinguish the three different operating conditions of normal disturbance, abnormal disturbance and potential risk disturbance, and provide a reliable basis for the accurate marking of abnormal monitoring nodes.

[0045] (2) The dual-channel and decision fusion mechanism of this invention is reflected in achieving comprehensive feature extraction and improving the reliability of classification results. The time-domain path 1D-AlexNet-T focuses on mining the dynamic change patterns in the time series, while the frequency-domain path 1D-AlexNet-F focuses on capturing frequency distribution characteristics. The two are adapted to feature attributes of different dimensions, forming complementary advantages. Decision fusion integrates the classification results of the two paths by weight allocation based on the path classification accuracy and the majority voting principle, effectively reducing the bias of single-dimensional analysis, avoiding misjudgment caused by the omission of local features, and significantly improving the objectivity and accuracy of perturbation classification.

[0046] (3) The significance of the dynamic time warping method of this invention lies in its ability to flexibly handle the offset of the power flow increment trajectory between the abnormal monitoring node and the neighboring node on the time axis, and to accurately measure the morphological similarity of the disturbance characteristics of different nodes. By quantifying the matching degree of feature sequences, this method successfully screens out a set of spatiotemporally related nodes with highly similar disturbance patterns, providing a key quantitative basis for reconstructing the disturbance propagation path based on the topology, and solving the problem of comparing the mismatch between the disturbance propagation time sequence and the physical connection of nodes in the distribution network.

[0047] (4) The significance of using local extremum shapes in this invention lies in achieving efficient dimensionality reduction of perturbation features and preservation of key information. By extracting key points such as local maxima and minima of the perturbation source scalar sequence, redundant data is eliminated. At the same time, the HOG-1D algorithm is combined to encode local shape information, forming a shape descriptor that is both representative and discriminative. This process reduces the computational load of similarity calculation, improves analysis efficiency, highlights the core morphological features of the perturbation pattern, avoids interference from irrelevant information, ensures the targeting and accuracy of neighborhood node similarity comparison, and helps to quickly locate key related nodes of perturbation propagation;

[0048] (5) This invention calculates the feature weights of disturbance propagation path data using the entropy weight method. Combining the disturbance propagation path data and feature weights, it calculates the risk indicators of distribution network operation and classifies the risk levels. By using the entropy weight method to adaptively allocate weights based on the volatility of historical data, it ensures the scientificity and adaptability of weight allocation. This comprehensively quantifies the multi-dimensional characteristics of disturbance propagation, such as the range, speed, electrical impact, and operational status, and generates comprehensive and accurate risk indicators. Through the classification of levels (such as low, medium, and high risk), it provides clear decision-making basis for operators, effectively supports early warning, resource optimization and allocation, and graded response, and ultimately improves the active defense capability and operational resilience of the distribution network. Attached Figure Description

[0049] The technical solutions of the present invention will be further illustrated below using embodiments and specific implementation methods. In addition, some accompanying drawings are used in the description of the technical solutions. Those skilled in the art can obtain other drawings and the intent of the present invention from these drawings without any creative effort.

[0050] Figure 1 is a flowchart of a method according to an embodiment of the present invention;

[0051] Figure 2 is a structural diagram of a convolutional neural network model based on dual-channel and decision fusion according to an embodiment of the present invention;

[0052] Figure 3 is a flowchart of obtaining the neighborhood node set in an embodiment of the present invention;

[0053] Figure 4 is a step-by-step diagram of obtaining distribution network operation risk indicators in an embodiment of the present invention. Detailed Implementation

[0054] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0055] Please refer to Figure 1. This embodiment provides a method for monitoring the power distribution status of a power distribution network, which includes the following steps:

[0056] Step 1: Collect data from distributed nodes in the power distribution network using sensors to obtain distributed node data;

[0057] In one implementation, firstly, at nodes in the distribution network where distributed generation sources are connected, upstream transformer outgoing nodes, and important load concentration nodes exist, these nodes are collectively referred to as distributed nodes. Multifunctional power flow monitoring units with synchronous measurement capabilities are deployed at these distributed nodes. These monitoring units must synchronously collect and record the node's three-phase active power, three-phase reactive power, RMS voltage, RMS current, and power flow direction at sampling intervals of at least 100 milliseconds. To form a continuous monitoring chain, the monitoring units must cover all main lines from the distributed generation source connection point to the upstream substation outgoing line, and additional monitoring points must be installed at the entrances of key branch lines connected to distributed generation sources with a capacity exceeding 500 kilowatts or important loads.

[0058] To achieve standardized management and mapping of monitoring data, a centralized transmission channel for synchronous phasor measurement data was established using the IEEE C37.118.2 protocol as the communication standard. Each monitoring unit uploads time-stamped measurement data streams in real time through a dedicated power communication network, forming distributed node data in the distribution network with a unified time scale.

