Power distribution area load prediction method and device, electronic equipment and storage medium

Abstract

The application discloses a power distribution area load prediction method and device, electronic equipment and a storage medium, and belongs to the technical field of electric power. The method comprises the following steps: determining model configuration information, and determining at least one candidate model for load prediction of a power distribution area according to the model configuration information; loading each candidate model according to a model parameter file of a preset data structure of each candidate model; respectively performing a hot start on each candidate model based on historical load data of the power distribution area, so that each candidate model satisfies a load prediction execution condition; performing load prediction by using each candidate model to obtain a load prediction result corresponding to each candidate model; and determining a load prediction result of the power distribution area according to the load prediction result of each candidate model. The application improves the accuracy of load prediction of the power distribution area.

Description

Technical Field

[0001] This application belongs to the field of power technology, and in particular relates to a method, apparatus, electronic device and storage medium for predicting load in a distribution substation. Background Technology

[0002] With the development of new power systems, intelligent distribution network areas and edge computing have become development trends. As core devices in the distribution IoT, intelligent converged terminals need to perform extensive data analysis and load forecasting locally. Currently, edge-end load forecasting (such as ARM-based intelligent terminals) often uses a single, fixed model. Typically, the model is trained in the cloud, and the fixed parameters and code are then distributed to the terminal, which is only responsible for periodically collecting data and performing inference calculations. However, due to the single and fixed nature of the model, it cannot adapt to the differentiated electricity consumption characteristics of different distribution areas (such as residential areas, commercial areas, and industrial areas), resulting in poor forecasting performance due to a lack of environmental adaptability. Summary of the Invention

[0003] This application aims to address at least one of the technical problems existing in the related art. To this end, this application proposes a method, apparatus, electronic device, and storage medium for load forecasting of distribution substations, thereby improving the accuracy of load forecasting for distribution substations.

[0004] Firstly, this application provides a method for load forecasting in a distribution radio area, applied to a smart converged terminal in a distribution radio area. The method includes: Determine the model configuration information, and determine at least one candidate model for load forecasting of the distribution area based on the model configuration information; Load each candidate model according to the model parameter file of the preset data structure of each candidate model; Based on the historical load data of the distribution area, each candidate model is warm-started to ensure that each candidate model meets the load forecasting execution conditions. Load forecasting is performed using each of the candidate models, and the load forecasting results corresponding to each candidate model are obtained; Based on the load forecast results of each candidate model, the load forecast result of the distribution radio area is determined.

[0005] In the above technical solution, at least one candidate model is determined through model configuration information, multiple models are safely loaded through parameter files with preset data structures, and after the model is warm-started by combining historical load data of the distribution area, multiple prediction data are generated based on multi-model inference, resulting in the final load prediction result of the distribution area. Compared with the traditional single fixed model architecture, the backup of multiple candidate models can adapt to the differentiated electricity consumption characteristics of different distribution areas. In the case of multiple candidate models, the prediction result is obtained by comparing and evaluating multiple models and fusing the results, which can effectively reduce the error of a single model and improve the accuracy of load prediction of the distribution area.

[0006] Secondly, this application provides a distribution area load forecasting device, the device comprising: A determination module is used to determine model configuration information and, based on the model configuration information, determine at least one candidate model for load forecasting of the distribution area; The loading module is used to load each candidate model according to the model parameter file of the preset data structure of each candidate model; The startup module is used to perform a hot start on each of the candidate models based on the historical load data of the distribution area, so that each of the candidate models meets the load forecasting execution conditions. The prediction module is used to perform load prediction using each of the candidate models to obtain the load prediction results corresponding to each candidate model. The prediction module is further configured to determine the load prediction result of the distribution radio area based on the load prediction results of each candidate model.

[0007] Thirdly, this application provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the distribution area load forecasting method as described in the first aspect above.

[0008] Fourthly, this application provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the distribution area load forecasting method as described in the first aspect above.

[0009] Fifthly, this application provides a chip including a processor and a communication interface, the communication interface being coupled to the processor, the processor being used to run programs or instructions to implement the distribution area load forecasting method as described in the first aspect.

[0010] Sixthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the distribution area load forecasting method as described in the first aspect above.

[0011] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0012] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is one of the flowcharts illustrating the distribution area load forecasting method provided in some embodiments of this application; Figure 2 This is a second schematic flowchart of the distribution area load forecasting method provided in some embodiments of this application; Figure 3 This is a schematic diagram of the structure of a distribution area load forecasting device provided in some embodiments of this application; Figure 4 These are schematic diagrams of the structure of electronic devices provided in some embodiments of this application.

[0013] Explanation of reference numerals in the attached figures: 300: Distribution area load forecasting device; 301: Determination module; 302: Loading module; 303: Startup module; 304: Prediction module; 400: Electronic device; 401: Processor; 402: Memory. Detailed Implementation

[0014] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0015] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0016] The distribution area load forecasting method, apparatus, electronic equipment, and storage medium provided in this application will be described in detail below with reference to the accompanying drawings and through specific embodiments and application scenarios.

