Block area load forecasting method and system adaptive to microcontroller

Abstract

This application discloses a method and system for predicting transformer area loads adapted to microcontrollers, relating to the field of distribution network edge load prediction technology. The solution is as follows: real-time acquisition and preprocessing of transformer area power load data and environmental data to obtain a time-series current load data sequence; extraction of features from three dimensions—time, period, and external environment—to construct a multi-dimensional load feature vector; processing of the feature vector using a preset state machine model to determine the target standard operating condition corresponding to the current operating condition of the transformer area. This target standard operating condition is one of several progressively layered operating conditions divided according to the stability of load operation; selection of a target prediction model adapted to the target standard operating condition from a preset hierarchical progressive prediction model set. This model set contains multiple lightweight time-series prediction models that correspond one-to-one with the standard operating conditions at each level and have increasing computational complexity; and calling the target prediction model to perform load prediction on the multi-dimensional load feature vector to obtain the future load prediction result for the transformer area.

Description

Technical Field

[0001] This application relates to the field of distribution network edge load forecasting technology, and in particular to a method and system for forecasting distribution area loads adapted to a microcontroller. Background Technology

[0002] With the rapid development of new power systems, distribution substations are gradually evolving towards intelligence and edge computing. The traditional centralized load forecasting model, relying on main stations or the cloud, is no longer adequate for on-site engineering needs. Distribution substations have a large number of terminal devices distributed over a wide area. If all forecasting calculations are performed in the cloud, it will continuously increase the communication bandwidth load, and long-distance data transmission will introduce latency issues, significantly reducing the real-time performance of forecasts. The industry urgently needs to implement load forecasting capabilities on local equipment in distribution substations, especially on microcontroller-based edge terminals, to improve the response speed and overall operational reliability of the distribution network system.

[0003] Current mainstream load forecasting solutions for distribution transformer areas fall into two categories: traditional time-series and regression analysis algorithms, which have low computational overhead but insufficient prediction accuracy under complex operating conditions; and deep learning prediction algorithms, represented by recurrent neural networks, which offer higher prediction accuracy but require significant computing and storage resources, making them difficult to deploy directly on microcontroller terminals. Therefore, there is an urgent need for a solution that balances lightweight algorithms with high load forecasting accuracy, enabling high-precision, real-time adaptive load forecasting on lightweight microcontroller terminals located locally within distribution transformer areas. Summary of the Invention

[0004] This application provides a method and system for load forecasting of distribution transformers adapted to microcontrollers. The main purpose is to complete high-precision calculations on local microcontroller terminals in distribution transformer areas with limited hardware resources, forming a complete solution that balances computing power and forecasting performance. It can achieve high-precision, real-time adaptive load forecasting on lightweight local microcontroller terminals in distribution transformer areas.

[0005] To achieve the above objectives, this application mainly provides the following technical solutions: The first aspect of this application provides a method for predicting transformer loads in a distribution area that is adapted to a microcontroller, the method comprising: Preprocessing is performed on the real-time collected power load data and environmental data of the transformer area to obtain a time-series current load data sequence. Each sequence point in the current load data sequence stores the power load data and environmental data collected at the same time. Features corresponding to the time dimension, period dimension, and external environment dimension are extracted from the current load data sequence to construct a multi-dimensional load feature vector that matches the current load data sequence; The multidimensional load feature vector is processed using a preset state machine model to determine the target standard operating condition corresponding to the current operating condition of the transformer area. The target standard operating condition is a standard operating condition in one of multiple progressive levels divided according to the stability of load operation. The preset state machine model is used to map the multidimensional load feature vector to the standard operating condition. Select a target prediction model that is adapted to the target standard working condition from a preset hierarchical progressive prediction model set. The hierarchical progressive prediction model set includes multiple lightweight time series prediction models that correspond one-to-one with the multiple progressive hierarchical standard working conditions and whose computational complexity increases. The lightweight time series prediction model is a time series prediction model whose computational and storage requirements are adapted to the operation of the microcontroller. The target prediction model is invoked to perform load prediction on the multidimensional load feature vector to obtain the prediction results corresponding to the future load of the transformer area.

[0006] A second aspect of this application provides a transformer area load forecasting system adapted to a microcontroller, the system comprising: The data acquisition and preprocessing module is used to preprocess the real-time acquired power load data and environmental data of the transformer area to obtain a time-series current load data sequence. Each sequence point in the current load data sequence stores the power load data and environmental data acquired at the same time. The data acquisition and preprocessing module is also used to extract features corresponding to the time dimension, period dimension, and external environment dimension from the current load data sequence, and construct a multi-dimensional load feature vector that matches the current load data sequence. The state machine condition identification module is used to process the multidimensional load feature vector using a preset state machine model, and determine the target standard condition corresponding to the current operating condition of the transformer area. The target standard condition is a standard condition in one of multiple progressive levels divided according to the stability of load operation. The preset state machine model is used to map the multidimensional load feature vector to the standard condition. The hierarchical progressive prediction module is used to select a target prediction model that is adapted to the target standard working condition from a preset hierarchical progressive prediction model set. The hierarchical progressive prediction model set includes multiple lightweight time series prediction models that correspond one-to-one with the multiple progressive level standard working conditions and whose computational complexity increases. The lightweight time series prediction model is a time series prediction model whose computational and storage requirements are adapted to the operation of the microcontroller. The prediction result output module is used to call the target prediction model to perform load prediction on the multidimensional load feature vector and obtain the prediction result corresponding to the future load of the transformer area.

[0007] A third aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the microcontroller-adapted substation load forecasting method described above.

[0008] A fourth aspect of this application provides an electronic device, the device including at least one processor, and at least one memory and bus connected to the processor; The processor and the memory communicate with each other via the bus. The processor is used to call program instructions in the memory to execute the transformer load forecasting method adapted to the microcontroller described above.

[0009] By employing the above-described technical solution, the technical solution provided in this application has at least the following advantages: This application provides a method and system for predicting transformer loads in a microcontroller-compatible area. First, it collects real-time data on transformer load and its correlation with the external environment, and performs standardized preprocessing to obtain a time-aligned current load data sequence. Then, it extracts load correlation features from three dimensions: time, period, and external environment, constructing a standardized multidimensional load feature vector that can be directly used for model calculation. Next, it uses a preset state machine model to process the multidimensional load feature vector, determining the target standard operating condition corresponding to the current operating condition of the transformer area. The target standard operating condition is one of several progressively different operating conditions divided according to the stability of load operation. Then, it selects a target prediction model adapted to the target standard operating condition from a preset hierarchical progressive prediction model set. This set contains multiple lightweight time-series prediction models that correspond one-to-one with the multiple progressively different standard operating conditions and have increasing computational complexity. Finally, it calls the target prediction model to perform load prediction on the multidimensional load feature vector, obtaining the prediction result corresponding to the future load of the transformer area.

[0010] Compared to existing technologies that struggle to adapt to the deployment requirements of low-computing-power microcontroller terminals in local distribution areas, this application adopts an overall architecture combining hierarchical operating condition identification with differentiated lightweight model matching, achieving a balance between lightweight computing power and high load forecasting accuracy. The solution accurately identifies load operating states based on a pre-set state machine model and calls lightweight time-series forecasting models of corresponding complexity as needed. This enables high-precision calculations to be performed on local microcontroller terminals with limited hardware resources, forming a complete solution that balances computing power overhead and forecasting performance. It allows for high-precision, real-time adaptive load forecasting on lightweight microcontroller terminals in local distribution areas.

[0011] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0012] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart of a transformer area load prediction method adapted to a microcontroller is provided in an embodiment of this application; Figure 2 A hierarchical progressive prediction structure diagram provided in the embodiments of this application; Figure 3 The state machine transition diagram provided for the embodiments of this application; Figure 4 A diagram illustrating the computing power-aware dynamic scheduling mechanism provided in this application embodiment; Figure 5 A complete method flowchart illustrating parallel operation of dual closed loops is provided for embodiments of this application; Figure 6 This is a flowchart illustrating the overall method of an embodiment of this application; Figure 7 The overall architecture block diagram of the transformer load prediction system adapted to the microcontroller provided in the embodiments of this application is shown. Detailed Implementation

[0013] Exemplary embodiments of the present application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application will be thorough and complete, and will fully convey the scope of the present application to those skilled in the art.

[0014] This application provides a method for predicting transformer loads in a distribution area that is adapted to a microcontroller, such as... Figure 1 As shown, the following specific steps are provided in this embodiment of the application: Step 100: Perform preprocessing on the real-time collected power load data and environmental data of the transformer area to obtain a time-series current load data sequence. Each sequence point in the current load data sequence stores the power load data and environmental data collected at the same time.

[0015] Among them, the power load data of the transformer area includes electrical quantity data such as current, voltage, and load rate that directly reflect the power consumption status of the transformer area; the external environment data includes environmental and social factors that affect users' power consumption behavior, such as temperature, humidity, season, holidays, and extreme weather events.

[0016] It should be noted that this step involves simultaneously collecting and storing two types of data. On the one hand, it addresses the industry pain point of dynamic fluctuations and uncertainties in transformer load caused by multiple factors such as user behavior, climate change, and holidays. For scenarios involving sudden load changes and drastic fluctuations caused by high-temperature peak electricity consumption, extreme weather, and special events, it completely retains all key factors affecting load changes, providing comprehensive input support for subsequent operating condition identification and load forecasting calculations. This allows the overall forecasting method to dynamically adjust in real time according to load changes, exhibiting excellent multi-operating condition adaptability. On the other hand, after data collection, the raw data undergoes unified preprocessing, transforming the discrete and chaotic collected information into a current load data sequence with a temporal order. Within the sequence, electricity load data and environmental data collected simultaneously at the same time are bound at sequence points, achieving one-to-one matching of load information with corresponding environmental impact information for the time period. This generates a time-aligned and standardized data carrier, providing well-organized and time-matched basic input data for the subsequent 200+ dimension feature extraction steps.

[0017] Step 200: Extract the features corresponding to the time dimension, period dimension, and external environment dimension from the current load data sequence, and construct a multi-dimensional load feature vector that matches the current load data sequence.

