A power distribution network edge gateway computing decision system based on a digital intelligent multi-level model

Abstract

The application provides a power distribution network edge gateway computing decision system based on a digital intelligent multi-level model, and relates to the technical field of intelligent power systems.The system comprises a perception control layer for collecting perception data of a low-voltage power distribution network, an edge intelligent layer comprising a plurality of sequentially connected multi-level models, the multi-level models comprising a data processing platform, a digital intelligent model library, a cross-validation layer and a collaborative decision layer, and a cloud platform for storing full data, globally coordinating and scheduling and visually displaying, and performing bidirectional parameter interaction with the edge intelligent layer.The application can significantly improve the edge decision-making capability of new power distribution networks, reduce cloud computing load and provide a strong guarantee for the safe operation of power systems.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system intelligence, and particularly relates to a power distribution network edge gateway computing decision system based on a digital intelligent multi-level model. BACKGROUND

[0002] Under the background of large-scale access of new energy and increasing complexity of power systems, traditional power distribution networks face a series of severe challenges. For example, there is a significant shortcoming in real-time performance, the computing pressure is highly concentrated, and the edge device lacks effective security mechanisms in data transmission and storage.

[0003] At present, the existing technologies such as centralized systems, edge independent systems and conventional cloud-edge collaborative architectures all have certain defects. The centralized system highly depends on cloud global optimization, but has poor real-time performance, high bandwidth cost, and great difficulty in processing edge heterogeneous data. The edge independent system can achieve real-time response, but lacks global collaboration, so that the decision is easy to fall into local optimum rather than global optimum. The conventional cloud-edge collaborative architecture has a relatively simple structure and insufficient intelligent processing capability, and it is difficult to compensate for the decrease in prediction accuracy caused by external environmental factors in real time.

[0004] Therefore, it is of great practical significance to develop a decision system that is suitable for edge gateway and has the characteristics of lightweight and intelligence. SUMMARY

[0005] To this end, the present application provides a power distribution network edge gateway computing decision system based on a digital intelligent multi-level model, which is used to solve the problems of insufficient real-time performance, concentrated computing pressure, lack of data transmission and storage security mechanisms of edge devices in the traditional power distribution network, and poor real-time performance, lack of global collaboration and insufficient intelligent processing of the existing centralized system, edge independent system and cloud-edge collaborative architecture.

[0006] To solve the above problems, the present application provides a power distribution network edge gateway computing decision system based on a digital intelligent multi-level model, which comprises:

[0007] A perception control layer is used to collect perception data of a low-voltage power distribution network, including operation data, device state data and external environment data;

[0008] An edge intelligent layer comprises a plurality of connected multi-level models, and the multi-level models comprise a data processing center, a digital intelligent model library, a cross-validation layer and a collaborative decision layer, wherein:

[0009] The data processing center is used for pre-processing the perception data, including error processing, anomaly detection, data conversion and feature extraction, data fusion and protocol packaging;

[0010] The digital intelligence model library is used for real-time analysis and prediction of the operating status of the power distribution network based on preprocessed data;

[0011] The cross-validation layer is used to dynamically correct and cross-validate the output results of the digital intelligence model library based on the adaptive kernel recursive least squares algorithm.

[0012] The collaborative decision-making layer is used to combine cloud instructions with local key data to generate dynamic decisions at the edge, enabling cloud-edge collaboration.

[0013] The cloud platform is used for full data storage, global collaborative scheduling and visualization, and interacts with the edge intelligence layer in two directions.

[0014] Preferably, the system further includes a dynamic adaptation mechanism for adjusting multi-level model parameters and cloud-edge collaboration strategies in real time based on the power distribution network operation status, equipment resource status, and task urgency. The dynamic adaptation mechanism supports switching between multiple versions of lightweight models, including ultra-short-term, short-term, and long-term load forecasting models, and sets trigger conditions based on time scales.

[0015] Preferably, the data processing platform includes a data acquisition layer, a data processing layer, and a data transmission layer, wherein the data acquisition layer includes an event-triggered sampling mechanism, and its trigger function is:

[0016] ;

[0017] In the formula, For the current moment Monitoring parameters, The sampling interval is... The standard deviation of historical data;

[0018] The data processing layer includes anomaly detection methods, such as outlier detection formulas based on statistical features:

[0019] ;

[0020] In the formula, This is the average of historical data. Normal distribution of Quantiles Indicates the significance level.

[0021] Preferably, the digital intelligence model library is constructed based on partial least squares and improved temporal convolutional networks, and the steps include:

[0022] Let the input matrix be and the corresponding output matrix , For the sample size, The number of input features for each sample, is the output dimension;

[0023] extracting the latent variables of the input and output data by partial least squares ;

[0024] the latent variables extracted by partial least squares as the input of the improved time convolution network, constructing the time convolution network, finding the nonlinear relationship mapping of the output vector of the hidden layer as and the output as , randomly giving the input weight and the residual , selecting as the activation function, setting the number of hidden layer neurons as , calculating the output matrix of the hidden layer :

[0025] ;

[0026] calculating the output weight matrix and the predicted output matrix :

[0027] ;

[0028] wherein, is the Moore-Penrose generalized inverse matrix of , and is the real output matrix;

[0029] finally obtaining the PLS-ITCN data-driven model, and the model output is:

[0030] ;

[0031] wherein, are the projection weights of the latent variables to the original output space, is the residual matrix.

[0032] Preferably, the latent variable extraction formula is:

[0033] ;

[0034] wherein, and are the latent variables of the input matrix and the output matrix, and are the load matrices of the input matrix and the output matrix, are the projection weights of the original input features to the latent variables, are the projection weights of the latent variables to the original output space, The number of implicit variables, and is a residual matrix.