[0017] Among them, the distribution area load forecasting method can be applied to the terminal, and can be executed by the hardware or software in the terminal.

[0018] The distribution transformer area load forecasting method provided in this application is applied to a distribution transformer area smart converged terminal. The executing entity of this distribution transformer area load forecasting method can be the distribution transformer area smart converged terminal or a functional module or entity within the distribution transformer area smart converged terminal capable of implementing the distribution transformer area load forecasting method. The following description uses the distribution transformer area smart converged terminal as the executing entity as an example to illustrate the distribution transformer area load forecasting method provided in this application.

[0019] A smart distribution transformer combine terminal unit (STU) is an intelligent data acquisition and control terminal installed in a distribution transformer area. It meets the needs of concurrency, storage, data acquisition, localized analysis and decision-making, and collaborative computing. It has functions such as data acquisition, equipment operation status monitoring, and power metering, and supports the development needs of marketing, power distribution, and emerging businesses.

[0020] Figure 1 This is one of the flowcharts illustrating the distribution area load forecasting method provided in some embodiments of this application. For example... Figure 1 As shown, the distribution area load forecasting method includes steps 110, 120, 130, 140 and 150.

[0021] Step 110: Determine the model configuration information and, based on the model configuration information, determine at least one candidate model for load forecasting of the distribution area.

[0022] Model configuration information refers to the data used by the intelligent converged terminal in the distribution area to manage the model. Specifically, it may include model type, network structure, operating parameters, input and output specifications, scheduling strategy, resource thresholds, start and stop rules, etc.

[0023] In some embodiments, the model configuration information is structured configuration information set according to a preset structure. For example, the intelligent converged terminal in the distribution area can receive structured configuration messages sent by the server, parse them, and verify their legality locally to generate model configuration information, then match them with a model library to select suitable candidate models. Determining the configuration method of candidate models based on the configuration information facilitates subsequent model iteration, addition, or replacement, enabling load forecasting models to be selected according to specific scenarios (such as different distribution areas like residential areas, commercial areas, and industrial areas); thus improving the scalability of model management.

[0024] Candidate models refer to models that can be deployed on intelligent converged terminals in distribution areas to perform load forecasting. Specifically, they can be one or more forecasting models specified in the model configuration information, including models with different network complexities, computational loads, and applicable scenarios.

[0025] Step 120: Load each candidate model according to the model parameter file of the preset data structure of each candidate model.

[0026] The model parameter file stores parameters such as model weights, biases, and network hyperparameters, serving as the basis for model loading and inference execution. The default data structure is a pre-defined storage format, which can be binary, structured byte sequences, etc., used to define the storage, parsing, and verification methods of the model parameter file to ensure compatibility across different regions or terminals.

[0027] In some embodiments, the model parameter file is issued by the server, received and verified by the intelligent fusion terminal of the distribution area, and then used; it can also be stored locally on the intelligent fusion terminal of the distribution area, called on demand when loading candidate models, and the corresponding model parameter file can be dynamically updated by training the candidate models.

[0028] Understandably, a pre-defined data structure helps standardize the storage, parsing, and verification of model parameter files, reducing operational failures caused by file format inconsistencies. When multiple candidate models exist, multi-model synchronous loading can be performed based on the model parameter files corresponding to each model, followed by parallel inference and horizontal comparison of multiple models. This allows for the determination of the load forecast results for the distribution area based on the load forecast results of each candidate model.

[0029] Step 130: Based on the historical load data of the distribution substation, perform a hot start for each candidate model to ensure that each candidate model meets the load forecasting execution conditions.

[0030] Model hot start refers to using historical real load data of the transformer area for pre-inference warm-up, quickly converging the model's operating state. This is different from cold start, which initializes from zero. It can eliminate the initial inference delay and enable the model to quickly reach the load forecast execution conditions.

[0031] For example, historical load time-series data stored locally in the intelligent converged terminal of the distribution area can be retrieved, missing values ​​can be filled in, and normalization preprocessing can be performed. Then, the data can be batch-input into the model to perform multiple rounds of pre-inference and complete the model warm-up. Alternatively, recent continuous steady-state load data can be extracted, packaged according to the business standard input format, and subjected to lightweight inference in a single batch to quickly calibrate the model's operating status and meet the business execution requirements.

[0032] Step 140: Perform load forecasting using each candidate model to obtain the load forecasting results corresponding to each candidate model.

[0033] After completing the warm start of each candidate model and ensuring that each candidate model meets the load prediction execution conditions, load prediction is performed using each candidate model. For example, real-time collected load data can be encapsulated in a unified data channel and input into each candidate model in parallel, with multiple independent prediction results output synchronously. Alternatively, a time-sharing scheduling method can be used to schedule each candidate model to complete inference in sequence, reducing the terminal's instantaneous computing power and memory usage, and improving adaptability to low-computing-power scenarios on edge terminals.

[0034] Step 150: Determine the load forecast results for the distribution substation based on the load forecast results of each candidate model.

