Distributed power supply management and control method and system based on time division multiplexing

Abstract

The application relates to the technical field of distributed power supply management and control, and discloses a distributed power supply management and control method and system based on time division multiplexing. The method comprises the following steps: collecting distributed power supply parameters and performing missing detection, abnormality identification and redundancy filtering to obtain a data quality score; allocating a communication time slot according to the quality score and positioning high abnormality rate parameters and improving the sampling frequency thereof to obtain multi-frequency heterogeneous data when the quality score is lower than a threshold value; inputting the high-frequency data and the low-frequency data into a time sequence convolution network and a full connection network respectively to extract feature vectors, and then calculating a correlation weight matrix through a cross attention mechanism to obtain time scale decoupling mixed features; separating the mixed features, decoding and reconstructing the mixed features respectively, and requesting original data correction according to the correlation weight matrix to obtain a fault diagnosis result and generate a control instruction when the reconstruction error is higher than a threshold value. The application realizes second-level transparent perception of the operation state of the distributed power supply and significantly improves the accuracy of fault diagnosis.

Description

Technical Field

[0001] This application relates to the field of distributed power management technology, and in particular to a distributed power management method and system based on time-sharing multiplexing. Background Technology

[0002] As an important component of the new power system, distributed generation places higher demands on the grid monitoring and control capabilities due to its large-scale integration. Existing technologies generally adopt time-division multiplexing-based communication methods to collect data and monitor the status of massive distributed generation devices. By dividing the communication time into multiple discrete time slots, each device has exclusive access to the channel to upload operating data within its allocated time slot, avoiding channel conflicts caused by multiple devices sending data simultaneously. After receiving the operating parameters such as voltage, current, and power uploaded by each device, the main control center performs centralized processing such as data quality detection, anomaly identification, and status assessment, and issues control commands based on the analysis results to achieve coordinated control of distributed generation. This time-division multiplexing architecture can meet basic data collection needs when the number of devices is limited. However, with the rapid growth of the installed capacity of distributed generation, the limitations of traditional methods are gradually becoming apparent.

[0003] The existing time-division multiplexing control methods have the following shortcomings: First, they adopt a static and fixed time slot allocation strategy, where all distributed power devices occupy the same communication time slot regardless of their operating status. This results in stable devices transmitting a large amount of redundant data, wasting communication resources, while critical devices experiencing anomalies cannot upload high-frequency diagnostic data in a timely manner due to time slot limitations. This leads to low efficiency in communication resource utilization and an inability to meet real-time monitoring needs under abnormal conditions. Second, the data quality assessment and processing are all concentrated on the master station. The raw data collected by the distributed power terminal must go through a long link of sampling, buffering, waiting for time slots, channel transmission, master station reception, and queuing processing. The total delay from the generation to availability of a single data point can accumulate to several seconds or even tens of seconds, which may cause the best control opportunity to be missed during rapid transient processes such as power grid frequency fluctuations. Third, when data quality deterioration is detected and deep acquisition needs to be initiated, existing methods blindly increase the sampling frequency of all parameters simultaneously, causing the data volume to surge by tens to hundreds of times. This severely impacts the limited communication bandwidth, causing channel congestion and exacerbating the computing load on the master station. Summary of the Invention

[0004] This application provides a distributed power supply management method and system based on time-division multiplexing, which solves the technical problems in existing time-division multiplexing management methods, such as low communication resource utilization efficiency caused by static time slot allocation, data explosion caused by full-parameter deep acquisition, information distortion caused by forced alignment of multi-frequency heterogeneous data, and increased reconstruction error caused by cross-timescale feature hybrid encoding. Through a full-process closed-loop mechanism of dynamic time slot allocation driven by edge-side quality assessment, parameter dimension adaptive deep acquisition, time-scale decoupled dual-stream hierarchical compression, and master station collaborative correction and feedback optimization, it achieves second-level transparent perception of distributed power supply operating status and significantly improves the accuracy of fault diagnosis.

[0005] Firstly, this application provides a distributed power supply management method based on time-division multiplexing, the distributed power supply management method based on time-division multiplexing comprising: Step S1: Collect the electrical and status parameters of the distributed power source, and perform missing detection, anomaly identification, and redundancy filtering on the electrical and status parameters to obtain a data quality score; Step S2: Calculate the time decay weighted value based on the data quality score and sort it. Allocate the communication time slot length according to the sorting result. When the data quality score is lower than the preset threshold, decompose the quality score formula to locate the high anomaly rate parameter. Only increase the sampling frequency of the high anomaly rate parameter to obtain multi-frequency heterogeneous data containing high-frequency data and low-frequency data. Step S3: Input the high-frequency data into a temporal convolutional network to extract high-frequency feature vectors, input the low-frequency data into a fully connected network to extract low-frequency feature vectors, and calculate the correlation weight matrix between the high-frequency feature vectors and the low-frequency feature vectors through a cross-attention mechanism to obtain time-scale decoupled hybrid features; Step S4: Separate the time-scale decoupled hybrid features into high-frequency branches and low-frequency branches, and input them into the high-frequency decoder and low-frequency decoder respectively to reconstruct the data. When the reconstruction error exceeds the threshold, identify the causal relationship according to the correlation weight matrix and request the original data segment for correction to obtain the fault diagnosis result. Based on the fault diagnosis result, query the mapping database to generate the operation parameter control instruction.

[0006] Secondly, this application provides a distributed power management system based on time-division multiplexing, the distributed power management system based on time-division multiplexing comprising: The filtering module is used to collect the electrical and status parameters of the distributed power source, perform missing detection, anomaly identification and redundancy filtering on the electrical and status parameters, and obtain a data quality score. The sorting module is used to calculate the time decay weighted value based on the data quality score and sort it, allocate the communication time slot length according to the sorting result, and decompose the quality score formula to locate the high anomaly rate parameter when the data quality score is lower than the preset threshold. The sampling frequency is increased only for the high anomaly rate parameter to obtain multi-frequency heterogeneous data containing high-frequency data and low-frequency data. The input module is used to input the high-frequency data into a temporal convolutional network to extract high-frequency feature vectors, input the low-frequency data into a fully connected network to extract low-frequency feature vectors, and calculate the correlation weight matrix between the high-frequency feature vectors and the low-frequency feature vectors through a cross-attention mechanism to obtain time-scale decoupled hybrid features. The correction module is used to separate the time-scale decoupled hybrid features into high-frequency branches and low-frequency branches, respectively inputting them into the high-frequency decoder and the low-frequency decoder to reconstruct the data. When the reconstruction error exceeds the threshold, the causal relationship is identified according to the correlation weight matrix and the original data segment is requested for correction to obtain the fault diagnosis result. Based on the fault diagnosis result, the mapping database is queried to generate the operation parameter control instruction.

