Intelligent prediction and management method for load change trend of low-voltage distribution network

Abstract

The invention provides a low-voltage power distribution network load change trend intelligent prediction and management method, belongs to the field of low-voltage power distribution network load prediction, and is used for solving the problems of large load fluctuation, insufficient prediction precision and high cloud deployment delay of a hybrid industry transformer area in related technologies. The method is deployed at an edge node of a transformer area, high-quality data is output through multi-modal data anomaly detection and scene completion, a four-dimensional dynamic load portrait is constructed based on the high-quality data, model super-parameter self-adaptive parameter adjustment is realized by combining transfer learning and Bayesian optimization, and accurate load data is output through three-dimensional linkage resource scheduling and dynamic fusion residual error correction. The method improves the load prediction precision and efficiency, reduces the response delay, and can effectively support the real-time scheduling of the power distribution network.

Description

Technical Field

[0001] This application relates to the field of low-voltage distribution network load forecasting, and in particular to an intelligent forecasting and management method for low-voltage distribution network load change trends. Background Technology

[0002] Low-voltage distribution networks are the final link connecting the power system to users, and their load changes directly affect the stability of power supply. With the widespread adoption of charging piles and the integration of industrial and residential electricity consumption, the load on distribution networks presents a complex situation with multiple overlapping characteristics, making load forecasting a core support for power grid dispatching and planning.

[0003] Currently, load forecasting for low-voltage distribution networks mostly employs traditional time series models or general machine learning algorithms, with some reliance on cloud computing power for data processing and model training. These methods have been applied in single-business scenarios, providing technical references for basic load forecasting.

[0004] However, existing technologies have significant drawbacks: First, they are difficult to adapt to the load differences of mixed business formats such as "charging piles-industry-residential", and are not well adapted to the characteristics of short-term high fluctuations of charging piles and the stability of industrial loads; second, cloud deployment leads to high prediction delays, which cannot meet the local dispatching needs of distribution areas; third, data processing is crude and model parameter tuning efficiency is low, making it difficult to guarantee the overall prediction accuracy, which restricts the improvement of the level of lean management of distribution networks. Summary of the Invention

[0005] This application provides an intelligent prediction and management method for load change trends in low-voltage distribution networks, which can adapt to the characteristics of mixed business loads, improve prediction accuracy and efficiency, and meet the needs of local dispatching of distribution networks.

[0006] Firstly, this application provides an intelligent prediction and management method for load change trends in low-voltage distribution networks. This method, applied to edge computing nodes in distribution areas, includes: collecting multi-source data from a distribution network encompassing "charging piles - industry - residential" mixed-use scenarios; outputting high-quality data through anomaly detection, scenario-based completion, and feature enhancement processing; calculating four-dimensional load characteristic indicators based on this data, constructing a dynamic user load profile, and adaptively evolving it according to load fluctuation levels; combining profile feature labels, initializing the hyperparameters of the long short-term memory network model using transfer learning, and performing lightweight optimization to adapt the model to the characteristics of mixed-use loads; dynamically allocating computing power according to resource scarcity, user priority, and scenario urgency, synchronizing the scheduling results to the distribution automation system in the distribution area; dynamically fusing the prediction results of the model and association rules, outputting load data adapted for scheduling after residual correction, and feeding back updated profile parameters.

[0007] By adopting the above technical solutions and using edge nodes as deployment carriers, a closed-loop system of "data processing - profile building - model optimization - resource scheduling - fusion prediction" is constructed. This system is specifically adapted to the load characteristics of mixed business formats and reduces cloud transmission latency. The combination of four-dimensional profiling and transfer learning improves model adaptability and parameter tuning efficiency. Dynamic fusion and residual correction further ensure prediction accuracy and provide precise data support for local dispatching of distribution areas.

[0008] Furthermore, anomaly detection and completion include: first, marking suspected anomalies according to the daily periodicity of the distribution network load; then, eliminating non-global anomalies caused by local faults by combining the branch topology of the distribution area; and finally, determining valid anomalies by using a model trained with typical anomaly data of the distribution network. Single-point missing data caused by communication jitter is completed using time-attenuation weighted data; short-to-medium-term missing data caused by equipment failure is replaced by loads of the same type of equipment; and long-term missing data caused by link interruption is completed using weighted data of neighboring users within the same branch.

[0009] By adopting the above technical solutions, and combining the distribution network topology and load characteristics, anomalies can be accurately identified. The differentiated completion strategy can be adapted to different data missing scenarios, effectively improving data quality, laying a reliable data foundation for subsequent prediction, and reducing prediction deviations caused by data anomalies.

[0010] Furthermore, the calculation and evolution of the four-dimensional load characteristic indicators include: combining the regularity of electricity consumption with the weighting of daily and weekly load fluctuations; adjusting the load elasticity according to the seasonal adjustment of electricity price and temperature response weights; introducing the industry-specific contribution coefficient into the equipment weights; combining the spatiotemporal correlation with the load similarity and topological distance attenuation characteristics; and the profile evolves according to the load fluctuation levels, prioritizing the updating of load elasticity indicators in high fluctuation scenarios, with different levels corresponding to different characteristic indicator update ranges.

[0011] By adopting the above technical solutions, the four-dimensional indicators accurately quantify the load characteristics of mixed business formats, and the business format-specific coefficients enhance the adaptability of scenarios. The hierarchical evolution mechanism enables the profile to dynamically match load changes, and the targeted updates for high-fluctuation scenarios further improve the fit between the profile and the real-time load, providing accurate feature support for model optimization.

[0012] Furthermore, the transfer learning initialization and lightweight optimization include: constructing a distribution network user profile-model hyperparameter mapping library with multiple business types and load levels, associating each data with a four-dimensional profile and the optimal hyperparameter; calculating the load characteristics and topological dual-dimensional similarity between the target user and the profile in the library, prioritizing matching similar profiles within the same transformer area and filtering the results, and generating initial hyperparameters by combining similarity and historical accuracy weights.

[0013] By adopting the above technical solutions, the multi-dimensional mapping library provides a rich basis for transfer learning, the dual-dimensional matching of topology and load ensures the adaptability of initial hyperparameters, and the priority matching rule for the same transformer area fits the topological isolation characteristics of the distribution network, effectively reducing the cost of parameter tuning and improving the accuracy of model initialization.