[0035] The load forecast results include the electricity load values ​​of the smart converged terminals in the distribution area for a preset period of time in the future. For example, the output of the best candidate model can be selected as the final load forecast result for the distribution area by combining the error indicators of each model and the weighted scoring results; alternatively, the forecast values ​​of multiple models can be fused and weighted to comprehensively correct the deviation of a single model and generate the final load forecast result after fusion.

[0036] In some embodiments, each candidate model includes a primary model and a backup model. The primary model is the model that takes priority in undertaking the task of load prediction and inference for the distribution area among multiple candidate models; the backup model is a model that is deployed in advance and can be switched on at any time as a redundant backup in scenarios such as malfunction of the primary model or limited terminal resources.

[0037] In some embodiments, when the candidate models include a primary model and a backup model, the model configuration information includes at least: a primary model identifier for indicating the primary model, a model error evaluation threshold for serving as a trigger condition for model evaluation or primary / backup model switching, a maximum CPU usage value, a model input window length for limiting the amount of time-series data input to the model, and a model update cycle for setting the primary model for retraining or cloud synchronization.

[0038] By configuring key parameters such as the primary model identifier, model error assessment threshold, CPU usage limit, model input window length, and model update cycle, unified constraints and dynamic management can be exercised over various aspects such as primary and backup model scheduling, model accuracy determination, and terminal resource protection. While ensuring the accuracy of load forecasting, it adapts to the limited computing power and storage conditions of the intelligent converged terminal in the distribution area, effectively improving the stability, scenario adaptability, and overall controllability of load forecasting services.

[0039] The distribution substation load forecasting method provided in this application determines at least one candidate model through model configuration information, achieves safe loading of multiple models through parameter files with preset data structures, completes model hot start-up by combining historical load data of the substation, generates multiple forecast data based on multi-model inference, and finally obtains the load forecasting result of the distribution substation. Compared with the traditional single fixed model architecture, the backup of multiple candidate models can adapt to the differentiated electricity consumption characteristics of different substations. In the case of multiple candidate models, the prediction result is obtained by comparing and evaluating multiple models and fusing the results, which can effectively reduce the error of a single model and improve the accuracy of distribution substation load forecasting.

[0040] In some embodiments of this application, based on historical load data of the distribution substation, each candidate model is warm-started to ensure that each candidate model meets the load forecasting execution conditions, including: Obtain historical load data for each candidate model; For each candidate model, the corresponding historical load data is input into the loaded candidate model, and model inference is performed to ensure that the candidate model meets the load forecasting execution conditions.

[0041] To obtain historical load data for each candidate model, the smart converged terminal in the distribution area can locally access the historical load data interface. Based on the preset time length for each candidate model, it can extract the corresponding time-series load data. For missing data, linear interpolation is used to fill in the gaps. Specifically, it uses adjacent valid data points to fit and calculate the missing values, ensuring data continuity. Finally, the data is standardized to the range of the model input values ​​to meet the input requirements. The preset time length can be dynamically set according to the input time series requirements of each candidate model (e.g., 1 hour, 24 hours, 7 days).

[0042] The input requirements for candidate models can be determined by combining model configuration information or model parameter files, including data dimensions, time series length, numerical range, and data format (e.g., the input must be continuous time series load values, without missing values ​​or abnormal fluctuations).

[0043] The candidate model has been loaded, indicating that the model has been instantiated in the memory of the smart converged terminal in the distribution area and can perform inference operations normally. Inputting preprocessed historical load data into the model for inference can be achieved by simultaneously distributing the preprocessed historical load data to all loaded candidate models. Each model independently performs pre-inference operations, calibrating model parameters through multiple rounds of inference. Once the model output accuracy fluctuation is within a preset threshold and the inference latency stabilizes, the model is deemed to have met the conditions for load forecasting execution. For example, the conditions for load forecasting execution can be that the candidate model can stably and accurately execute formal load forecasting, which can be reflected in model parameter convergence, inference latency within a preset threshold, and output accuracy reaching business standards.

[0044] The distribution area load forecasting method provided in this application obtains historical load data of the distribution area corresponding to each candidate model and ensures that the data meets the input standards of the candidate models through missing value filling preprocessing; performs pre-operation on each loaded candidate model to complete the model state calibration so that all candidate models meet the execution conditions for formal load forecasting; the influence of missing data can be eliminated through data preprocessing and the model state can be calibrated through pre-inference, effectively avoiding various problems caused by cold start and ensuring the continuous and reliable operation of load forecasting services.

[0045] In some embodiments of this application, historical load data is stored in the intelligent converged terminal of the distribution area using a variable time granularity strategy; the variable time granularity strategy includes: The sampling frequency of storing historical load data is adjusted in stages based on the time elapsed between the historical load data and the current time. The time elapsed is negatively correlated with the sampling frequency.