[0007] The technical solution provided in this application deploys a real-time data quality assessment mechanism at the edge of the distributed power supply, moving the traditional centralized quality detection from the main station to the data acquisition source. This allows data missingness, anomalies, and redundancy to be quantified into a comprehensive quality score at the moment of acquisition. This score drives the main station to dynamically adjust the communication time slot length of each device. Devices with low quality scores receive longer time slots to upload in-depth diagnostic data, while devices with high quality scores release time slot resources. This breaks the rigid configuration of communication resources caused by traditional fixed time slot allocation and achieves adaptive resource scheduling based on real-time operating status. When a quality score is detected to be lower than a preset threshold, the quality score is decomposed... By precisely identifying the dominant factors causing quality degradation through formula analysis, and further statistically analyzing the cumulative values ​​of abnormal flags for each parameter to identify parameters with high anomaly rates, this application increases the sampling frequency to a deep sampling frequency only for these high anomaly rate parameters and their physically strongly correlated diagnostic parameters, while maintaining the normal sampling frequency for other parameters. This forms a parameter dimension adaptive deep acquisition strategy. Compared to traditional methods that blindly increase the sampling frequency of all parameters simultaneously, resulting in a data volume increase of tens of times, this application precisely selects a subset of parameters that require deep monitoring. While ensuring the integrity of the information required for anomaly diagnosis, it controls the amount of deep acquisition data within a reasonable range, avoiding communication channel congestion and main station computational overload.

[0008] This application addresses the multi-frequency heterogeneous data characteristics generated by parameter-dimensional adaptive depth acquisition by innovatively employing a dual-stream hierarchical coding architecture. It uses a temporal convolutional network to process high-frequency data and extract fast dynamic features, while a fully connected network processes low-frequency data and extracts slow trend features. This avoids the false information or loss of detail introduced by traditional methods that forcibly align different frequency data to a unified time axis through interpolation or downsampling. More importantly, it introduces a cross-attention mechanism to calculate the association weight matrix between high-frequency and low-frequency feature vectors. This matrix explicitly models the semantic similarity and causal driving relationship between fast dynamics and slow trends, solving the problem of cross-timescale feature semantic breakage caused by independent dual-stream coding. In the ground station decoding and reconstruction stage, when the reconstruction error exceeds a threshold, the causal relationship between high-weight elements and low-frequency parameters is identified based on the association weight matrix. If the reconstruction errors of both high- and low-frequency parameters exceed the threshold, the low-frequency parameters are prioritized for reconstruction. The original parameter data is used to correct the underlying slow trend prediction. Accurate trend constraints are used to improve rapid dynamic reconstruction. If only the high-frequency parameter error exceeds the threshold, the original data is requested for targeted correction based on the gradient weighted activation mapping algorithm to select key time points. This hierarchical correction mechanism, which is a collaboration between the main station and the edge, not only retains the communication efficiency advantage of intelligent compression on the edge side, but also ensures the accuracy of fault diagnosis by minimizing the amount of original data returned. At the same time, the ground station feeds back the reconstruction error assessment and diagnostic confidence to the edge side, driving the online optimization of the cross-attention weight matrix of the time-scale decoupled attention network. This allows the compression model to gradually learn the optimal cross-scale feature alignment strategy for different fault scenarios, forming a full-process adaptive closed loop of data acquisition, quality assessment, dynamic time slot allocation, adaptive depth acquisition, hierarchical compression, multi-scale decoding, collaborative correction, and feedback optimization. This significantly improves the transparency perception capability of the distributed power supply's operating status and the accuracy of fault diagnosis. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of one embodiment of the distributed power management method based on time-division multiplexing in this application. Figure 2 This is a schematic diagram of the convergence curve of the reconstruction error iterative optimization in an embodiment of this application; Figure 3 This is a schematic diagram comparing time slot allocation strategies in the embodiments of this application. Detailed Implementation

[0011] This application provides a distributed power management method and system based on time-division multiplexing. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0012] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the distributed power management method based on time-division multiplexing in this application includes: Step S1: Collect the electrical and status parameters of the distributed power source, perform missing detection, anomaly identification, and redundancy filtering on the electrical and status parameters, and obtain a data quality score; Among them, when the difference between two consecutive timestamps exceeds 1.2 times the normal sampling period, it is judged as missing data. The third-order Lagrange interpolation algorithm uses the three historical sampling points before the missing time to construct an interpolation polynomial for prediction. Anomaly identification includes two categories: statistical anomalies and regular anomalies. Statistical anomalies are calculated based on the mean and standard deviation of the parameters within the sliding window to calculate the deviation multiple. Regular anomalies are judged based on the physical constraints of the distributed power source to determine whether the parameters exceed the limits. Redundancy filtering identifies redundant sampling points with basically unchanged values ​​by calculating the rate of change of parameters between adjacent times. The data quality score is obtained by inverting the three indicators of missing degree, anomaly degree, and redundancy degree and then weighting and summing them.

[0013] Step S2: Calculate the time decay weighted value based on the data quality score and sort it. Allocate the communication time slot length according to the sorting result. When the data quality score is lower than the preset threshold, decompose the quality score formula to locate the high anomaly rate parameter. Only increase the sampling frequency for the high anomaly rate parameter to obtain multi-frequency heterogeneous data containing high-frequency data and low-frequency data. An exponential decay function is applied to the historical quality score sequence of each device, with lower weights for historical scores from more distant time intervals. After sorting, the device ranking is mapped to the time slot length range using a linear interpolation formula. The quality score formula decomposition involves calculating the difference between the current value and the historical benchmark value of each of the three components in the weighted summation as the deviation. The component with the largest absolute value of the deviation is used to determine the dominant factor in quality deterioration. High anomaly rate parameters are located by accumulating the number of anomaly flags of each parameter within a statistical sliding window. Only the parameters with the highest cumulative number of anomalies are increased in sampling frequency to the deep frequency, while the remaining parameters are collected at the regular frequency. This results in a multi-frequency heterogeneous data structure with different sampling frequencies for different parameters within the same device.