[0014] Furthermore, the three-dimensional linkage resource scheduling includes: dynamically adjusting the weights of computing, memory, and storage resources according to the load period based on resource stress; combining user priority with spatiotemporal correlation, load elasticity, electricity consumption patterns, and emergency coefficients of scheduling procedures; prioritizing high-priority users when computing power is sufficient, restricting low-priority computing power when stressed, and prioritizing support for the fault handling needs of the power distribution automation system when overloaded.

[0015] By adopting the above technical solutions, dynamic weight and priority design enables on-demand resource allocation, the combination of load periods and scheduling procedures enhances scenario adaptability, the fault handling priority mechanism during overload ensures the emergency needs of the power grid, and improves the resource utilization efficiency and scheduling coordination capabilities of edge nodes.

[0016] Furthermore, dynamic fusion includes: model weights are calculated by combining time-series attention, credibility, historical error, and business type coefficients; attention is adjusted according to the degree of load fluctuation; credibility is determined based on historical performance in different scenarios; the fusion results are converted into daily load data at 15-minute intervals and short-term load data at 1-hour intervals, with the format conforming to the communication standards of distribution automation systems.

[0017] By adopting the above technical solutions, multi-factor weight calculation enables the fusion strategy to adapt to load fluctuations and differences in business types, while time-series attention enhances the prediction accuracy during critical periods; the standardized output format enables seamless integration with the scheduling system, reduces data conversion steps, and improves the engineering practicality of the prediction results.

[0018] Furthermore, residual correction includes: splitting historical prediction residuals according to the peak and valley periods of the distribution network, filtering non-periodic fault residuals through fault labels; constructing residual models for peak and valley periods respectively, with peak periods focusing on short-term load fluctuation residuals and valley periods focusing on stable load residuals, and using residual prediction values ​​to correct the fusion results.

[0019] By adopting the above technical solutions, peak-valley segmentation and fault filtering ensure the periodicity and effectiveness of residual data, targeted residual modeling improves correction accuracy, effectively offsets system errors, further optimizes prediction results, and reduces prediction deviation during peak-valley periods.

[0020] Furthermore, lightweight optimization includes: in the coarse search stage, the hyperparameter space is screened according to distribution network indicators such as peak-valley error, and areas with substandard accuracy are eliminated; the fine search adopts a function iteration that balances exploration and utilization, with the number of rounds adapted to the computing power of edge nodes, and real-time monitoring of GPU memory during training, triggering model quantization when the memory usage is too high.

[0021] By adopting the above technical solutions, the index-oriented coarse search narrows the hyperparameter range, the balanced iteration and computing power adaptation ensure optimization efficiency, and the memory monitoring and quantization mechanism controls resource consumption, so that the model optimization process can fully adapt to the computing power constraints of edge nodes and improve optimization feasibility.

[0022] Furthermore, the method implementation relies on a three-dimensional dynamic parameter library, which includes: a library of equipment contribution coefficients determined based on the measured load of the equipment, a library of scenario emergency coefficients based on the scheduling procedures, and a library of user attribute weights based on the load ratio of business types; and adopts an iterative mechanism that combines quarterly full updates with incremental updates triggered by changes in equipment in the distribution area.

[0023] By adopting the above technical solutions, the three-dimensional parameter library provides accurate parameter support for the implementation of the method, and the reliability of the parameters is ensured by actual measurements and procedures; the dynamic iteration mechanism adapts to equipment changes, incremental updates reduce maintenance costs, and ensure the timeliness and accuracy of the parameter library.

[0024] Furthermore, the entire process is deployed at outdoor edge nodes of the distribution network. The adaptation strategies include: the hardware adopts a wide temperature design and IP65 protection level, and the power supply supports a wide voltage input; the software optimizes resources through process isolation and memory pools, triggers a low-power mode when the load of the transformer area is lower than 20% of the rated load for 1 hour, and the communication adopts the power industry standard protocol.

[0025] By adopting the above technical solutions, the hardware design is adapted to harsh outdoor environments, and the wide voltage input can cope with power supply fluctuations; software optimization and low power consumption mechanisms reduce resource consumption, and standard communication protocols ensure system compatibility, thus realizing the engineering implementation and stable operation of the method.

[0026] In summary, this application has at least the following beneficial effects: It provides a load forecasting method that adapts to mixed business formats, improves accuracy and efficiency, and meets the needs of local dispatching; Topology constraint data processing and dynamic profiling technology enhance scenario adaptability; edge node resource optimization and standardized output improve engineering practicality.

[0027] It should be understood that the description in the Summary Section is not intended to limit the key or essential features of the embodiments of this application, nor is it intended to restrict the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0028] The above and other features, advantages, and aspects of the embodiments of this application will become more apparent from the accompanying drawings and the following detailed description. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein: Figure 1 A schematic diagram of an exemplary operating environment in which embodiments of this application can be implemented is shown.

[0029] Figure 2 A flowchart of an intelligent prediction and management method for load change trends in a low-voltage distribution network, as described in an embodiment of this application, is shown. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0031] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0032] This application provides an intelligent prediction and management method for load change trends in low-voltage distribution networks. It is adapted to the characteristics of mixed business loads, improves prediction accuracy and efficiency, reduces latency through edge deployment, and can accurately support local dispatching and lean management of distribution networks.

[0033] Figure 1 A schematic diagram of an exemplary operating environment in which embodiments of this application can be implemented is shown.

[0034] Reference Figure 1 The operating environment includes a hardware system that supports the entire process of intelligent prediction and management of load change trends in low-voltage distribution networks. This system uses data acquisition as the input, edge computing as the core processing unit, communication transmission as the link, scheduling system as the collaborative object, and data storage as the support. All parts work together to form a closed-loop operating architecture, ensuring the efficient implementation of all aspects of the method, such as data processing, profile construction, and model optimization.

[0035] The core components of this operating environment include multimodal data acquisition hardware, which is deployed at the main meter of the transformer substation, branch line nodes, and user power consumption side. Specifically, it includes electrical quantity acquisition equipment, non-electrical quantity acquisition equipment, and spatiotemporal topology acquisition equipment, which are used to collect electrical parameters such as active power and voltage, environmental and operating condition parameters such as temperature, humidity, and electricity price, as well as GIS-based transformer substation branch topology information. All acquisition equipment establishes a data transmission link with the core processing unit through a communication interface.

