A multi-modal aggregation digital power supply service command terminal and method

Abstract

The present invention relates to the technical field of power distribution network, and discloses a multi-modal aggregation digital power supply service command terminal and method, wherein the multi-modal aggregation digital power supply service command terminal comprises: a data acquisition module, which acquires observation signals of transmission lines in real time; a data processing module, which estimates a mixed matrix and an independent source signal by using an independent component analysis algorithm; an event discrimination module, which is used to output the probability of an event category of a corresponding independent source signal; and an early warning scheme formulation module, which formulates an early warning scheme. The present invention can obtain bird activity signals appearing on power grid lines within a set time period, judge whether the corresponding bird activities will cause damage to the power grid lines, and formulate corresponding prevention schemes based on the judgment information, so as to prevent faults that may be caused by bird activities in a targeted manner, improve power grid stability, and reduce power grid maintenance costs.

Description

Technical Field

[0001] The present invention relates to the technical field of power distribution networks, and more specifically, to a multi-modal aggregated digital power supply service command terminal and method. Background Art

[0002] In order to improve the efficiency and reliability of power supply and optimize customer service experience, it is necessary to monitor, analyze and dispatch grid resources in real time to ensure a stable supply of electricity and support rapid response in emergencies.

[0003] There are many kinds of emergencies, among which the conflicts between power transmission lines and bird activities are becoming increasingly prominent. In open areas with sparse trees (especially grasslands, plateaus, etc.), power transmission lines and distribution towers can provide birds with scarce "high branches" with good views, which are crucial for the safety of nesting and the convenience of foraging. Even in forest areas, tall electric towers stand out in the woods and are favored. Birds' nesting, defecation, resting, fighting and other activities on high-voltage transmission lines may cause high-voltage transmission lines to trip and power outages. Common bird-repelling facilities cover a small area and cannot fully cover all power grid areas. This also causes power grid lines to often fail due to bird activities. How to specifically avoid power grid failures caused by bird activities is an urgent problem to be solved. Summary of the invention

[0004] The purpose of the present invention is to provide a multi-modal aggregated digital power supply service command terminal and method in order to solve the above-mentioned problems.

[0005] The present invention provides a multi-modal aggregated digital power supply service command method, which is characterized by comprising the following steps:

[0006] Step 100: collecting observation signals of the transmission line in real time. M observation points are set on the transmission line. Each observation point collects observation signals of a line segment on the transmission line. The observation signals include current signals and voltage signals.

[0007] Step 200: construct an observation signal model;

[0008] The observed signal model is expressed as:

[0009] ;

[0010] in, represents the mixing matrix, which is an M×M matrix, representing the propagation characteristics from the signal source to the observation point. represents M independent source signals at the tth moment;

[0011] ;

[0012] in, , and Respectively represent the 1st, 2nd and Mth independent source signals at the tth moment;

[0013] represents the observation signals collected from M observation points;

[0014] ;

[0015] in, , and They represent the first, second and Mth observation signals of the transmission line at the tth moment respectively;

[0016] Step 300: Use independent component analysis algorithm to estimate the mixing matrix A and independent source signals ;

[0017] Step 400: Input the independent source signal within the set time period into the event discrimination model, and the event discrimination model outputs the probability of the event category of the corresponding independent source signal, and the event categories include bird electric shock, bird nesting and bird resting on the power line;

[0018] Step 500: determine the fault type of the line segment of the transmission line where the corresponding observation point is located based on the probability of the event category of the independent source signal output by the event discrimination model, and take the event category with the highest probability of the i-th independent source signal as the fault type of the i-th line segment, and then formulate an early warning plan for the fault type of the line segment.

[0019] As a further optimization scheme of the present invention, the independent component analysis algorithm is used to estimate the mixing matrix A and the independent source signal The method comprises the following steps:

[0020] Step S301: observe the signal Centralized processing;

[0021] Step S302: Observation signal after centralization Perform whitening treatment;

[0022] Step S303: Randomly initialize the separation matrix , separation matrix It is an M×M matrix, each row represents an independent component The initial weight of

[0023] Step S304: Separation matrix Perform fixed point iteration, the iteration process is:

[0024] a. Separation matrix The i-th independent component in The weight vector Perform standardization, the standardization formula is:

[0025] ;

[0026] in, represents the i-th column vector of the separation matrix, Represents the weight vector The norm of Represents the weight vector after normalization;

[0027] b. Weight vector Perform fixed point iteration, the iteration formula is:

[0028] ;