[0014] Step S3: Input high-frequency data into a temporal convolutional network to extract high-frequency feature vectors, input low-frequency data into a fully connected network to extract low-frequency feature vectors, and calculate the correlation weight matrix between high-frequency and low-frequency feature vectors through a cross-attention mechanism to obtain time-scale decoupled hybrid features; The temporal convolutional network adopts a three-layer convolutional structure. The kernel size of each layer is fixed, but the dilation rate increases exponentially. The first layer has a dilation rate of 1 to capture local temporal patterns, the second layer has a dilation rate of 2 to expand the receptive field, and the third layer has a dilation rate of 4 to cover a longer time span. Global average pooling compresses the temporal dimension of the convolutional output into a single value to obtain a fixed-dimensional feature vector. The fully connected network performs two linear transformations and nonlinear activations on the low-frequency data to map it to the target feature space. The association weight matrix of the cross-attention mechanism is obtained by multiplying the transpose of the high-frequency embedded features and the low-frequency embedded features, dividing by the square root of the dimension, scaling, and then normalizing by softmax. Each element in the matrix represents the semantic similarity between a certain dimension of the high-frequency feature and a certain dimension of the low-frequency feature. The weight matrix is ​​multiplied by the low-frequency embedded features to generate cross-association features. These features encode the association pattern between high-frequency rapid changes and low-frequency slow trends. Finally, the high-frequency original features, low-frequency original features, and cross-association features are concatenated and compressed to a unified dimension through a dimensionality reduction convolutional layer to form a time-scale decoupled hybrid feature.

[0015] Step S4: Separate the time-scale decoupled hybrid features into high-frequency branches and low-frequency branches, and input them into the high-frequency decoder and low-frequency decoder respectively to reconstruct the data. When the reconstruction error exceeds the threshold, identify the causal relationship according to the correlation weight matrix and request the original data segment for correction to obtain the fault diagnosis result. Based on the fault diagnosis result, query the mapping database to generate the operation parameter control instruction.

[0016] Based on a predefined index range, different segments of the mixed feature vector are extracted and fed into the high-frequency decoder and low-frequency decoder respectively. The three-layer deconvolutional network of the high-frequency decoder upsamples layer by layer according to a structure mirror-symmetric to the encoder to recover the temporal length. The fully connected network of the low-frequency decoder maps the feature vector back to the product dimension of the number of parameters and the number of time points, and then reshapes it into a two-dimensional temporal matrix. The reconstruction error is quantified by calculating the mean square error between the original data and the reconstructed data to determine the reconstruction quality. Causal correlation identification selects high-weight elements with values ​​greater than a threshold from the correlation weight matrix. The row index of this element corresponds to the high-frequency parameter dimension, and the column index corresponds to the low-frequency parameter dimension. When the reconstruction errors of both the high-frequency and low-frequency parameters exceed the threshold, the reconstruction is considered complete. The system prioritizes requesting all original sampling points of low-frequency parameters. When only the error of high-frequency parameters exceeds the threshold, it calculates the gradient contribution of each time point to the reconstruction error based on the gradient weighted activation mapping algorithm and selects several time points with the largest absolute gradient values ​​to request original data. The received original data directly replaces the corresponding reconstruction values ​​and is then sent back to the decoder to improve reconstruction accuracy. The fault diagnosis classification network extracts bidirectional long short-term memory network features from the spliced ​​complete time series data and outputs the probability values ​​of each fault type through a multi-layer fully connected classification head. The mapping database stores the correspondence between fault type codes and control parameter adjustment schemes. Based on the diagnosed fault type, the database is queried to obtain recommended parameter settings and generate control instructions.

[0017] In one specific embodiment, step S1 includes: The electrical and status parameters of the distributed power supply are acquired through analog acquisition interfaces and digital communication interfaces, and timestamps are added. The collected electrical and status parameters are time-stamped for continuity verification. When a timestamp jump or a data validity flag of 0 is detected, a missing flag is marked. The missing value is predicted using historical data through a third-order Lagrange interpolation algorithm, and the data missing degree within the time window is calculated. Within the sliding time window, the mean and standard deviation of each parameter are calculated. When the deviation of the parameter value from the mean exceeds three times the standard deviation, it is marked as a statistical anomaly. At the same time, rule-based anomaly identification is performed based on physical constraints, and the comprehensive anomaly degree is calculated. Calculate the rate of change of each parameter at adjacent sampling times. When the rate of change of a preset number of consecutive sampling points is less than a preset threshold, it is marked as redundant data. Calculate the redundancy, and weight and fuse the data missingness, comprehensive anomaly, and redundancy to obtain a data quality score.

[0018] Specifically, timestamp continuity verification is achieved by calculating the difference between the timestamps of the current sampling time and the previous sampling time. Under normal circumstances, this difference should be equal to the preset sampling period. When the difference exceeds 1.2 times the sampling period, data loss is determined. The third-order Lagrange interpolation algorithm constructs a third-order polynomial function using the three historical sampling points before the missing time. The function value of this polynomial at the three known times is exactly equal to the corresponding historical sampling value. The polynomial function value is calculated by substituting the time variable at the missing time as the predicted value of the missing data. The data missing degree is obtained by dividing the total number of missing flag bits of all parameters within the statistical time window by the total number of data points within the window.

[0019] The mean calculation within the sliding time window involves summing all sampled values ​​of each parameter within the window and dividing by the number of sampling points. The standard deviation calculation first calculates the sum of squares of the deviations of each sampled value from the mean, then divides by the number of sampling points and takes the square root. Statistical anomaly detection divides the absolute value of the difference between the current parameter value and the mean by the standard deviation to obtain the deviation multiple. Regular anomaly identification sets upper and lower limit thresholds for each parameter based on the rated operating range of the distributed power supply equipment. When a parameter value exceeds the upper or lower limit or violates electrical and physical constraints, it is marked as a regular anomaly. The overall anomaly degree is determined by the combination of statistical anomaly flags and regular anomaly flags. Alternatively, the cumulative number of results within the time window can be divided by the total number of data points. The rate of change is obtained by subtracting the parameter value from the previous time point from the current time point and then dividing by the sampling time interval. Redundancy is calculated as the cumulative proportion of consecutive low rate of change sampling points. The data quality score is calculated by multiplying the missing value, outlier, and redundancy by their respective weight coefficients and then summing them. Since the larger the values ​​of missing value, outlier, and redundancy, the worse the quality, the three indicators need to be inverted before weighted fusion, i.e., by subtracting each indicator value from 1, so that the larger the final quality score, the better the data quality.

[0020] In one specific embodiment, step S3 includes: The high-frequency data is preprocessed by normalization. The normalized high-frequency data is then input into a temporal convolutional network. Temporal features are extracted by dimensionality reduction layer by layer through the first, second and third convolutional layers. The temporal dimension is then compressed by a global average pooling layer to obtain the high-frequency feature vector. The low-frequency data is preprocessed by normalization. The normalized low-frequency data is then input into a fully connected network. The data is mapped to the hidden layer through the first fully connected layer and an activation function is applied. Finally, the data is mapped to the target dimension through the second fully connected layer to obtain the low-frequency feature vector. High-frequency embedding features are obtained by adding a fast dynamic scaling embedding vector to the high-frequency feature vector, and low-frequency embedding features are obtained by adding a slow trend scaling embedding vector to the low-frequency feature vector. High-frequency embedded features are used as query vectors, and low-frequency embedded features are used as key and value vectors. The association weight matrix is ​​calculated using the attention weight calculation formula. Based on the association weight matrix and value vector, cross-association features are calculated. The high-frequency feature vector, low-frequency feature vector and cross-association features are concatenated and compressed through a dimensionality reduction layer to obtain time-scale decoupled hybrid features.