[0018] This step is based on the current load data sequence in time series form output from step 100 preprocessing, and completes feature extraction in three independent dimensions: time dimension features depict the trend of load change over time; period dimension features capture the periodic fluctuation pattern of electricity consumption in the transformer area; and external environment dimension features reflect the impact of external conditions such as temperature and holidays on users' electricity consumption behavior.

[0019] The features extracted from the three dimensions are fused and integrated to generate a multi-dimensional load feature vector that matches the current load data sequence one-to-one. This vector is a structured and vectorized representation of three key types of information: time series, periodicity, and environment. It integrates the various types of information scattered in the time series data into a standardized input that can be directly read by the subsequent load condition identification model, using a fixed-dimensional numerical vector form. For example, if the current load data sequence contains load and environmental data for N consecutive time points, this step will extract three types of features from each sequence point: timestamp, daily / weekly period, and environmental impact, and concatenate them to form a feature vector in the form of […]. f time , f cycle , f envA multidimensional vector, where f time For time dimension feature values, f cycle For periodic dimension feature values, f env The environmental dimension feature value is the multidimensional load feature vector matched with the corresponding sequence points.

[0020] Load changes in a distribution area are driven by multiple factors, including time sequence, electricity consumption cycle, and external environment. A single-dimensional feature cannot fully depict the load change pattern. This step simultaneously extracts three types of features and integrates them into a unified multi-dimensional load feature vector, fully covering all key factors affecting load fluctuations and avoiding operational condition identification errors caused by missing feature dimensions. At the same time, this multi-dimensional load feature vector serves as the standard input to the preset state machine model in step 300, providing comprehensive and sufficient data support for the three-level progressive operational condition determination. This ensures the identification accuracy of steady-state operation, fluctuating operation, and strong disturbance operation, laying a complete feature input foundation for the entire load forecasting system adapted to multiple operational conditions.

[0021] Step 300: Use a preset state machine model to process the multidimensional load feature vector and determine the target standard operating condition corresponding to the current operating condition of the transformer area. The target standard operating condition is a standard operating condition in one of the multiple progressive levels divided according to the stability of load operation. The preset state machine model is used to map the multidimensional load feature vector to the standard operating condition.

[0022] In this step, the multidimensional load feature vector obtained in the previous steps is input into the preset state machine model for processing, so as to determine the target standard operating condition corresponding to the current operating condition of the transformer area.

[0023] The target standard operating condition is one of several progressive levels divided according to the stability of load operation: that is, based on the stability of the load operation in the transformer area, several progressive levels are divided from stable to disturbed, and each progressive level corresponds to a type of standard operating condition; the target standard operating condition obtained by the actual operation of the transformer area is one of the types of standard operating conditions matched among these progressive levels.

[0024] As an exemplary embodiment, the standard operating conditions corresponding to multiple progressive levels can be divided into three mutually exclusive operating conditions. The definitions and characteristics of each type of operating condition are as follows: (1) Steady-state operating condition: The load of the transformer area changes slowly, the fluctuation amplitude is at a low level for a long time, the load change trend is stable, there is no obvious sudden change or violent fluctuation, and the electricity consumption behavior is regular and highly predictable; (2) Fluctuating operating condition: The load of the transformer area produces periodic or non-periodic moderate fluctuations, the fluctuation amplitude does not exceed the normal range, and the electricity consumption behavior is affected by factors such as time period and season, showing regular changes; (3) Strong disturbance operating condition: The load of the transformer area undergoes large sudden changes and irregular violent fluctuations, the load change rate and fluctuation intensity significantly exceed the normal range, mostly induced by extreme weather, sudden events, equipment failures, etc., and the uncertainty of electricity consumption behavior increases significantly.

[0025] Among them, the preset state machine model is used to map multidimensional load feature vectors to standard operating conditions: mapping means taking multidimensional load feature vectors as input, and according to the judgment logic inside the preset state machine model, completing the assignment matching of the input feature vectors to a certain type of standard operating condition, and outputting the matched target standard operating condition.

[0026] Specifically, the preset state machine model contains multiple progressive levels corresponding one-to-one with the aforementioned standard operating conditions. After receiving multi-dimensional load feature vectors, it verifies the load fluctuation level level by level according to the progressive order of load operation stability, completes the operating condition boundary matching layer by layer, and finally outputs the target standard operating condition corresponding to the current load of the transformer area. The core function of this preset state machine model is to complete the refined identification and classification of the operating condition of the transformer area load based on multi-dimensional load feature vectors, providing a clear basis for the operating condition attribution for subsequent matching and prediction models.

[0027] Traditional load forecasting schemes rely on a single, fixed model for load condition assessment, making it difficult to adapt to diverse scenarios such as dynamic load changes and drastic disturbances in distribution transformer areas. Prediction accuracy significantly decreases under load abrupt changes and extreme conditions. This step utilizes a pre-set state machine model to finely differentiate between multiple operating conditions, accurately classifying different load states. The pre-set state machine model verifies load fluctuation levels hierarchically, avoiding confusion and identification errors caused by single threshold judgments. It accurately captures dynamic load changes of varying magnitudes, providing a reliable basis for selecting a suitable lightweight time-series forecasting model. The entire forecasting system can differentiate forecasting models for different load fluctuation scenarios while considering the hardware constraints of low-computing-power microcontrollers and the prediction accuracy requirements under multiple operating conditions, achieving real-time response to dynamic load changes in distribution transformer areas.

[0028] Step 400: Select a target prediction model that is adapted to the target standard operating condition from the preset hierarchical progressive prediction model set. The hierarchical progressive prediction model set contains multiple lightweight timing prediction models that correspond one-to-one with multiple progressive level standard operating conditions and whose computational complexity increases. The lightweight timing prediction model is a timing prediction model whose computational and storage requirements are adapted to the operation of the microcontroller.

[0029] Among them, the hierarchical progressive prediction model set is a pre-built set of models, which includes multiple lightweight time series prediction models that correspond one-to-one with multiple progressive level standard operating conditions and have increasing computational complexity; that is, the lower the operating condition level of load operation stability, the higher the computational complexity of the corresponding lightweight time series prediction model, thereby achieving dynamic matching between prediction accuracy and computing resources under different operating conditions.

[0030] Among them, the lightweight timing prediction model refers to the timing prediction model whose computational and storage requirements are adapted to the operation of the microcontroller. Each lightweight timing prediction model has been optimized by lightweight trimming, and has the characteristics of small parameter size and low computational overhead, which can adapt to the operating conditions of the limited computing and storage resources of the microcontroller in the distribution area.

[0031] As an exemplary embodiment, the hierarchical progressive prediction model set may include three lightweight time-series prediction models, each corresponding to one of the three standard operating conditions: steady-state operation, fluctuating operation, and strong disturbance operation. Specifically, the matching relationship is as follows: steady-state operation is matched with a low-complexity Holt-Winters lightweight time-series prediction model; fluctuating operation is matched with a medium-complexity unidirectional GRU lightweight time-series prediction model; and strong disturbance operation is matched with a high-complexity bidirectional GRU + attention mechanism lightweight time-series prediction model. When step 300 determines the target standard operating condition to which the current production area belongs, it retrieves the pre-bound lightweight time-series prediction model for that condition from the hierarchical progressive prediction model set as the target prediction model and performs load prediction.

[0032] For example, combined Figure 2 The diagram showing the hierarchical and progressive prediction structure is expanded for further explanation. Figure 2 This document visually demonstrates the matching relationship and overall computational logic between the working condition identification process and the lightweight time series prediction model in this application.

[0033] Figure 2 The left side shows the three types of operating conditions output by the preset state machine, namely steady state S1, fluctuation S2, and strong disturbance S3, which serve as the basis for selecting the corresponding lightweight time series prediction model from the hierarchical progressive prediction model set in this step. Figure 2 On the right, three independent lightweight time-series prediction model computation branches are set up, with each of the three branches corresponding to one of the three operating conditions: The steady-state S1 condition corresponds to a low-complexity, lightweight model branch, employing the Holt-Winters lightweight time-series forecasting model. This model offers advantages such as low latency and low computational cost, making it suitable for steady-state operation scenarios with stable loads and no significant fluctuations. Figure 2 The "short-term rapid prediction (low latency, low computing power)" marked in the text is a typical application scenario where this model has advantages.

[0034] The fluctuating S2 operating condition corresponds to a medium-complexity model branch, employing a unidirectional GRU lightweight time-series forecasting model. This model balances computational power and prediction accuracy, and is suitable for fluctuating operating scenarios where the load exhibits moderate-amplitude periodic fluctuations. Figure 2 The “medium-term fluctuation prediction (balance accuracy and computing power)” marked in the text is a typical application scenario where this model has advantages.

[0035] The high-precision, high-complexity model branch corresponding to the strong disturbance S3 operating condition adopts a lightweight time series prediction model with bidirectional GRU + attention mechanism. This model has the advantages of high accuracy and high robustness, and is suitable for strong disturbance operating scenarios with sudden load changes and irregular and violent fluctuations. Figure 2 The "long-term / strongly perturbed prediction (high accuracy, high robustness)" marked in the text represents a typical application scenario where this model has advantages.

[0036] Each model branch is activated as needed, and different branches independently complete the load forecasting calculation under the corresponding operating conditions, ultimately outputting the load forecasting results in a unified manner. Figure 2 The overall structure clearly reflects the core design logic of this application: "dynamically matching the corresponding lightweight time-series prediction model based on the operating conditions".

[0037] Step 500: Call the target prediction model to perform load prediction on the multi-dimensional load feature vector to obtain the prediction results corresponding to the future load of the transformer area.

[0038] This step follows step 400, which matches the selected target prediction model with the multidimensional load feature vector constructed in step 200. The multidimensional load feature vector is used as the input data for the target prediction model, and load prediction calculation is performed to obtain the prediction results corresponding to the future load of the transformer area.

[0039] Specifically, the target prediction model takes a multi-dimensional load feature vector as input, calculates the future load of the transformer area based on its internal prediction logic, and outputs the load prediction results in the corresponding prediction time domain. The target prediction model selected under different operating conditions is different, and its prediction time domain and prediction accuracy are also adapted to the load fluctuation degree of the current operating condition.