[0029] in, represents the observed signal after whitening processing, represents the expectation operator, represents transpose, represents the projection value of the i-th independent component, Represents the iterative update ;

[0030] Repeat steps a and b until Convergence, the convergence condition is:

[0031] ;

[0032] in, Represents the convergence threshold. The value range of the convergence threshold is [10 -6 , 10 -3 ];

[0033] Step S305: As a separation matrix A row of ;

[0034] ;

[0035] Step S306, repeat steps S303-S305 to extract all independent components;

[0036] The extracted i-th independent component is used as the i-th independent source signal at the t-th moment;

[0037] ;

[0038] Where W represents the separation matrix after iteration, Represents the inverse matrix of the iterated separating matrix.

[0039] As a further optimization scheme of the present invention, the centralized observation signal is whitened, specifically:

[0040] ;

[0041] in, represents the observed signal after centralized processing. The covariance matrix of , E is The eigenvector matrix of The eigenvalue diagonal matrix of .

[0042] As a further optimization solution of the present invention, the event discrimination model includes a CNN model;

[0043] The CNN model includes:

[0044] Input layer: The ith independent source signal within the set time period is sampled in time, and each sampling forms a one-dimensional vector sequence;

[0045] Convolution layer: Through one-dimensional convolution operation, the local features of the signal are extracted and the feature map is generated;

[0046] Pooling layer: downsample the feature map through maximum pooling or average pooling operations;

[0047] Fully connected layer: Flattens the feature map after convolution and pooling to form a feature vector, and passes it through one or more fully connected layers to achieve nonlinear combination and classification of features;

[0048] Output layer: Through the softmax activation function, the probability of the event category of the i-th independent source signal is output.

[0049] As a further optimization solution of the present invention, the event discrimination model includes an RNN model;

[0050] The RNN model includes:

[0051] Input layer: The ith independent source signal within the set time period is sampled in time to form a sequence. A sequence unit of the sequence represents the ith independent source signal at a sampling moment. A sequence unit is used as an input vector, and each time step corresponds to an input vector.

[0052] Recurrent layer: Through recurrent connections, the hidden state is updated at each time step to capture the temporal dynamic characteristics of the signal;

[0053] Output layer: At the last time step of the sequence, the hidden state is converted into output through a fully connected layer and a softmax activation function.

[0054] As a further optimization solution of the present invention, the event discrimination model includes a CNN model, an RNN model, a feature fusion layer and an output layer;

[0055] Feature fusion layer:

[0056] Concatenate the output features of the CNN model and the RNN model;

[0057] ;

[0058] represents the hidden state of the last time step of the RNN, Represents the output features of the pooling layer of the CNN model, Represents vector concatenation;

[0059] Output layer:

[0060] Nonlinear transformation is performed through fully connected layers and activation functions:

[0061] ;

[0062] Among them, FC(o) represents the concatenated features after the transformation of the fully connected layer, and h represents the intermediate features of the concatenated features FC(o) after the transformation of the fully connected layer after the transformation of the ReLU activation function;

[0063] The probability of event category is output through the Softmax layer:

[0064] ;

[0065] Among them, FC(h) represents the intermediate features after the full connection layer and activation function transformation, and p represents the probability of each event category output by the Softmax layer.

[0066] As a further optimization scheme of the present invention, a method for formulating an early warning scheme includes: if the event category of the i-th independent source signal is bird electric shock, it means that a bird electric shock fault event occurs on the corresponding line section of the transmission line where the i-th observation point is located, and a early warning scheme is formulated to add bird repellent equipment in the corresponding line section area;

[0067] If the event category of the i-th independent source signal is bird nesting, it means that a bird nesting fault event occurs on the corresponding line section of the transmission line where the i-th observation point is located, and an early warning plan is formulated to add artificial bird nests in the corresponding line section area;

[0068] As a further optimization scheme of the present invention, the method for formulating an early warning scheme also includes: if the event category of the i-th independent source signal is birds resting on the transmission line, it means that a fault event of birds resting on the transmission line has occurred on the corresponding line segment of the transmission line where the i-th observation point is located, and an early warning scheme is formulated to inspect and clean the corresponding line segment.

[0069] A multi-modal aggregated digital power supply service command terminal, which is used to execute a multi-modal aggregated digital power supply service command method as described above, comprising:

[0070] The data acquisition module collects the observation signals of the transmission line in real time. M observation points are set on the transmission line. Each observation point collects the observation signals of a line section on the transmission line. The observation signals include current signals and voltage signals.