[0021] Specifically, the normalization preprocessing calculates the mean and standard deviation of each parameter dimension for both high-frequency and low-frequency data. The original value of each parameter is subtracted from its corresponding mean and then divided by its corresponding standard deviation, mapping the normalized data distribution to a standard normal distribution with zero mean and unit variance. The first convolutional layer of the temporal convolutional network applies a convolutional kernel to the input multi-parameter temporal data for sliding window feature extraction. The kernel size determines the number of time steps covered in a single convolution, the stride determines the interval of kernel sliding, and the dilation rate expands the receptive field by inserting holes between kernel elements. The number of feature map channels output by the first convolutional layer equals the number of kernels. The temporal length is calculated using the formula relating the input length, kernel size, and stride. After the length is reduced, the second and third convolutional layers repeat the convolution operation on the feature map output by the previous layer, increasing the number of feature map channels layer by layer and reducing the temporal length. The global average pooling layer averages the values ​​of all time steps for each channel of the final convolutional layer output, compressing the two-dimensional feature map into a one-dimensional feature vector. The vector dimension is equal to the number of channels in the final convolutional layer. The first layer of the fully connected network for low-frequency data maps the flattened vector, which is the product of the number of input parameters and the number of time points, to the hidden layer dimension through a linear transformation of the weight matrix. The ReLU activation function is applied to zero out the negative values ​​of the hidden layer output and retain the positive values ​​to introduce nonlinearity. The second fully connected layer maps the hidden layer output to the target feature vector dimension again through a linear transformation of the weight matrix.

[0022] The fast dynamic scaling embedding vector and the slow trend scaling embedding vector are pre-trained or randomly initialized learnable parameter vectors. An additional operation merges the scaling embedding vectors with the original feature vectors in the dimensional direction through vector concatenation. The dimension of the high-frequency embedding feature equals the dimension of the high-frequency feature vector plus the dimension of the fast dynamic scaling embedding vector; the dimension of the low-frequency embedding feature equals the dimension of the low-frequency feature vector plus the dimension of the slow trend scaling embedding vector. The attention weight calculation formula first performs matrix multiplication on the query vector (i.e., the high-frequency embedding feature) and the transpose of the key vector (i.e., the low-frequency embedding feature), resulting in a matrix whose dimension is the high-frequency embedding feature dimension multiplied by the low-frequency embedding feature dimension. Then, each element in the matrix is ​​divided by the square root of the low-frequency embedding feature dimension for numerical scaling to avoid gradient vanishing. Finally, a softmax function is applied to each row of the matrix to scale the in-row weights. Elements are normalized to a probability distribution, and the normalized matrix is ​​the association weight matrix. The element in the i-th row and j-th column of the matrix represents the attention weight of the i-th dimension of the high-frequency feature on the j-th dimension of the low-frequency feature. The cross-association feature is calculated by matrix multiplication of the association weight matrix and the value vector, i.e., the low-frequency embedding feature. The output dimension is the same as the high-frequency embedding feature dimension. The concatenation operation connects the high-frequency feature vector, the low-frequency feature vector, and the cross-association feature vector end to end in the dimensional direction to form a higher-dimensional joint vector. The dimensionality reduction layer uses a one-dimensional convolution operation with a kernel size of 1. By adjusting the number of convolution kernels, the dimension of the joint vector is compressed to the preset hybrid feature dimension. The compressed vector is the time-scale decoupled hybrid feature, which simultaneously contains high-frequency fast dynamic information, low-frequency slow trend information, and the association pattern information between the two.

[0023] In one specific embodiment, the first convolutional layer of the temporal convolutional network is configured with a first convolutional kernel size, a first stride, and a first dilation rate; the second convolutional layer is configured with a second convolutional kernel size, a second stride, and a second dilation rate; and the third convolutional layer is configured with a third convolutional kernel size, a third stride, and a third dilation rate. The first dilation rate, the second dilation rate, and the third dilation rate are exponentially increasing, so that the receptive field of the third convolutional layer covers the temporal data of a preset duration.

[0024] Specifically, the first, second, and third convolutional kernel sizes are typically set to the same value, such as 8. The kernel size determines the number of consecutive time steps covered by a single convolution operation. The first, second, and third strides are typically set to 2. The stride determines the interval distance when the convolutional kernel slides in the temporal direction. A stride of 2 means that the convolutional kernel skips one time step each time it slides, making the output temporal length approximately half the input temporal length. The dilation rate achieves exponential expansion of the receptive field by inserting several zero-value positions between adjacent elements of the convolutional kernel. A first dilation rate of 1 indicates that the convolutional kernel elements are arranged continuously without gaps, at which point the receptive field range is equal to the kernel size. A second dilation rate of 2 indicates that the convolutional kernel elements are adjacent... One zero-value position is inserted between elements. The actual receptive field is calculated as the kernel size plus the kernel size minus 1, multiplied by the dilation rate minus 1. The third dilation rate is set to 4, which means that three zero-value positions are inserted between adjacent elements. The dilation rates of the three convolutional layers increase exponentially in the order of 1, 2, and 4. The effective receptive field of the third convolutional layer is calculated by accumulating the receptive field contributions of the first two convolutional layers. When the kernel size is 8 and the dilation rate is 1, 2, and 4, the cumulative receptive field of the third convolutional layer can cover about 55 time steps. For high-frequency data with a sampling frequency of 20Hz, 55 time steps correspond to a time series span of 2.75 seconds. This span is sufficient to capture the complete time series evolution process of typical anomalies such as distributed power supply voltage fluctuations and power oscillations.

[0025] In one specific embodiment, the attention weight calculation process is as follows: Multiply the high-frequency embedding features and the transpose of the low-frequency embedding features by a matrix, divide by the square root of the dimension of the low-frequency embedding features for scaling, and normalize by the softmax function to obtain the association weight matrix. The weight elements in the association weight matrix represent the semantic similarity between each dimension of high-frequency features and each dimension of low-frequency features. When the weight element is greater than the preset similarity threshold, it is identified as a strong causal relationship.