[0040] The complete single-transformer load forecasting process is executed sequentially from step 100 to step 500, which sequentially completes load and environmental data preprocessing, multi-dimensional feature extraction, operating condition category determination, adaptation model matching, and the entire load forecasting process calculation. In actual transformer online monitoring scenarios, the complete process from step 100 to step 500 can be executed iteratively to continuously collect real-time transformer load and environmental data and dynamically update the load forecasting results, thereby achieving continuous rolling iterative forecasting of transformer load.

[0041] After obtaining the final prediction results, standardized data encapsulation is performed on the prediction results, and the encapsulated data is sent to the local microcontroller hardware. Based on the local hardware computing power, various business decisions can be made based on the prediction results, such as distribution area load allocation, power grid operation optimization and control, and local anomaly alarm judgment. The entire intelligent analysis work is completed locally on the edge device, realizing the implementation of the distribution area load prediction solution on low-computing-power embedded devices.

[0042] Based on the entire processing chain from steps 100 to 500 above, the complete load forecasting process can be completed serially in a single step. This application's embodiment adopts an overall architecture combining hierarchical operating condition identification with differentiated lightweight model matching, achieving a balance between lightweight computing power and high load forecasting accuracy. The solution accurately identifies the load operating state based on a preset state machine model and calls lightweight time-series forecasting models of corresponding complexity as needed. This enables high-precision calculations to be performed on local microcontroller terminals in power distribution areas with limited hardware resources, forming a complete solution that balances computing power overhead and forecasting performance. High-precision, real-time adaptive load forecasting can be achieved on local lightweight microcontroller terminals in power distribution areas.

[0043] Furthermore, in an optional embodiment, the operation of preprocessing the real-time collected power load data and environmental data of the transformer area to obtain the current load data sequence in time-series form in step 100 is specifically detailed to include the following sub-steps (a1) to (a3): (a1) Collect raw data of the transformer area at fixed time intervals. The raw data includes current, voltage, load rate corresponding to power load data and environmental data.

[0044] This step continuously collects raw data from the transformer substation at fixed time intervals, collecting two types of data dimensions: firstly, data related to the power load of the substation, including current, voltage, and load rate indicators; and secondly, supporting environmental data. These are uniformly collected as the raw data required for this step, ensuring that subsequent feature extraction and operating condition identification have a complete and synchronous basic data source, and realizing the synchronous collection of load and environmental correlation information.

[0045] (a2) The original data is denoised by using a filtering method based on statistical features to remove outliers in the data and obtain the pre-processed transformer area data.

[0046] This step employs a filtering method based on statistical characteristics to perform denoising operations on the raw data of the transformer substations collected in the previous step. By identifying and removing abnormal and sudden values ​​that deviate from the normal range through the statistical distribution patterns of the data, it eliminates data distortion caused by acquisition interference and equipment failure, and outputs preliminary pre-processed data of the transformer substations after removing outliers, thus avoiding interference from dirty data on subsequent operating condition judgment and load prediction results.

[0047] (a3) Introduce a sliding window mechanism to segment and organize the data of the transformer area. According to the fixed window length and fixed sliding step size, the data collected at the same time are segmented and arranged as sequence points to construct the current load data sequence corresponding to the transformer area data.

[0048] This step introduces a sliding window processing mechanism to segment and time-series regularize the denoised transformer area data. Based on a preset fixed window length and fixed sliding step size, all data collected at the same time are systematically extracted, and the data collected at a single time moment are arranged into standard sequence points. Finally, the data is regularized to generate a current load data sequence in time-series form, unifying the data time-series format and providing standardized time-series input materials for subsequent multi-dimensional feature vector construction.

[0049] Based on the explanations in (a1) to (a3) ​​above, the embodiments of this application complete time-sharing synchronous acquisition, statistical filtering and noise reduction, and sliding window time-series normalization in three subdivided steps. This allows for the simultaneous and complete collection of multi-dimensional raw data on the load and environment of the transformer area, accurately eliminating abnormal data to ensure data reliability, and uniformly outputting a load data sequence with normalized time and standard format. The entire preprocessing process is lightweight and has low computational overhead, adapting to the hardware computing power constraints of microcontrollers. It provides high-quality, standardized basic input data for the subsequent feature extraction, operating condition identification, and load forecasting processes, reducing prediction errors caused by abnormal data from the source.

[0050] Furthermore, in an optional embodiment, step 300, which uses a preset state machine model to process the multi-dimensional load feature vector and determine the target standard operating condition corresponding to the current operating condition of the transformer area, specifically includes the following sub-steps (b1) to (b4): (b1) For the three standard operating conditions of steady-state operation, fluctuating operation and strong disturbance operation predefined in the preset state machine model, configure the switching rules between the standard operating conditions. The switching rules include at least: switching threshold and duration threshold. The switching threshold is used to distinguish the difference between the operating states of the standard operating conditions at adjacent levels. The duration threshold is used to constrain the stable duration required for the operating condition determination.

[0051] This step configures the switching rules between three standard operating conditions predefined in the preset state machine model: steady-state operation, fluctuating operation, and strong disturbance operation.

[0052] The switching rules include at least a switching threshold and a duration threshold. The switching threshold is used to distinguish the differences between the operating states of standard operating conditions at adjacent levels. It is a numerical boundary standard for dividing the fluctuation boundary between different load operating conditions, and is used to define the boundary between the three standard operating conditions: steady-state operation, fluctuating operation, and strong disturbance operation. The duration threshold is used to constrain the stable duration required for operating condition determination. That is, when the operating condition changes, the new operating condition must be maintained for the duration limited by the duration threshold before it can be confirmed, so as to avoid frequent changes in operating conditions caused by instantaneous load fluctuations.

[0053] By configuring the above two types of thresholds, a complete parameter benchmark is provided for subsequent determination of judgment parameters, calculation of comprehensive working condition judgment index values, and step-by-step matching judgment.

[0054] (b2) Determine four types of judgment parameters, namely prediction error, error change rate, load fluctuation intensity, and load change gradient. Prediction error is the deviation between the predicted load value of the historical cycle and the corresponding actual load value. Error change rate is the amount of change between the prediction errors of adjacent historical cycles. Load fluctuation intensity is the degree of fluctuation of the real-time load data sequence. Load change gradient is the rate of change of the real-time load data in the time dimension. Load fluctuation intensity and load change gradient are calculated based on the real-time load data in the multi-dimensional load feature vector.

[0055] This step identifies four types of judgment parameters: prediction error, error change rate, load fluctuation intensity, and load change gradient.

[0056] Among them, the prediction error is the deviation between the predicted load value of the historical cycle and the corresponding actual load value, and the error change rate is the change between the prediction errors of adjacent historical cycles. The prediction error and the error change rate are extracted from the historical prediction statistics of the transformer area and are used to characterize the historical prediction deviation and its change pattern.

[0057] Load fluctuation intensity is the degree of fluctuation in the real-time load data sequence, and load change gradient is the rate of change of the real-time load data in the time dimension. Load fluctuation intensity and load change gradient are calculated based on the real-time load data in the multi-dimensional load feature vector and are used to characterize the real-time fluctuation characteristics of the current transformer area load.

[0058] The above four types of judgment parameters, combined with historical forecast data and real-time load data, comprehensively characterize the load operation status of the transformer area from four dimensions: forecast deviation, deviation change trend, load fluctuation degree, and load change rate, providing a parameter basis for the calculation of subsequent comprehensive operating condition judgment indicators.

[0059] (b3) Substitute the four types of judgment parameters that have been normalized into the preset comprehensive judgment function to calculate and output the comprehensive operating condition judgment index value. The preset comprehensive judgment function is used to integrate the four dimensions of data represented by the four types of judgment parameters to make a comprehensive quantitative judgment on the operating condition of the transformer area.

[0060] First, normalization is performed on the four types of judgment parameters output by (b2) to map parameters with different dimensions and different value ranges to a unified numerical range, eliminating the calculation deviation caused by the difference in the dimensions of different parameters; then, the normalized four types of judgment parameters are substituted into the pre-set comprehensive judgment function for unified calculation, and finally a single comprehensive working condition judgment index value is output, which integrates the multi-dimensional load fluctuation characteristics into a unified value, simplifying the subsequent hierarchical working condition matching and comparison process.

[0061] For example, this application abandons the traditional single threshold judgment mode and constructs a comprehensive operating condition judgment index function based on the above four-dimensional parameters to realize the refined dynamic identification of the operating status of the transformer area. For example, the preset comprehensive judgment function can adopt a weighted summation form, and the expression of the comprehensive operating condition judgment index is as follows (1): Formula (1); In formula (1): This is a comprehensive working condition assessment index used for working condition classification; Here are the weights for each feature, where Weights for prediction errors Weights for the rate of change of error Weights for load fluctuation intensity Let these be the weights of the load change gradient, and all weights satisfy... ; The normalized prediction error is... The normalized rate of change of error. The normalized load fluctuation intensity This represents the normalized load change gradient; the symbol "~" indicates that the corresponding parameter has undergone normalization.

[0062] (b4) Using a hierarchical progressive comparison logic, the comprehensive working condition judgment index value, switching threshold and duration threshold are matched step by step to complete the judgment of the target standard working condition of the transformer area.

[0063] This step employs a hierarchical, progressive comparison logic, matching the comprehensive operating condition judgment index value, switching threshold, and duration threshold level by level to complete the determination of the target standard operating condition of the transformer area.

[0064] Specifically, matching calculations are performed step by step according to the progressive order of steady-state operation, fluctuating operation, and strong disturbance operation: the comprehensive operating condition judgment index value output by (b3) is compared with the switching threshold configured by (b1) step by step, and at the same time, it is checked whether the duration of the operating condition meets the constraint of the duration threshold. The current load fluctuation amplitude and stable duration are judged layer by layer to see if they meet the judgment conditions of the corresponding level standard operating condition. If the judgment conditions of the current level are met, it is determined to be the standard operating condition of that level. If not, it enters the next level to continue the comparison, and the matching results are filtered step by step to finally determine the target standard operating condition to which the transformer area belongs.

[0065] By using a hierarchical and progressive comparison logic, the confusion and identification bias caused by a single threshold judgment can be avoided, and the standard operating conditions corresponding to different load fluctuation levels can be accurately classified, providing a clear basis for the subsequent selection of a suitable target prediction model.