[0035] Preferably, the cross-validation layer comprises:

[0036] The online sparse modeling method comprises the following steps: initializing a time series data sequence and parameters; calculating a neighbor distance and a neighbor vector of a new sample, and determining the new sample as an isolated point and discarding the new sample if the neighbor distance and the neighbor vector exceed an expected range; and dynamically updating online model parameters according to a prediction error and a minimum distance of a data dictionary.

[0037] An online model parameter updating formula is as follows:

[0038] ;

[0039] In the formula, x is a model parameter value of the i th model in the j th iteration, is a model parameter value of the i th model in the j th iteration, is a step length parameter, is a prediction error. Preferably, the online sparse modeling method calculates a neighbor distance and a neighbor vector of a new sample, and determines the new sample as an isolated point and discards the new sample if the neighbor distance and the neighbor vector exceed an expected range, and the specific range is defined as follows:

[0040] ;

[0041] ;

[0042] ;

[0043] In the formula, x is a model parameter value of the i th model in the j th iteration, and are quantization factors, and are standard deviations, is an expected neighbor distance at a time point t, is an expected neighbor vector at a time point t. Preferably, the collaborative decision layer comprises a cloud-edge collaborative mechanism and a decision fusion strategy, wherein the cloud-edge collaborative mechanism comprises SM2 signature verification of a cloud-side instruction, local alarm triggering and execution rejection when verification fails, local data uploading according to a cloud-side instruction based on a priority, or cloud-side instruction issuing to an edge online model, and dynamic adjustment of model parameters; the decision fusion strategy adopts a hybrid inference mode, and simultaneously has a historical strategy cache and a network interruption decision logic.

[0044] Preferably, the collaborative decision layer comprises a cloud-edge collaborative mechanism and a decision fusion strategy, wherein the cloud-edge collaborative mechanism comprises SM2 signature verification of a cloud-side instruction, verification failure triggering local alarm and execution rejection, edge-side uploading of local data according to a cloud-side instruction based on a priority, or issuing of a cloud-side instruction to an edge online model, and dynamic adjustment of model parameters; the decision fusion strategy adopts a hybrid inference mode, and simultaneously has a historical strategy cache and a network interruption decision logic.

[0045] ​​Preferably, the devices of the perception control layer include existing devices and new devices, wherein the existing devices include smart meters, power distribution terminals, power monitoring instruments and distribution network SCADA systems, etc., and are read through IEC61850 protocol, MQTT protocol or Modbus-TCP; the new devices are external environment monitoring devices for collecting irradiance, wind speed and temperature / humidity of the distributed photovoltaic, and the installation mode adopts a spatial interpolation method to ensure that the distance from any position to the nearest monitoring point is less than a spatial threshold , and the minimum number of devices is calculated based on a Voronoi diagram coverage method :

[0046] ;

[0047] wherein, 、 are the length and width of the power station respectively, is a ceiling function.

[0048] Preferably, the feature extraction of the data processing platform adopts an adaptive weighted algorithm and a kernel principal component analysis algorithm, and the weight calculation formula is:

[0049] ;

[0050] wherein, is a same feature sequence set, is a target variable, 、 、 are information entropy, cross-correlation coefficient and volatility of respectively, is a hyperparameter, and satisfies .

[0051] As can be seen from the above technical solutions, the present application has the following beneficial effects:

[0052] The power distribution network edge gateway computing decision system based on the digital multi-level model provided in the present application realizes real-time collection of multi-source data through the perception control layer, realizes local real-time analysis and cloud-edge collaborative decision through the edge intelligent layer, optimizes resource allocation in combination with a dynamic adaptation mechanism, effectively solves the problems of insufficient real-time performance, concentrated computing pressure and difficulty in global collaboration of the traditional system, and has the advantages of improving response speed, reducing bandwidth cost and enhancing decision adaptability. BRIEF DESCRIPTION OF DRAWINGS

[0053] In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the drawings needed in the embodiments will be briefly described below. The features and advantages of the present application can be more clearly understood by referring to the drawings. The drawings are schematic and should not be construed as any limitation to the present application. Those skilled in the art can obtain other drawings according to these drawings without any creative effort. Among them:

[0054] Figure 1 A block diagram of a power distribution network edge gateway computing decision system based on a digital intelligent multi-level model provided by the present application;

[0055] Figure 2 A cloud-edge collaborative computing architecture diagram of the present application;

[0056] Figure 3 An edge gateway computing architecture diagram of the present application;

[0057] Figure 4 A flowchart of the digital intelligent model algorithm of the present application;

[0058] Figure 5 An online sparse modeling method diagram of the present application;

[0059] Figure 6 A multi-level model prediction diagram of user load and photovoltaic output of the present application. DETAILED DESCRIPTION

[0060] In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the drawings needed in the embodiments will be briefly described below. The features and advantages of the present application can be more clearly understood by referring to the drawings. The drawings are schematic and should not be construed as any limitation to the present application. Those skilled in the art can obtain other drawings according to these drawings without any creative effort. Among them:

[0061] Embodiment one:

[0062] In order to solve the problems in the prior art, such as the lack of real-time performance, the concentration of computing pressure, the lack of data transmission and storage security mechanism of edge devices, and the poor real-time performance, lack of global collaboration, and insufficient intelligent processing of existing centralized systems, edge independent systems, and cloud-edge collaborative architecture, as shown in Figure 1 The present application proposes a power distribution network edge gateway computing decision system based on a digital intelligent multi-level model. The system includes:

[0063] A perception control layer for collecting perception data of a low-voltage power distribution network, including operation data, device state data, and external environment data;

[0064] The edge intelligent layer comprises a plurality of hierarchical models connected in sequence, and the plurality of hierarchical models comprise a data processing center, a digital intelligent model library, a cross-validation layer, and a collaborative decision-making layer, wherein: the data processing center is configured to preprocess the sensing data, including error processing, anomaly detection, data conversion and feature extraction, data fusion, and protocol encapsulation; the digital intelligent model library is configured to analyze and predict the operation state of the power distribution network in real time based on the preprocessed data; the cross-validation layer is configured to dynamically correct and cross-validate the output results of the digital intelligent model library based on an adaptive kernel recursive least squares algorithm; and the collaborative decision-making layer is configured to generate edge dynamic decisions in combination with cloud instructions and local key data, and realize cloud-edge collaboration.