[0026] Specifically, both high-frequency and low-frequency embedding features are one-dimensional vectors. The transpose operation converts the low-frequency embedding features from a horizontal arrangement to a vertical arrangement. Matrix multiplication multiplies the horizontal high-frequency embedding features with the vertical low-frequency embedding features, resulting in a two-dimensional matrix. The number of rows in the matrix equals the number of dimensions of the high-frequency embedding features, and the number of columns equals the number of dimensions of the low-frequency embedding features. The value of each element in the matrix is ​​equal to the product of the corresponding high-frequency dimension component and the low-frequency dimension component. The square root scaling operation divides each element in the matrix by the square root of the number of dimensions of the low-frequency embedding features. Scaling avoids excessively large matrix element values, which would cause the subsequent softmax function output to concentrate on extreme values. Softmax normalization processes the matrix row by row. For all elements in each row, the natural exponential function value is calculated first, and then each exponential value is divided by the sum of all exponential values ​​in that row. After normalization, the sum of the elements in each row equals 1, and all elements are positive numbers between 0 and 1. The normalized matrix is ​​the correlation weight matrix.

[0027] In the association weight matrix, each weight element represents the semantic similarity between the corresponding high-frequency feature dimension and the low-frequency feature dimension. The closer the weight value is to 1, the higher the similarity; the closer it is to 0, the lower the similarity. When the value of a certain weight element is greater than a preset similarity threshold, such as 0.6, it is determined that there is a strong causal relationship between the high-frequency dimension and the low-frequency dimension corresponding to that weight element. A strong causal relationship means that there is an inherent driving relationship between the rapid change pattern of the high-frequency parameter and the slow trend change of the low-frequency parameter. For example, the slow upward trend of the active power of the distributed power source may drive the rapid fluctuation of the output voltage. When the weight elements corresponding to these two parameters exceed the threshold, if the reconstruction error of the high-frequency voltage parameter is too large in the subsequent reconstruction correction stage, the original data of the low-frequency power parameter is requested first for underlying trend correction. By improving the prediction accuracy of the slow trend, the accuracy of the fast dynamic reconstruction is improved.

[0028] Figure 2 This is a schematic diagram of the convergence curve of the reconstruction error iterative optimization in an embodiment of this application; Figure 2The diagram illustrates the convergence of reconstruction error with and without time-scale decoupling as the number of iterations increases. The gray curve represents the reconstruction error without time-scale decoupling, which gradually decreases from an initial 0.48 to approximately 0.21 and then stabilizes. The black curve represents the reconstruction error after employing the time-scale decoupling attention mechanism, which rapidly decreases from an initial 0.31 and converges to around 0.08 by the 40th iteration. The two dashed lines mark the high-frequency error threshold of 0.08 and the low-frequency error threshold of 0.05, respectively. The improved arrow at the 40th iteration indicates that the reconstruction error of the time-scale decoupling method is reduced by 66.8% compared to the method without decoupling. This fully verifies the technical solution of this application, which calculates the correlation weight matrix between high-frequency and low-frequency features through a cross-attention mechanism and explicitly models the semantic relationship between fast and slow time scales. This effectively solves the reconstruction distortion problem caused by the semantic breakage of cross-level features after hierarchical compression of multi-frequency heterogeneous data.

[0029] In one specific embodiment, step S2 includes: The system receives data quality scores uploaded by each distributed power device, and applies an exponential decay kernel to calculate a time decay weighted value for the historical quality score sequence of each device. The weight coefficient of the current quality score is 1, and the weight coefficient of the historical quality score decays exponentially with the time interval. The distributed power devices are sorted in descending order based on the time decay weighting value. The communication time slot length of each device is calculated by linear interpolation formula according to the sorting position. The device with the lowest quality score gets the longest time slot length, and the device with the highest quality score gets the shortest time slot length. When a device's data quality score is detected to be lower than a preset threshold and the rate of decrease in the quality score exceeds the preset decrease threshold, the quality score formula is decomposed to calculate the deviation of the data missingness component, the outlier component, and the redundancy component, and the component with the largest deviation is taken as the dominant factor. The cumulative values ​​of the abnormal flag bits of each parameter are counted, and the parameters with the highest cumulative number of abnormalities are identified as high abnormality rate parameters. Only the high abnormality rate parameters and their physically strongly correlated diagnostic parameters are increased in sampling frequency to deep sampling frequency, while the other parameters are kept at the normal sampling frequency, resulting in multi-frequency heterogeneous data containing high-frequency and low-frequency data.

[0030] Specifically, the time decay weighted value calculation involves multiplying the quality score of a device over a certain number of past moments by the corresponding decay weight and then summing the results. The decay weight is calculated using the negative exponential form of the natural exponential function, with the formula W = Q0 × e 0 + Q1×e -0 · 15 + Q2×e -0 · 3 + Q3×e -0 · 45 + Q4×e -0 ·6 Where W is the time decay weighted value, Q0 is the quality score at the current moment, Q1 to Q4 are the historical quality scores of the previous 1 to 4 seconds, and e is the natural constant approximately equal to 2.718. The larger the time interval, the larger the absolute value of the negative exponent and the smaller the decay weight. The linear interpolation formula calculates the time slot length based on the device ranking position, and the formula is T = 5 - 4.5 × (i-1) / (N-1), where T is the allocated time slot length in milliseconds, i is the device ranking position, and N is the total number of devices. For example, when there are a total of 100 devices, the time slot length of the device ranked 1st is T = 5 - 4.5 × 0 / 99 = 5 milliseconds, and the time slot length of the device ranked 100th is T = 5 - 4.5 × 99 / 99 = 0.5 milliseconds. The intermediate positions are allocated proportionally by linear interpolation.

[0031] The quality scoring formula decomposes the three components of the weighted summation into separate values: missing value C1 = 0.5 × (1 - δ), outlier value C2 = 0.3 × (1 - α), and redundancy value C3 = 0.2 × (1 - ρ), where δ is the missing value, α is the outlier value, and ρ is the redundancy value. The deviation is calculated as D = Ccurrent - Cbaseline, where Ccurrent is the current value of the component, and Cbaseline is the median value of that component over the past 24 hours. For example, if the current missing value C1 = 0.45 and the historical median is 0.48, then the deviation D1 = 0.45 - 0.48 = -0.03; if the current outlier value C2 = 0.21 and the historical median is 0.27, then the deviation D2 = 0.21 - 0.27 = -0.06; and if the current redundancy value C3 = 0.18 and the historical median is 0.19, then the deviation D3 = 0.18 - 0.19 = -0.01, take the component with the largest absolute value of deviation. In this example, |D2| = 0.06, the maximum judgment abnormality is the dominant factor. The cumulative value of the abnormality flag counts the number of times each parameter is marked as abnormal within the sliding time window. The parameters with the highest cumulative number of times are identified as high abnormality rate parameters. Only these high abnormality rate parameters and their physically strongly correlated diagnostic parameters have their sampling frequency increased from the normal 1Hz to the deep sampling frequency of 20Hz. The other parameters are kept at 1Hz sampling, forming a multi-frequency heterogeneous data structure with different sampling frequencies for different parameters within the same device.