[0066] Based on the above explanations of (b1) to (b4), the embodiments of this application pre-set a state machine model with matching switching thresholds and duration thresholds, multi-dimensional load judgment parameters, comprehensive judgment functions, and hierarchical progressive comparison logic to complete the working condition classification judgment. It integrates two types of information, historical prediction deviation and real-time load fluctuation, for comprehensive evaluation. The duration threshold suppresses the incorrect switching of working conditions caused by instantaneous disturbances. The hierarchical progressive comparison logic can accurately distinguish the three standard working condition boundaries of steady-state operation, fluctuating operation, and strong disturbance operation at each level. The entire judgment process has lightweight operation logic, adapts to the computing power constraints of microcontrollers, and can accurately and stably output the target standard working conditions of the transformer area. It provides a reliable and stable basis for the subsequent matching of the corresponding lightweight time series prediction model, ensuring the adaptive adaptability of the entire prediction process from the working condition identification stage.

[0067] Furthermore, in an optional embodiment, step (b4) completes the operating condition determination by comparing indicators and thresholds in a hierarchical and progressive manner, specifically including the following sub-steps (b41), (b42) and (b43); wherein the switching threshold includes a first switching threshold and a second switching threshold with smaller values, the steady-state operating condition corresponds to the first duration threshold, the fluctuating operating condition corresponds to the second duration threshold, and the strong disturbance operating condition corresponds to the third duration threshold; the words "first", "second" and "third" are only used as distinguishing identifiers and do not represent the numerical size or the order of execution.

[0068] (b41) If the comprehensive operating condition judgment index value is not greater than the first switching threshold and the holding time reaches the first duration threshold, then the target standard operating condition of the transformer area is determined to be steady-state operation.

[0069] This step corresponds to the steady-state operation condition determination rule. When the comprehensive operation condition determination index value is less than or equal to the first switching threshold, and the index meets the requirement that the holding time of the above interval reaches the pre-configured first duration threshold, the current target standard operation condition of the transformer area is determined to be steady-state operation. This rule is used to identify stable transformer area operation scenarios with small load fluctuation amplitude and stable prediction deviation. It filters out instantaneous small disturbances by relying on the duration constraint to avoid misjudging the operation condition due to short-term index fluctuations.

[0070] (b42) If the comprehensive operating condition judgment index value is greater than the first switching threshold and not greater than the second switching threshold, and the holding time reaches the second duration threshold, then the target standard operating condition of the transformer area is determined to be fluctuating operation.

[0071] This step corresponds to the judgment rule for fluctuating operating conditions. When the comprehensive operating condition judgment index value is greater than the first switching threshold but not greater than the second switching threshold, and the duration of the index remaining in this range reaches the second duration threshold, the target standard operating condition of the transformer area is judged to be fluctuating. This rule covers moderate fluctuation scenarios where the load has periodic small fluctuations and the prediction deviation changes slightly, serving as a transitional judgment interval between steady-state and strong disturbance operating conditions.

[0072] (b43) If the comprehensive operating condition judgment index value is greater than the second switching threshold and the holding time reaches the third duration threshold, then the target standard operating condition of the transformer area is determined to be strong disturbance operation.

[0073] This step corresponds to the judgment rule for strong disturbance operating conditions. When the comprehensive operating condition judgment index value is greater than the second switching threshold, and the duration of the index in this range reaches the third duration threshold, the target standard operating condition of the transformer area is judged to be strong disturbance operation. This rule is designed for extreme disturbance scenarios such as sudden load changes, drastic fluctuations, and a significant increase in prediction errors, capturing the state of drastic load changes in the transformer area.

[0074] For example, combined Figure 3 The state machine transition diagram shown is further explained below: Figure 3 The complete decision logic corresponds intuitively to (b41), (b42), and (b43). Figure 3 middle θ 1 This is the first switching threshold of this application. θ 2 This is the second switching threshold of this application; T 1 , T 2 , T 3 Corresponding to the first, second, and third duration thresholds, respectively. S t It serves as an indicator for comprehensive working condition assessment. Figure 3The three state branches in the middle are matched with three sets of sub-step decision conditions respectively: (1) When S t ≤ θ 1 And continuously meet the duration T 1 Execute rule (b41) to switch to the S1 steady-state operation; (2) When θ 1 < S t ≤ θ 2 And continuously meet the duration T 2 Execute rule (b42) to switch to S2 fluctuation operation state; (3) When S t > θ 2 And continuously meet the duration T 3 Execute rule (b43) to switch to S3 strong disturbance operation state.

[0075] at the same time Figure 3 It shows that bidirectional state switching is supported between the three types of working conditions. The switching process also needs to meet the corresponding threshold range and duration constraints, and fully reproduces the quantitative judgment logic of hierarchical progressive comparison.

[0076] Based on the above explanation of (b41), (b42) and (b43), the embodiments of this application divide three non-overlapping index intervals by two fixed thresholds, and combine them with the specific duration thresholds for each working condition to form a quantitative and clear working condition judgment standard, decomposing the abstract hierarchical progressive comparison logic into three judgment branches that can be directly and lightweightly calculated on the microcontroller. Figure 3 The system visually presents all threshold boundaries, duration constraints, and bidirectional state switching relationships, offering intuitive logic and facilitating embedded program deployment. The entire set of judgment rules strictly distinguishes between three types of load operating boundaries: steady-state, fluctuating, and strong disturbance. It filters out instantaneous interference using duration thresholds, significantly reducing the probability of frequent operating condition jumps and misjudgments, and outputting stable and reliable operating condition classification results. This provides accurate quantitative basis for the adaptive matching of the hierarchical lightweight prediction model.

[0077] Furthermore, in an optional embodiment, step 500 calls the target prediction model to perform load forecasting and obtain the future load forecast results for the transformer area, which is further refined into the following sub-steps (c1) to (c5); wherein the words "first", "second" and "third" are only used as distinguishing identifiers and do not represent numerical values ​​or execution order limitations.

[0078] (c1) Determine the short-term prediction layer, medium-term prediction layer and long-term prediction layer contained in the target prediction model; the short-term prediction layer is matched with the first prediction step size, the medium-term prediction layer is matched with the second prediction step size, and the long-term prediction layer is matched with the third prediction step size; the first prediction step size is smaller than the second prediction step size, and the second prediction step size is smaller than the third prediction step size.

[0079] This step clarifies the three-layer prediction hierarchy within the target prediction model: short-term, medium-term, and long-term prediction layers. Each of the three prediction layers is matched with its own prediction step size, namely the first prediction step size, the second prediction step size, and the third prediction step size. The step size values ​​of the three layers increase sequentially, that is, the first prediction step size is smaller than the second prediction step size, and the second prediction step size is smaller than the third prediction step size. This hierarchical recursive prediction divides the time scale into different time dimensions to adapt to the load change patterns of different time dimensions.

[0080] (c2) Input the multidimensional load feature vector into the short-term prediction layer, perform prediction calculations based on the first prediction step size, and obtain the short-term load prediction value.

[0081] This step takes the multidimensional load feature vector output in the previous text as the bottom-level input data and sends it into the short-term prediction layer. Based on the first prediction step size of this layer, the load prediction calculation is performed. The short-term load extrapolation is completed first, and the short-term load prediction value is output, providing basic input data for the subsequent medium- and long-term hierarchical recursive calculation.

[0082] (c3) The short-term load forecast value is sent to the medium-term forecast layer as a recursive input parameter, and the forecast calculation is performed according to the second forecast step size to obtain the medium-term load forecast value.

[0083] This step uses the short-term load forecast value obtained in (c2) as a recursive input parameter to pass to the medium-term forecast layer. Based on the second forecast step size, a medium-term load forecast calculation is performed to deduce the medium-term load forecast value, thus realizing the progressive deduction from the short-term forecast result to the medium-term forecast.

[0084] (c4) The medium-term load forecast value is sent to the long-term forecast layer as a recursive input parameter, and the forecast calculation is performed according to the third forecast step size to obtain the long-term load forecast value.

[0085] This step uses the medium-term load forecast value output by (c3) as the recursive input parameter to the long-term forecast layer. Based on the third forecast step size, the long-term load forecast operation is performed to calculate the long-term load forecast value, thus completing the hierarchical recursion from medium-term to long-term forecast.

[0086] (c5) Based on the short-term load forecast, medium-term load forecast and long-term load forecast, generate the complete multi-period load forecast results for a single transformer area and complete the output.

[0087] This step collects forecast data of three different time periods: short-term load forecast, medium-term load forecast, and long-term load forecast. It integrates and summarizes the forecast values ​​from multiple time periods to generate a single forecast process that fully covers the short, medium, and long-term load forecast results for transformer areas and outputs them uniformly.

[0088] It should be noted that, as mentioned above Figure 2 compared to, Figure 2 The diagram illustrates the horizontal selection relationship between different lightweight time-series prediction models for different operating conditions. Specifically, it selects a suitable target prediction model from a hierarchical progressive prediction model set based on the target standard operating condition. In contrast, the optional embodiment (c1) to (c5) further divides the target prediction model into short-term, medium-term, and long-term prediction layers within it after selection, outputting multi-period prediction results in a step-by-step recursive manner. These two approaches belong to the model selection level and the internal prediction layer level, respectively, and are different in dimension, thus not conflicting with each other.

[0089] Based on the above explanations of (c1) to (c5), the embodiments of this application construct a short-term prediction layer, a medium-term prediction layer, and a long-term prediction layer within the target prediction model, progressively advancing from short to long. By passing the prediction results of the previous layer as the input of the next layer, continuous load extrapolation across multiple time periods is achieved. Each layer is configured with a differentiated incremental prediction step size, which aligns with the operating characteristics of rapid short-term load fluctuations and gradual long-term load changes in the distribution area. The long-term prediction task is decomposed into layers, reducing the computational load of a single model and adapting to the low computing power hardware constraints of the microcontroller. The entire process can output a complete set of short-term, medium-term, and long-term load prediction data simultaneously in a single operation, without the need for multiple independent modeling and predictions. While controlling the computational overhead of embedded devices, it fully covers the load prediction requirements of multiple time scales and outputs complete and unified multi-time period load prediction results.