[0036] The edge computing node hardware, which serves as the core processing unit, is deployed locally within the distribution area and is equipped with adapted CPU, GPU computing modules, memory modules, and local storage modules. Its core functions are to receive raw data transmitted by the acquisition hardware, execute core steps of methods such as anomaly detection, model training, and resource scheduling, and output control commands and data requests to other hardware units. It is the central hub of the entire operating environment.

[0037] The communication transmission hardware connecting all hardware covers three levels of links: short-range communication links between the acquisition devices and edge computing nodes, wide-area communication links between edge computing nodes and power distribution automation systems, and is built using 4G / 5G communication modules or industrial Ethernet interfaces to ensure the real-time performance (latency ≤500ms) and reliability (transmission success rate ≥99.5%) of data transmission.

[0038] The hardware of the distribution automation system is the scheduling and coordination object of the method, which includes a scheduling host, a fault identification module and a data interaction interface. It is used to receive resource scheduling results transmitted by edge computing nodes, feed back distribution network fault tags and operating status data, and realize the time-series coordination of predictive calculation and power grid scheduling instructions.

[0039] The data storage hardware is divided into two parts: local cache at the edge nodes and historical storage in the background. The local cache uses a Redis module to store real-time collected data and temporary calculation results, while the historical storage uses a PostgreSQL database server to store historical load data, model hyperparameter mapping library and 3D dynamic parameter library. Both of them establish bidirectional data interaction with the edge computing nodes to meet the dual requirements of real-time processing and historical data retrieval in the method.

[0040] Each hardware component works together through a pre-defined communication protocol and data interface. The acquisition hardware uploads multi-dimensional data to the edge computing node in real time. The edge computing node calls the historical data and parameters in the storage hardware to complete the processing and outputs the scheduling adaptation data to the power distribution automation system. At the same time, it saves the updated profile parameters and model hyperparameters back to the storage hardware, forming a closed loop of "acquisition-processing-coordination-storage", providing stable and reliable hardware support for the entire method.

[0041] This application discloses an intelligent prediction and management method for load change trends in low-voltage distribution networks.

[0042] Figure 2 A flowchart of an intelligent prediction and management method for load change trends in a low-voltage distribution network, as described in an embodiment of this application, is shown.

[0043] Reference Figure 2This method is specifically applied to edge computing nodes in distribution networks, with the entire process deployed on outdoor edge nodes of the power distribution network. The hardware adopts a wide-temperature design and IP65 protection level, and the power supply supports a wide voltage input. The software optimizes resources through process isolation and memory pools. When the load in the distribution area is lower than 20% of the rated load for 1 hour, a low-power mode is triggered. Communication adopts the power industry standard protocol. The method specifically includes the following steps: S1: Collects multi-source data of power distribution networks from mixed business formats including charging piles, industry and residential, and outputs high-quality data through anomaly detection, scenario-based completion and feature enhancement processing.

[0044] The specific methods in this step include: the implementation of the method relies on a three-dimensional dynamic parameter library, which includes a library of equipment contribution coefficients determined based on the measured load of the equipment, a library of scenario emergency coefficients based on the dispatching procedures, and a library of user attribute weights based on the load ratio of different business types. An iterative mechanism combining quarterly full updates and incremental updates triggered by changes in transformer area equipment is employed. Anomaly detection and completion include first periodically marking suspected anomalies according to the daily load of the distribution network, specifically by calculating the average load data at the same time over the past 7 days. and standard deviation ,in accordance with Criterion mark satisfies The load data is suspected to be abnormal, among which This is the load data at the current moment. and All data are statistical results from the same period in the historical load database of the distribution network.

[0045] The update process for the 3D dynamic parameter library is as follows: a full update is triggered at the end of each quarter, synchronizing the device contribution coefficient, scenario urgency coefficient, and user attribute weight; when a device is added or removed from a distribution area, an incremental update is triggered, updating only the contribution coefficient of the corresponding device and the attribute weight of the associated user.

[0046] By combining the branch topology of the transformer area, non-global anomalies caused by local faults are eliminated. That is, by using the GIS topology data of the transformer area to determine whether the suspected anomaly is only concentrated in a certain branch line, if it is a local branch anomaly, it is determined to be a non-global anomaly and is not included in the subsequent processing procedure.

[0047] Then, a model trained with typical distribution network anomaly data is used to determine valid anomalies. This model employs the Isolation Forest algorithm, and the training samples contain 100,000 real distribution network anomaly data points, covering three typical scenarios: communication jitter, equipment failure, and link interruption. The model outputs anomaly scores, and when the score is... Anomalies are identified as valid anomalies, with the anomaly score calculated using path length normalization of the isolated forest.

[0048] Single-point missing data caused by communication jitter is filled using time-attenuation weighted filling, and the filling formula is as follows: Where X is the completed data, The time decay coefficient is set to 0.6 (determined based on actual measurements of the time-series correlation of the distribution network load). , These are normal load data adjacent to the abnormal time, respectively, and the data comes from the load sequence collected in real time.

[0049] For short- to medium-term missing data caused by equipment failure, load replacement with the same type of equipment is used to fill the gap. This means selecting load data of equipment of the same type, power level and operating scenario as the failed equipment as the replacement source. The replacement source data comes from the correlation and matching results between the distribution network equipment ledger and the real-time operation database.

[0050] Long-term data loss due to link interruption is filled using weighted imputation of neighboring user data within the same branch. The imputation formula is as follows: ,in For the concurrent load data of the i-th neighboring user, Let be the weight coefficient of the i-th neighboring user, satisfying , The spatiotemporal correlation coefficient between the target user and the i-th neighboring user (derived from the user attribute weight library). Let be the topological distance attenuation coefficient, satisfying , The branch line topology distance (in meters) between the target user and the i-th neighboring user is defined. Neighboring users are limited to those on the same branch line and whose topology distance is specified. Within a range of meters, a maximum of 3 users can be selected as the effective neighborhood.