[0059] Step 2: Extract the time-domain and frequency-domain features of the distributed node data to obtain the power flow incremental trajectory feature vector, including time-domain features, frequency-domain features, and power flow direction features;

[0060] In one implementation, a historical steady-state data sample library is constructed, consisting of long-term collected data from various monitoring nodes. Each data record is labeled with corresponding operational background information, including season (spring, summer, autumn, winter), date type (weekday or holiday), time period, and weather conditions (sunny, cloudy, rainy, etc.). Node data for each monitoring node is collected under stable conditions, where the voltage deviation of the monitoring node does not exceed ±5% of the rated value, the power fluctuation rate is less than 10%, and there are no fault warning records. Next, a background matching rule is established to match the current operational background with historical data records, prioritizing historical data with the same season, date type, and time period, and the closest weather conditions, resulting in a filtered data subset. For the data subset matched at the current analysis time t, the statistical characteristic value of the corresponding monitored quantity (e.g., active power) within the subset is calculated. If the data distribution within the subset exhibits a normal distribution, the arithmetic mean is taken as the baseline value for the current time. If a skewed distribution exists, the median of that subset is taken as the benchmark value to ensure that the benchmark value represents the normal steady-state level of the node under a specific operating context. Finally, the benchmark value of the data of the i-th distributed node at time t for the x-th feature is obtained. .

[0061] For each distributed node i, the measured value at each sampling time t within the current time window win and its benchmark value Perform a subtraction operation to generate the power flow increment trajectory for that monitoring node. :

[0062] ;

[0063] in, This represents the power flow increment trajectory of the Xth feature of the data from the i-th distributed node. Let X represent the measured value of the Xth feature of the i-th distributed node data at time t. (t) represents the baseline value of the Xth feature of the i-th distributed node data at time t;

[0064] It should be noted that the real-time monitoring values ​​collected by the monitoring unit specifically include three-phase active power P, three-phase reactive power Q, effective voltage U, and effective current DL, as well as discrete state quantities, namely, power flow direction D (defined as: +1 for positive direction and -1 for negative direction). The power flow increment trajectory comprises two parts: one is the increment trajectory of electrical quantities, calculated by subtracting the measured values ​​of each electrical quantity from its corresponding reference values. For example, the increment trajectory of three-phase active power is... Secondly, the state sequence of the power flow direction directly records the original state change sequence of this discrete quantity within the time window. This is used to characterize the direction reversal process. Together, these two parts constitute the complete power flow increment trajectory used for subsequent feature extraction and disturbance analysis. =[ , , , , ].

[0065] After obtaining the power flow increment trajectories of each monitoring node, multi-dimensional features need to be extracted for disturbance classification. Firstly, for... , , , Time-domain and frequency-domain features are extracted: time-domain features include fluctuation amplitude (difference between maximum and minimum values), average rate of change, and energy integral over a specific period; frequency-domain features are extracted using Fast Fourier Transform to obtain the dominant oscillation frequency and its energy proportion. Simultaneously, the power flow direction state sequence is analyzed to extract its reversal frequency per unit time, state duration, and the temporal correlation between direction changes and power abrupt changes.

[0066] For each distributed node, within a given time window win, the power flow increment trajectory is calculated based on all sampled points within that window. Extract the following feature values ​​to describe the overall dynamic characteristics of this window:

[0067] First, for each distributed node, extract the power flow increment trajectory. Within the time window win, the feature values ​​are first extracted in the time domain from the electrical quantity increment trajectory to calculate the fluctuation amplitude:

[0068] ;

[0069] in, Feature vector of power flow increment trajectory The fluctuation range of the Xth feature, Represents the feature vector of power flow increment trajectory The maximum value of the Xth feature. This represents the minimum value of the Xth feature in the power flow increment trajectory;

[0070] Calculate the average rate of change for each feature:

[0071] ;

[0072] in, Let N be the average rate of change of the Xth feature of the power flow increment trajectory, N be the total number of sampling points within the time window win, and k be the index of the sampling point. Indicates the k-th sampling time The increment value of the Xth feature in the +1 power flow increment trajectory. Indicates the k-th sampling time The increment value of the Xth feature in the power flow increment trajectory. The sampling time interval;

[0073] Calculate the absolute and integral quantities:

[0074] ;

[0075] in, Let X be the absolute sum and integral of the Xth feature of the power flow increment trajectory;

[0076] Frequency domain features are extracted from the tidal current increment trajectory. First, a discrete Fourier transform is performed on the tidal current increment trajectory to obtain the spectrum X(f); based on the spectrum X(f), the dominant oscillation frequency is calculated. and the proportion of dominant frequency energy Where f represents frequency.

[0077] Feature extraction is performed on the power flow direction state sequence in the power flow increment trajectory. Suppose that the state sequence D(t) undergoes M direction flips within the time window (i.e., D( ) D( )).

[0078] Then the frequency of flipping:

[0079] ;

[0080] in, M represents the frequency of rotations, where M is the number of directional rotations. This represents the length of the time window.

[0081] Extract the duration of the power flow direction state sequence, and identify continuous time periods within the time window win where all power flow directions remain unchanged; these time periods are called the steady state duration. The specific process is as follows: traverse the direction state sequence; when a change in direction value is detected, mark the end point of the previous stable state and calculate its duration (i.e., the difference between the current time and the start time of the previous state), thus obtaining a set containing M+1 duration values ​​{ , ,..., } where M is the number of direction flips, and M flips divide the time window into M+1 stable segments. Calculate the average duration in the set of duration values. , It reflects the overall stability of the directional state; and obtains the shortest duration in the set of duration values. .

[0082] In one implementation, the average duration For each stable segment, calculate its duration. For example, if a stable period starts at t=1s and ends at t=3s, then... =2s, ultimately yielding a set of duration values ​​{ , ,..., } and calculate the arithmetic mean of the combined duration values: , This reflects the average duration during which the power flow direction remains stable throughout the entire analysis window. The larger the value, the more stable the direction and the less frequent the reversals; the smaller the value, the more frequent the direction switching, and the more likely it is in an oscillating or unstable state.