[0046] Variable time-granularity storage strategy means dynamically adjusting data storage precision. Specifically, it adjusts the data sampling frequency based on the time elapsed since the load data was generated. Unlike traditional fixed-time-frequency storage, this approach balances data storage precision and storage resource consumption. For example, the terminal can calculate the duration of each load data entry in real time and automatically adjust the storage granularity according to the correspondence between time length and storage granularity (which can be adjusted through model configuration information). Alternatively, storage areas can be divided based on time length, such as defining two ring buffers. The short-term buffer stores the data of the most recent day at a high frequency (e.g., one sampling point every 5 minutes), while the long-term buffer stores the data of the most recent 30 days at a low frequency (e.g., one sampling point every hour) through automatic downsampling (e.g., mean pooling or sliding window).

[0047] Time length refers to the interval between the acquisition and generation of a certain load data or batch of load data and the current moment; while time granularity refers to the storage interval of load data, that is, the time interval between two adjacent load data. In the embodiments of this application, the time length of the load data to date is negatively correlated with the sampling frequency, that is, the longer the time length (the further back the load data is from now), the larger the storage time granularity (the longer the interval); the shorter the time length (the more recent the load data is from now), the smaller the storage time granularity (the shorter the interval).

[0048] The variable time granularity strategy for storing historical load data provided in this application also helps improve the efficiency and timeliness of warm-starting candidate models. Specifically, compared with the conventional storage strategy using fixed time granularity, this embodiment can store load data of the same duration using less storage space.

[0049] As an example, under the same retention of 30 days of historical load data, if the conventional fixed-granularity strategy uses a 5-minute sampling precision, it needs to retain approximately 8640 data points. During model warm-start, a large amount of data needs to be read and parsed, resulting in long startup time. However, the variable resolution storage solution of this application only needs to retain 720 data points, which compresses the amount of historical data storage. Compared with the conventional fixed-granularity strategy, which requires waiting for the data acquisition window to fill (several hours), it solves the problems of excessive storage space or loss of short-term load details and long cold-start time in conventional solutions. While reducing storage resource consumption, it also reduces the computational and time overhead of reading historical load data during model warm-start.

[0050] The distribution area load forecasting method provided in this application adopts a variable time granularity strategy, which makes the length of time of load data to date negatively correlated with the sampling frequency of storing historical load data. This avoids the problem of excessive terminal storage space occupied by long-term data due to fine-grained storage alone, and also avoids the problem of insufficient accuracy of recent data due to coarse-grained storage. The variable time granularity strategy can save storage space significantly. Compared with fixed time granularity storage in related technologies, it helps to achieve a balance between the accuracy and cost of historical data storage, ensuring the characteristics of long short-term memory while saving a lot of storage space.

[0051] In some embodiments of this application, a short-time buffer and a long-time buffer are preset; the short-time buffer is used to store historical load data within a first duration from the current time at a first sampling frequency, and the long-time buffer is used to store historical load data exceeding the first duration from the current time at a second sampling frequency; the first sampling frequency is higher than the second sampling frequency; The step of obtaining the historical load data corresponding to each of the candidate models includes: For each candidate model, historical load data corresponding to the time window length is obtained from the short-time buffer and / or the long-time buffer, based on the time window length corresponding to the candidate model. In cases where the time window length exceeds the first duration, historical load data exceeding the first duration is obtained from the long-term buffer, and missing value imputation is performed to make the historical load data meet the input requirements of the candidate model.

[0052] The first duration is the time boundary threshold for dividing high- and low-frequency storage, which can be set according to the actual application scenario. The time window length corresponding to the candidate model refers to the total duration of continuous historical data required for the candidate model to meet the load forecasting execution conditions during a warm start. In this embodiment, if the time window length is greater than the first duration, data from both the short-term buffer and the long-term buffer storage areas need to be retrieved simultaneously.

[0053] In some embodiments, both the short-term and long-term buffers are circular time-series caches, with fixed storage space pre-allocated in the intelligent converged terminal of the distribution area. Typically, only the latest time period data is retained, and older data is automatically overwritten in a cyclical manner to avoid unlimited storage space occupation.

[0054] Imputation of missing values ​​helps address the potential incompleteness of historical load data, preventing hot start failures and abnormal model states caused by missing data input, thus ensuring a smooth hot start process. Preprocessed historical data matches the input requirements of candidate models, eliminating the need for additional data adaptation and reducing computing power consumption at edge terminals, thereby improving hot start efficiency. Specific methods for imputation include cubic spline interpolation upsampling imputation.

[0055] Understandably, when the candidate model is a time series model (such as Long Short-Term Memory (LSTM) and Temporal Convolutional Network (TCN)), the time series model's requirement for the temporal resolution of the input data exhibits a bimodal distribution. Specifically, the model needs high-frequency sampling to capture short-term fluctuations (such as minute-level load changes and start-up / shutdown shocks of industrial equipment) in order to avoid key inflection points being smoothed out by low-pass filtering; the model needs long-term spans to model long-term trends (such as daily cycles, weekly cycles, and seasonal electricity consumption patterns), and high-frequency storage of long-term spans will exponentially consume storage resources.