[0051] The data cleaning stage also requires removing the total active power. ,Voltage or Invalid outliers; if user behavior tags are missing, use... The KNN algorithm, based on the similarity of load curves, selects 5 similar users for label completion, improving data integrity to [percentage missing]. Simultaneously, the data needs to undergo dual-mode standardization processing, including electrical quantities ( Z-score standardization is used, and the formula is as follows: Where x is the original electrical quantity, The mean of this parameter for the training set. The standard deviation of this parameter is used for the training set. and Based on 8760 hours of historical data throughout 2024; non-electrical quantities (temperature, electricity price) were standardized using the Min-Max method, with the formula as follows: ,in The parameter's historical minimum value. The values ​​represent the historical maximum values ​​of the parameters, all derived from statistical data collected throughout 2024.

[0052] Feature enhancement processing is completed based on the aforementioned high-quality data to support subsequent steps. Specifically, it is achieved by calculating spatiotemporal-attribute fusion features, with the core formula being: , where H is the spatiotemporal entropy (characterizing load concentration). The weights of neighboring user attributes (derived from the user attribute weight library). Let t be the load ratio between the target user and the nth neighboring user, satisfying , For target user load data at time t, For the load data of the nth neighboring user at time t, To avoid logarithmically meaningless correction coefficients, this feature enhances the spatiotemporal distribution characteristics of quantifiable loads, thereby improving the accuracy of subsequent profile construction.

[0053] S2: Based on this data, calculate the four-dimensional load characteristic index, construct a dynamic profile of user load, and adaptively evolve according to the degree of load fluctuation.

[0054] This step specifically includes: calculation and evolution of four-dimensional load characteristic indicators, which combines electricity consumption regularity with daily and weekly load fluctuation weighting. The core of this indicator is used to quantify the temporal stability of user load, and the calculation formula is as follows: R is the electricity consumption regularity index (ranging from 0 to 1, with values ​​closer to 1 indicating stronger regularity), and 0.6 and 0.4 are weighting coefficients based on actual measurements of 100 mixed-business transformer areas (stored in the user attribute weight library). The standard deviation of the 24-hour load data for a single day. denoted as the standard deviation of load data for the same period within a week, and μ as the average load within the calculation period. All load data are derived from high-quality data processed by S1. Typical R value for industrial users. Typical R value for residential users This threshold is determined by statistical analysis of the load characteristics of different business types in the equipment contribution coefficient database.

[0055] Load resilience, with seasonally adjusted electricity prices and temperature response weights, is used to characterize the load's sensitivity to external factors. The calculation formula is as follows: Where E is the load elasticity index, and Electricity price and temperature response weights (summer) In winter, the weight values ​​are stored in the scenario emergency coefficient library. This represents the load change. This represents the change in electricity price (derived from electricity price data collected by S1). This represents the temperature change (derived from ambient temperature and humidity data collected by S1). Typical E value for a concentrated charging station area. This feature is used for targeted optimization of subsequent model parameters.

[0056] Equipment weights incorporate business-specific contribution coefficients from the business-specific contribution coefficient library to differentiate the impact of different equipment on the overall load. The calculation formula is as follows: Where W is the equipment weight, Pi is the rated power of a single equipment, and Ti is the average daily operating time of the equipment (derived from equipment operating data collected by S1). Contribution coefficient specific to the business type (charging piles) Industrial motors Residential lighting The coefficient value is stored in the equipment contribution coefficient library. The numerator is the sum of the load contribution of a single device, and the denominator is the total load of the transformer area. This indicator can highlight the impact of high-contribution devices on load forecasting.

[0057] Spatiotemporal correlation, combining load similarity and topological distance attenuation characteristics, is used to quantify the load correlation between a target user and its neighboring users. The calculation formula is as follows: Where S is the spatiotemporal correlation coefficient (ranging from 0 to 1, with values ​​closer to 1 indicating stronger correlation), t is the time scale within a day (1-24 hours), and n is the selected neighboring user IDs (maximum of 5). and d represents the load data of the target user and the nth neighboring user at time t, respectively, and d represents the branch topology distance between the target user and the neighboring user (unit: meters, derived from the spatiotemporal topology data collected by S1). This is a topological distance decay factor, ensuring that users who are closer to each other have higher weights.

[0058] The load profile evolves adaptively based on the degree of load fluctuation, which is quantified by load variance and calculated using the following formula: ,in Let N be the load variance, and N be the sample size of the load data. For single-sample loading values, This is the sample mean. When... Time-triggered micro-evolution updates only the spatiotemporal correlation coefficients. ;when Time-triggered evolution updates electricity consumption patterns , load elasticity Spatiotemporal correlation coefficient S; when When the full evolution is triggered (such as during peak charging periods for electric vehicles), the load elasticity E is updated first, and then other three-dimensional indicators are updated synchronously. The evolution time is controlled within 2 seconds to adapt to the real-time processing needs of edge nodes.

[0059] Different levels correspond to different feature index update ranges. During the calculation process, relevant parameters from the three-dimensional dynamic parameter library are called to ensure the accuracy of the indicators. The updated four-dimensional feature indicators will serve as the core content of the user load dynamic profile, providing accurate feature label support for subsequent model training.

[0060] S3: Combining portrait feature labels, the hyperparameters of the long short-term memory network model are initialized using transfer learning, and then the model is optimized to adapt to the load characteristics of mixed business formats.

[0061] The specific methods in this step include: transfer learning initialization and lightweight optimization, including the construction of a distribution network user profile-model hyperparameter mapping library with multiple business types and load levels. This mapping library contains 1,500 typical user data of the distribution network, covering three business types: industrial, commercial, and residential; five load levels: below 10kW to above 50kW; and four scenarios: weekdays, weekends, holidays, and extreme weather. Each data point is associated with the "four-dimensional feature index (R, E, W, S) output from step S2 - optimal hyperparameter of the long short-term memory network" pair. The core hyperparameters are the hidden layer dimension n_h and the learning rate η. The data source is the actual test training results of 20 pilot transformer areas in a provincial distribution network over 6 months.

[0062] In this step, initial hyperparameters are first generated through transfer learning, then the hyperparameter space is filtered based on peak-valley error, and finally lightweight optimization is completed through fine search using the EI function.