[0090] Below, the entire processing chain for single load forecasting based on steps 100 to 500 can be further extended to a cyclic iterative continuous execution mode to achieve rolling real-time forecasting of transformer area loads. Specifically, a processing time period is set for the real-time collected transformer area power load data and environmental data, and the single load forecasting operation for the transformer area is repeatedly executed according to this time period. In an optional embodiment, when performing transformer area load forecasting iteratively based on multiple processing time periods, a dynamic scheduling process is added, which includes sub-steps (d1) to (d3): (d1) During the real-time monitoring of microcontroller resource status, a comprehensive computing power load function is constructed based on CPU utilization and memory usage. The comprehensive computing power load function is used to measure computing power consumption.

[0091] This step continuously monitors the operating status of microcontroller hardware resources in real time, continuously collecting two core hardware resource indicators: CPU usage and memory usage. Based on these two types of collected resource indicators, a comprehensive computing load evaluation rule is constructed to quantify the computing power consumption of the current load prediction computing task on the hardware, thus completing a unified quantitative representation of the device's real-time computing load and providing a basic evaluation basis for subsequent multi-objective optimization in terms of hardware resources.

[0092] For example, to quantify the real-time computing power consumption of the microcontroller, the module continuously collects two types of hardware indicators: CPU utilization and memory usage, and constructs a weighted comprehensive computing power load function as shown in the following formula (2): Formula (2); In formula (2): C t This represents the overall computing load at the current moment, used to uniformly quantify the overall hardware resource consumption of the prediction task. U cpu The CPU utilization rate of the microcontroller is obtained in real time. U mem This refers to the real-time acquisition of microcontroller memory usage. α The weighting coefficient corresponding to CPU utilization; β This refers to the weighting coefficient corresponding to memory occupancy. This application embodiment uses a weighted fusion of two types of hardware occupancy indicators to transform multi-dimensional hardware load data into a single quantifiable value, intuitively representing the real-time computing power pressure of the device.

[0093] (d2) Incorporate both the computing power consumption and the prediction error corresponding to the current period into the optimization objective to construct an optimization objective function that takes into account both prediction error and computing power consumption.

[0094] This step introduces two core optimization evaluation indicators: the device computing power consumption level quantified by (d1) and the prediction error level corresponding to the output result after the current complete prediction cycle is completed. These two indicators are incorporated into a unified optimization evaluation system to build a comprehensive optimization evaluation standard that can simultaneously take into account the accuracy of prediction results and the constraints of hardware computing power consumption, thus forming a unified optimization benchmark that balances prediction accuracy and hardware resource consumption.

[0095] For example, based on the quantification results of computing load, prediction accuracy constraints are introduced simultaneously, and the two core indicators of computing power consumption and prediction error are integrated to construct a minimization optimization objective function, which simultaneously takes into account prediction accuracy and hardware resource limitations. The minimization optimization objective function is as follows (3): Formula (3); In formula (3): JTo comprehensively optimize the target value, the smaller the value, the better the overall effect of prediction accuracy and computing power consumption; E t The forecast error corresponding to the output result of the current forecast period represents the accuracy of the load forecast. C t The aforementioned comprehensive computing load represents the hardware computing power consumption brought about by the current prediction task; λ 1 The weighting coefficients corresponding to the prediction error are used to adjust the proportion of prediction accuracy in the optimization objective; λ 2 This is the weighting coefficient corresponding to the overall computing load, used to adjust the proportion of hardware resource consumption in the optimization objective.

[0096] After solving the above optimization objective function to obtain the optimal control scheme, the system synchronously and dynamically adjusts the three types of prediction operation parameters to achieve optimal allocation of hardware resources: (1) Prediction step size: adjust the time span corresponding to a single inference of the hierarchical prediction layer to balance the computational load of short-term and long-term prediction; (2) Model call frequency: control the startup frequency of the lightweight time series prediction model in the continuous period to reduce the computational density during high-load periods; (3) Inference computation intensity: constrain the computational complexity of each prediction level in the short-term, medium-term and long-term within the model, and simplify or increase the computational overhead of model inference as needed.

[0097] The entire scheduling logic is based on real-time hardware monitoring and multi-objective collaborative optimization. In multi-cycle rolling iterative prediction scenarios, it continuously balances prediction accuracy and embedded terminal computing load, avoiding long-term overload operation of the microcontroller.

[0098] (d3) By solving the optimization objective function, the prediction step size, inference computation intensity and model call frequency used by the lightweight time series prediction model called in the next cycle are dynamically adjusted. The model call frequency is used to control the startup frequency of the hierarchical lightweight time series model, and the inference computation intensity is used to constrain the computational complexity of the prediction layers contained in the lightweight time series prediction model.

[0099] This step solves the comprehensive optimization evaluation criteria established above. Based on the optimal scheduling scheme obtained from the solution, the three operating parameters of the lightweight time series prediction model to be run in the next cycle are dynamically adapted and adjusted. The three parameters to be adjusted are the prediction step size used by the model, the complexity and intensity of the model inference operation, and the startup call frequency of the model in subsequent cycles. Among them, the model call frequency is used to control the number of times the hierarchical lightweight time series model is started and run, the inference operation intensity is used to constrain the computational complexity of each prediction level within the model, and the prediction step size is used to control the time scale corresponding to a single load simulation at each level. The adaptive scheduling and control of multi-cycle iterative prediction is completed by coordinating the adjustment of the three types of parameters.

[0100] For example, using Figure 4 The diagram of the computing power-aware dynamic scheduling mechanism fully presents the complete method flow of computing power adaptive control during multi-cycle iterative prediction: The process begins by collecting two types of hardware operation data, namely CPU utilization and memory occupancy, and substituting them into the comprehensive computing power load calculation formula labeled as formula (2) to complete the quantification of computing power consumption; the quantified comprehensive computing power load value and the current cycle prediction error are substituted into the optimization target calculation formula labeled as formula (3) for comprehensive solution; based on the solution results, the adjustment decision of three types of parameters is completed simultaneously, and the prediction step size, model call frequency, and inference calculation intensity are adapted and modified respectively; the adjusted parameters are sent to the load prediction process for execution; after the single cycle prediction is completed, the operation feedback data such as task time consumption, power consumption, and queue length are collected and sent back to the front-end data collection link to form a continuous loop control logic, and the prediction operation strategy is continuously and dynamically optimized during the continuous multi-cycle prediction process.

[0101] Based on the above explanations of (d1) to (d3), this application embodiment achieves adaptive dynamic adjustment of the prediction strategy during multi-cycle rolling prediction by real-time sensing of hardware computing load and constructing multi-objective optimization logic in conjunction with prediction accuracy. This method simultaneously considers the two core constraints of prediction error and hardware resource consumption, flexibly adjusting the prediction step size, model call frequency, and inference computation complexity according to the real-time operating load of the device, thus rationally allocating limited hardware resources while ensuring load prediction accuracy. It effectively avoids problems such as microcontroller computing overload, power consumption surges, and program lag under long-term iterative prediction, continuously balancing prediction accuracy and terminal computing power consumption, significantly improving the long-term uninterrupted stable operation capability of embedded terminals in distribution centers, and adapting to the actual implementation needs of long-term online rolling load prediction for low-computing-power microcontrollers.

[0102] Below, the entire processing chain for single load forecasting based on steps 100 to 500 can be further extended to a cyclic iterative continuous execution mode to achieve rolling real-time forecasting of transformer area loads. Specifically, a processing time period is set for the real-time collected transformer area power load data and environmental data, and the single load forecasting operation for the transformer area is repeatedly executed according to this time period. In an optional embodiment, when performing transformer area load forecasting based on multiple processing time periods, a threshold adaptive update process is added, which includes sub-steps (e1) to (e3): (e1) Determine the prediction error obtained from a single load prediction operation of the transformer area in the current cycle, and retrieve the historical switching threshold and historical duration threshold configured for the preset state machine model in the operating condition identification operation of the previous cycle.

[0103] This step extracts two types of key data: first, it calculates the prediction error corresponding to the output of the single load prediction operation of the current complete cycle area, thereby characterizing the matching degree of the current cycle condition judgment threshold; second, it retrieves two types of historical parameters, namely the switching threshold and the duration threshold, which were pre-configured in the state machine model in the previous cycle condition identification stage, and retains the historical threshold benchmark to provide historical reference for subsequent parameter updates.

[0104] (e2) Combining the historical switching threshold, the historical duration threshold and the prediction error corresponding to the current period, a weighted fusion update method is used to generate a new switching threshold and a new duration threshold. The weighted fusion update method takes into account both the continuity of the historical threshold and the correction effect of the current prediction error.

[0105] This step integrates the historical switching thresholds, historical duration thresholds, and current cycle prediction errors obtained through data fusion, and uses a weighted fusion update logic to generate new switching thresholds and duration thresholds. This weighted fusion update method takes into account two layers of constraints: on the one hand, it preserves the changing trends of historical thresholds to maintain the continuous stability of the operating condition judgment criteria; on the other hand, it incorporates the current cycle prediction error to complete the correction, adjusting the threshold boundaries in reverse to adapt to the real-time load change characteristics of the distribution area based on the current prediction deviation.

[0106] (e3) Configure the new switching threshold and the new duration threshold into the preset state machine model for use in the next cycle when performing condition identification.

[0107] This step reconfigures the new switching threshold and new duration threshold obtained through weighted fusion update into the preset state machine model, completing the iterative update of the model's judgment parameters. The updated set of threshold parameters will be directly called when the working condition identification process is executed in the next cycle, realizing the automatic update of the working condition judgment benchmark after each prediction cycle is completed, forming a cycle closed-loop adaptive adjustment.

[0108] For example, traditional fixed thresholds cannot adapt to long-term dynamic changes in transformer area load and are prone to operating condition identification errors. Therefore, this application is equipped with an adaptive threshold update mechanism, which dynamically iterates and updates the judgment threshold based on the periodic prediction error. The update adopts a weighted fusion calculation method, as shown in the following formula (4): Formula (4); In formula (4): θ t The updated threshold values ​​include the updated switching threshold and duration threshold. θ t-1 The historical threshold retained from the previous period corresponds to the historical switching threshold and historical duration threshold retrieved by (e1); γThis is the historical threshold weighting coefficient, with a value between 0 and 1. It is used to control the proportion of historical thresholds in the fusion update and to ensure the continuity of threshold changes. e t The current period prediction error is the current period prediction error extracted from (e1), which is used to correct the threshold boundary.