[0083] In one implementation, the longest duration was not included in the feature extraction process of the power flow direction state sequence. As a core feature, the longest duration can only reflect the longest steady-state segment within the window. However, in scenarios with a high proportion of distributed power generation, a longer steady state may correspond to normal operating conditions (such as continuous photovoltaic power generation or stable load consumption), and its value is constrained by the length of the time window and cannot effectively characterize disturbance characteristics; conversely, the shortest duration It can capture instantaneous reversal behavior of orientation state, which has significant indicative significance for identifying key disturbance modes such as oscillatory faults or equipment anomalies.

[0084] Finally, the feature vector FX of the power flow increment trajectory is obtained. , , , , , , , ].

[0085] Step 3: Obtain the one-hot encoded vector and environmental data corresponding to each distributed node, and obtain the environmental feature vector through the environmental data; input the power flow incremental trajectory feature vector into the convolutional neural network model based on dual-channel and decision fusion, and output the disturbance classification label; mark the distributed nodes with the disturbance classification label as abnormal as abnormal monitoring nodes.

[0086] In one implementation, the tidal current increment trajectory characteristics essentially consist of three parts, with the time-domain component including: fluctuation amplitude. Average rate of change , The frequency domain component of the power flow increment trajectory characteristics includes the dominant oscillation frequency. and the proportion of dominant frequency energy The directional characteristics of the incremental tidal current trajectory include the frequency of reversals. Average duration and shortest duration The temporal component of the tidal current increment trajectory features, along with tidal current direction features (such as flip frequency and duration, which are also temporal behavior features), are combined into a temporal feature sub-vector. The frequency domain component of the tidal current increment trajectory features is called the frequency domain feature sub-vector.

[0087] The time-domain feature vector and frequency-domain feature vector are both Z-score standardized. One-hot encoding is then performed based on the type identifier of the distribution network monitoring node (e.g., photovoltaic access point, important load point, main line node, etc.) to obtain the one-hot encoded vector.

[0088] It should be noted that the types of monitoring nodes include distributed nodes with typical dynamic characteristics (such as photovoltaic access points affected by rapid changes in sunlight, electric vehicle charging stations with high-frequency randomness in charging and discharging behavior, and residential inverter nodes with frequent start-stop phenomena) and key nodes characterizing the main structure of the power grid (such as upstream transformer outgoing line nodes and important load concentration nodes). Although the latter two types of nodes do not belong to distributed power sources, their operating status has a decisive impact on the overall power flow distribution and stability of the distribution network, and therefore they are also included in the classification system as important monitoring types. For the above monitoring node types, one-hot encoding is used to generate a unique binary one-hot encoding vector for each monitoring node type. If the total number of types is 5, then the photovoltaic access point is encoded as [1,0,0,0,0], and the electric vehicle charging station is encoded as [0,1,0,0,0]. Transforming the abstract monitoring node types into numerical forms eliminates the implicit mathematical relationships between types, enabling the model to learn the unique operating rules and disturbance response patterns of different types of nodes without bias.

[0089] Environmental data is collected for each distributed node. For photovoltaic (PV) access points, the average solar irradiance and rate of change of solar irradiance within a time window are collected to quantify the overall level and volatility of solar energy. For electric vehicle charging stations, the average real-time electricity price and rate of change of electricity price within a time window are collected to capture the driving effect of electricity price signals on charging and discharging behavior. For residential inverter nodes, both the average solar irradiance and the average local low-voltage total load are collected. For upstream transformer outgoing nodes and important load concentration nodes, the average ambient temperature, average ambient humidity, and date type flags (1 for weekdays, 0 for holidays) are collected to characterize the basic operating environment conditions of the power grid. If a certain type of node lacks specific environmental data (e.g., transformer nodes lack direct solar irradiance data), the global average value of other types of nodes with that feature data in the training set is used to fill in that feature dimension to maintain the consistency of the input vector dimension. All environmental data are standardized to obtain an environmental feature vector.

[0090] The standardized temporal feature vectors, one-hot encoded vectors, and environmental feature vectors are combined into a temporal hybrid vector and input into the temporal path 1D-AlexNet-T of the convolutional neural network model based on dual-channel and decision fusion; the standardized frequency domain feature vectors, one-hot encoded vectors, and environmental feature vectors are combined into a frequency domain hybrid vector and input into the frequency domain path 1D-AlexNet-F of the convolutional neural network model based on dual-channel and decision fusion.

[0091] In one implementation, as shown in Figure 2, the convolutional neural network model based on dual-channel and decision fusion includes a time-domain path and a frequency-domain path.

[0092] Both the time-domain and frequency-domain paths include a first convolutional layer, a first max-pooling layer, a second convolutional layer, a second max-pooling layer, a third convolutional layer, and a fully connected layer, all connected in sequence. The first convolutional layer uses a kernel size of 3, 16 output channels, a stride of 1, and padding of 1, followed by a batch normalization layer and a ReLU activation function. The first max-pooling layer uses a kernel size of 2 and a stride of 2. The second convolutional layer uses a kernel size of 3, 32 output channels, a stride of 1, and padding of 1, followed by a batch normalization layer and a ReLU activation function. The second max-pooling layer uses a kernel size of 2 and a stride of 2. The third convolutional layer uses a kernel size of 3, 64 output channels, a stride of 1, and padding of 0, followed by a flattening layer and a batch normalization layer and a ReLU activation function. The fully connected layer includes a first fully connected layer and a second fully connected layer; the output layer uses a Softmax activation function to output a probability distribution vector.