[0032] Figure 3 This is a schematic diagram comparing the time slot allocation strategies in the embodiments of this application. Figure 3This paper compares the effects of traditional fixed time slot allocation and quality-driven dynamic time slot allocation. The left side shows that in the traditional method, all eight distributed power devices are allocated the same 1.25 ms time slot length. The right side shows that the method in this application dynamically adjusts the time slot length based on the data quality score of each device. Devices 1, 3, and 6, with quality scores below the 0.6 threshold, are identified as low-quality devices and receive extended time slots of 5.00 ms, 1.79 ms, and 2.43 ms, respectively. Device 7, with the highest quality score, receives only the shortest time slot of 0.50 ms. The remaining devices with intermediate quality scores are allocated medium-length time slots using a linear interpolation formula. Dynamic time slot allocation allows low-quality devices to obtain more communication resources to upload deep acquisition data, while high-quality devices release time slot resources for other devices to use, fully demonstrating the adaptive communication resource allocation mechanism based on quality feedback in this application.

[0033] In one specific embodiment, step S4 includes: The time-scale decoupled hybrid features are separated into high-frequency branch features and low-frequency branch features according to a predefined index mapping relationship; The high-frequency branch features are input into the high-frequency decoder, and the high-frequency time-series data is reconstructed by upsampling layer by layer through the first deconvolution layer, the second deconvolution layer and the third deconvolution layer. Normalized metadata is applied for denormalization to obtain the high-frequency data reconstruction value. The low-frequency branch features are input into the low-frequency decoder, mapped and reshaped into a low-frequency time matrix through a multi-layer fully connected network, and then denormalized using normalized metadata to obtain the reconstructed low-frequency data values. The reconstruction error of high-frequency data reconstruction value and low-frequency data reconstruction value are calculated. When the reconstruction error exceeds the preset threshold, the causal relationship between the high-frequency feature dimension and the low-frequency feature dimension corresponding to the high-weight element is identified according to the correlation weight matrix. If the reconstruction errors of both the high-frequency and low-frequency parameters exceed the threshold, all original data points of the low-frequency parameters are requested. If only the reconstruction error of the high-frequency parameters exceeds the threshold, the key time points of the high-frequency parameters are screened based on the gradient weighted activation mapping algorithm and the corresponding original data is requested. The reconstruction value is replaced with the original data and the decoding is re-executed. The corrected high-frequency data reconstruction values ​​are concatenated with the low-frequency data reconstruction values ​​to form complete time-series data. This data is then input into a fault diagnosis classification network to extract time-series dependent features. The classification head outputs the probability distribution of fault types, and the category with the highest probability is taken as the fault diagnosis result. Based on the fault diagnosis result, the corresponding control parameter adjustment scheme is queried in the mapping database to generate operating parameter control instructions.

[0034] Specifically, a predefined index mapping relationship divides the hybrid feature vector into different segments according to its dimensional position. For example, when the total dimension of the hybrid feature is 128, the first 80 dimensions are mapped to high-frequency branch features, and the last 48 dimensions are mapped to low-frequency branch features. The separation operation extracts the feature values ​​of the corresponding segments through index slicing. The deconvolution layer achieves upsampling by inserting zero-value positions between adjacent elements of the input feature before performing a convolution operation. The first deconvolution layer expands the temporal length of the input feature from 19 to 45, the second deconvolution layer expands it from 45 to 97, and the third deconvolution layer expands it from 97... Expanding to 200, each deconvolution layer simultaneously adjusts the number of feature channels, gradually reducing them until they are equal to the number of parameters. The inverse normalization operation multiplies the reconstructed standardized value by the standard deviation parameter saved during normalization, plus the mean parameter, to restore the original dimensions. The fully connected network of the low-frequency decoder maps the low-frequency branch features to the product dimension of the number of parameters and the number of time points through two layers of linear transformation. The reshape operation reorganizes the one-dimensional vector into a two-dimensional matrix, with the number of rows equal to the number of parameters and the number of columns equal to the number of time points. After inverse normalization, the reconstructed low-frequency data value in the original physical unit is obtained.

[0035] The reconstruction error is calculated by summing the squared differences between the original and reconstructed data and dividing by the total number of data points to obtain the mean square error. The threshold for high-frequency reconstruction error is set to 0.08, and the threshold for low-frequency reconstruction error is set to 0.05. Weights greater than 0.6 in the correlation weight matrix are identified as strong causal relationships. When the reconstruction error of the high-frequency parameter is 0.12 (exceeding the threshold of 0.08) and the reconstruction error of the low-frequency parameter is 0.08 (exceeding the threshold of 0.05), both are considered to exceed the thresholds and a causal relationship exists. All 10 original sampling points of the low-frequency parameter within the deep acquisition period are requested for correction. The gradient-weighted activation mapping algorithm calculates the partial derivatives of the reconstruction error with respect to the input data at each time point. Time points with an absolute value of the partial derivative greater than 0.05 are identified as key points; for example, the top 15 time points with the highest absolute values ​​of their partial derivatives out of 200 time points. The time points are selected as key points. Only the original data of these 15 time points are requested to replace the reconstructed values ​​at the corresponding positions and then sent back to the decoder. The fault diagnosis classification network uses a bidirectional long short-term memory network to extract time-series dependent features and capture the correlation patterns between time points before and after. The probability distribution of 16 types of faults is output through a three-layer fully connected classification head. The fault type corresponding to the category with the largest value in the probability distribution is used as the diagnosis result. The mapping database stores the correspondence between fault type codes and control parameter adjustment schemes. For example, when fault type code 3 corresponds to an inverter IGBT fault, the query obtains the adjustment scheme of reducing the power setpoint to 50% of the rated value and the switching frequency to 5 kHz. Based on the query result, a control instruction data packet containing parameter type codes and new setpoints is generated and sent to the distributed power equipment to perform parameter updates.