[0109] The weighted fusion calculation method is completely consistent with the weighted fusion update logic recorded in (e2). It retains the stable benchmark of historical thresholds and introduces the current prediction error to complete dynamic correction. It realizes automatic updating of the operating condition judgment boundary after each cycle, continuously adapts to the load change characteristics of the transformer area, and improves the robustness and environmental adaptability of operating condition identification in the long-term iterative prediction process.

[0110] Based on the explanations of (e1) to (e3) above, the embodiments of this application adopt a periodic closed-loop threshold adaptive update, abandoning the fixed judgment threshold. The operating condition judgment boundary is dynamically corrected based on the error generated in each round of prediction. Simultaneously, weighted fusion is used to retain the continuity of historical thresholds, avoiding significant threshold jumps that could cause operational condition identification disorder. This allows for continuous adaptation to the dynamic changes in load distribution and fluctuation patterns during long-term operation of the transformer substation, significantly improving the adaptability and stability of multi-layer progressive state machine operational condition identification, reducing misjudgments and incorrect judgments of operating conditions during long-term iterative prediction, and ensuring the adaptation accuracy of the hierarchical matching prediction model from the source.

[0111] Furthermore, the preferred embodiment provided in this application is: a multi-cycle iterative prediction process, which includes two sets of parallel closed-loop control logics, forming a dual closed-loop feedback mechanism. The resource closed loop corresponds to the (d1) to (d3) computing power dynamic scheduling process, and the prediction closed loop corresponds to the (e1) to (e3) adaptive threshold update process. The two closed loops can coexist and run synchronously, and work together to complete the adaptive control of the entire rolling prediction process.

[0112] For the dual-loop feedback mechanism, the entire control logic is divided into two types of closed-loop links: prediction closed loop and resource closed loop. The prediction closed loop completes feedback control based on prediction error. The complete process is as follows: after completing load prediction and outputting the predicted value, the predicted value is compared with the actual load value to calculate the error. Then, the error data is fed back to the operating condition judgment logic for dynamically updating the operating condition judgment criteria and the hierarchical prediction model matching strategy, forming a continuously iteratively optimized error correction link. The resource closed loop completes feedback control based on hardware computing power. It continuously collects relevant operating data on microcontroller CPU and memory usage, evaluates computing power load in real time, and dynamically corrects the prediction operation scheduling strategy to avoid hardware resource overload or computing power idle waste during long-term iterative prediction, forming a continuous self-balancing control link for computing power. The complete flow logic of the prediction closed loop is prediction → error calculation → state adjustment → model update; the complete flow logic of the resource closed loop is computing power monitoring → scheduling optimization → model execution → resource feedback.

[0113] For example, such as Figure 5 This document fully demonstrates the complete methodology for parallel operation of a dual-loop system. The system is divided into two independent closed loops, which interact and cooperate with each other. The upper loop is the prediction loop: it collects real load data to predict the load and outputs the prediction results. It then compares the predictions with the real load to calculate the error. Based on the error values, it updates and adjusts the operating condition judgment thresholds and model matching rules. The updated judgment rules are then passed to the operating condition identification stage for the next cycle's load stratification prediction. The lower loop is the resource loop: it continuously collects hardware usage data such as CPU and memory to assess computing power load. It then generates a scheduling adjustment plan based on the load assessment results, manages the prediction operation parameters using the scheduling plan, and simultaneously feeds back the newly updated judgment thresholds obtained through adaptive threshold updates to the resource monitoring front end, enabling data exchange between the two closed loops. After both closed loops complete operation, the complete load prediction results for a single cycle are aggregated and output. All operational data is returned to the beginning of the process, forming a continuous iterative control logic.

[0114] The preceding sections described the core process of single-cycle load forecasting, the computing power scheduling process for multi-cycle iterations, and the adaptive update process of operating condition thresholds step by step. To fully present the complete closed-loop execution logic for multi-cycle rolling load forecasting in this application, the following section will combine... Figure 6 The overall method flowchart shown provides a unified explanation of the complete process S101 to S109 of this application.

[0115] Specifically, the overall execution flow of the method in this application is as follows: Figure 6 The overall method flowchart shows the complete sequence of steps to be executed: S101: Real-time collection of load data in the transformer area and preprocessing, corresponding to step 100, and data cleaning, missing value filling and feature standardization can be completed by refining the logic according to sub-steps (a1) to (a3); S102: Construct a multi-dimensional load feature vector, which integrates the pre-processed power load data and environmental data of the transformer area into a standardized feature vector to provide input for subsequent operating condition identification; S103: Based on the state machine model, identify the working condition. In step 300, the judgment logic can be refined by combining sub-steps (b1) to (b4), sub-steps (b41), (b42) and (b43) to complete the load working condition division. S104: Select the corresponding prediction model based on the operating conditions and match a lightweight time-series prediction model that is compatible with the current load fluctuation conditions of the transformer area. S105: Perform hierarchical progressive load forecasting, corresponding to step 500. Short-term and medium-to-long-term load synchronous forecasting can be completed based on the hierarchical deduction logic of sub-steps (c1) to (c5). S106: Dynamic scheduling control based on computing power status. The above sub-steps (d1) to (d3) are used to sense the computing power load of the microcontroller hardware in real time, solve the optimization target, and then dynamically adjust the predicted operating parameters for the next cycle. S107: Calculate the prediction error and perform feedback update. Compare the prediction output of this round with the actual load data of the distribution area to quantify the prediction error of the current cycle, which will serve as the benchmark data for subsequent adaptive adjustment. S108: Update model parameters and operating condition thresholds. Corresponding to the above sub-steps (e1) to (e3), retrieve historical judgment thresholds, combine them with the current prediction error to generate new switching thresholds and duration thresholds through weighted fusion, and update them to the state machine model. S109: Output the final prediction result, outputting the multi-scale load prediction values ​​obtained from the current period's hierarchical progressive load prediction.

[0116] After all steps in a single cycle are completed, it is determined whether to continue load forecasting for the distribution area. If the determination is to continue forecasting, the process returns to S101 and repeats the entire process, achieving multi-cycle rolling iterative forecasting. If the determination is to terminate forecasting, the entire process ends. All steps from S101 to S109 work together to form a closed-loop adaptive predictive control system, fully realizing the "prediction-feedback-optimization" cyclical operation mechanism. Based on the dual control logic of resource closed loop and prediction closed loop, it continuously balances the forecasting accuracy and the computing power overhead of the embedded terminal, adapting to the scenario of long-term uninterrupted online load forecasting for distribution areas.

[0117] Furthermore, this application also provides a preferred implementation method. The layered progressive load forecasting method for distribution areas described above adopts a cyclical mode, repeatedly iterating the complete load forecasting process according to a fixed time cycle, possessing the basic operating conditions for multi-cycle continuous operation and continuous accumulation of historical operating data. Based on this multi-cycle iterative operation characteristic, in order to further optimize the edge terminal operating condition switching response speed and reduce hardware storage read / write losses and computing resource consumption, this application also adds a preferred implementation method, namely, an operating condition prediction model pre-caching scheduling mechanism, as a supplementary optimization means to the original basic scheme. The specific explanation includes the following: The pre-caching scheduling mechanism of the working condition prediction model is as follows: Based on the historical working condition data accumulated over multiple cycles of iteration, the load working conditions that are likely to occur in the next cycle are predicted in advance, and the corresponding lightweight models are pre-stored in a hierarchical manner to form a candidate set. This avoids the delay, storage loss, and memory resource contention issues caused by loading the model after the original basic solution has completed the working condition identification. It fully adapts to the basic rule that only one working condition is determined in a single cycle.

[0118] For example, this application can only determine and output a single target operating condition among steady state, fluctuation, and strong disturbance in each operating cycle. On this basis, cross-cycle operating condition prediction pre-preparation logic is added. The specific implementation steps include the following: Step (1): Summarize the working condition switching records of multiple historical operating cycles, statistically analyze the occurrence pattern of mutual conversion between the three types of working conditions, predict the probability of each type of working condition appearing in the next cycle, and select the working conditions with the highest probability as candidate working conditions. Step (2): Construct a lightweight model candidate set based on the predicted probability, load the complete lightweight model corresponding to the highest probability case into high-speed memory in advance; cache only the core inference parameters of the model for the second highest probability case; and retain only the basic structure information for the low probability case without occupying high-speed running memory. Step (3): After entering the next cycle, first execute the standard working condition identification process to determine the unique target working condition for the current period, and directly retrieve the matching model from the pre-prepared model candidate set for hierarchical progressive prediction; Step (4): After the prediction is completed, the current actual working conditions and prediction error are used to update the working condition switching rules, and the model pre-caching strategy for the next cycle is iteratively corrected to form a self-optimizing cycle. Step (5): Incorporate model cache usage into the computing load assessment, dynamically constrain cache size, and prevent pre-caching operations from preempting the computing resources required for predictive inference.

[0119] In addition to the aforementioned pre-caching scheduling mechanism for the predictive model, based on the operational foundation of this application—multi-cycle iterative and continuous collection and updating of transformer load and environmental time-series data—this application also provides another preferred implementation method to address the problems of power consumption characteristic drift, interference from old and failed samples in determining operating conditions, and continuous decay of prediction accuracy during long-term operation of transformer areas. This method is a dynamic forgetting and adaptive correction mechanism for load time-series data, serving as a supplementary optimization to the original data preprocessing, operating condition identification, and threshold update processes. The specific explanation is as follows: The load time-series dynamic forgetting and feature drift adaptive correction mechanism is applied to the data preprocessing stage. When constructing the load time-series sequence, it distinguishes the reference weights of new and old data, assigns differentiated reference weights according to the distance of the sample from the current collection time, and adjusts the data forgetting decay rate in conjunction with the unique target operating condition determined in the current period. It synchronously quantifies and monitors the degree of long-term power consumption feature deviation and links it to correct the operating condition judgment rules. When the feature deviation exceeds the preset range, it automatically reduces the weight of drift failure features in the operating condition judgment. It can also synchronously transmit the drift quantification results to the operating condition threshold update stage, and synchronously adjust the update magnitude of the switching threshold and duration threshold. It adapts to three types of differentiated load operating conditions: steady state, fluctuation, and strong disturbance. It reduces the probability of operating condition misjudgment from the data source. The entire process only uses lightweight weighted calculations and will not add extra computing power burden to the microcontroller.