[0093] The input to the time-domain path is a time-domain mixture vector with a dimension of 30, and the input to the frequency-domain path is a frequency-domain mixture vector with a dimension of 26. Finally, the perturbation classification label is output based on the convolutional neural network model with dual channels and decision fusion.

[0094] In one implementation, the temporal feature vector has a dimension of 12 (4 physical quantities × 3 features), the node type OneHot encoding has a dimension of 5, and the environment feature vector has a dimension of 13, resulting in a total dimension of 12 + 5 + 13 = 30 after concatenation. The frequency domain feature vector has a dimension of 8 (4 physical quantities × 2 features), which is concatenated with the same OneHot encoding (5 dimensions) and environment feature vector (13 dimensions), resulting in a total dimension of 8 + 5 + 13 = 26.

[0095] The concatenated temporal blending vector is input into the 1D-AlexNet-T path. This network learns deep temporal patterns from the combination of temporal features and contextual information through one-dimensional convolutional operations. The concatenated frequency-domain blending vector is input into the 1D-AlexNet-F path. This network focuses on mining the frequency distribution characteristics from the combination of frequency-domain features and contextual information.

[0096] For the 1D-AlexNet-T path (focusing on temporal features), the input layer receives data with a batch size of (32, 30, 1). The first convolutional layer uses a kernel size of 3, 16 output channels, a stride of 1, and padding of 1, outputting data with a batch size of (32, 30, 16), followed by a batch normalization layer and a ReLU activation function. Next is a max-pooling layer with a kernel size of 2 and a stride of 2, outputting data with a batch size of (32, 15, 16). The second convolutional layer uses a kernel size of 3, 32 output channels, a stride of 1, and padding of 1, outputting data with a batch size of (32, 15, 32), followed by a batch normalization layer and a ReLU activation function. Another max-pooling layer uses a kernel size of 2 and a stride of 2, outputting data with a batch size of (32, 7, 32). The third convolutional layer uses a kernel size of 3, 64 output channels, a stride of 1, and padding of 0, resulting in an output shape of (batch size 32, 5, 64). This is followed by a batch normalization layer and a ReLU activation function. Then, a flattening layer transforms the output to (batch size 32, 320). The fully connected layers consist of a first fully connected layer (units 320 to 128, ReLU activation, Dropout ratio 0.5), a second fully connected layer (units 128 to 64, ReLU activation, Dropout ratio 0.5), and an output layer using the Softmax activation function, outputting a probability distribution vector. c=1, 2, 3 correspond to three categories of labels: normal disturbance, abnormal disturbance, and potential risk disturbance, respectively. This represents the probability that the input sample belongs to the c-th class in the 1D-AlexNet-T path, where C is the total number of classes.

[0097] For the 1D-AlexNet-F path (focusing on frequency domain features), the structure is similar but adjusted to accommodate shorter input sequences (length 26). The input layer receives data with a batch size of (32, 26, 1). The first convolutional layer uses a kernel size of 3, 16 output channels, a stride of 1, and padding of 1, outputting a batch size of (32, 26, 16), followed by a batch normalization layer and a ReLU activation function. The max pooling layer uses a kernel size of 2 and a stride of 2, outputting a batch size of (32, 13, 16). The second convolutional layer uses a kernel size of 3, 32 output channels, a stride of 1, and padding of 1, outputting a batch size of (32, 13, 32), followed by a batch normalization layer and a ReLU activation function. Another max pooling layer uses a kernel size of 2 and a stride of 2, outputting a batch size of (32, 6, 32). The third convolutional layer uses a kernel size of 3, 64 output channels, a stride of 1, and padding of 0, resulting in an output shape of (batch size 32, 4, 64). This is followed by a batch normalization layer and a ReLU activation function. Then, a flattening layer transforms the output to (batch size 32, 256). The fully connected layers consist of a first fully connected layer (units 256 to 128, ReLU activation, Dropout ratio 0.5), a second fully connected layer (units 128 to 64, ReLU activation, Dropout ratio 0.5), and an output layer using a Softmax activation function, outputting a probability distribution vector. ,in This represents the probability that the input sample belongs to the c-th class in the 1D-AlexNet-F path.

[0098] Finally, based on the weighted probability voting principle, the final perturbation classification label is calculated. :

[0099] ;

[0100] in, Indicates the disturbance classification label. This represents the probability that a sample output from the time-domain path belongs to class c. This represents the probability that a sample output from the frequency domain path belongs to class c. Represents the time-domain path weights. This represents the frequency domain path weight.