[0071] The data processing module builds the observation signal model based on the real-time acquisition of the observation signal on each transmission line, and then uses the independent component analysis algorithm to estimate the mixing matrix A and the independent source signal ;

[0072] An event discrimination module is used to construct an event discrimination model, and then all independent source signals are input into the event discrimination model. The event discrimination model outputs the probability of the event category of the corresponding independent source signal. The event categories include birds being electrocuted, birds building nests, and birds resting on power lines.

[0073] The early warning plan formulation module determines the fault type of the line section of the transmission line where the corresponding observation point is located based on the probability of the event category of the corresponding independent source signal output by the event discrimination model, and formulates an early warning plan.

[0074] A storage medium storing non-temporary computer-readable instructions, which, when executed by a computer, can execute the steps in a multi-modal aggregated digital power supply service command method as described above.

[0075] The beneficial effects of the present invention are as follows: the present invention can extract electrical parameter change signals related to bird activities from electrical signals with a large amount of noise interference and background fluctuations, and judge the corresponding bird activity classification according to the change characteristics, obtain the bird activity signals appearing on the power grid line within a set time period, judge whether the corresponding bird activities will cause damage to the power grid line, and formulate corresponding prevention plans based on the judgment information, so as to prevent faults that may be caused by bird activities in a targeted manner, improve the stability of the power grid, and reduce the maintenance cost of the power grid. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 It is a module block diagram of a multi-modal aggregated digital power supply service command terminal of the present invention;

[0077] Figure 2 It is a flow chart of a multi-modal aggregated digital power supply service command method of the present invention.

[0078] In the figure: 1. Data acquisition module; 2. Data processing module; 3. Event identification module; 4. Early warning plan formulation module. DETAILED DESCRIPTION

[0079] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that the discussion of these embodiments is only intended to enable those skilled in the art to better understand and implement the subject matter described herein. In addition, the features described relative to some examples may also be combined in other examples.

[0080] like Figure 1 As shown, a multi-modal aggregated digital power supply service command terminal includes:

[0081] The data acquisition module 1 collects the observation signals of the transmission line in real time. M observation points are set on the transmission line. Each observation point collects the observation signals of a line section on the transmission line. The observation signals include current signals and voltage signals.

[0082] Data processing module 2 builds an observation signal model based on real-time acquisition of observation signals on each transmission line, and then uses independent component analysis algorithm to estimate the mixing matrix A and independent source signals. ;

[0083] The event discrimination module 3 is used to construct an event discrimination model, and then all independent source signals are input into the event discrimination model, and the event discrimination model outputs the probability of the event category of the corresponding independent source signal, and the event categories include birds being electrocuted, birds building nests, and birds resting on power lines;

[0084] The early warning plan formulation module 4 determines the fault type of the line section of the transmission line where the corresponding observation point is located based on the probability of the event category of the corresponding independent source signal output by the event discrimination model, and formulates an early warning plan.

[0085] like Figure 2 As shown, a multi-modal aggregated digital power supply service command method includes the following steps:

[0086] Step 100: collecting observation signals of the transmission line in real time. M observation points are set on the transmission line. Each observation point collects observation signals of a line segment on the transmission line. The observation signals include current signals and voltage signals.

[0087] The observation signal of a transmission line contains the observation signals of M observation points. Then the observation signal of M observation points on a transmission line is , , and They respectively represent the first, second and Mth observation signals of the transmission line at the tth moment.

[0088] An example of an observed signal is: ;

[0089] in, represents the voltage signal observed at the first observation point at the tth moment, Represents the current signal observed at the first observation point at the tth moment;

[0090] The observation signal may include more types of signals besides voltage signals and current signals.

[0091] The equipment for observing current and voltage signals on transmission lines usually includes current transformers and voltage transformers, which are used to collect current and voltage signals on transmission lines. The collected signals will be transmitted to a monitoring center or data center through data transmission equipment for real-time monitoring of the operating status, load conditions and fault detection of the power grid.

[0092] Step 200: construct an observation signal model;

[0093] The observed signal model is expressed as:

[0094] ;

[0095] in, represents the mixing matrix, which is an M×M matrix, representing the propagation characteristics from the signal source to the observation point. represents M independent source signals at the tth moment;

[0096] ;

[0097] in, , and Respectively represent the 1st, 2nd and Mth independent source signals at the tth moment;

[0098] represents the observation signals collected from M observation points;

[0099] ;

[0100] in, , and They respectively represent the first, second and Mth observation signals of the transmission line at the tth moment.