[0065] The cloud platform is configured to store full-amount data, globally collaborate and schedule, and visually display, and perform bidirectional parameter interaction with the edge intelligent layer.

[0066] In addition, the system further comprises a dynamic adaptation mechanism configured to adjust the hierarchical model parameters and the cloud-edge collaboration strategy in real time according to the operation state of the power distribution network, the state of the equipment resources, and the task urgency; the dynamic adaptation mechanism supports switching of multiple versions of lightweight models, including ultra-short-time, short-time, and long-time load prediction models, and sets a trigger condition based on a time scale.

[0067] As can be seen from the above technical solution, the present application proposes a power distribution network edge gateway computing decision system based on a digital intelligent hierarchical model. The sensing control layer refers to a set of sensing devices deployed at the end of the power distribution network, which can be implemented by using intelligent electric meters, environmental monitors, and the like. The sensing control layer is configured to build a global sensing network to capture multi-dimensional operation information. The edge intelligent layer refers to a cluster of computing modules deployed in the gateway device, which realizes the conversion from data to decision-making by building a hierarchical model, and solves the problem of single processing flow in the traditional architecture. The data processing center refers to a computing unit with data cleaning and feature extraction functions, which adopts an event-triggered sampling mechanism to reduce the amount of redundant data transmission. The digital intelligent model library refers to a prediction model that combines statistical methods and deep learning, which improves the prediction accuracy by combining partial least squares and time convolution networks. The cross-validation layer refers to a parameter correction module based on an adaptive algorithm, which eliminates the interference of abnormal data by online sparse modeling. The collaborative decision-making layer refers to a strategy generation module that integrates cloud instructions and local data, which adopts a hybrid reasoning mode to ensure the robustness of the decision-making, and uses an SM2 signature verification mechanism to ensure the security of the instruction transmission. The cloud platform refers to a resource scheduling center deployed in a remote server. The dynamic adaptation mechanism refers to a regulation and control module that switches models according to the task urgency, which realizes resource optimization configuration by managing lightweight model versions.

[0068] Compared with the prior art, the traditional centralized system relies on cloud backhaul instructions, which causes decision delay. The present scheme realizes millisecond-level response by constructing a complete decision chain on the edge side. The conventional edge-independent system lacks a model correction mechanism and is prone to accumulate prediction errors. The present scheme establishes a parameter dynamic adjustment mechanism through cross-validation layers to suppress bias diffusion. The existing cloud-edge collaborative architecture adopts a fixed strategy, which is difficult to adapt to environmental changes. The present scheme realizes real-time matching of model versions and resource states through a dynamic adaptation mechanism.

[0069] The present application further proposes a perception control layer device including existing devices and new devices. The existing devices include smart meters, power distribution terminals, power monitoring instruments, and distribution network SCADA systems, which read through IEC61850 protocol, MQTT protocol, or Modbus-TCP. The new devices are external environment monitoring devices for collecting irradiance, wind speed, and temperature and humidity of the distributed photovoltaic. The installation method adopts a spatial interpolation method to ensure that the distance from any position to the nearest monitoring point is less than a spatial threshold δ. The minimum number of devices is calculated based on the Voronoi diagram coverage method. The expression is where L and W are the length and width of the power station, respectively, is the ceiling function.

[0070] The spatial interpolation method refers to constructing a monitoring point network in the spatial dimension through mathematical methods, which can be implemented by Kriging interpolation or inverse distance weighted interpolation algorithm. It is used to generate the optimal monitoring point position when deploying new devices. The Voronoi diagram coverage method refers to dividing the planar region into several units, each containing a monitoring point. It can generate a polygon coverage area through computational geometry algorithms to determine the minimum number of devices that meet the distance constraint. IEC61850 protocol, MQTT protocol, or Modbus-TCP refers to the power system communication standard, which can be used for substation automation, Internet of Things data transmission, and industrial device communication, respectively, to realize data acquisition and integration of existing devices.

[0071] Specifically, in the distributed photovoltaic scenario, first, according to the length L and width W of the power station area, combined with the set spatial threshold δ (for example, it can be set to 50 meters), the minimum number of required monitoring devices is calculated through the formula. The new environmental monitoring devices are deployed according to the coordinates determined by the spatial interpolation method, for example, in the photovoltaic panel array, based on the historical irradiance data distribution law to select the installation point. The existing devices access the system through standard communication protocols, for example, the power distribution terminal uploads voltage data through the IEC61850 protocol, and the smart meter transmits power consumption information through the MQTT protocol. The monitoring area is divided by the Voronoi diagram to ensure that the distance between devices in each polygon area does not exceed , thereby avoiding monitoring blind spots.

[0072] Through the above technical solution, this application effectively solves the problem of low deployment efficiency of environmental monitoring equipment in distributed photovoltaic scenarios. By using mathematical modeling methods to accurately calculate the installation location and quantity of equipment, it avoids monitoring blind spots and reduces waste of hardware resources. The integration of standard communication protocols enables rapid access for multi-source heterogeneous devices, providing a complete and reliable data foundation for subsequent data processing.

[0073] This application further proposes a data processing platform (first-level model) comprising a data acquisition layer, a data processing layer, and a data transmission layer:

[0074] Data Acquisition Layer: The data acquisition layer communicates directly with physical devices, comprehensively collecting various parameters from the sensing and control layer devices. For data from external environmental monitoring devices, an event-triggered sampling mechanism is used, with the trigger function being... .in, For the current moment Monitoring parameters, The sampling interval is... This represents the standard deviation of historical data. When the variation of the monitored parameter exceeds three times the standard deviation, data should be collected promptly to reduce the amount of invalid data collected and improve data collection efficiency.