[0101] In one implementation, the convolutional neural network model based on dual-channel and decision fusion is a pre-trained model. During training, to construct a training dataset suitable for supervised learning, historical monitoring data needs to be accurately labeled. The labeling process is as follows: When the electrical quantity (such as active power) of a monitoring node experiences a sharp increase or decrease in a short period of time, and its trend is highly consistent with predictable external environmental driving factors (such as solar intensity change curves, preset electric vehicle charging and discharging plans), and no system protection alarm is triggered, the sample is labeled as "normal disturbance"; when the electrical quantity exhibits violent oscillations, continuous over-limits, or frequent reversals in power flow direction, and there are corresponding fault action records (such as protection tripping, equipment abnormal signals) in the distribution automation system event log, it is labeled as "abnormal disturbance"; when the electrical quantity (such as voltage or reactive power) shows a continuous and slow deterioration trend (such as long-term low voltage, continuous expansion of reactive power deficit), and there is no clear environmental factor or instantaneous fault corresponding to it, it is labeled as "potential risk disturbance". This labeling process needs to be cross-validated with real-time event logs and domain expert knowledge bases to ensure the objectivity and accuracy of the labels.

[0102] The time-domain path weight Frequency domain path weights Based on the independent validation set statistics of each path classifier, the training set, independent validation set, and test set are divided in a 7:2:1 ratio to ensure that the validation set nodes do not overlap with the training set. The validation set samples are then input into two paths of the convolutional neural network model, and the category with the highest probability for each path is taken as the classification result. The number of matches between each path classification result and the true label is counted to determine the temporal path classification accuracy. = Frequency domain path classification accuracy = Then, weights are assigned based on the classification accuracy ratio of the time-domain path (1D-AlexNet-T) and the frequency-domain path (1D-AlexNet-F), i.e., the time-domain path weights. Frequency domain path weights .

[0103] When the output disturbance classification label is "abnormal disturbance" ( =2) and "potential risk disturbances" When the value is 3), the disturbance classification label of the distributed node is considered to be abnormal, and the distributed node is marked as an abnormal monitoring node.

[0104] Step 4: Construct a distribution network topology based on the distributed nodes of the distribution network; perform neighborhood analysis on the abnormal monitoring nodes based on the distribution network topology to obtain a set of neighboring nodes; calculate the similarity between the power flow increment trajectory feature vector of the abnormal monitoring node and the power flow increment trajectory feature vector of each distributed node in the set of neighboring nodes using a dynamic time warping algorithm based on local extremum shape to obtain the disturbance propagation path.

[0105] In one implementation, as shown in Figure 3, complete network topology data of the distribution network is obtained, including the location information and connection relationships of all distributed nodes and designated key electrical nodes (outgoing nodes of upstream transformers and nodes where important loads are concentrated; the criteria for judgment are whether the node is located at the structural hub of the power grid (such as power outlets or main lines) or whether it is connected to equipment / loads that have a significant impact on the operation of the system (such as large-capacity distributed power sources or important loads)). Based on this, a topology graph of the distribution network G=(V, BB) is constructed, where the vertex set V represents all distributed nodes and key electrical nodes, and the edge set BB represents the electrical connection relationships between nodes.

[0106] Centered on the identified abnormal monitoring node, a neighborhood analysis is performed along the topology graph based on its feeder location: searching upstream (power supply side, such as towards the substation), downstream (load side), and through bypass feeders connected by closed tie switches. The physical neighborhood is defined using topological distance (the number of edges traversed by the shortest path in the graph) as the primary metric. Typically, all monitoring nodes with a topological distance within the LY hop range (e.g., LY=3) are defined as the direct neighbors of the abnormal node, serving as the set of neighboring nodes for subsequent feature similarity comparison and propagation analysis.

[0107] In one implementation, the anomaly monitoring node obtained in step 3 is used as the initial anomaly node. Starting from its initial anomaly window, feature data from subsequent windows (WIN windows) are continuously extracted to form a sequence Source. After standardization, the calculation The L2 norm of the vector under each time window yields the initial scalar sequence of outlier nodes over WIN time windows. =[ The process involves extracting key points (local maxima and local minima) from the initial scalar sequence of outlier nodes, reducing the sequence length while preserving morphological features. Local maxima points are... > and The local minimum point is < and The final output is the index sequence corresponding to the extreme points in the initial scalar sequence of outlier nodes. .

[0108] In one implementation, the combination of "strict inequality and equals signs" in the determination of local extreme points aims to handle the plateau period (with consecutive sampling point values ​​being equal) in the power flow scalar sequence FEPS of the distribution network, ensuring the capture of key morphological turning points without redundancy: the first part uses strict inequality signs (> or <): ensuring that the determined extreme point is "the starting point of entering the plateau from the rise / fall", avoiding subsequent identical value points within the plateau period being repeatedly determined as extreme points (e.g., in FEPS=[2,3,3,2], only the first 3 is determined as a maximum value); the second part uses equals signs (≥ or ≤): ensuring that the starting point of the plateau period is included in the extreme point (e.g., in FEPS=[5,2,2,4], the first 2 is determined as a minimum value), avoiding the omission of key morphological features of "entering the stable segment after the fall / rise"; this design can improve the completeness of extreme point extraction, reduce redundancy rate (repeated determination), and balance feature representativeness and computational efficiency.