[0101] Step 300: Use independent component analysis algorithm to estimate the mixing matrix A and independent source signals ;

[0102] The specific steps include:

[0103] Step S301: observe the signal Perform centralization, that is, subtract its mean; this step is to simplify subsequent calculations;

[0104] Step S302: performing whitening processing on the centralized observation signal;

[0105] The commonly used whitening method is principal component analysis (PCA). Assume that the observed signal after centering is , whose covariance matrix is ;

[0106] The signal after whitening is:

[0107]

[0108] Where E is The eigenvector matrix of The eigenvalue diagonal matrix of ;

[0109] Step S303: Randomly initialize the separation matrix , separation matrix It is an M×M matrix, each row represents an independent component The initial weight of

[0110] Step S304: Separation matrix Perform fixed point iteration, the iteration process is:

[0111] a. Separation matrix The i-th independent component in The weight vector Perform standardization, the standardization formula is:

[0112] ;

[0113] in, represents the i-th column vector of the separation matrix, Represents the weight vector The norm of Represents the weight vector after normalization;

[0114] b. Weight vector Perform fixed point iteration, the iteration formula is:

[0115] ;

[0116] in, represents the observed signal after whitening processing, represents the expectation operator, represents transpose, represents the projection value of the i-th independent component, Represents the iterative update ;

[0117] Repeat steps a and b until Convergence, the convergence condition is:

[0118] ;

[0119] in, Represents the convergence threshold. The value range of the convergence threshold is [10 -6 , 10 -3 ];

[0120] Step S305: As a separation matrix A row of ;

[0121] ;

[0122] Step S306, repeat steps S303-S305 to extract all independent components;

[0123] The extracted i-th independent component is used as the i-th independent source signal at the t-th moment;

[0124] ;

[0125] Where W represents the separation matrix after iteration, Represents the inverse matrix of the iterated separating matrix.

[0126] Here is a simple example:

[0127] Suppose we observe two mixed signals and , which are composed of two independent source signals and The mixing process can be expressed as:

[0128]

[0129]

[0130] in, is an unknown mixing matrix;

[0131] Our goal is to use the independent component analysis (ICA) algorithm to and Separate and ;

[0132] First of all, and After centralization and whitening, we get and ;

[0133] Then, randomly initialize the separation matrix , separation matrix It is an M×M matrix, each row represents an independent component The initial weights of the separation matrix Perform fixed point iteration, the iteration process is:

[0134] a. Separation matrix The i-th independent component in The weight vector Perform standardization, the standardization formula is:

[0135] ;

[0136] in, represents the i-th column vector of the separation matrix, Represents the weight vector The norm of Represents the weight vector after normalization;

[0137] b. Weight vector Perform fixed point iteration, the iteration formula is:

[0138] ;

[0139] in, represents the observed signal after whitening processing, , represents the expectation operator, represents transpose, represents the projection value of the i-th independent component, Represents the iterative update ;

[0140] Repeat steps a and b until Convergence, the convergence condition is:

[0141] ;

[0142] in, Represents the convergence threshold. The value range of the convergence threshold is [10 -6 , 10-3 ];

[0143] After the convergence As a separation matrix A row of ;

[0144] ;

[0145] Repeat the above process until all independent components are extracted, and take the first independent component extracted as , the second independent component extracted is taken as , so we have a mixed signal and Separate the independent source signals and .

[0146] In one embodiment of the present invention, if the independent source signal is sparse in a certain transform domain (such as wavelet domain, short-time Fourier transform domain, etc.), sparse decomposition can be used to perform blind source separation.

[0147] Commonly used sparse decomposition algorithms include matching pursuit (MP), basis pursuit (BP), etc. Taking MP as an example, its basic steps are:

[0148] a. Construct an over-complete atomic dictionary (such as Gabor dictionary) to represent independent source signals;

[0149] b. Find the atom that best matches the residual signal in the dictionary and calculate its coefficient;

[0150] c. Approximate the independent source signal with matching atoms and update the residual;

[0151] d. Repeat steps b and c until the sparsity requirement is met.

[0152] Through sparse decomposition, sparse representation of independent source signals can be obtained, thereby achieving separation of independent source signals.

[0153] Step 400: Input the independent source signal within the set time period into the event discrimination model, and the event discrimination model outputs the probability of the event category of the corresponding independent source signal, and the event categories include bird electric shock, bird nesting and bird resting on the power line;

[0154] Set the length of the time period to 10 to 120 seconds, the default value is 30 seconds;

[0155] The event discrimination model includes:

[0156] CNN branches:

[0157] In the recognition of electric shock and landing signals, we can regard the estimated independent source signals as a one-dimensional time series and classify them using a one-dimensional CNN. A typical one-dimensional CNN model can include the following parts:

[0158] a. Input layer: The i-th independent source signal within the set time period is sampled in time, and each sampling forms a one-dimensional vector sequence.