[0036] The above describes the distributed power management method based on time-division multiplexing in the embodiments of this application. The following describes the distributed power management system based on time-division multiplexing in the embodiments of this application. One embodiment of the distributed power management system based on time-division multiplexing in the embodiments of this application includes: The filtering module is used to collect the electrical and status parameters of the distributed power source, perform missing detection, anomaly identification and redundancy filtering on the electrical and status parameters, and obtain a data quality score. The sorting module is used to calculate the time decay weighted value based on the data quality score and sort it, allocate the communication time slot length according to the sorting result, and decompose the quality score formula to locate the high anomaly rate parameter when the data quality score is lower than the preset threshold. The sampling frequency is increased only for the high anomaly rate parameter to obtain multi-frequency heterogeneous data containing high-frequency data and low-frequency data. The input module is used to input the high-frequency data into a temporal convolutional network to extract high-frequency feature vectors, input the low-frequency data into a fully connected network to extract low-frequency feature vectors, and calculate the correlation weight matrix between the high-frequency feature vectors and the low-frequency feature vectors through a cross-attention mechanism to obtain time-scale decoupled hybrid features. The correction module is used to separate the time-scale decoupled hybrid features into high-frequency branches and low-frequency branches, respectively inputting them into the high-frequency decoder and the low-frequency decoder to reconstruct the data. When the reconstruction error exceeds the threshold, the causal relationship is identified according to the correlation weight matrix and the original data segment is requested for correction to obtain the fault diagnosis result. Based on the fault diagnosis result, the mapping database is queried to generate the operation parameter control instruction.

[0037] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A distributed power supply management method based on time-division multiplexing, characterized in that, The method includes: Step S1: Collect the electrical and status parameters of the distributed power source, and perform missing detection, anomaly identification, and redundancy filtering on the electrical and status parameters to obtain a data quality score; Step S2: Calculate the time decay weighted value based on the data quality score and sort it. Allocate the communication time slot length according to the sorting result. When the data quality score is lower than the preset threshold, decompose the quality score formula to locate the high anomaly rate parameter. Only increase the sampling frequency of the high anomaly rate parameter to obtain multi-frequency heterogeneous data containing high-frequency data and low-frequency data. Step S3: Input the high-frequency data into a temporal convolutional network to extract high-frequency feature vectors, input the low-frequency data into a fully connected network to extract low-frequency feature vectors, and calculate the correlation weight matrix between the high-frequency feature vectors and the low-frequency feature vectors through a cross-attention mechanism to obtain time-scale decoupled hybrid features; Step S4: Separate the time-scale decoupled hybrid features into high-frequency branches and low-frequency branches, and input them into the high-frequency decoder and low-frequency decoder respectively to reconstruct the data. When the reconstruction error exceeds the threshold, identify the causal relationship according to the correlation weight matrix and request the original data segment for correction to obtain the fault diagnosis result. Based on the fault diagnosis result, query the mapping database to generate the operation parameter control instruction.

2. The distributed power supply management method based on time-division multiplexing according to claim 1, characterized in that, Step S1 includes: The electrical and status parameters of the distributed power supply are acquired through analog acquisition interfaces and digital communication interfaces, and timestamps are added. The collected electrical and status parameters are time-stamped for continuity verification. When a timestamp jump or a data validity flag of 0 is detected, a missing flag is marked. The missing value is predicted using historical data through a third-order Lagrange interpolation algorithm, and the data missing degree within the time window is calculated. Within the sliding time window, the mean and standard deviation of each parameter are calculated. When the deviation of the parameter value from the mean exceeds three times the standard deviation, it is marked as a statistical anomaly. At the same time, rule-based anomaly identification is performed based on physical constraints, and the comprehensive anomaly degree is calculated. Calculate the rate of change of each parameter at adjacent sampling times. When the rate of change of a consecutive preset number of sampling points is less than a preset threshold, it is marked as redundant data. Calculate the redundancy. Weight and fuse the data missingness, the comprehensive anomaly, and the redundancy to obtain a data quality score.

3. The distributed power supply management method based on time-division multiplexing according to claim 1, characterized in that, Step S3 includes: The high-frequency data is preprocessed by normalization. The normalized high-frequency data is then input into a temporal convolutional network. Temporal features are extracted by dimensionality reduction layer by layer through the first, second and third convolutional layers. The temporal dimension is then compressed by a global average pooling layer to obtain the high-frequency feature vector. The low-frequency data is preprocessed by normalization. The normalized low-frequency data is then input into a fully connected network. The data is mapped to the hidden layer through the first fully connected layer and an activation function is applied. The data is then mapped to the target dimension through the second fully connected layer to obtain the low-frequency feature vector. A fast dynamic scaling embedding vector is added to the high-frequency feature vector to obtain high-frequency embedding features, and a slow trend scaling embedding vector is added to the low-frequency feature vector to obtain low-frequency embedding features. The high-frequency embedded features are used as query vectors, and the low-frequency embedded features are used as key vectors and value vectors. The association weight matrix is ​​calculated using the attention weight calculation formula. Based on the association weight matrix and the value vector, the cross-association features are calculated. The high-frequency feature vector, the low-frequency feature vector, and the cross-association features are concatenated and compressed through a dimensionality reduction layer to obtain time-scale decoupled hybrid features.

4. The distributed power supply management method based on time-division multiplexing according to claim 3, characterized in that, The first convolutional layer of the temporal convolutional network is configured with a first convolutional kernel size, a first stride, and a first dilation rate; the second convolutional layer is configured with a second convolutional kernel size, a second stride, and a second dilation rate; and the third convolutional layer is configured with a third convolutional kernel size, a third stride, and a third dilation rate. The first dilation rate, the second dilation rate, and the third dilation rate are in an exponentially increasing relationship, so that the receptive field of the third convolutional layer covers temporal data of a preset duration.

5. The distributed power supply management method based on time-division multiplexing according to claim 4, characterized in that, The attention weight calculation process is as follows: The high-frequency embedding feature and the transpose of the low-frequency embedding feature are multiplied by a matrix, and then scaled by the square root of the dimension of the low-frequency embedding feature. The matrix is ​​then normalized using the softmax function to obtain the association weight matrix. The weight elements in the association weight matrix represent the semantic similarity between each dimension of high-frequency features and each dimension of low-frequency features. When the weight element is greater than a preset similarity threshold, it is identified as a strong causal relationship.

6. The distributed power supply management method based on time-division multiplexing according to claim 1, characterized in that, Step S2 includes: The system receives data quality scores uploaded by each distributed power device, and applies an exponential decay kernel to calculate a time decay weighted value for the historical quality score sequence of each device. The weight coefficient of the current quality score is 1, and the weight coefficient of the historical quality score decays exponentially with the time interval. The distributed power devices are sorted in descending order according to the time decay weighting value. The communication time slot length of each device is calculated by linear interpolation formula according to the sorting position. The device with the lowest quality score gets the longest time slot length, and the device with the highest quality score gets the shortest time slot length. When a device's data quality score is detected to be lower than a preset threshold and the rate of decrease in the quality score exceeds the preset decrease threshold, the quality score formula is decomposed to calculate the deviation of the data missingness component, the outlier component, and the redundancy component, and the component with the largest deviation is taken as the dominant factor. The cumulative values ​​of the abnormal flag bits of each parameter are counted, and the parameters with the highest cumulative number of abnormalities are identified as high abnormality rate parameters. Only the sampling frequency of the high abnormality rate parameters and their physically strongly correlated diagnostic parameters is increased to the deep sampling frequency, while the sampling frequency of the other parameters is maintained at the normal sampling frequency, so as to obtain multi-frequency heterogeneous data containing high-frequency data and low-frequency data.