[0063] The similarity between the target user and the profiles in the database is calculated using both load feature and topological dimensions. Load feature similarity is calculated using cosine similarity, with the formula being: [feature vector, numerator = vector dot product, denominator = vector magnitude product]. Topological similarity is calculated using Mahalanobis distance correction, with the formula: [formula omitted]. d represents the branch topology distance between the target user and the user corresponding to the profile in the database (unit: meters, derived from the spatiotemporal topology data collected by S1). Meters (the standard deviation of distance determined by actual measurements based on the load isolation characteristics of the distribution network branches). The formula for calculating the two-dimensional hybrid similarity is: 0.6 and 0.4 are weighting coefficients verified by load forecasting accuracy (stored in the 3D dynamic parameter library). During calculation, priority is given to matching profile data within the same transformer area—due to the correlation of user load within the same transformer area. Far higher than cross-regional Finally, the matching results of the Top-2 in the same transformer area and the Top-1 in the adjacent transformer area are selected to form a candidate set.

[0064] The initial hyperparameters are generated by weighting similarity and historical accuracy, as shown in the formula: ,in For the initialized hyperparameters (n_h or ), Let be the mixed similarity of the i-th candidate image. The historical prediction accuracy of the hyperparameters corresponding to this candidate profile (industrial user scenario) (Data sourced from mapping library history) Hyperparameter values ​​for candidate profiles. These hyperparameter values ​​are tailored to the complexity of the user load; for example, industrial users require high load stability. The initial value is 128. Residential users have highly random load characteristics. The initial value is 64, and the learning rate n is uniformly initialized to the interval [0.001, 0.01].

[0065] Lightweight optimization includes a coarse-search phase that filters the hyperparameter space based on distribution network indicators such as peak-valley error. The core indicator for peak-valley error is the "peak segment prediction error rate". , This represents the number of data samples during the peak period (8:00-22:00). The measured load at time t is derived from high-quality data from S1. These are the model predictions, discarded during the coarse search. The hyperparameter region will Limited to [64, 128], Limiting the search to [0.001, 0.01] reduces the range of invalid searches.

[0066] The fine search employs an iterative approach using the Expected Improvement (EI) function, which balances exploration and exploitation. The formula is as follows: ,in For hyperparameters to be optimized, (The objective function is the weighted sum of peak-to-valley error rate and mean absolute error.) For hyperparameters The corresponding objective function prediction value, To predict the standard deviation, ( U is the load fluctuation coefficient calculated in step S2. (used to balance exploration intensity) and These are the cumulative distribution function and probability density function of the standard normal distribution, respectively. The number of iterations for the refined search is adapted to the computing power of the edge nodes—based on a commonly used NVIDIA Jetson Xavier NX edge device (15 TOPS computing power), the time consumed in a single round of hyperparameter evaluation is measured. The total number of iterations is set to 15, ensuring that the total optimization time is controlled within 4.5 hours.

[0067] During training, GPU memory usage is monitored in real time. GPU memory utilization is obtained through the edge node hardware monitoring interface. When the utilization rate... Model quantization is triggered on time, using INT8 quantization to convert model parameters from 32-bit floating-point to 8-bit integer. The quantization formula is as follows: These are the original floating-point parameters. Here is the range of parameter values, and q is the quantized integer. Quantization ensures optimal model memory usage. And the prediction accuracy loss is controlled within Within this range, it adapts to the resource constraints of edge nodes.

[0068] Prioritize matching similar profiles within the same distribution area and filter the results. Combine similarity with historical accuracy to generate initial hyperparameters. Lightweight optimization includes screening the hyperparameter space based on distribution network indicators such as peak-valley error during the coarse search stage and eliminating areas with substandard accuracy. Fine search adopts a function iteration that balances exploration and utilization, with the number of rounds adapted to the computing power of edge nodes. During training, GPU memory is monitored in real time, and model quantization is triggered when the memory usage is too high.

[0069] Model training requires output from a Bayesian LSTM. Confidence intervals quantify prediction uncertainty; the formula for correcting the uncertainty coefficient is as follows: Where U' is the corrected uncertainty and U is the original uncertainty (calculated using formula) ), This is the error trend coefficient (1.3 when the error is rising, 1.0 when it is stable, and 0.8 when it is falling). Time-triggered model orientation fine-tuning can reduce prediction error by 15%.

[0070] S4: Dynamically allocate computing power based on resource scarcity, user priority, and scenario urgency, and synchronize the scheduling results to the distribution automation system of the transformer area.

[0071] The specific methods in this step include: three-dimensional linkage resource scheduling, which involves dynamically adjusting resource stress according to load periods, calculating memory and storage resource weights. Load periods are divided into peak periods (8:00-22:00), flat periods (6:00-8:00, 22:00-24:00), and valley periods (0:00-6:00) according to the distribution network dispatching regulations. The resource weights for different periods are determined through a three-dimensional dynamic parameter library, and the resource weights for peak periods are calculated. Memory weight Storage weight flat section Valley section The weight values ​​are based on the resource usage statistics of 10 pilot transformer areas over one month to ensure that they match the load calculation requirements for each time period.

[0072] Resource scarcity is quantified using a comprehensive weighted index, calculated using the following formula: The CPU is the edge node computing resource utilization rate, which is the ratio of the number of CPU cores currently used by the running process to the total number of cores; MEM is the memory resource utilization rate, calculated as the ratio of used memory capacity to total memory capacity; STO is the storage resource utilization rate, calculated as the ratio of used storage capacity to total storage capacity. All three data are collected in real time through the edge node hardware monitoring interface (sampling period of 10 seconds). The range of values ​​is The closer it is to 1, the more scarce the resources are.

[0073] User priority is calculated using a weighted scoring method, which combines spatiotemporal correlation, load elasticity, electricity consumption patterns, and the emergency coefficient of scheduling procedures in the scenario emergency coefficient library. The formula is as follows: Where P is the user priority score (ranging from 0 to 10, with higher scores indicating higher priority); K is the scenario urgency coefficient, derived from a scenario urgency coefficient database, and is associated with the user in the fault repair process. Areas with concentrated charging stations Industrial users Ordinary residential users ; for The load elasticity index calculated step by step (normalized to) (interval); for Electricity consumption regularity index calculated step by step (range of values) ); for The spatiotemporal correlation coefficients calculated in the steps (normalized to) ); weighting coefficient , The parameters are determined by the priority verification of the distribution network dispatch and stored in the three-dimensional dynamic parameter library.