[0075] Among them, the event-triggered sampling mechanism refers to the technical means of dynamically adjusting the sampling frequency according to the change of the monitoring parameters. Specifically, it can be implemented by the standard deviation threshold comparison method. When the change of the monitoring parameters exceeds three times the standard deviation, data acquisition is triggered; otherwise, a low sampling frequency is maintained.

[0076] Specifically, the data acquisition layer connects to devices such as smart meters via the IEC61850 protocol or Modbus-TCP protocol to continuously acquire raw monitoring data streams. The data processing layer dynamically processes the raw data streams: when the fluctuation range of parameters such as current and voltage is less than three times the historical standard deviation, the event-triggered sampling mechanism reduces the sampling frequency to 1 / 3 of the base frequency, for example, from 1Hz to 0.33Hz.

[0077] Data Processing Layer: The collected data suffers from massive volume, redundancy, and strong correlation. In the error handling stage, for gross errors, statistical analysis is used to identify and remove obvious outliers; for random errors, filtering algorithms are used for smoothing; and for missing data, interpolation is performed based on data characteristics and historical patterns to complete the data. Anomaly detection is based on outlier detection formulas using statistical features. ,in This is the average of historical data. Normal distribution of Quantiles The significance level is used to accurately identify abnormal data points. Data conversion uses normalization methods to unify data of different magnitudes and distributions to a specific interval, facilitating subsequent analysis and processing.

[0078] The anomaly detection method refers to a technical means for identifying data outliers through statistical methods. Specifically, it can be implemented using the outlier determination rule based on the normal distribution confidence interval. The anomaly detection sensitivity is controlled by setting the significance level parameter a.

[0079] Specifically, when the temperature parameter mutation exceeds the range of ±2.58 times the standard deviation of the historical mean, the anomaly detection method automatically marks the data as an outlier and triggers the data cleaning process. The processed standardized data is uploaded to the edge intelligence layer using the MQTT protocol through the data transmission layer, while retaining the last 24 hours of data cache.

[0080] Feature extraction uses an adaptive weighting algorithm and kernel principal component analysis algorithm. The weight calculation formula is where is the same feature sequence set, is the target variable, , , are the information entropy, cross-correlation coefficient, and volatility of , respectively, is a hyperparameter that satisfies , effectively extracting key features of the data.

[0081] The adaptive weighting algorithm refers to a method of dynamically adjusting feature weights. Specifically, it can achieve multi-dimensional feature evaluation by measuring feature diversity through information entropy, evaluating feature correlation through cross-correlation coefficient, and reflecting feature stability through volatility. The kernel principal component analysis algorithm refers to a nonlinear dimensionality reduction method. Specifically, it can achieve data denoising by mapping the original features to a high-dimensional space and extracting principal components. Information entropy is used to quantify the information richness of features. Specifically, it can calculate the uncertainty of feature distribution using the Shannon entropy formula. Cross-correlation coefficient is used to represent the linear correlation strength between features and target variables. Specifically, it can be calculated using the Pearson correlation coefficient method. Volatility is used to measure the stability of feature data. Specifically, it can be achieved by calculating the ratio of the standard deviation to the mean of the feature value.

[0082] Specifically, in the feature extraction process of power distribution network operation data, first, a feature sequence set is constructed from the collected multi-dimensional data such as current, voltage, power, etc. The information content of each feature is evaluated by calculating the information entropy, for example, the information entropy of the current harmonic distortion rate feature is higher than that of the steady-state current feature. Then, the cross-correlation coefficient between each feature and the target variable is calculated, for example, the correlation coefficient of the voltage sag feature and the device fault state is higher than that of the temperature feature. At the same time, the volatility of each feature in the time dimension is calculated, for example, the volatility of the photovoltaic output data is significantly higher than that of the distribution transformer load rate data. After weighting and summing the three indicators according to the preset hyperparameters, the feature weight is normalized to form a feature selection mechanism with dynamic adaptability. Finally, the high-dimensional feature space is nonlinearly reduced by kernel principal component analysis, for example, the original features are mapped to the reproducing kernel Hilbert space using a Gaussian kernel function to extract the main components.

[0083] Compared with the prior art, the traditional feature selection method mainly uses fixed weights or single-dimensional evaluation indicators, which cannot dynamically adapt to the non-stationary characteristics of power distribution network operation data. For example, the conventional principal component analysis only considers the linear dimension reduction method of maximizing variance, which cannot effectively process nonlinear features such as photovoltaic output and load fluctuation. The feature selection method based on static threshold is prone to ignore the dynamic correlation between features, resulting in the loss of key features.

[0084] Through the above technical solutions, the application effectively solves the problem of poor adaptability of heterogeneous data feature selection in power distribution networks. By integrating multi-dimensional evaluation indicators and dynamic weight mechanisms, key features in different operating scenarios can be accurately identified, such as automatically increasing the weight of transient features in fault states and focusing on energy efficiency feature analysis in steady-state operation. Combined with the dimension reduction processing of kernel principal component analysis, the edge computing efficiency can be significantly improved while preserving the nonlinear characteristics of the data, providing reliable feature input for subsequent state prediction.

[0085] Data transmission layer: aggregate the processed data, remove duplicates and redundant parts, and encapsulate according to a specific protocol. According to the multi-level model selected by the edge gateway, the processed time series data is transmitted in real time and efficiently, ensuring smooth data flow within the system.