[0056] If a single sampling rate is used alone, it is difficult to meet the model's requirements for the differences in long and short resolutions of historical load data, and to solve the problem that a single sampling rate cannot take into account both long-term electricity consumption patterns and short-term load fluctuations. However, the embodiments of this application use a hierarchical storage architecture with short-term and long-term buffers, which helps to solve the problem that the storage capacity of all high-frequency historical load data is insufficient, while the storage of all low-frequency data is prone to losing short-term features, thus achieving a balance between storage cost and prediction accuracy.

[0057] In some embodiments of this application, the input data for performing load forecasting using each of the candidate models includes three-phase load data of the distribution substation; In this process, the load data of each phase in the three-phase load data is processed and used as an independent channel input for the candidate model. The imbalance between the load data of each phase is calculated and used as a new channel input for the candidate model, and is input into the candidate model in conjunction with the load data of each phase.

[0058] Three-phase load data refers to the electrical load data corresponding to the three-phase power supply in the distribution substation, which may include load-related parameters such as voltage, current, and power of each phase.

[0059] In some embodiments, the intelligent fusion terminal of the distribution substation collects load data (voltage, current, and active power) of the three phases on the low-voltage side of the distribution substation in real time through a built-in acquisition module; and performs standardized preprocessing operations on the collected raw three-phase load data to eliminate data noise and unify the data format, ensuring that the data meets the input requirements of the candidate model; compared with the traditional single model that only uses total load data as input, the independent introduction of three-phase load data can make the model input data more consistent with the actual power supply scenario of the distribution substation.

[0060] For example, the collected three-phase raw load data can be standardized (normalized to the 0-1 range) and noise filtered (such as by using the moving average method to eliminate acquisition errors). After processing, the data of each phase can be input into the model through the three independent input channels preset by the candidate model. Each channel transmits independently and does not interfere with each other. The model extracts the load characteristics of each phase separately.

[0061] Unbalance refers to the degree of difference between the three-phase load data of a distribution substation, and can be used to reflect the deviation of the load values ​​of each phase. In some embodiments, the average value of the load data of each phase is calculated separately, and the unbalance value is obtained by formula (e.g., maximum phase load - minimum phase load) / average phase load). The calculated unbalance value is then standardized and input through an independent channel added to the model, synchronously transmitted with the independent channel of the load data of each phase, and collaboratively input into the candidate model for comprehensive inference. In some embodiments, the unbalance calculation weights can also be adaptively adjusted to obtain the unbalance value based on the load characteristics of different substations.

[0062] It is understandable that three-phase load imbalance is a common problem in distribution transformer areas. This characteristic directly affects the accuracy of load forecasting. Adding an imbalance input channel can enrich the input dimensions of the model, enabling the model to simultaneously capture the recurring fluctuations of single phases and the overall balance of the three phases, thereby improving the accuracy and generalization ability of forecasts and adapting to the load characteristics of different types of distribution transformer areas. This helps to improve the accuracy of load forecasting and solve the problem of large prediction deviations in traditional models.

[0063] The distribution substation load forecasting method provided in this application standardizes the load data of each phase and inputs them into the model through independent channels to avoid interference from phase characteristics. It calculates the three-phase load imbalance and inputs it in conjunction with the data of each phase through a new channel. By enriching the model input dimensions, it can solve the problems of low prediction accuracy and poor scenario adaptability caused by traditional single input features, making the model input more consistent with the actual power supply scenario of the distribution substation. The model can simultaneously capture the repeated fluctuation pattern of single phase and the overall balance state of three phases, thereby effectively reducing the load forecasting error.

[0064] In some embodiments of this application, the candidate models include a primary model and a backup model; the primary model and the backup model are determined through the following steps: Obtain the load forecast results of each candidate model, and calculate the root mean square error, mean absolute error, and mean absolute percentage error for each candidate model by combining the actual load data of the distribution area. The root mean square error, mean absolute error, and mean absolute percentage error are fused and calculated according to the preset weighting rules to obtain the score results of each candidate model. The primary and backup models among the candidate models are determined based on the scoring results.

[0065] Actual load data refers to the real-time electricity load data collected on-site in the distribution substation. It can be used as a benchmark reference for evaluating the prediction accuracy of the model and to quantify the deviation of the prediction results of each model.

[0066] It is understood that the primary model refers to the model that takes priority in undertaking the load prediction and inference task for the distribution area among multiple candidate models; the backup model refers to the model that is deployed in advance as a redundant backup and can be switched on and activated at any time. The primary model and the backup model together constitute the multi-model operation architecture of the intelligent converged terminal for the distribution area in the embodiments of this application. In some embodiments, when only one candidate model exists, that candidate model is used as the primary model.

[0067] For example, in some embodiments, the candidate model with the best overall accuracy is set as the primary model, and the other models with different accuracy and lightweight properties are uniformly listed as backup models. The terminal defaults to scheduling the primary model to complete real-time load prediction. Alternatively, primary and backup models can be pre-configured based on model configuration information. Under normal operating conditions, the terminal will always call the primary model for inference, while the backup model remains mounted.