[0159] b. Convolution layer: Through one-dimensional convolution operation, the local features of the signal are extracted and feature maps are generated. The size of the convolution kernel determines the range of the receptive field, and the number of convolution kernels determines the number of feature maps. By setting the appropriate size and number of convolution kernels, the features of the signal at different scales and in different aspects can be captured.

[0160] The local characteristics of the signal include:

[0161] Amplitude characteristics:

[0162] Maximum value: the maximum amplitude of the signal;

[0163] Minimum value: the minimum amplitude of the signal;

[0164] Peak-to-peak value: the dynamic range of the signal;

[0165] Mean: DC component of the signal;

[0166] Absolute mean: reflects the average amplitude of the signal;

[0167] Root mean square: reflects the effective value of the signal.

[0168] Duration Features:

[0169] Duration: The length of time that a signal exceeds a certain threshold, which can be calculated using a threshold function;

[0170] Rise time: The time required for a signal to rise from 10% of its maximum value to 90% of its maximum value;

[0171] Fall Time: The time it takes for a signal to fall from 90% of its maximum value to 10% of its maximum value.

[0172] Energy characteristics:

[0173] Energy: the total energy of the signal;

[0174] Average Energy: The average power of the signal.

[0175] High-level statistical features:

[0176] Variance: reflects the degree of fluctuation of the signal;

[0177] Skewness: reflects the asymmetry of the signal;

[0178] Kurtosis: reflects the non-Gaussian nature of the signal;

[0179] Power Spectral Density:

[0180] The power spectral density (PSD) reflects the power distribution of the signal at different frequencies and can be calculated by the square of the modulus of the spectrum;

[0181] Frequency domain energy characteristics:

[0182] Total energy: The total energy of the signal, which should be equal to the time domain energy;

[0183] Band energy: the energy of the signal in a specific frequency band;

[0184] Frequency domain entropy: reflects the uniformity of the spectrum;

[0185] Frequency domain statistical characteristics:

[0186] Center frequency: Normalize the PSD to get the probability distribution, and then calculate its mean, which reflects the main frequency components of the signal;

[0187] Bandwidth: Normalize the PSD to get the probability distribution, and then calculate its variance to reflect the dispersion of the signal;

[0188] c. Pooling layer: Through maximum pooling or average pooling operations, the feature map is downsampled to reduce the size of the data while retaining the most significant features.

[0189] d. Fully connected layer: Flatten the feature map after convolution and pooling to form a feature vector, and pass it through one or more fully connected layers to achieve nonlinear combination and classification of features.

[0190] e. Output layer: Through the softmax activation function, the probability of each category is output.

[0191] Methods for training event discrimination models:

[0192] Different types of events will show different physical characteristics in independent source signals; this prior knowledge can be used to classify the local features of the estimated independent source signals.

[0193] Electric shock signal: When a bird is electrocuted, it comes into direct contact with a high-voltage wire, which generates a transient high-amplitude current pulse. This pulse usually lasts for a short time (milliseconds), but has a high amplitude, much higher than the current level during normal operation. In the frequency domain, the energy of the electric shock signal is mainly concentrated in the high frequency band.

[0194] Landing signal: When a bird lands on a line, the equivalent capacitance of the line changes due to the conductivity of the bird's body, which in turn causes a slow change in voltage and current. This change usually lasts for a long time (seconds) and has a small amplitude. In the frequency domain, the energy of the landing signal is mainly concentrated in the low frequency band.

[0195] Other signals: In addition to electric shock and grounding, other events may occur on the line, such as wind, rain, etc. These events will also cause changes in electrical parameters, but the change pattern is different from that of electric shock and grounding. For example, the vibration caused by wind is usually periodic, and the change caused by rain is usually related to the intensity of rainfall;

[0196] Based on these physical characteristics, the local features of the estimated independent source signals can be preliminarily classified, such as electric shock and landing. Based on the number of occurrences of local features classified as bird landing within a set time period, bird behaviors can be further judged, such as bird nesting and bird resting on power lines.