[0086] The application further proposes a digital model library (secondary model) based on partial least squares (PLS) and improved time convolution network (ITCN), which is composed of data-driven models and artificial intelligence-driven models, including load prediction models, photovoltaic output models, etc. The steps include:

[0087] Let the input matrix be and the corresponding output matrix be , n is the number of samples, m is the number of input features for each sample, for output dimension;

[0088] Extracting the implicit variables of input and output data by partial least squares ;

[0089] Extracting the implicit variables of input and output data by partial least squares as the input of improved time convolution network, constructing time convolution network, finding the nonlinear relationship mapping of implicit layer output vector as and output as , randomly giving input weight and residual , selecting as activation function, setting the number of implicit layer neurons as , calculating the output matrix of implicit layer :

[0090] ;

[0091] Calculating the output weight matrix and the predicted output matrix :

[0092] ;

[0093] wherein, is the Moore-Penrose generalized inverse matrix of , and is the real output matrix.

[0094] Finally, the PLS-ITCN data-driven model is obtained, and the model output is:

[0095] ;

[0096] wherein, are the projection weights of latent variables to the original output space, respectively, is the residual matrix.

[0097] Among them, partial least squares is a statistical method for reducing the dimension of high-dimensional data by projection. Specifically, implicit variables can be extracted by calculating the covariance structure of the input matrix and the output matrix to eliminate multicollinearity and reduce data redundancy. Improved time convolution network is a deep learning model with dilated causal convolution structure, which can be implemented by stacked convolution layers and asymmetric residual connection to capture long-term dependencies in time series. The activation function It belongs to the Sigmoid function type, and specifically, the input can be mapped to the 0-1 interval through a nonlinear transformation, which is used to enhance the model's fitting ability for nonlinear relationships. The number of hidden layer neurons L can be set according to the computing resources or data size, for example, set to an integer multiple of the number of input features, to balance the complexity of the model and the generalization performance.

[0098] Specifically, the input matrix and the output matrix are decomposed into a linear combination of hidden variables and residuals, and the hidden variables are extracted by the partial least squares method and input into the improved time convolution network. Inside the time convolution network, the hidden variables are subjected to multi-level convolution operations, and the output of each layer is subjected to nonlinear transformation through an activation function, and finally outputs a prediction result matching the dimension of the hidden variables. The output weight matrix is calculated by the generalized inverse matrix, so that the model can approximate the true output with the least square error. Thus, the model combines the linear dimension reduction advantage of partial least squares and the nonlinear time series modeling capability of the time convolution network, forming a composite data-driven architecture.

[0099] Compared with the prior art, the traditional method usually separately uses partial least squares regression or a single time series model, the former is difficult to handle nonlinear relationships, and the latter lacks dimension reduction capability for high-dimensional data. The present scheme cascades partial least squares and an improved time convolution network, retaining the advantages of linear feature extraction while enhancing the model's ability to capture complex time series patterns. In the prior art, when recursive neural networks are used to process time series data, there are problems of gradient disappearance and low computational efficiency. The present scheme uses the dilated convolution structure of the improved time convolution network, which can process time series data in parallel and reduce the time-consuming of calculation.

[0100] Through the above technical solutions, the present application can effectively handle the coupling relationship between high-dimensional heterogeneous data and complex time series characteristics in power distribution network operation, and improve the state prediction accuracy. At the same time, the model processes data in stages through linear and nonlinear modules, reducing the requirement for prior assumptions about data distribution and enhancing adaptability in different operating scenarios. In addition, the introduction of hidden variables reduces the interference of redundant features on model training, enabling edge gateways to achieve efficient modeling under limited computing resources.

[0101] The present application further proposes a hidden variable extraction formula, which is specifically expressed as the combination relationship of input matrix X and output matrix Y being decomposed into hidden variable matrix T, U and load matrix P, Q, and introducing residual matrix E, F to represent the decomposition error. The formula is specifically: . Wherein, and are the hidden variables of the input matrix and the output matrix, and are the load matrices of the input matrix and the output matrix, are the projection weights of the original input features to the latent variables, are projection weights of latent variables to original output space respectively, is the number of latent variables, and is a residual matrix.

[0102] wherein the latent variable refers to a low-dimensional core feature extracted from the original data through mathematical decomposition, which can be realized by a partial least squares projection algorithm, and is used to eliminate redundant information and retain the correlation features between the input and output. The load matrix refers to the projection coefficient matrix of the original data to the latent variable space, which can be solved by an iterative optimization algorithm, and is used to establish the mapping relationship between the original data and the latent variable. The residual matrix refers to the data residual that is not explained by the model in the decomposition process, which can be obtained by matrix subtraction operation, and is used to represent the decomposition error and guide the model parameter correction.

[0103] Specifically, in the distribution network edge gateway, the input matrix contains real-time monitoring data such as voltage and current, and the output matrix corresponds to the load prediction result. By decomposing the input and output data into a linear combination of latent variables and load matrices, the potential correlation between the two can be effectively captured, for example, in the scenario of distributed photovoltaic output fluctuation, the latent variable can extract the nonlinear coupling characteristics of light intensity and power output. The residual matrix is used to quantify the model fitting deviation, and when the residual exceeds the preset threshold, the model dynamic correction mechanism is triggered, for example, in the case of thunderstorm weather causing data anomaly, the sudden increase of residual can be identified as an external disturbance event.

[0104] Compared with the prior art, the traditional principal component analysis only reduces the dimension of the input data, without considering the correlation of the output variable, resulting in deviation of feature extraction from the actual prediction target. The present scheme decomposes the input and output matrices jointly, so that the latent variable reflects the coupling relationship between the device state and the operation target at the same time, for example, in the voltage sag event, the latent variable can simultaneously represent the fault characteristics and the protection action demand. In addition, the prior art often ignores the influence of residual on the robustness of the model, while the present scheme monitors the model error in real time through the residual matrix, for example, when the communication interruption causes data loss, the residual change can trigger the local caching strategy.