[0120] Below, to fully present the complete closed-loop execution logic for multi-cycle rolling load forecasting in this application, an embodiment of this application provides a microcontroller-adapted distribution area load forecasting system. This system embodiment corresponds to the aforementioned method embodiment. For ease of reading, this system embodiment will not repeat the details of the aforementioned method embodiment, but it should be understood that the system in this application embodiment can correspondingly implement all the contents of the aforementioned method embodiment. This system is applied to form a complete solution that balances computing power overhead and forecasting performance, enabling high-precision, real-time adaptive load forecasting on a lightweight microcontroller terminal located in the distribution area. Specifically, as shown below... Figure 7 The overall system structure diagram shown below includes: The data acquisition and preprocessing module 110 is used to perform preprocessing on the real-time acquired power load data and environmental data of the transformer area to obtain a time-series current load data sequence. Each sequence point in the current load data sequence stores the power load data and environmental data acquired at the same time. The data acquisition and preprocessing module 110 is also used to extract features corresponding to the time dimension, period dimension, and external environment dimension from the current load data sequence, and construct a multi-dimensional load feature vector that matches the current load data sequence. The state machine condition identification module 120 is used to process the multidimensional load feature vector using a preset state machine model, and determine the target standard condition corresponding to the current operating condition of the transformer area. The target standard condition is a standard condition in one of multiple progressive levels divided according to the stability of load operation. The preset state machine model is used to map the multidimensional load feature vector to the standard condition. The hierarchical progressive prediction module 130 is used to select a target prediction model that is adapted to the target standard working condition from a preset hierarchical progressive prediction model set. The hierarchical progressive prediction model set includes multiple lightweight time series prediction models that correspond one-to-one with the multiple progressive level standard working conditions and whose computational complexity increases. The lightweight time series prediction model is a time series prediction model whose computational and storage requirements are adapted to the operation of the microcontroller. The prediction result output module 170 is used to call the target prediction model to perform load prediction on the multidimensional load feature vector and obtain the prediction result corresponding to the future load of the transformer area.

[0121] Furthermore, in some optional embodiments, the multiple progressively hierarchical standard operating conditions divided according to the degree of load operation stability include: steady-state operating condition, fluctuating operating condition, and strong disturbance operating condition. The state machine operating condition identification module 120 is also specifically used for: For three standard operating conditions—steady-state operation, fluctuating operation, and strong disturbance operation—predefined in the preset state machine model, switching rules are configured between each standard operating condition. The switching rules include at least a switching threshold and a duration threshold. The switching threshold is used to distinguish the differences between the operating states of standard operating conditions at adjacent levels, and the duration threshold is used to constrain the stable duration required for the operating condition determination. Four types of judgment parameters are determined: prediction error, error change rate, load fluctuation intensity, and load change gradient. The prediction error is the deviation between the predicted load value of a historical load cycle and the corresponding actual load value. The error change rate is the change in the prediction error between adjacent historical cycles. The load fluctuation intensity is the degree of fluctuation in the real-time load data sequence. The load change gradient is the rate of change of the real-time load data in the time dimension. The load fluctuation intensity and the load change gradient are calculated based on the real-time load data in the multi-dimensional load feature vector. The four types of judgment parameters that have been normalized are substituted into the preset comprehensive judgment function for calculation, and the comprehensive operating condition judgment index value is output. The preset comprehensive judgment function is used to integrate the four dimensions of data represented by the four types of judgment parameters to make a comprehensive quantitative judgment on the operating condition of the transformer area. A hierarchical and progressive comparison logic is adopted to match the comprehensive operating condition judgment index value, the switching threshold and the duration threshold level by level to complete the determination of the target standard operating condition of the transformer area.

[0122] Furthermore, in some optional embodiments, the switching threshold includes a first switching threshold and a second switching threshold, the first switching threshold being less than the second switching threshold, the steady-state operating condition corresponding to a first duration threshold, the fluctuating operating condition corresponding to a second duration threshold, and the strong disturbance operating condition corresponding to a third duration threshold. Therefore, during the operation of the state machine condition identification module 120, a hierarchical progressive comparison logic is adopted to match the comprehensive condition judgment index, the switching threshold, and the duration threshold level by level to complete the determination of the target standard condition of the transformer area, including: If the comprehensive operating condition judgment index value is not greater than the first switching threshold, and the holding time reaches the first duration threshold, then the target standard operating condition of the transformer area is determined to be steady-state operation; or, If the comprehensive operating condition judgment index value is greater than the first switching threshold and not greater than the second switching threshold, and the duration of maintenance reaches the second duration threshold, then the target standard operating condition of the transformer area is determined to be fluctuating operation; or... If the comprehensive operating condition judgment index value is greater than the second switching threshold and the holding time reaches the third duration threshold, then the target standard operating condition of the transformer area is determined to be strong disturbance operation.

[0123] Furthermore, in some optional embodiments, the prediction result output module 170 is also specifically used for: The target prediction model is defined to include a short-term prediction layer, a medium-term prediction layer, and a long-term prediction layer; the short-term prediction layer is matched with a first prediction step size, the medium-term prediction layer is matched with a second prediction step size, and the long-term prediction layer is matched with a third prediction step size; the first prediction step size is smaller than the second prediction step size, and the second prediction step size is smaller than the third prediction step size. The multidimensional load feature vector is input into the short-term prediction layer, and a prediction operation is performed based on the first prediction step size to obtain the short-term load prediction value. The short-term load forecast value is sent as a recursive input parameter to the medium-term forecast layer, and a forecast calculation is performed based on the second forecast step size to obtain the medium-term load forecast value. The medium-term load forecast value is sent to the long-term forecast layer as a recursive input parameter, and a forecast calculation is performed based on the third forecast step size to obtain the long-term load forecast value. Based on the short-term load forecast, the medium-term load forecast, and the long-term load forecast, a complete multi-period load forecast result for a single transformer area is generated and output.

[0124] Furthermore, in some optional embodiments, the data acquisition and preprocessing module 110 is also specifically used for: Raw data of the transformer area is collected at fixed time intervals. The raw data includes power load data corresponding to current, voltage, and load rate, as well as environmental data. The original data is denoised using a filtering method based on statistical features to remove outliers and obtain pre-processed transformer area data. A sliding window mechanism is introduced to segment and organize the data of the transformer area. Data collected at the same time are extracted in an orderly manner according to a fixed window length and a fixed sliding step size and arranged as sequence points to construct the current load data sequence corresponding to the transformer area data.

[0125] Furthermore, in some optional embodiments, a processing time period is set for the real-time collected power load data and environmental data of the transformer area, so as to perform a single load prediction operation for the transformer area according to the time period. In the process of realizing load prediction of the transformer area based on multiple time periods, the system also includes a computing power perception dynamic scheduling module 140, which is specifically used for: The CPU utilization and memory usage of the microcontroller are monitored in real time, and a comprehensive computing power load function is constructed based on the monitoring results (based on CPU utilization and memory usage). The comprehensive computing power load function is used to measure computing power consumption. The computing power consumption and the prediction error corresponding to the current period are both incorporated into the optimization objective to construct an optimization objective function that takes into account both prediction error and computing power consumption. By solving the optimization objective function, the prediction step size, inference computation intensity, and model call frequency used by the target prediction model in the next cycle are dynamically adjusted. The model call frequency is used to control the startup frequency of the hierarchical lightweight time series model, and the inference computation intensity is used to constrain the computational complexity of the prediction layers contained in the target prediction model.

[0126] Furthermore, in some optional embodiments, the system sets a processing time period for the real-time collected power load data and environmental data of the transformer area, and performs a single load prediction operation for the transformer area according to the time period. In the process of performing load prediction for the transformer area based on multiple time periods, the system further includes an adaptive threshold update module 160, which is specifically used for: Determine the prediction error obtained from a single load prediction operation of the transformer area in the current cycle, and retrieve the historical switching threshold and historical duration threshold configured for the preset state machine model in the operating condition identification operation of the previous cycle. Combining the historical switching threshold, the historical duration threshold, and the prediction error corresponding to the current period, a weighted fusion update method is used to generate a new switching threshold and a new duration threshold. The weighted fusion update method takes into account both the continuity of historical thresholds and the correction effect of the current prediction error. The new switching threshold and the new duration threshold are configured into a preset state machine model for use in the next cycle when identifying the operating condition.

[0127] Furthermore, in some optional embodiments, the multi-cycle iterative prediction process includes two sets of parallel closed-loop control logics, forming a dual closed-loop feedback mechanism. The resource closed loop corresponds to the dynamic scheduling process of computing power, and the prediction closed loop corresponds to the adaptive threshold update process. The two closed loops can coexist and operate synchronously, collaboratively completing the adaptive control of the entire rolling prediction process. Therefore, the system provided in this application embodiment also includes a dual closed-loop feedback control module 150, used to coordinate the respective work of the computing power-aware dynamic scheduling module 140 and the adaptive threshold update module 160.

[0128] Furthermore, in some optional embodiments, the system provided in this application embodiment also includes a microcontroller execution unit 180, which is used to carry out the overall operation of the entire process of the "microcontroller-adapted transformer area load prediction method".

[0129] In summary, the microcontroller-adapted substation load forecasting system provided in this application includes a processor and a memory. The aforementioned data acquisition and preprocessing module, state machine condition identification module, hierarchical progressive forecasting module, computing power perception dynamic scheduling module, dual closed-loop feedback control module, adaptive threshold update module, forecast result output module, and microcontroller execution unit are all stored as program units in the memory. The processor executes the aforementioned program units stored in the memory to achieve the corresponding functions.

[0130] The processor contains a kernel, which retrieves the corresponding program unit from memory. One or more kernels can be configured, and by adjusting the kernel parameters, high-precision calculations can be performed on local microcontroller terminals in areas with limited hardware resources. This forms a complete solution that balances computing power overhead and prediction performance, enabling high-precision, real-time adaptive load prediction on lightweight local microcontroller terminals in the distribution area.