[0197] For example, ornithologists have made precise records that a pair of gray magpies took nesting materials 666 times in four or five days of nesting, including 253 dead branches, 154 green leaves, 123 grass roots, 82 cow and wool, and 54 mud balls. After averaging, the minimum number of nesting materials taken per day is 133 times. The set time period is defined as 24 hours. If the local features related to bird activities appear more than 100 times in the independent source signals collected by a certain collection point within 24 hours, it is highly likely that birds have built nests at the power grid line corresponding to the collection point, and a prevention plan can be formulated for this section of the line.

[0198] If a single landing time exceeds 20 minutes, and the number of landing signals is less than 30 times within the set time period, it can be determined that birds have landed to rest on the route;

[0199] Whether building nests or resting on transmission lines, birds may defecate, and the feces excreted by birds will corrode the transmission lines. If they are not cleaned and intervened manually, the transmission lines may be damaged due to corrosion and discharge short circuit failures may occur.

[0200] Example: Assume we have a set of independent source signals of electric shock and landing, each signal contains 1000 sampling points. We can design a CNN model as follows:

[0201] Input layer: 1000 nodes, corresponding to 1000 sampling points of independent source signals.

[0202] Convolutional layer 1: 32 convolution kernels of size 10, stride 1, ReLU activation function.

[0203] Pooling layer 1: max pooling of size 5 and stride 2.

[0204] Convolutional layer 2: 64 convolution kernels of size 5, stride 1, ReLU activation function.

[0205] Pooling layer 2: max pooling of size 5 and stride 2.

[0206] Fully connected layer: 128 nodes, ReLU activation function.

[0207] Output layer: 2 nodes, softmax activation function, corresponding to the two categories of electric shock and landing.

[0208] By training this CNN model, it can automatically learn the distinguishing features of electric shock and landing signals, and effectively classify new independent source signals. The training set can be obtained from the power grid fault database.

[0209] RNN is a deep learning model specifically designed to process sequence data. Unlike CNN, RNN can process variable-length sequences and capture long-term dependencies in sequences. RNN introduces a recurrent connection at each time step of the sequence, so that the output of the current time step depends not only on the current input, but also on the state of the previous time step.

[0210] In the recognition of electric shock and landing signals, we can use RNN to model independent source signals and capture their temporal dynamic characteristics. A typical RNN model can include the following parts:

[0211] a. Input layer: The i-th independent source signal within the set time period is sampled in time to form a sequence. A sequence unit of the sequence represents the i-th independent source signal at a sampling moment. A sequence unit serves as an input vector, and each time step corresponds to an input vector.

[0212] b. Recurrent layer: Through recurrent connections, the hidden state is updated at each time step to capture the temporal dynamic characteristics of the signal. Commonly used recurrent units include simple RNN, LSTM, GRU, etc. They can effectively alleviate the gradient vanishing and gradient exploding problems and achieve long-term memory.

[0213] c. Output layer: At the end of each time step or sequence, the hidden state is converted into output through a fully connected layer and an activation function, which can be the category probability of each time step or the category probability of the entire sequence.

[0214] Example: Assume that we have a set of independent source signals of electric shock and landing, each signal contains sampling points of variable length. We can design an RNN model as follows:

[0215] Input layer: Each time step corresponds to a sampling point, and the dimension is 1.

[0216] Recurrent layer: Use LSTM units with a hidden state dimension of 64.

[0217] Output layer: At the last time step of the sequence, the probabilities of the two categories of electric shock and landing are output through the fully connected layer and softmax activation function.

[0218] By training this RNN model, it can automatically learn the temporal evolution patterns of electric shock and landing signals, and effectively classify independent source signals of indefinite length.

[0219] In one embodiment of the present invention, the aforementioned CNN model and RNN model are combined to construct a combined model, and the combined model includes, in addition to the CNN model and the RNN model:

[0220] Feature fusion layer:

[0221] Concatenate the output features of CNN and RNN.

[0222]

[0223] represents the hidden state of the last time step of the RNN, represents the output features of the pooling layer 2 of the CNN model, Represents vector concatenation.

[0224] Output layer:

[0225] Nonlinear transformation is performed through fully connected layers and activation functions:

[0226] Among them, FC(o) represents the concatenated features after the transformation of the fully connected layer, and h represents the intermediate features of the concatenated features FC(o) after the transformation of the fully connected layer after the transformation of the ReLU activation function;

[0227] The probability of event category is output through the Softmax layer:

[0228] Among them, FC(h) represents the intermediate features after the full connection layer and activation function transformation, and p represents the probability of each event category output by the Softmax layer.