7. The distributed power supply management method based on time-division multiplexing according to claim 1, characterized in that, Step S4 includes: The time-scale decoupled hybrid features are separated into high-frequency branch features and low-frequency branch features according to a predefined index mapping relationship; The high-frequency branch features are input into the high-frequency decoder, and the high-frequency time-series data is reconstructed by upsampling layer by layer through the first deconvolution layer, the second deconvolution layer and the third deconvolution layer. Normalized metadata is applied to perform denormalization to obtain the high-frequency data reconstruction value. The low-frequency branch features are input into the low-frequency decoder, mapped and reshaped into a low-frequency time matrix through a multi-layer fully connected network, and then denormalized using normalized metadata to obtain the low-frequency data reconstruction value. The reconstruction error of the high-frequency data reconstruction value and the reconstruction error of the low-frequency data reconstruction value are calculated. When the reconstruction error exceeds a preset threshold, the causal relationship between the high-frequency feature dimension and the low-frequency feature dimension corresponding to the high-weight element is identified according to the correlation weight matrix. If the reconstruction errors of both the high-frequency and low-frequency parameters exceed the threshold, all original data points of the low-frequency parameters are requested. If only the reconstruction error of the high-frequency parameters exceeds the threshold, the key time points of the high-frequency parameters are screened based on the gradient weighted activation mapping algorithm and the corresponding original data is requested. The reconstruction value is replaced with the original data and the decoding is re-executed. The corrected high-frequency data reconstruction values ​​are concatenated with the low-frequency data reconstruction values ​​to form complete time-series data. This data is then input into a fault diagnosis classification network to extract time-series dependent features. The classification head outputs the probability distribution of fault types, and the category with the highest probability is taken as the fault diagnosis result. Based on the fault diagnosis result, the corresponding control parameter adjustment scheme is queried in the mapping database to generate operating parameter control instructions.

8. A distributed power management system based on time-division multiplexing, characterized in that, For implementing the distributed power management method based on time-division multiplexing as described in any one of claims 1-7, the distributed power management system based on time-division multiplexing comprises: The filtering module is used to collect the electrical and status parameters of the distributed power source, perform missing detection, anomaly identification and redundancy filtering on the electrical and status parameters, and obtain a data quality score. The sorting module is used to calculate the time decay weighted value based on the data quality score and sort it, allocate the communication time slot length according to the sorting result, and decompose the quality score formula to locate the high anomaly rate parameter when the data quality score is lower than the preset threshold. The sampling frequency is increased only for the high anomaly rate parameter to obtain multi-frequency heterogeneous data containing high-frequency data and low-frequency data. The input module is used to input the high-frequency data into a temporal convolutional network to extract high-frequency feature vectors, input the low-frequency data into a fully connected network to extract low-frequency feature vectors, and calculate the correlation weight matrix between the high-frequency feature vectors and the low-frequency feature vectors through a cross-attention mechanism to obtain time-scale decoupled hybrid features. The correction module is used to separate the time-scale decoupled hybrid features into high-frequency branches and low-frequency branches, respectively inputting them into the high-frequency decoder and the low-frequency decoder to reconstruct the data. When the reconstruction error exceeds the threshold, the causal relationship is identified according to the correlation weight matrix and the original data segment is requested for correction to obtain the fault diagnosis result. Based on the fault diagnosis result, the mapping database is queried to generate the operation parameter control instruction.

9. The system according to claim 8, characterized in that, The electrical and status parameters of the distributed power source are collected, and the electrical and status parameters are subjected to missing parameter detection, anomaly identification, and redundancy filtering to obtain a data quality score, including: The electrical and status parameters of the distributed power supply are acquired through analog acquisition interfaces and digital communication interfaces, and timestamps are added. The collected electrical and status parameters are time-stamped for continuity verification. When a timestamp jump or a data validity flag of 0 is detected, a missing flag is marked. The missing value is predicted using historical data through a third-order Lagrange interpolation algorithm, and the data missing degree within the time window is calculated. Within the sliding time window, the mean and standard deviation of each parameter are calculated. When the deviation of the parameter value from the mean exceeds three times the standard deviation, it is marked as a statistical anomaly. At the same time, rule-based anomaly identification is performed based on physical constraints, and the comprehensive anomaly degree is calculated. Calculate the rate of change of each parameter at adjacent sampling times. When the rate of change of a consecutive preset number of sampling points is less than a preset threshold, it is marked as redundant data. Calculate the redundancy. Weight and fuse the data missingness, the comprehensive anomaly, and the redundancy to obtain a data quality score.

10. The system according to claim 8, characterized in that, The high-frequency data is input into a temporal convolutional network to extract high-frequency feature vectors, and the low-frequency data is input into a fully connected network to extract low-frequency feature vectors. A cross-attention mechanism is used to calculate the association weight matrix between the high-frequency and low-frequency feature vectors, resulting in a time-scale decoupled hybrid feature, including: The high-frequency data is preprocessed by normalization. The normalized high-frequency data is then input into a temporal convolutional network. Temporal features are extracted by dimensionality reduction layer by layer through the first, second and third convolutional layers. The temporal dimension is then compressed by a global average pooling layer to obtain the high-frequency feature vector. The low-frequency data is preprocessed by normalization. The normalized low-frequency data is then input into a fully connected network. The data is mapped to the hidden layer through the first fully connected layer and an activation function is applied. The data is then mapped to the target dimension through the second fully connected layer to obtain the low-frequency feature vector. A fast dynamic scaling embedding vector is added to the high-frequency feature vector to obtain high-frequency embedding features, and a slow trend scaling embedding vector is added to the low-frequency feature vector to obtain low-frequency embedding features. The high-frequency embedded features are used as query vectors, and the low-frequency embedded features are used as key vectors and value vectors. The association weight matrix is ​​calculated using the attention weight calculation formula. Based on the association weight matrix and the value vector, the cross-association features are calculated. The high-frequency feature vector, the low-frequency feature vector, and the cross-association features are concatenated and compressed through a dimensionality reduction layer to obtain time-scale decoupled hybrid features.

Images