[0068] Root Mean Squared Error (RMSE) is a quantitative metric used to measure the deviation between model predictions and actual values; Mean Absolute Deviation (MAE) reflects the average absolute deviation between predictions and actual values; Mean Absolute Percentage Error (MAPE) is a relative metric used to measure the error between predictions and actual values, expressed as a percentage of prediction accuracy, with smaller values ​​indicating more accurate model predictions.

[0069] In some embodiments, for each candidate model, the load forecast results output by each candidate model in the same period are collected synchronously, matched with the true value of the actual load of the transformer area at the same timestamp, and the single-time error is calculated point by point. Then, through periodic statistical averaging, the root mean square error, mean absolute error, and mean absolute percentage error corresponding to each candidate model are obtained respectively. Alternatively, the time-series forecast data and historical actual load data of each model can be collected in batches with a daily or weekly statistical period, and then the three types of error indicators are calculated in batches to form the long-term steady-state error data of each model. The use of multi-indicator joint evaluation comprehensively measures the model's predictive ability from multiple dimensions such as absolute deviation, relative deviation, and extreme fluctuations, which can make up for the one-sidedness of evaluation by a single error indicator.

[0070] In some embodiments, fixed weight coefficients are configured for root mean square error, mean absolute error, and mean absolute percentage error according to business needs. The normalized error indicators are multiplied by their corresponding weights and summed to obtain the model's comprehensive score. The smaller the error, the higher the final score. Alternatively, a dynamic weighting rule can be adopted, which adaptively adjusts the weights of each error indicator according to the load fluctuation status of the transformer area, seasonal periods, and operating conditions. For example, the weight of root mean square error is increased when the load fluctuation is large, and the weights of mean absolute error and mean absolute percentage error are increased when the operation is steady, so as to achieve differentiated and accurate evaluation.

[0071] The following is an example of a candidate model's score obtained through fusion calculation: Score=0.4 RMSE +0.3 MAE+0.3 MAPE Wherein, Score is the score of the candidate model; RMSE is the root mean square error of the candidate model; MAE is the mean absolute error of the candidate model; and MAPE is the mean absolute percentage error of the candidate model.

[0072] In some embodiments, candidate models are sorted by score from highest to lowest, and the candidate model with the highest score is selected as the primary model, while the remaining models with lower ranking, lower computational cost, and satisfactory accuracy are selected as backup models. Alternatively, a minimum passing score threshold can be set to retain only qualified models with satisfactory scores, and the primary model can be selected from among the qualified models.

[0073] The distribution area load forecasting method provided in this application calculates three types of error indicators simultaneously: root mean square error, mean absolute error, and mean absolute percentage error, based on the forecast results synchronously output by each candidate model and the true value of the actual load in the distribution area. It then uses preset weighting rules to perform multi-indicator fusion and quantitative scoring, automatically classifying the primary and backup models based on the scoring results. By combining the comprehensive evaluation of multiple indicators such as absolute error, relative error, and fluctuation error, the method can fully reflect the model's prediction accuracy, stability, and generalization ability, making the model selection results more closely aligned with actual field needs.

[0074] In some embodiments of this application, when load forecasting is performed using a primary model, the method further includes: If the load rate of the intelligent converged terminal in the distribution area exceeds the overload limit and / or the operating temperature exceeds the overheat limit, a sleep delay will be inserted during the load prediction process. If the memory usage of the intelligent converged terminal in the distribution area exceeds the usage limit and / or the load prediction time exceeds the time limit, a target candidate model with a lower load prediction calculation amount than the primary model will be selected from the backup models as the new primary model.

[0075] Overload limits are pre-configured thresholds for the utilization rate of the central processing unit (CPU) or computing core of the intelligent converged terminal in the distribution area. Overheat limits refer to the maximum safe operating temperature allowed by the hardware of the intelligent converged terminal in the distribution area. During the load forecasting process using the primary model, the load rate and hardware operating temperature of the CPU of the intelligent converged terminal in the distribution area are collected and compared with the corresponding limits. If either indicator exceeds the limit, a single forecast inference is split into multiple execution segments by inserting fixed or dynamic sleep delays, which can lengthen the overall inference cycle and reduce instantaneous computing power consumption.

[0076] The memory usage limit is the safe threshold for the operating memory of the intelligent converged terminal in the distribution area. The prediction time limit is the maximum allowable time for a single load prediction inference.

[0077] During the load forecasting process using the primary model, the terminal memory usage and single inference time are monitored in real time. When the corresponding preset limit is exceeded, the parameter scale and inference time of all backup models are automatically compared, and the backup model with the least computational cost is selected to complete the hot model switch and replace the original primary model to continue to perform forecasting. Alternatively, all backup models can be marked with computational cost levels in advance and sorted from high to low. When resource or time consumption exceeds the limit, the model complexity is reduced level by level, and the low-consumption backup model is switched in turn to achieve step-by-step degradation.