[0074] In another example, resource stress is quantified using the K-value, calculated as follows: Where K is the resource stress level (value...) ), They are respectively Real-time memory and bandwidth utilization. For the corresponding total amount of resources, The dynamic weights are positively correlated with the 1-hour load variance and satisfy the following conditions: All resource data is collected through the Prometheus monitoring agent (sampling period 500 ms).

[0075] User priority is quantified using a Priority value, calculated using the formula: Priority Where Priority is the priority level (values) ), The spatiotemporal correlation coefficient, For load elasticity, To ensure regular electricity consumption, The scenario urgency coefficient (derived from the scenario urgency coefficient library) is used, and the weights are determined based on the fitting of distribution network dispatching requirements.

[0076] Computing power allocation is divided into three scenarios based on resource stress T: when If the system determines that computing power is sufficient, it allocates computing power according to the user's priority score ratio, i.e., the ratio of computing power allocation to a particular user. ,in This is assigned a priority score to the user, where N is the total number of users in the area, ensuring that high-priority users receive more computing resources to improve prediction accuracy; when The time was determined to be a time of computing power shortage, for Low-priority users will have their computing power limited, ensuring that their computing power usage does not exceed the total computing power. The released resources will be prioritized for replenishment. High-priority users; when If the system is detected as overloaded, it will be immediately suspended. For each user's non-essential computing tasks, all the freed computing power will be used to support the fault handling needs of the distribution automation system. At this time, the resource usage priority of fault handling-related computing tasks is the highest, ensuring the real-time response of the distribution network fault.

[0077] During scheduling, relevant coefficients are updated in real time based on a 3D dynamic parameter library. The scenario urgency coefficient K is updated every 5 minutes, synchronizing the scheduling procedure results. Resource weights are also considered. The update will be completed 1 minute before the load shift, with user priority. The system recalculates every 15 minutes based on the latest load data to ensure that the scheduling strategy matches the real-time operating status. The scheduling results are encapsulated in JSON format, containing core information such as user ID, computing power allocation percentage, and scheduling effective time. This information is synchronized to the distribution automation system of the transformer substation via the IEC 61850 standard communication protocol, with synchronization latency controlled within 100 ms and a synchronization success rate of [missing information]. This ensures that dispatching instructions and the operation instructions of the power distribution automation system are coordinated and consistent.

[0078] The three-dimensional linkage resource scheduling includes dynamically adjusting the weights of computing, memory, and storage resources according to the resource tension during the load period, combining user priority with spatiotemporal correlation, load elasticity, electricity consumption patterns, and emergency coefficients in the scenario emergency coefficient library. When computing power is sufficient, high-priority users are prioritized, low-priority computing power is restricted when there is a shortage, and the fault handling needs of the power distribution automation system are prioritized when there is an overload. During the scheduling process, relevant coefficients are updated in real time based on the three-dimensional dynamic parameter library.

[0079] S5: The prediction results of the dynamic fusion model and association rules are output after residual correction to provide load data for scheduling adaptation, and the updated profile parameters are fed back.

[0080] The specific methods in this step include: the core of dynamic fusion is... The prediction results of the optimized long short-term memory network model are "and " based on The association rule prediction results extracted from four-dimensional features are weighted and integrated. These association rules are constructed by mining the temporal correlations (e.g., stable load for industrial users from 9-5 PM on weekdays) and spatiotemporal correlations (e.g., synchronous load fluctuations among neighboring users) of users in the same business type. Each rule includes a trigger condition (e.g., E≥0.7) and a load prediction range. The output of the Long Short-Term Memory network model is denoted as... The output of the association rule is denoted as t is the prediction time (at 15-minute intervals). Up to 96, at 1-hour intervals (Up to 24), both are generated based on the high-quality data processed by S1 and the profile features of S2.

[0081] Dynamic fusion includes the calculation of model weights by combining temporal attention, credibility, historical error, and business format coefficients, with temporal attention weights being a key component. Adjusted according to the degree of load fluctuation, the calculation formula is as follows: ,in for The load fluctuation coefficient at time (i.e., calculated in step S2) (P)), the denominator is the maximum value of the fluctuation coefficients of the first 24 time periods, to ensure that the greater the load fluctuation, the higher the attention weight the model receives due to its stronger ability to capture time-series features. The range of values ​​is normalized to [ ] ] .

[0082] Credibility and Determined based on historical performance in different scenarios, the calculation formula is: N is the number of historical scene samples (data of the same type of scene taken in the last 30 days). The single-sample prediction error rate represents the reliability of the Long Short-Term Memory network during the peak period (8:00–22:00). Reliability of association rules in the valley segment (0:00-6:00) The data comes from historical accuracy records in the 3D dynamic parameter library.

[0083] Historical error weights The inverse error weighting method is used, and the formula is as follows: , The mean absolute error of the model over the last M samples ( ), The mean absolute error of the association rule is denoted by , and both are calculated using historical prediction data.

[0084] Business format coefficient Sourced from the equipment contribution coefficient database, charging pile concentration areas Industrial users , ordinary residents The advantages of different forecasting methods for adapting to different business formats vary.

[0085] The formula for calculating the overall weight is as follows: , The final fusion result is .

[0086] In another example, the formula for calculating the overall weight is: Where w is the LSTM prediction weight and A is the temporal attention. , The reliability of the two types of prediction results are respectively. , These represent the historical errors of the two types of prediction results. (Characteristic contribution constant) To represent the corrected uncertainty, S is the spatiotemporal correlation degree. W represents the weight of the business scenario (0.8 for residential scenarios), and W represents the weight of the device. , It is calculated using historical prediction data based on association rules.

[0087] Attention is adjusted according to the degree of load fluctuation, and the reliability is determined based on the historical performance of different scenarios. The fusion result is converted into daily load data with a 15-minute interval and short-term load data with a 1-hour interval. The format conforms to the communication standard of distribution automation system, specifically adopting the JSON structure specified by the IEC 61850-90-7 protocol, which includes fields such as "user ID - prediction time - load value (kW) - data reliability". The load value is retained to two decimal places to meet the scheduling accuracy requirements.

[0088] Residual correction includes splitting historical forecast residuals according to the peak and valley periods of the distribution network. The peak period is defined as 8:00-22:00, the valley period as 0:00-6:00, and the flat periods as 6:00-8:00 and 22:00-24:00. ( (This refers to the measured load value). Non-periodic fault residuals are filtered by fault labeling; that is, when the three-dimensional dynamic parameter library marks that moment as a "fault state," the residual... The residuals under normal operating conditions are not included in the training set.