[0229] Step 500: Determine the fault type of the line section of the transmission line where the corresponding observation point is located based on the probability of the event category of the corresponding independent source signal output by the event discrimination model, and formulate an early warning plan;

[0230] Specifically, the event category of the i-th independent source signal represents the event category of the i-th observation point, and the event category of the i-th observation point represents the fault type of the line segment of the transmission line where the observation point is located; thus, the fault position and fault type on the transmission line can be located;

[0231] If the event category of the i-th independent source signal is bird electrocution, it means that a bird electrocution fault event has occurred on the corresponding line section of the transmission line where the i-th observation point is located, and bird-repelling equipment, such as bird-repelling audio playback equipment, should be added in the corresponding line section area;

[0232] If the event category of the i-th independent source signal is bird nesting, it means that a bird nesting fault event occurs on the corresponding line section of the transmission line where the i-th observation point is located. Artificial bird nests should be added in the corresponding line section area to prevent the transmission line discharge short circuit fault caused by birds nesting on the transmission line;

[0233] The causes of line faults caused by nesting behavior are: the nesting materials brought back by birds are wet or rain wets the nesting materials, causing discharge of the transmission line;

[0234] At the same time, the transmission line can be cleaned to prevent the droppings discharged by birds from falling on the transmission line and corroding the transmission line, thereby preventing the discharge short circuit fault caused by the corrosion of the transmission line;

[0235] If the event category of the i-th independent source signal is birds resting on the transmission line, it means that a fault event of birds resting on the transmission line has occurred on the corresponding line section of the transmission line where the i-th observation point is located. The corresponding line section should be inspected and cleaned to prevent the droppings discharged by the birds from falling on the transmission line and causing corrosion to the transmission line.

[0236] Based on event judgment and corresponding to the corresponding observation points, the maintenance accuracy and efficiency of power grid lines can be effectively improved, the failure rate can be reduced, and the cost of large-scale manual maintenance can also be reduced.

[0237] The present invention also provides a storage medium storing non-temporary computer-readable instructions. When the non-temporary computer-readable instructions are executed by a computer, the steps in the multi-modal aggregated digital power supply service command method as described above can be executed.

[0238] The present embodiment is described above, but the present embodiment is not limited to the above-mentioned specific implementation manner. The above-mentioned specific implementation manner is merely illustrative and not restrictive. Under the guidance of the present embodiment, ordinary technicians in this field can also make many forms, which are all within the protection of the present embodiment.