[0105] Through the above technical scheme, the present application can solve the problem of low model efficiency caused by high dimension of edge data and limited computing resources, for example, in the substation topology analysis task, the latent variable compresses the original 128-dimensional data to 8-dimensional, significantly reducing the computational complexity. At the same time, the dynamic monitoring mechanism of the residual matrix can improve the adaptability of the model to abnormal data, for example, when the device fault causes data jump, the residual analysis can avoid misjudgment as normal load fluctuation, thereby improving the decision reliability.

[0106] The present application further proposes an online sparse modeling method for cross-validation layer (three-level model) including:

[0107] Initialization of time series data sequence and parameters: initialization of time series data sequence and step parameter η, kernel width σ, quantization factor threshold θ, quantization factor , and other parameters, initial data dictionary , coefficients of elements in the data dictionary

[0108] Calculate the standard deviation of the sequence composed of at time ; calculate the standard deviation of the sequence composed of at time .

[0109] Calculate the neighbor distance of the new sample and the neighbor vector , and its corresponding range: and . If both are not within the above expected range, the new data sample is an isolated point, which will be discarded by the iterative learning process and not used for parameter updating.

[0110] wherein the neighbor distance range refers to the data acceptance interval set by the dynamic standard deviation and the quantization factor, which can be updated in real time by using the sliding window statistical method to realize, for eliminating the adaptability problem caused by fixed threshold. The quantization factor refers to the adjustment coefficient that controls the range boundary, which can be determined by offline training or online adaptive algorithm, for balancing the strictness of data screening and the model generalization ability. The expected neighbor vector refers to the reference vector sequence generated based on the historical data distribution, which can be generated by using the time series prediction model, for representing the data correlation characteristics under normal working conditions.

[0111] Specifically, when processing the edge data stream of the power distribution network in real time, the method first initializes the time series data sequence and the model parameters. For each newly arrived sample, calculate its Euclidean distance with the samples in the current data dictionary to obtain the neighbor distance statistics, and extract the neighbor vector direction feature. When the neighbor distance or vector direction of the new sample exceeds the dynamic range, it is determined as an abnormal point and discarded directly to avoid abnormal data entering the model updating process. For data within the range, further calculate the prediction error, and dynamically adjust the model parameters according to the error gradient. By continuously updating the standard deviation and the expected value, the method can adapt to the changes of the power distribution network operation state, ensuring the robustness of the sparse modeling process.

[0112] ​​​Compared to existing technologies, traditional outlier detection methods typically employ fixed thresholds or static statistical models, which are prone to misjudgment or missed detection in dynamic power distribution network scenarios, leading to a decrease in model accuracy. This solution introduces a time-series-based dynamic range calculation mechanism, enabling the anomaly detection boundary to automatically adjust as data distribution changes, effectively addressing the problem of insufficient model adaptability caused by environmental disturbances.

[0113] Through the above technical solution, this application can achieve real-time filtering of abnormal data and online optimization of model parameters under the condition of limited edge computing resources. For example, in scenarios of sudden changes in distributed photovoltaic output or load fluctuations, abnormal data caused by sensor failures or communication interference can be quickly identified, avoiding interference with the load prediction model, while reducing the consumption of ineffective computing resources and improving the decision-making efficiency and reliability of the edge gateway.

[0114] Calculate data samples Prediction error ,calculate Minimum distance to the current data dictionary Dynamically update online model parameters, if Keep the current data dictionary unchanged; otherwise, update the current data dictionary as follows: The updated formula is as follows ,in For the first The iteration of the ... Each model parameter value, For step size parameters, This represents the prediction error.

[0115] Among these, the online sparsity modeling method refers to actively removing outlier samples during time-series data processing through a nearest neighbor vector screening mechanism. Specifically, this can be achieved using a sliding time window combined with kernel density estimation. This technique can solve the model overfitting problem caused by data redundancy. The online model parameter update formula refers to a gradient descent algorithm based on prediction error feedback, specifically implemented using an adaptive learning rate adjustment strategy. This technique can dynamically adjust the parameter update step size according to real-time errors, avoiding oscillations and divergence during model iteration. Nearest neighbor distance refers to the Euclidean distance or cosine similarity of sample vectors in the feature space, specifically implemented using the k-nearest neighbor algorithm. This technique can effectively identify outlier data points. The expected range refers to a dynamic confidence interval constructed based on historical statistics, specifically implemented using the moving average method and standard deviation calculation. This technique can adapt to the characteristics of data distribution changing over time.

[0116] Specifically, the cross-validation layer performs real-time cleaning on the time series data stream through an online sparsification modeling method, for example, when the newly collected voltage fluctuation data exceeds the preset threshold range, it is determined as noise data and automatically filtered. The online model parameter update formula continuously optimizes the model parameters through an error feedback mechanism, for example, when the load prediction error exceeds the set threshold, the step size parameter is automatically increased to speed up the convergence speed. Through online data screening and dynamic parameter adjustment, the technical solution effectively solves the problem of prediction accuracy decline caused by data quality fluctuation in traditional edge computing models.

[0117] The application further proposes a collaborative decision-making layer (fourth-level model) including a cloud-edge collaboration mechanism and a decision fusion strategy. The cloud-edge collaboration mechanism includes SM2 signature verification of cloud instructions, local alarm triggering and execution rejection when verification fails, edge-side uploading of local data according to cloud instructions, or issuing of cloud instructions to edge online models, and dynamic adjustment of model parameters according to priority. The decision fusion strategy adopts a hybrid inference mode, has historical strategy caching and offline decision logic.