[0089] Residual models were constructed for peak and valley periods respectively. Due to frequent load fluctuations during the peak period, a gated cyclic unit (GRU) model was adopted, with the input being the residual sequence of the previous three time periods. The output is the predicted residual value at time t. The valley section, due to its stable load, employs an autoregressive integral moving average (ARIMA) model. The model formula is ,in These are the autoregressive coefficients. The moving average coefficient is... for The time-series model residuals, whose coefficients are obtained through training with nearly 1,000 valley residual samples, are stored in a three-dimensional dynamic parameter library.

[0090] In another example, residual prediction uses the ARIMA(1,1,1) model, and the residual calculation formula is: ,in ε represents the residual from the previous time step, and ε represents white noise (within the range of [-0.01, 0.01]).

[0091] Peak load periods focus on short-term load fluctuation residuals, while valley load periods focus on stable load residuals. The fusion results are corrected using the predicted residual values. The correction formula is as follows: ,in Take in the peak segment Valley section For the flat sections, a weighted average of the peak-valley model predictions is used (each weight is 0.5).

[0092] The output load data is synchronously fed back to step S2 to update the profile parameters. ( When the user's rated load (derived from the equipment ledger) is reached, the profile parameters are updated, specifically updating the weekly fluctuation component of the electricity consumption regularity R. Temperature response weighting with load elasticity E Temperature, the updated parameters are saved back to the 3D dynamic parameter library, and the update cycle is consistent with the portrait evolution cycle (micro-evolution 1 hour / time, full evolution 15 minutes / time).

[0093] Dynamic fusion includes model weights combined with time-series attention, credibility, historical error, and business type coefficients. Attention is adjusted according to the degree of load fluctuation, and credibility is determined based on historical performance in different scenarios. The fusion result is converted into daily load data at 15-minute intervals and short-term load data at 1-hour intervals, with the format conforming to the communication standards of distribution automation systems. Residual correction includes splitting historical prediction residuals according to the peak and valley periods of the distribution network, filtering non-periodic fault residuals through fault labels, constructing residual models for peak and valley periods respectively, focusing on short-term load fluctuation residuals for peak periods and focusing on stable load residuals for valley periods, correcting the fusion result with residual prediction values, and synchronously feeding the output load data back to step S2 to update the profile parameters. The updated parameters are then stored back into the three-dimensional dynamic parameter library.

[0094] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to the embodiments of this application, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0095] This method utilizes a progressive technical chain of "data perception → spatiotemporal profiling → lightweight model → resource scheduling → intelligent fusion" to achieve end-to-end optimization from data input to predictive output through the coordinated adaptation of technical means at each stage. The derivation logic from its technical means to its technical effects is as follows: Multimodal data quality enhancement lays a reliable foundation for the entire process: Comprehensive load-related data is ensured through multi-terminal collection covering electrical, non-electrical, and spatiotemporal quantities; a series anomaly detection mechanism employing the "3σ criterion for initial screening + isolated forest for precise judgment" accurately identifies different types of data anomalies to avoid interference from invalid information; for differentiated missing scenarios such as communication jitter and link interruptions, scenario-based completion methods such as time decay weighting and neighboring user weighting are used to ensure data integrity based on data temporal and spatial correlation; combined with spatiotemporal entropy feature enhancement based on user attribute weights, the spatiotemporal concentration of load is quantified to highlight business type differences, ultimately outputting high-quality data with comprehensiveness, integrity, and feature discrimination, addressing the core pain point of "single data dimension and poor quality" in traditional forecasting.

[0096] Four-dimensional dynamic profiling enables precise characterization of load characteristics: Based on enhanced features output from the data perception layer, standardized mathematical formulas are used to quantify electricity consumption regularity (R), load elasticity (E), equipment weight (W), and spatiotemporal correlation coefficient (S), forming multi-dimensional load feature labels. Coupled with a dynamic evolution mechanism triggered by load variance, the profile update range is adaptively adjusted according to the degree of load fluctuation, ensuring that the profile is synchronized with real-time load changes. This quantitative characterization method overcomes the ambiguity of traditional load descriptions, accurately distinguishing differences in business types such as the high elasticity of charging piles and the high regularity of industrial users, providing concrete and quantifiable feature basis for subsequent model parameter tuning.

[0097] Adaptive hyperparameter tuning improves prediction adaptability and efficiency: Using four-dimensional features of spatiotemporal profiles as indexes, a hybrid similarity-weighted matching historical hyperparameter mapping library is employed to reuse high-quality hyperparameters from similar scenarios, shortening the initialization cycle. A Bayesian optimization strategy of "gradient-weighted coarse search + EI function fine search" is used to efficiently select optimal hyperparameters under edge node computing power constraints, while memory monitoring and model quantization mechanisms control resource consumption. Uncertainty quantization and directional fine-tuning mechanisms are introduced to dynamically correct model bias based on prediction confidence intervals. This deep integration of profile features and hyperparameter tuning strategies enables the model to quickly adapt to the load fluctuation characteristics of mixed business formats, avoiding redundant hyperparameter tuning "from scratch," while ensuring efficient model operation at edge nodes.

[0098] Three-dimensional linked resource scheduling ensures real-time prediction and priority: Resource stress is quantified through dynamic weighting, and the status of resources such as CPU and memory at edge nodes is perceived in real time. Combined with spatiotemporal correlations and load elasticity features output from the profiling layer, and overlaid with a scenario urgency coefficient to calculate user priority, a linked scheduling logic of "resource status - user demand - scenario urgency" is formed. This scheduling method can prioritize the allocation of computing power to critical objects such as high-temperature charging pile scenarios and users associated with faults, avoiding unnecessary computing power consumption. While ensuring the real-time prediction of critical scenarios, it improves resource utilization and provides stable resource support for model operation and result output.

[0099] Intelligent fusion and residual correction output reliable prediction results: A multi-factor dynamic weighting strategy is adopted to fuse the prediction results of the LSTM model and association rules. The weight calculation incorporates multi-dimensional indicators such as time series attention, credibility, and uncertainty, so that the fusion result can balance the advantages of different prediction methods—emphasizing the time series capture capability of LSTM when the load fluctuates greatly, and relying on the stability of association rules when the load is stable. The historical residuals are predicted and corrected through the ARIMA model to eliminate systematic errors. This fusion correction mechanism further improves the accuracy and reliability of the prediction results. The output data format conforms to the distribution automation system standard and can directly support real-time dispatching decisions of the distribution network.