[0078] In some embodiments, after selecting a target candidate model with lower load forecasting computational cost than the primary model from the backup models as the new primary model, the method further includes: replacing the primary model with the target candidate model through memory atomic pointers, so that the input / output interface of the load forecasting results remains unchanged after the target candidate model becomes the primary model.

[0079] Understandably, unlike conventional model switching, this embodiment achieves seamless switching between primary and backup models through a memory atomic pointer replacement mechanism. This maintains the model's prediction input / output interfaces and the format of the prediction results unchanged throughout the process, ensuring that downstream power services are unaware of the model switch and guaranteeing uninterrupted load forecasting. Interface specifications remain consistent, enabling flexible degradation in scenarios with limited computing power.

[0080] Inserting a hibernation delay does not require shutting down forecasting services or downgrading and switching models, and can achieve flexible management of resource load without affecting the forecasting output.

[0081] Understandably, the inference and incremental training of the model for load forecasting in distribution substations are both computationally intensive, involving numerous matrix operations and gradient calculations. This can easily lead to continuous high loads on the intelligent fusion terminals in the substations during operation. Meanwhile, power protection services (such as fault detection, overcurrent protection, and switch control) have the highest scheduling priority. Prolonged load forecasting by the model can easily preempt CPU resources, causing a series of stability issues such as power protection response timeouts, watchdog timer malfunctions triggering system resets, and CPU overheating leading to hardware frequency reduction. Existing scheduling methods, such as overall process start-stop and global frequency reduction, are task-level scheduling schemes, which generally suffer from excessively large scheduling granularity. These methods cannot maintain continuous operation of the model for load forecasting in distribution substations while ensuring the strong real-time requirements of power services.

[0082] In this application, scheduling is performed using operator-level delay insertion. Taking a Long Short-Term Memory (LSTM) network as the primary model, some embodiments insert sleep delays during load prediction. This includes embedding sleep delays into nodes that have completed gating calculations at each time step of the LSTM but have not yet updated their cell states. The sleep duration can be adaptively adjusted based on the difference between the current CPU utilization and a preset overload threshold; the higher the load, the longer the sleep time. Since the gating parameters have been calculated at this node, pausing the operation will not destroy the cell state or hidden state of the LSTM, allowing for safe temporary pauses. Using a single time step as the smallest unit of control, scheduling accuracy can reach sub-millisecond levels.

[0083] In this embodiment, the single sleep latency can be set from 10 milliseconds (ms) to 50 ms, which is less than the maximum tolerable latency of 100 ms to 200 ms commonly used in power protection services. This ensures that power tasks can obtain CPU scheduling resources in a timely manner. The intermediate running state of the model is resident in memory, and the recovery operation does not require recalculation of the completed time sequence step, thus taking into account the continuity of load prediction. Adaptive sleep can stably control the CPU load below the overload threshold and avoid the device overheating and triggering hardware protection.

[0084] If operator-level control cannot achieve load reduction under prolonged high load, the method further includes: Training is paused after saving the current training checkpoint; once hardware resources are sufficient, training can resume from where it left off, based on the checkpoint. This hierarchical control model combines fine-grained operator current limiting with resume training from where it left off, balancing the real-time nature of power business operations, the continuity of load forecasting in model execution, and the reliability of terminal hardware operation.

[0085] In some embodiments, after operator-level sleep delay insertion is completed in the load prediction related training process, if the CPU load rate of the intelligent converged terminal in the distribution area is still higher than the preset overload limit, and / or the hardware operating temperature exceeds the set overheat limit, the following degradation processing steps are performed: Store the gradient information, model weights, and LSTM cell state corresponding to the current batch to the terminal flash memory, and generate a training checkpoint. Pause the model training loop process, proactively release CPU computing resources, and ensure the normal scheduling of high-priority real-time services such as power protection. Once the terminal load and temperature return to normal, load all training states based on the stored checkpoints and continue training.

[0086] The distribution area load forecasting method provided in this application embodiment achieves flexible load reduction during the inference process by inserting a sleep delay when the load rate or temperature of the distribution area's intelligent converged terminal is abnormal; it automatically switches to a low-computational-load backup model to complete degraded operation when the memory of the distribution area's intelligent converged terminal is insufficient or the forecast times out; it realizes dynamic matching between terminal hardware status, resource consumption and load forecasting services, which can effectively prevent the model from preempting the core business resources of the distribution area's intelligent converged terminal during load forecasting and causing system crashes, and can effectively improve the long-term stability of edge terminal operation and equipment lifespan.

[0087] In some embodiments of this application, the method further includes: The primary model is trained when the training trigger conditions set according to the preset time and / or preset data volume are met. Update the model parameter file based on the model parameters obtained from training, and load the trained main model based on the updated model parameter file; During the training process of the primary model: According to the preset iteration data volume rules, the model parameters during the training process are stored in the intelligent fusion terminal of the distribution area to generate checkpoint files. The checkpoint files are used to resume training after a break in the training process. If the computing power of the intelligent fusion terminal in the distribution area does not meet the training requirements, the load data of the distribution area will be uploaded to the server in a preset format for collaborative training, so as to load the trained main model according to the model parameter file issued by the server. If the load rate of the intelligent converged terminal in the distribution area exceeds the overload limit and / or the operating temperature exceeds the overheat limit, a sleep delay will be inserted during the training process.