[0109] In one implementation, to enhance the semantic meaning of alignment and encode local shape information for each extremum point, subsequence sampling is first performed: for each extremum point index ,by Centered on the initial abnormal node scalar sequence Extract a subsequence of length b. , The boundary handling ensures sequence integrity through duplication or padding. The window length b is set to 30 by default to balance detail capture and computational efficiency. Next, the HOG-1D (One-Dimensional Oriented Gradient Histogram) function is used to process the subsequences. First, calculate the gradient, then calculate the subsequence. The gradient of each point in the matrix is ​​calculated (i.e., the difference from the previous point). Then, binning statistics are performed, dividing the gradient direction (e.g., sign) into BIN intervals (usually BIN is 8 to 16), and the sum of the gradient magnitudes in each interval is calculated. Finally, the shape descriptor vector MS=[h1,h2,...,hb,...,hB] is obtained through normalization, where hb represents the normalized value of the b-th interval. This yields the shape descriptor vector for each extreme point. The final output is a comprehensive shape descriptor for the anomaly monitoring points. =[ , ,..., ,…, ].

[0110] For any node i in the neighborhood node set, if the node is also marked as anomalous within any WIN windows following the initial anomalous window of the initial anomalous node, then the power flow increment trajectory feature vector is taken from the window corresponding to the initial anomalous node as the starting point and then for WIN time windows. If the node is not marked as anomalous, then the initial time window corresponding to the initial anomalous node is used as the unified starting point, and WIN consecutive power flow increment trajectory feature vectors are also extracted. Finally, the comprehensive shape descriptor for each distributed node in the neighborhood node set is also calculated. .

[0111] Calculate the similarity between the anomaly monitoring node and each distributed node in the neighboring node set:

[0112]

[0113] in, This represents the similarity between the initial abnormal node and the i-th distributed node in the set of neighboring nodes. This is the composite shape descriptor for the initial abnormal node. It is the comprehensive shape descriptor of the i-th distributed node in the neighborhood node set.

[0114] If the similarity between a distributed node in the neighborhood node set and an anomaly monitoring node is greater than the similarity threshold θ (e.g., θ=0.85), then the node is included in the spatiotemporally associated node set. This set is a group of nodes with highly similar perturbation patterns to the initial anomaly node.

[0115] After obtaining the set of spatiotemporally associated nodes, each node is sorted chronologically based on the window in which it was first identified as an anomaly, forming a time series of node anomaly occurrences. For nodes in the set that are not marked as anomalies but have high similarity, since they lack a clear anomaly time window, the time window with the highest morphological similarity to the initial anomaly node is used as the initial anomaly time window for that node, and it is included in the above time series for sorting. Subsequently, the physical direct or indirect connectivity of temporally adjacent nodes is verified by combining the connection relationships of the distribution network topology (such as upstream and downstream logical links and impedance parameters) to ensure that the propagation path conforms to the actual electrical constraints. For node pairs that do not meet the direct connectivity requirement, the shortest path algorithm is used to complete the intermediate nodes in the topology. These completed intermediate nodes, even if not included in the spatiotemporally associated set, will be introduced to form a continuous electrical path, thereby reconstructing a complete propagation chain from the suspected disturbance source node, spreading along the temporal direction to subsequent nodes. Finally, an ordered node sequence is output as the disturbance propagation path, which may include marked anomaly nodes, unmarked but similarly behaving non-anomaly nodes, and intermediate nodes introduced to ensure connectivity.

[0116] Step 5: Collect all distributed node data of the disturbance propagation path to obtain disturbance propagation path data. Calculate the feature weights of the disturbance propagation path data using the entropy weight method. Combine the disturbance propagation path data and feature weights to calculate the distribution network operation risk index and classify the risk level, thereby realizing the status monitoring of the distribution network.

[0117] In one implementation, as shown in Figure 4, all distributed node data of the disturbance propagation path are collected to obtain disturbance propagation path data RD, which includes propagation range factor, propagation speed factor, electrical influence factor, and situation factor.

[0118] Among them, the propagation range factor Propagation speed factor ,in, It represents the sum of the electrical lengths of each line segment along the disturbance propagation path, in kilometers or per-unit (pu). and These represent the initial anomaly time window numbers for the end node and the start node, respectively. Indicates the length of the time window, i.e., the physical duration. Electrical Influence Factor The number of voltage-over-limit nodes refers to the distributed nodes that exceed the voltage limit (e.g., exceed ±5%); situation factor. During peak load periods, the value is 1.0; during normal periods, it is 0.7; and during off-peak periods, it is 0.4.

[0119] collect We used historical disturbance propagation path data and performed Z-score standardization on each feature (propagation range factor, propagation speed factor, electrical influence factor, and situation factor) of the historical disturbance propagation path data to eliminate the influence of dimensional differences and value ranges, and determined the entropy value of the feature in the disturbance propagation path data:

[0120] ;

[0121] in, It is the entropy value of the r-th feature in the historical disturbance propagation path data. It is the data sample size of historical disturbance propagation path data. It refers to the first The proportion of each sample data in the r-th feature column.

[0122] Calculate the feature weights of the disturbance propagation path data:

[0123] ;

[0124] in, R is the feature weight of the r-th feature in the historical disturbance propagation path data, and R is the total number of features in the historical disturbance propagation path data.

[0125] After normalizing the disturbance propagation path data, combining the disturbance propagation path data and the feature weights, the operation risk index of the distribution network is calculated as follows:

[0126] ;

[0127] where, I is the operation risk index of the distribution network, is the r-th eigenvalue of the disturbance propagation path data; R' is the total number of features of the disturbance propagation path data.