[0118] The SM2 signature verification refers to the process of verifying the digital signature of cloud instructions based on the SM2 algorithm, which can be implemented through elliptic curve encryption technology to ensure the legality and integrity of the instruction source and prevent malicious instruction injection. The local alarm triggering refers to automatically activating the alarm module when signature verification fails, which can be implemented by combining a preset alarm threshold with an audible and visual prompt to avoid system abnormalities caused by illegal instruction execution. The priority structured uploading refers to classified transmission of local data according to data criticality and timeliness, which can be implemented by combining data tags with bandwidth allocation strategies to optimize data transmission efficiency and reduce bandwidth occupation. The dynamic adjustment of model parameters refers to real-time updating of edge model weights or structure according to cloud instructions, which can be implemented by an online learning algorithm to adapt to changes in power grid operation state. The hybrid inference mode refers to a comprehensive decision-making method combining real-time data and historical strategies, which can be implemented by combining a rule engine with a machine learning model to improve the robustness of decision-making. The historical strategy caching refers to storing historical optimal decision-making schemes as backups, which can be implemented by a time series database to ensure that backup schemes can be called when real-time analysis fails. The offline decision logic refers to enabling local autonomous decision-making processes when the network is interrupted, which can be implemented by combining pre-set emergency rules with offline models to maintain basic system functions.

[0119] Specifically, when the cloud sends control instructions to the edge, the instructions are first verified by the SM2 algorithm for digital signature, and if the verification fails, a local alarm is triggered and the instruction is refused to be executed, preventing unauthorized operations. After successful verification, the edge uploads local data according to the preset priority, for example, device fault alarm data is transmitted first, while regular monitoring data is delayed, thereby optimizing network resource allocation. At the same time, the parameter adjustment instructions issued by the cloud are synchronized to the edge online model, for example, dynamically correcting the weight coefficient of the load prediction model, so that the model adapts to the fluctuation of photovoltaic output. In the decision-making process, the hybrid inference mode calls real-time analysis results and historical cache strategies, for example, combining the current grid load state with the historical optimal scheme of similar scenarios to generate decisions. When the network connection is interrupted, the offline decision logic immediately enables the historical strategy and preset rules cached locally, for example, based on the last effective cloud instruction and local data to maintain basic control functions.

[0120] Compared with the prior art, the conventional cloud-edge collaborative architecture lacks effective security verification mechanism and cannot identify illegal instructions, while the present scheme strengthens the credibility of the instruction source through the SM2 signature verification technology. The data upload strategy of the prior art usually adopts a fixed transmission mode, which is easy to cause bandwidth waste or delay of critical data, while the present scheme significantly improves data transmission efficiency through priority classification upload. In addition, the existing edge system is prone to functional paralysis when the network is interrupted, while the offline decision logic of the present scheme realizes degraded operation through local cache strategy and emergency rules, ensuring the basic availability of the system.

[0121] The present application further proposes that the cloud platform is responsible for full data storage, global collaborative scheduling and visual display. By establishing an efficient data storage architecture, massive power data is classified and stored for easy retrieval and analysis. Artificial intelligence algorithms are used to realize global collaborative scheduling to optimize power resource allocation. Visualization technology is used to present the power system operation state to users in the form of intuitive charts and graphs to assist decision-making. At the same time, the cloud platform and the edge intelligent layer perform bidirectional parameter interaction to realize continuous optimization of the model and improvement of system performance.

[0122] Taking the distributed photovoltaic and user load model on June 1, 2022 as an example, the user load data is collected by the intelligent electric meter in the transformer area, and the prediction result is output every 15 minutes. The distributed photovoltaic data, including historical data and real-time environmental data, is collected, and the output is also output every 15 minutes. During the running process, the perception control layer continuously collects data and transmits to the edge intelligent layer. In the data processing, the data is preprocessed by the station, the load and photovoltaic output prediction is carried out by the digital model library, and the prediction result is corrected by the cross-validation layer. The collaborative decision-making layer makes decisions combined with cloud instructions and local data, and the cloud platform realizes data storage, scheduling and visual display. From the experimental results (such as the output result graph of user load and photovoltaic output), the output result of the three-level model has a correction effect on the output result of the two-level model, which effectively improves the accuracy of the model prediction result, and verifies the effectiveness and reliability of the system.

[0123] Through the above technical solutions, the application effectively solves the problem of lack of cloud instruction security verification, avoids system risks caused by illegal operations, optimizes data transmission efficiency, reduces network bandwidth pressure, enhances the autonomous decision-making ability of the edge in the case of network anomalies, and ensures the continuity and stability of the power distribution network operation.

[0124] Those skilled in the art will appreciate that embodiments of the application can be provided as methods, systems, or computer program products. Therefore, the application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0125] The application is described with reference to flowcharts and / or block diagrams according to the method, equipment (system), and computer program product of the embodiments of the application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing equipment to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing equipment produce a means for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one flow or multiple flows and / or blocks Figure 1 The device for implementing the functions specified in one block or multiple blocks.

[0126] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the Figure 1 function specified in the flow or flows and / or blocks of the block or blocks. These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the function specified in the flow or flows and / or blocks of the block or blocks. Figure 1 Figure 1 function specified in the flow or flows and / or blocks of the block or blocks. These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the function specified in the flow or flows and / or blocks of the block or blocks. Figure 1

[0127] Obviously, the above-described embodiments are only examples and are not intended to limit the present application. Based on the above description, other different forms of changes or modifications can be made by those skilled in the art. Here, it is not necessary or possible to exhaust all the embodiments. The obvious changes or modifications derived therefrom are still within the protection scope of the present application.​​