Claims

1. A multi-modal aggregated digital power supply service command method, characterized in that: The following steps are involved: Step 100: collecting observation signals of the transmission line in real time. M observation points are set on the transmission line. Each observation point collects observation signals of a line segment on the transmission line. The observation signals include current signals and voltage signals. Step 200: construct an observation signal model; The observed signal model is expressed as: ； in, represents the mixing matrix, which is an M×M matrix, representing the propagation characteristics from the signal source to the observation point. represents M independent source signals at the tth moment; ； in, , and Respectively represent the 1st, 2nd and Mth independent source signals at the tth moment; Represents the observation signals collected from M observation points; ； in, , and They represent the first, second and Mth observation signals of the transmission line at the tth moment respectively; Step 300: Use independent component analysis algorithm to estimate the mixing matrix A and independent source signals ; The following steps are involved: Step S301: observe the signal Centralized processing; Step S302: Observation signal after centralization Perform whitening treatment; Step S303: Randomly initialize the separation matrix , separation matrix It is an M×M matrix, each row represents an independent component The initial weight of Step S304: Separation matrix Perform fixed point iteration, the iteration process is: a. Separation matrix The i-th independent component in The weight vector Perform standardization, the standardization formula is: ； in, represents the i-th column vector of the separation matrix, Represents the weight vector The norm of Represents the weight vector after normalization; b. Weight vector Perform fixed point iteration, the iteration formula is: ； in, represents the observed signal after whitening processing, represents the expectation operator, represents transpose, represents the projection value of the i-th independent component, Represents the iterative update ; Repeat steps a and b until Convergence, the convergence condition is: ； in, Represents the convergence threshold, and the value range of the convergence threshold is [10 -6 , 10 -3 ]; Step S305: As a separation matrix A row of ; ； Step S306, repeat steps S303-S305 to extract all independent components; The extracted i-th independent component is used as the i-th independent source signal at the t-th moment; ； Where W represents the separation matrix after iteration, represents the inverse matrix of the iterative separation matrix; The centralized observation signal is whitened as follows: ； in, represents the observed signal after centralized processing. The covariance matrix of , E is The eigenvector matrix of The eigenvalue diagonal matrix of ; Step 400: Input the independent source signal within the set time period into the event discrimination model, and the event discrimination model outputs the probability of the event category of the corresponding independent source signal, and the event categories include bird electric shock, bird nesting and bird resting on the power line; Step 500: determine the fault type of the line segment of the power transmission line where the corresponding observation point is located based on the probability of the event category of the independent source signal output by the event discrimination model, and use the event category with the highest probability of the i-th independent source signal as the fault type of the i-th line segment, and then formulate an early warning plan for the fault type of the line segment; The event discrimination model includes CNN model, RNN model, feature fusion layer and output layer; The CNN model includes: Input layer: The ith independent source signal within the set time period is sampled in time, and each sampling forms a one-dimensional vector sequence; Convolution layer: Through one-dimensional convolution operation, the local features of the signal are extracted and the feature map is generated; Pooling layer: downsample the feature map through maximum pooling or average pooling operations; Fully connected layer: Flattens the feature map after convolution and pooling to form a feature vector, and passes it through one or more fully connected layers to achieve nonlinear combination and classification of features; Output layer: Through the softmax activation function, the probability of the event category of the i-th independent source signal is output; The RNN model includes: Input layer: The ith independent source signal within the set time period is sampled in time to form a sequence. A sequence unit of the sequence represents the ith independent source signal at a sampling moment. A sequence unit is used as an input vector, and each time step corresponds to an input vector. Recurrent layer: Through recurrent connections, the hidden state is updated at each time step to capture the temporal dynamic characteristics of the signal; Output layer: At the last time step of the sequence, the hidden state is converted into output through a fully connected layer and a softmax activation function; Feature fusion layer: Concatenate the output features of the CNN model and the RNN model; ； represents the hidden state of the last time step of the RNN, Represents the output features of the pooling layer of the CNN model, Represents vector concatenation; Output layer: Nonlinear transformation is performed through fully connected layers and activation functions: ; Among them, FC(o) represents the concatenated features after the transformation of the fully connected layer, and h represents the intermediate features of the concatenated features FC(o) after the transformation of the fully connected layer after the transformation of the ReLU activation function; The probability of event category is output through the Softmax layer: ； Among them, FC(h) represents the intermediate features after the full connection layer and activation function transformation, and p represents the probability of each event category output by the Softmax layer.

2. A multi-modal aggregated digital power supply service command method according to claim 1, characterized in that: The method for formulating an early warning plan includes: if the event category of the i-th independent source signal is bird electrocution, it means that a bird electrocution fault event occurs on the corresponding line section of the transmission line where the i-th observation point is located, and a warning plan is formulated to add bird repellent equipment in the corresponding line section area; If the event category of the i-th independent source signal is bird nesting, it means that a bird nesting fault event occurs on the corresponding line section of the transmission line where the i-th observation point is located, and an early warning plan is formulated to add artificial bird nests in the corresponding line section area.

3. A multi-modal aggregated digital power supply service command method according to claim 1, characterized in that: The method for formulating an early warning plan also includes: if the event category of the i-th independent source signal is birds resting on the transmission line, it means that a fault event of birds resting on the transmission line has occurred on the corresponding line segment of the transmission line where the i-th observation point is located, and an early warning plan is formulated to inspect and clean the corresponding line segment.

4. A multi-modal aggregated digital power supply service command terminal, characterized in that: It is used to execute a multi-modal aggregated digital power supply service command method as described in any one of claims 1 to 3, comprising: The data acquisition module collects the observation signals of the transmission line in real time. M observation points are set on the transmission line. Each observation point collects the observation signals of a line section on the transmission line. The observation signals include current signals and voltage signals. The data processing module builds the observation signal model based on the real-time acquisition of the observation signal on each transmission line, and then uses the independent component analysis algorithm to estimate the mixing matrix A and the independent source signal ; An event discrimination module is used to construct an event discrimination model, and then all independent source signals are input into the event discrimination model. The event discrimination model outputs the probability of the event category of the corresponding independent source signal. The event categories include birds being electrocuted, birds building nests, and birds resting on power lines. The early warning plan formulation module determines the fault type of the line section of the transmission line where the corresponding observation point is located based on the probability of the event category of the corresponding independent source signal output by the event discrimination model, and formulates an early warning plan.

5. A storage medium, characterized in that: It stores non-temporary computer-readable instructions, and when the non-temporary computer-readable instructions are executed by a computer, it can execute the steps in a multimodal aggregated digital power supply service command method as described in any one of claims 1-3.

Images