[0128] In the state monitoring of the distribution network, according to the calculated CRI value, the risk level is mapped to a preset classification: low risk (Level1, CRI ≤ 0.3) indicates normal state and continuous monitoring is required; medium risk (Level2, 0.3 < CRI ≤ 0.7) triggers a warning, and enhanced monitoring and preparation of control measures are required; high risk (Level3, CRI > 0.7) marks an emergency state, and control measures (such as load shedding or adjusting the output of distributed power sources) need to be immediately initiated, thus forming a closed-loop management from quantitative assessment to hierarchical response to enhance the resilience and security of the distribution network.

[0129] The following further elaborates on the present invention through specific experiments;

[0130] The experimental process is first based on a simulated distribution network environment, which includes multiple distributed nodes (such as photovoltaic access points, load nodes, etc.), and uses historical operation data (including three-phase active power, three-phase reactive power, RMS voltage, RMS current, and power flow direction) to generate a simulation dataset to simulate normal disturbances, abnormal disturbances, and potential risk disturbances; collect the power flow increment trajectory feature vectors of more than 1000 time windows, and use the labeled dataset (where normal disturbances account for 40%, abnormal disturbances account for 35%, and potential risk disturbances account for 25%) to train and test the convolutional neural network model based on dual-channel and decision fusion, and at the same time simulate the disturbance propagation path through topology graph construction and neighborhood analysis; the experimental results show that the overall accuracy rate of 92.5% is achieved in the disturbance classification task, the weights of the time-domain path and the frequency-domain path are 0.55 and 0.45 respectively, and the risk index calculation can accurately divide the risk levels (such as when the CRI value is greater than 0.7, it corresponds to a high-risk event, verifying its feasibility in early warning and hierarchical response in the actual distribution network).

[0131] It should be understood that the embodiments described above are only some, not all, of the embodiments of the present invention. Furthermore, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form feasible technical solutions. Such combinations are not constrained by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0132] It should be understood that the above description of the preferred embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A method for monitoring the distribution status of a power distribution network, characterized in that, The process includes the following steps: Step 1: Collect data from distributed nodes in the distribution network to obtain distributed node data; Step 2: Extract the time domain and frequency domain features of the distributed node data to obtain the power flow incremental trajectory feature vector, including time domain features, frequency domain features, and power flow direction features. Step 3: Obtain the one-hot encoded vector and environmental feature vector corresponding to each distributed node. Input the one-hot encoded vector, environmental feature vector, and power flow incremental trajectory feature vector into a convolutional neural network model based on dual-channel and decision fusion. Output disturbance classification labels and mark abnormal monitoring nodes. One-hot encoding is performed according to the type identifier of the distribution network monitoring node to obtain the one-hot encoded vector. The convolutional neural network model based on dual-channel and decision fusion includes a time-domain path and a frequency-domain path. The input of the time-domain path is a time-domain mixed vector of dimension K1, and the input of the frequency-domain path is a frequency-domain mixed vector of dimension K2. Finally, the disturbance classification label is output based on the convolutional neural network model based on dual-channel and decision fusion, where K1 and K2 are preset values. Step 4: Construct a distribution network topology based on the distributed nodes of the distribution network. Based on the distribution network topology, neighborhood analysis is performed on the abnormal monitoring nodes to obtain a set of neighboring nodes; the similarity between the power flow increment trajectory feature vector of the abnormal monitoring node and the power flow increment trajectory feature vector of each distributed node in the set of neighboring nodes is calculated to obtain the disturbance propagation path; Step 5: Collect all distributed node data of the disturbance propagation path to obtain the disturbance propagation path data, calculate the feature weight of the disturbance propagation path data, and combine the disturbance propagation path data and feature weights to calculate the distribution network operation risk index and classify the risk level, thereby realizing the state monitoring of the distribution network.

2. The power distribution status monitoring method for a power distribution network according to claim 1, characterized in that: In step 1, the three-phase active power, three-phase reactive power, effective voltage value, effective current value, and power flow direction of the distributed nodes are synchronously collected and recorded at sampling intervals no less than the preset value.

3. The method for monitoring the distribution status of a power distribution network according to claim 1, characterized in that: In step 2, the power flow incremental trajectory feature vector FX=[ 、 、 、 、 、 、 、 ]; where, the fluctuation amplitude of the Xth feature of the tidal current increment trajectory feature vector. ， Represents the feature vector of power flow increment trajectory The maximum value of the Xth feature. Represents the trajectory of power flow increment The minimum value of the Xth feature; the measured value of each distributed node i at each sampling time t within the current time window win. and its benchmark value Perform a subtraction operation to generate the power flow increment trajectory for that monitoring node. ;in, This represents the power flow increment trajectory of the Xth feature of the data from the i-th distributed node. Let X represent the measured value of the Xth feature of the i-th distributed node data at time t. (t) represents the baseline value of the Xth feature of the i-th distributed node data at time t; the average rate of change of each feature of the power flow increment trajectory. ;in, Let N be the average rate of change of the Xth feature of the power flow increment trajectory, N be the total number of sampling points within the time window win, and k be the index of the sampling point. Indicates the k-th sampling time The increment value of the Xth feature in the +1 power flow increment trajectory. Indicates the k-th sampling time The increment value of the Xth feature in the power flow increment trajectory. The sampling time interval; the absolute sum and integral of the Xth feature of the power flow increment trajectory. Frequency domain features are extracted from the incremental tidal current trajectory. First, a discrete Fourier transform is performed on the incremental tidal current trajectory to obtain the spectrum X(f). Based on the spectrum X(f), the dominant oscillation frequency is calculated. and the proportion of dominant frequency energy Feature extraction is performed on the tidal current direction state sequence in the tidal current increment trajectory to obtain the reversal frequency. Where M is the number of direction flips. The length of the time window; extract the set of stable state duration values ​​for all power flow directions remaining unchanged within the time window win. , ,..., } Calculate the average of the durations in the set of duration values. The shortest duration in the set of duration values 。 4. The method for monitoring the distribution status of a power distribution network according to claim 1, characterized in that: In step 3, the convolutional neural network model based on dual-channel and decision fusion includes a time-domain path and a frequency-domain path. Both the time-domain and frequency-domain paths include a first convolutional layer, a first max-pooling layer, a second convolutional layer, a second max-pooling layer, a third convolutional layer, a flattening layer, and a fully connected layer, connected in sequence. The first convolutional layer is followed by a batch normalization layer and a ReLU activation function. The second convolutional layer is followed by a batch normalization layer and a ReLU activation function. The third convolutional layer is followed by a batch normalization layer and a ReLU activation function. The fully connected layer includes a first fully connected layer and a second fully connected layer, and the output layer uses the Softmax activation function to output a probability distribution vector.