Claims

1. A power distribution network edge gateway computing decision system based on a digital intelligent multi-level model, characterized in that, The system comprises: a perception control layer for collecting perception data of a low-voltage power distribution network, including operation data, device state data and external environment data; an edge intelligence layer comprising a plurality of connected levels, including a data processing center, a digital model library, a cross-validation layer and a collaborative decision-making layer, wherein: the data processing center is used for pre-processing the perception data, including error processing, anomaly detection, data conversion and feature extraction, data fusion and protocol packaging; the digital model library is used for real-time analysis and prediction of the operation state of the power distribution network based on the pre-processed data, and is constructed based on partial least squares and improved time convolution network, including the following steps: Let input matrix X ∈ R n×m and corresponding output matrix Y ∈ R n×s , n is the number of samples, m is the number of input features of each sample, and s is the output dimension. The implicit variables (t i , u i ) of the input and output data are extracted by partial least squares, and an implicit variable extraction formula is: where R = {t1, t2, …, t r} T ∈R n×r and U = {u1, u2, …, u r} T ∈R n×r are the input and output matrices of hidden variables, respectively, P = {p1, p2, …, p r} T ∈R m×r and Q = {q1, q2, …, q r} T ∈R m×r are the input and output matrices of hidden variables, respectively, P = {p1, p2, …, p r are the projection weights of the original input features to the latent variables, q1, q2, …, q r are the projection weights of the latent variables to the original output space, r is the number of hidden variables, and E and F are the residual matrices. The implicit variable T extracted by the partial least squares is taken as an input of the improved time convolution network, a time convolution network is constructed, and a nonlinear relationship mapping is searched for an implicit layer output vector h1, h2, …, h L and an output u1, u2, …, u r , a random input weight w i and a residual error b i are given, a g(x) = 1 / (1+e -x ) is selected as an activation function, the number of implicit layer neurons is set to L, and an output matrix H of the implicit layer is calculated. computing an output weight matrix β and a prediction output matrix wherein is the molar-Penrose generalized inverse matrix of H, and u is the true output matrix. a PLS-ITCN data-driven model is finally obtained, and the model output is: In the formula, q1, q2, …, q r are respectively the projection weights of latent variables to the original output space, F r is a residual matrix; cross-validation layer, used for dynamic parameter correction and cross-validation of the output results of the intelligent model library based on the adaptive kernel recursive least square algorithm, including an online sparse modeling method: initializing the time series data sequence and parameters; calculating the neighbor distance and neighbor vector of a new sample, if the neighbor distance is outside the expected range [NDe(i)-λ d σ d (i), NDe(i)+λ d σ d (i)], the sample is determined as an isolated point and discarded; dynamically updating the online model parameters according to the prediction error and the minimum distance of the data dictionary, and the online model parameter updating formula is: v σ v (i), VDe(i)+λ v σ v (i)]; β j (i) = β j (i-1) + ηe(i); where β j (i) is the jth model parameter value of the ith iteration, η is a step size parameter, e(i) is the prediction error, λ d and λ v are quantization factors, σ d (i) and σ v (i) are standard deviations, NDe(i) is the expected nearest neighbor distance at time i, and VDe(i) is the expected nearest neighbor vector at time i. the collaborative decision-making layer is used to generate edge dynamic decisions by combining cloud instructions and local key data, and to realize cloud-edge collaboration; the cloud platform is used for full data storage, global collaborative scheduling and visual display, and performs bidirectional parameter interaction with the edge intelligence layer.

2. The digital multi-level model based power distribution grid edge gateway computing decision system of claim 1, wherein, The system also includes a dynamic adaptation mechanism for real-time adjustment of multi-level model parameters and cloud-edge collaboration strategies according to the operation state of the power distribution network, the state of the device resources and the task urgency; The dynamic adaptation mechanism supports switching of multiple versions of lightweight models, including ultra-short-time, short-time and long-time load prediction models, and sets trigger conditions based on time scales.

3. The digital multi-level model based power distribution grid edge gateway computing decision system of claim 1, wherein, The data processing center includes a data acquisition layer, a data processing layer and a data transmission layer, wherein the data acquisition layer includes an event-triggered sampling mechanism, and the trigger function is: In the formula, x(t) is the monitoring parameter at the current time t, Δt is the sampling interval, σ x is the standard deviation of the historical data; The data processing layer includes an anomaly detection method based on a statistical feature outlier detection formula: In the formula, μ x This is the average of historical data. The distribution is normally distributed N(μ) x ,σ x )of Quantiles, where α represents the significance level.

4. The digital multi-level model based power distribution grid edge gateway computing decision system of claim 1, wherein, The collaborative decision-making layer includes a cloud-edge collaboration mechanism and a decision fusion strategy, wherein the cloud-edge collaboration mechanism includes SM2 signature verification of cloud instructions, local alarm triggering and execution rejection when verification fails, edge data structured uploading according to priority or cloud instruction issuing to edge online model according to cloud instructions, and dynamic adjustment of model parameters; The decision fusion strategy adopts a hybrid reasoning mode, and has historical strategy cache and offline decision logic.

5. The digital multi-level model based power distribution grid edge gateway computing decision system of claim 1, wherein, The devices of the perception control layer include existing devices and new devices. The existing devices include smart meters, power distribution terminals, power monitoring instruments, and distribution network SCADA systems, etc., which are read through IEC61850 protocol, MQTT protocol or Modbus-TCP. The new devices are external environment monitoring devices for collecting irradiance, wind speed and temperature / humidity of the distributed photovoltaic, and the installation method adopts a spatial interpolation method to ensure that the distance from any position to the nearest monitoring point is less than a spatial threshold δ, and the minimum number of devices N is calculated based on a Voronoi diagram coverage method min : where L, W are length and width of the power plant, respectively, is rounded up.

6. The digital multi-level model based power distribution grid edge gateway computing decision system of claim 1, wherein, The feature extraction of the data processing center adopts an adaptive weighted algorithm and a kernel principal component analysis algorithm, and the weight calculation formula is: where {x i ,i = 1, …, N} is the same feature sequence set, y is the target variable, Info(x i ), Relevance(x i ,y), Stability(x i ) are the information entropy, the correlation coefficient and the volatility of x i , respectively, and α, β, γ are hyperparameters satisfying α + β + γ = 1.

Images