[0100] In summary, the technical means at each stage are progressive and mutually supportive: data perception provides high-quality input for profile construction, profiles provide accurate basis for model parameter tuning, resource scheduling provides stable guarantee for model operation, and fusion correction provides accuracy support for the final output, forming a complete logical closed loop of "high-quality data - accurate features - model adaptation - efficient resources - reliable results", ultimately achieving accurate prediction of load in mixed-use distribution areas and providing core technical support for lean scheduling of distribution networks.

[0101] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the foregoing disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A method for intelligent prediction and management of load change trends in low-voltage distribution networks, applied to edge computing nodes of distribution transformer areas, characterized in that, include: Collect multi-source data of power distribution networks from mixed business formats including "charging piles, industry and residential", and output high-quality data through anomaly detection, scenario-based completion and feature enhancement processing; Based on this data, a four-dimensional load characteristic index is calculated to construct a dynamic profile of user load and adaptively evolve according to the degree of load fluctuation. By combining portrait feature labels, transfer learning is used to initialize the hyperparameters of the long short-term memory network model, and lightweight optimization is performed to adapt the model to the load characteristics of mixed business formats. Computing power is dynamically allocated based on resource scarcity, user priority, and scenario urgency, and the scheduling results are synchronized to the distribution automation system of the transformer area. The prediction results of the dynamic fusion model and association rules are corrected by residuals and output to provide load data for scheduling adaptation, and the updated profile parameters are fed back.

2. The intelligent prediction and management method for load change trends in low-voltage distribution networks according to claim 1, characterized in that, Anomaly detection and completion include: First, mark suspected anomalies according to the daily periodicity of the distribution network load, combine the branch topology of the transformer area to eliminate non-global anomalies caused by local faults, and then determine the valid anomalies by using a model trained with typical anomaly data of the distribution network. Single-point missing data caused by communication jitter is filled with time-attenuation weighted data; short-to-medium-term missing data caused by equipment failure is filled with load replacement from the same type of equipment; and long-term missing data caused by link interruption is filled with weighted data from neighboring users within the same branch.

3. The intelligent prediction and management method for load change trends in low-voltage distribution networks according to claim 1, characterized in that, The calculation and evolution of four-dimensional load characteristic indicators include: Electricity consumption patterns are weighted by daily and weekly load fluctuations, load elasticity is adjusted seasonally by electricity price and temperature response weights, equipment weights incorporate business-specific contribution coefficients, and spatiotemporal correlations are combined with load similarity and topological distance attenuation characteristics. The profile evolves according to load fluctuation levels. In high-fluctuation scenarios, load elasticity indicators are updated first, and different levels correspond to different feature indicator update ranges.

4. The intelligent prediction and management method for load change trends in low-voltage distribution networks according to claim 1, characterized in that, Transfer learning initialization and lightweight optimization include: Construct a user profile-model hyperparameter mapping library for distribution networks with multiple business types and load levels, and associate each data point with a four-dimensional profile and the optimal hyperparameter; Calculate the load characteristics and topological similarity between the target user and the profiles in the database. Prioritize matching similar profiles within the same area and filter the results. Combine similarity with historical accuracy to generate initial hyperparameters.

5. The intelligent prediction and management method for load change trends in low-voltage distribution networks according to claim 1, characterized in that, Three-dimensional linkage resource scheduling includes: Resource stress is dynamically adjusted according to load periods, with weights for computing, memory, and storage resources, and user priority is combined with spatiotemporal correlation, load elasticity, electricity consumption patterns, and emergency coefficients of dispatching procedures. When computing power is sufficient, priority is given to high-priority users; when resources are strained, low-priority computing power is restricted; and when overloaded, priority is given to supporting the fault handling needs of the power distribution automation system.

6. The intelligent prediction and management method for load change trends in low-voltage distribution networks according to claim 1, characterized in that, Dynamic fusion includes: The model weights are calculated by combining temporal attention, credibility, historical error, and business type coefficient. Attention is adjusted according to the degree of load fluctuation, and credibility is determined based on historical performance in different scenarios. The fusion results are converted into daily load data at 15-minute intervals and short-term load data at 1-hour intervals, with the format conforming to the communication standards of distribution automation systems.

7. The intelligent prediction and management method for load change trends in low-voltage distribution networks according to claim 6, characterized in that, Residual corrections include: Historical prediction residuals are split according to peak and valley periods of the distribution network, and non-periodic fault residuals are filtered by fault labels; Residual models are constructed for peak and valley segments respectively. The peak segment focuses on the residual of short-term load fluctuations, while the valley segment focuses on the residual of stable loads. The fusion results are corrected using the residual prediction values.

8. The intelligent prediction and management method for load change trends in low-voltage distribution networks according to claim 4, characterized in that, Lightweight optimizations include: In the coarse search stage, the overparameter space is screened according to distribution network indicators such as peak-valley error, and areas with substandard accuracy are eliminated. The fine search employs a function iteration that balances exploration and utilization, with the number of rounds adapted to the computing power of edge nodes. During training, GPU memory is monitored in real time, and model quantization is triggered when memory usage is too high.

9. The intelligent prediction and management method for load change trends in low-voltage distribution networks according to claim 1, characterized in that, The method implementation relies on a three-dimensional dynamic parameter library, which includes: The system is based on a database of equipment contribution coefficients determined by actual equipment load, a database of scenario emergency coefficients based on scheduling procedures, and a database of user attribute weights based on the proportion of business load. An iterative mechanism combining quarterly full updates and incremental updates triggered by changes in equipment in the distribution area is adopted.

10. The intelligent prediction and management method for load change trends in low-voltage distribution networks according to claim 1, characterized in that, The entire process is deployed at outdoor edge nodes of the power distribution network, and the adaptation strategies include: The hardware features a wide temperature range design and IP65 protection rating, and the power supply supports a wide voltage input. The software optimizes resources through process isolation and memory pools. When the load in the distribution area is below 20% of the rated load for 1 hour, a low-power mode is triggered. Communication adopts the standard protocol of the power industry.

Images