5. The method for monitoring the distribution status of a power distribution network according to claim 1, characterized in that: In step 3, the final perturbation classification label is calculated based on the weighted probability voting principle. ;in, Indicates the disturbance classification label. This represents the probability that a sample output from the time-domain path belongs to class c. This represents the probability that a sample output from the frequency domain path belongs to class c. Represents the time-domain path weights. This represents the frequency domain path weight.

6. The method for monitoring the distribution status of a power distribution network according to claim 1, characterized in that: In step 4, complete network topology data of the distribution network is obtained, including the location information and connection relationships of all distributed nodes and designated key electrical nodes. Based on this, a topology graph G=(V, BB) of the distribution network is constructed, where the vertex set V represents all distributed nodes and key electrical nodes, and the edge set BB represents the electrical connection relationships between nodes. Taking the identified abnormal monitoring node as the center, a neighborhood analysis is performed along the topology graph structure according to its feeder location: searching in the electrical upstream direction, downstream direction, and bypass feeders connected through closed tie switches. The physical neighborhood range is defined using topological distance as a metric, and all monitoring nodes within a preset topological distance are defined as the direct neighbor of the abnormal node, finally obtaining the set of neighboring nodes.

7. The method for monitoring the distribution status of a power distribution network according to claim 1, characterized in that: In step 4, the similarity between the power flow incremental trajectory feature vector of the anomaly monitoring node and the power flow incremental trajectory feature vector of each distributed node in the neighborhood node set is calculated using a dynamic time warping algorithm based on local extremum shapes. Specifically, the anomaly monitoring node is used as the initial anomaly node, and its initial anomaly window is used as the starting point. Feature data from subsequent windows (WIN windows) are continuously extracted to form the sequence Source. After standardization, the calculation The L2 norm of the vector under each time window yields the initial scalar sequence of outlier nodes over WIN time windows. =[ This process extracts local maxima and local minima from the initial scalar sequence of outlier nodes; finally, it outputs the index sequence corresponding to the extreme points in the initial scalar sequence of outlier nodes. ； For each extreme point index ,by Centered on the initial abnormal node scalar sequence Extract a subsequence of length b. ， Use the HOG-1D function to process subsequences. Finally, the shape descriptor vector MS = [h1, h2, ..., hb, ..., hB], where hb represents the normalized value of the b-th interval, then the shape descriptor vector for each extreme point is... The final output is a comprehensive shape descriptor for the anomaly monitoring points. =[ , ,..., ,…, Similarly, calculate the comprehensive shape descriptor for each distributed node in the neighborhood node set. ; Calculate the similarity between the anomaly monitoring node and each distributed node in the neighboring node set: ;in, This represents the similarity between the initial abnormal node and the i-th distributed node in the set of neighboring nodes. This is the composite shape descriptor for the initial abnormal node. It is the comprehensive shape descriptor of the i-th distributed node in the neighborhood node set.

8. The method for monitoring the distribution status of a power distribution network according to claim 1, characterized in that: In step 5, collect After standardizing the historical perturbation propagation path data using Z-Score, the entropy values ​​of the features in the perturbation propagation path data are determined based on the entropy values: ;in, It is the entropy value of the r-th feature in the historical disturbance propagation path data. It is the data sample size of historical disturbance propagation path data. It refers to the first The proportion of each sample data point in the r-th feature column; calculate the feature weights of the perturbation propagation path data: ;in, R is the feature weight of the r-th feature in the historical disturbance propagation path data, and R is the total number of features in the historical disturbance propagation path data.

9. The method for monitoring the distribution status of a power distribution network according to any one of claims 1-8, characterized in that: In step 5, after standardizing the disturbance propagation path data, the disturbance propagation path data and feature weights are combined. The following operational risk indicators for the power distribution network were calculated: ;in, I represents the risk indicator for power distribution network operation. R' represents the r-th feature value of the disturbance propagation path data; R' is the total number of features in the disturbance propagation path data.

10. A power distribution network status monitoring system, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the power distribution status monitoring method for a power distribution network as described in any one of claims 1 to 9.

Images