[0131] This application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the microcontroller-adapted substation load forecasting method described above.

[0132] This application provides an electronic device, which includes at least one processor, at least one memory and a bus connected to the processor; wherein the processor and the memory communicate with each other through the bus; the processor is used to call program instructions in the memory to execute the transformer load forecasting method adapted to the microcontroller as described above.

[0133] This application also provides a computer program product that, when executed on a data processing device, is adapted to perform the steps of initializing a transformer area load forecasting method with an adapted microcontroller.

[0134] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A system that specifies functions in one or more boxes.

[0135] In a typical configuration, the device includes one or more processors (CPUs), memory, and a bus. The device may also include input / output interfaces, network interfaces, etc.

[0136] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, like read-only memory (ROM) or flash RAM, and memory includes at least one memory chip. Memory is an example of computer-readable media.

[0137] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0138] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0139] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied 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.

[0140] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for predicting transformer load in a distribution area adapted to a microcontroller, characterized in that, The method includes: Preprocessing is performed on the real-time collected power load data and environmental data of the transformer area to obtain a time-series current load data sequence. Each sequence point in the current load data sequence stores the power load data and environmental data collected at the same time. Features corresponding to the time dimension, period dimension, and external environment dimension are extracted from the current load data sequence to construct a multi-dimensional load feature vector that matches the current load data sequence; The multidimensional load feature vector is processed using a preset state machine model to determine the target standard operating condition corresponding to the current operating condition of the transformer area. The target standard operating condition is a standard operating condition in one of multiple progressive levels divided according to the stability of load operation. The preset state machine model is used to map the multidimensional load feature vector to the standard operating condition. Select a target prediction model that is adapted to the target standard working condition from a preset hierarchical progressive prediction model set. The hierarchical progressive prediction model set includes multiple lightweight time series prediction models that correspond one-to-one with the multiple progressive hierarchical standard working conditions and whose computational complexity increases. The lightweight time series prediction model is a time series prediction model whose computational and storage requirements are adapted to the operation of the microcontroller. The target prediction model is invoked to perform load prediction on the multidimensional load feature vector to obtain the prediction results corresponding to the future load of the transformer area.

2. The method according to claim 1, characterized in that, The standard operating conditions, categorized by load stability, include: steady-state operating condition, fluctuating operating condition, and strong disturbance operating condition. The process of using a preset state machine model to process the multi-dimensional load feature vector and determine the target standard operating condition corresponding to the current operating condition of the transformer substation includes: For three standard operating conditions—steady-state operation, fluctuating operation, and strong disturbance operation—predefined in the preset state machine model, switching rules are configured between each standard operating condition. The switching rules include at least a switching threshold and a duration threshold. The switching threshold is used to distinguish the differences between the operating states of standard operating conditions at adjacent levels, and the duration threshold is used to constrain the stable duration required for the operating condition determination. Four types of judgment parameters are determined: prediction error, error change rate, load fluctuation intensity, and load change gradient. The prediction error is the deviation between the predicted load value of a historical cycle and the corresponding actual load value. The error change rate is the change in the prediction error between adjacent historical cycles. The load fluctuation intensity is the degree of fluctuation in the real-time load data sequence. The load change gradient is the rate of change of the real-time load data in the time dimension. The load fluctuation intensity and the load change gradient are calculated based on the real-time load data in the multi-dimensional load feature vector. The four types of judgment parameters that have been normalized are substituted into the preset comprehensive judgment function for calculation, and the comprehensive operating condition judgment index value is output. The preset comprehensive judgment function is used to integrate the four dimensions of data represented by the four types of judgment parameters to make a comprehensive quantitative judgment on the operating condition of the transformer area. A hierarchical and progressive comparison logic is adopted to match the comprehensive operating condition judgment index value, the switching threshold and the duration threshold level by level to complete the determination of the target standard operating condition of the transformer area.

3. The method according to claim 2, characterized in that, The switching threshold includes a first switching threshold and a second switching threshold, where the first switching threshold is less than the second switching threshold. The steady-state operating condition corresponds to a first duration threshold, the fluctuating operating condition corresponds to a second duration threshold, and the strong disturbance operating condition corresponds to a third duration threshold. The hierarchical progressive comparison logic is then used to match the comprehensive operating condition judgment index, the switching threshold, and the duration threshold level by level to complete the determination of the target standard operating condition of the transformer area, including: If the comprehensive operating condition judgment index value is not greater than the first switching threshold and the holding time reaches the first duration threshold, then the target standard operating condition of the transformer area is determined to be steady-state operation. If the comprehensive operating condition judgment index value is greater than the first switching threshold and not greater than the second switching threshold, and the duration of the switch reaches the second duration threshold, then the target standard operating condition of the transformer area is determined to be fluctuating operation. If the comprehensive operating condition judgment index value is greater than the second switching threshold and the holding time reaches the third duration threshold, then the target standard operating condition of the transformer area is determined to be strong disturbance operation.

4. The method according to claim 1, characterized in that, The step of calling the target prediction model to perform load prediction on the multidimensional load feature vector to obtain the prediction result corresponding to the future load of the transformer area includes: The target prediction model is defined to include a short-term prediction layer, a medium-term prediction layer, and a long-term prediction layer. The short-term prediction layer is matched with a first prediction step size, the medium-term prediction layer is matched with a second prediction step size, and the long-term prediction layer is matched with a third prediction step size. The first prediction step size is smaller than the second prediction step size, and the second prediction step size is smaller than the third prediction step size. The multidimensional load feature vector is input into the short-term prediction layer, and a prediction operation is performed based on the first prediction step size to obtain the short-term load prediction value. The short-term load forecast value is sent as a recursive input parameter to the medium-term forecast layer, and a forecast calculation is performed based on the second forecast step size to obtain the medium-term load forecast value. The medium-term load forecast value is sent to the long-term forecast layer as a recursive input parameter, and a forecast calculation is performed based on the third forecast step size to obtain the long-term load forecast value. Based on the short-term load forecast, the medium-term load forecast, and the long-term load forecast, a complete multi-period load forecast result for a single distribution area is generated and output.

5. The method according to claim 1, characterized in that, The preprocessing of the real-time collected power load data and environmental data of the transformer substations yields a time-series sequence of current load data, including: Raw data of the transformer area is collected at fixed time intervals. The raw data includes power load data corresponding to current, voltage, and load rate, as well as environmental data. The original data is denoised using a filtering method based on statistical features to remove outliers and obtain pre-processed transformer area data. A sliding window mechanism is introduced to segment and organize the data of the transformer area. Data collected at the same time are extracted in an orderly manner according to a fixed window length and a fixed sliding step size and arranged as sequence points to construct the current load data sequence corresponding to the transformer area data.

6. The method according to any one of claims 1 to 5, characterized in that, For the real-time collected power load data and environmental data of the transformer substation, a processing time period is set to perform a single load forecasting operation for the transformer substation according to the time period. In the process of performing load forecasting for the transformer substation based on multiple time periods, the method further includes: The CPU utilization and memory usage of the microcontroller are monitored in real time, and a comprehensive computing power load function is constructed based on the monitoring results. The comprehensive computing power load function is used to measure computing power consumption. The computing power consumption and the prediction error corresponding to the current period are both incorporated into the optimization objective to construct an optimization objective function that takes into account both prediction error and computing power consumption. By solving the optimization objective function, the prediction step size, inference computation intensity, and model call frequency used by the target prediction model in the next cycle are dynamically adjusted. The model call frequency is used to control the startup frequency of the target prediction model, and the inference computation intensity is used to constrain the computational complexity of the prediction layers contained in the target prediction model.

7. The method according to any one of claims 2 to 5, characterized in that, For the real-time collected power load data and environmental data of the transformer substation, a processing time period is set to perform a single load forecasting operation for the transformer substation according to the time period. In the process of performing load forecasting for the transformer substation based on multiple time periods, the method further includes: Determine the prediction error obtained from a single load prediction operation of the transformer area in the current cycle, and retrieve the historical switching threshold and historical duration threshold configured for the preset state machine model in the operating condition identification operation of the previous cycle. Combining the historical switching threshold, the historical duration threshold, and the prediction error corresponding to the current period, a weighted fusion update method is used to generate a new switching threshold and a new duration threshold. The weighted fusion update method takes into account both the continuity of historical thresholds and the correction effect of the current prediction error. The new switching threshold and the new duration threshold are configured into the preset state machine model for use in the next cycle when the operating condition is identified.

8. A transformer area load forecasting system adapted to a microcontroller, characterized in that, The system includes: The data acquisition and preprocessing module is used to preprocess the real-time acquired power load data and environmental data of the transformer area to obtain a time-series current load data sequence. Each sequence point in the current load data sequence stores the power load data and environmental data acquired at the same time. The data acquisition and preprocessing module is also used to extract features corresponding to the time dimension, period dimension, and external environment dimension from the current load data sequence, and construct a multi-dimensional load feature vector that matches the current load data sequence. The state machine condition identification module is used to process the multidimensional load feature vector using a preset state machine model, and determine the target standard condition corresponding to the current operating condition of the transformer area. The target standard condition is a standard condition in one of multiple progressive levels divided according to the stability of load operation. The preset state machine model is used to map the multidimensional load feature vector to the standard condition. The hierarchical progressive prediction module is used to select a target prediction model that is adapted to the target standard working condition from a preset hierarchical progressive prediction model set. The hierarchical progressive prediction model set includes multiple lightweight time series prediction models that correspond one-to-one with the multiple progressive level standard working conditions and whose computational complexity increases. The lightweight time series prediction model is a time series prediction model whose computational and storage requirements are adapted to the operation of the microcontroller. The prediction result output module is used to call the target prediction model to perform load prediction on the multidimensional load feature vector and obtain the prediction result corresponding to the future load of the transformer area.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the transformer load forecasting method for an adaptive microcontroller as described in any one of claims 1-7.

10. An electronic device, characterized in that, The device includes at least one processor, and at least one memory and bus connected to the processor; The processor and the memory communicate with each other via the bus. The processor is used to call program instructions in the memory to execute the transformer load forecasting method for the adaptive microcontroller as described in any one of claims 1-7.

Images