[0088] Training trigger conditions are used to determine whether iterative training of the primary model is needed, enabling the model to update itself periodically. In this embodiment, the training trigger conditions are start conditions set individually or in combination based on two types of rules: fixed time period and cumulative data collection volume.

[0089] For example, training the primary model can be triggered by a time period, automatically triggering training tasks on a fixed daily or weekly basis; and can also be triggered by data volume, actively initiating iterative training of the model when the cumulative new load data of the distribution substation reaches a set threshold, or starting training as soon as either condition is met; alternatively, it can be based on a combination of conditions, starting primary model training only when a specified time period is reached and the cumulative incremental data reaches the target. Compared to the limitations of traditional models that are trained and fixed once by a cloud server, the embodiments of this application can achieve autonomous iterative updates on the edge side, enabling the primary model to continuously adapt to the time-based or seasonal changes in electricity consumption in the distribution substation; and the training tasks can be started on demand, effectively avoiding meaningless high-frequency training that occupies terminal resources, balancing the model optimization effect with the computing power consumption of edge devices.

[0090] In some embodiments, after local training is completed, the original model parameter file is overwritten to complete the parameter version update; and after the update is completed, the model loading process is automatically restarted, the new parameters are hot-loaded, and the optimized primary model is seamlessly switched.

[0091] The preset iteration data volume rule is a pre-defined iteration interval rule during training, used to limit when checkpoint files are generated. For example, it could be to automatically save parameters and generate a checkpoint file after each specified batch of data training is completed. The checkpoint file is a snapshot file of model parameters saved periodically during training, used to record the current training iteration progress and model state. Setting a fixed batch iteration interval, such as automatically saving the current parameters and iteration progress as a checkpoint file after each fixed number of training samples, allows for automatic retrieval of the latest checkpoint and continuation of the remaining training process after an abnormal interruption.

[0092] Understandably, generating checkpoint files helps solve training failures caused by power outages, program crashes, and network fluctuations on edge devices, eliminating the need for repeated full training and saving computing power and time costs. By using checkpoint files to periodically solidify the model training state, it is possible to effectively prevent the loss of all training results due to a single training failure and improve the stability of long-term training on the edge.

[0093] In some embodiments, the terminal monitors computing resource usage in real time. When the local computing power is insufficient to support complex model training, the cleaned distribution area load data is packaged and uploaded in a unified format. After the server completes the training, the terminal receives the new parameter file and loads it locally. By constructing a cloud-edge collaborative training architecture, the computing power shortage of the intelligent converged terminal in the distribution area is compensated for, taking into account both the model iteration requirements and the hardware limitations of edge devices.

[0094] In some embodiments, during the training of the primary model, the load rate and hardware operating temperature of the central processing unit (CPU) of the intelligent converged terminal in the distribution area are collected and compared with corresponding limits. If any indicator exceeds the limit, a single prediction inference is split into multiple execution segments by inserting a fixed or dynamic sleep delay, which can lengthen the overall inference cycle and instantly reduce computing power consumption. Inserting a sleep delay does not require shutting down the prediction service or downgrading and switching models, and can achieve flexible resource load management without affecting the prediction output results.

[0095] The distribution area load forecasting method provided in this application starts training through dual trigger conditions of time and data volume, and updates model parameters by local iteration or server collaborative training. For terminal training scenarios, a checkpoint breakpoint resume training mechanism and a load temperature control sleep strategy are set up to form a model autonomous iterative optimization mechanism adapted to the intelligent fusion terminal of the distribution area. Compared with the drawback of the fixed and unchanging traditional edge model, through periodic training on the terminal side and cloud collaborative optimization, the main model can continuously learn the load change pattern of the distribution area, effectively adapt to complex scenarios such as load fluctuation and seasonal changes, and improve the accuracy of long-term load forecasting.

[0096] In some embodiments of this application, the preset data structure is a binary data structure; Based on the model parameter file with the preset data structure for each candidate model, load each candidate model, including: For each candidate model, the verification fields in the model parameter file are determined according to the field order of the binary data structure, and the model parameter file is verified against tampering based on the verification fields. After the anti-tampering verification is passed, the model parameter file is parsed according to the field order to obtain the model parameters of the candidate model; Based on the model parameters, and combined with the storage and computing power information of the intelligent converged terminal in the distribution area, the candidate models are subject to admission verification. After the admission verification is passed, the candidate model is loaded according to the model parameter file.

[0097] In some embodiments, the model parameter file is encapsulated and stored as a binary byte stream. For example, the included model weights, configuration parameters, verification information, etc., can be uniformly encoded into a binary byte array, encapsulated and packaged according to a fixed field format to generate a binary parameter file. It is understood that the binary format is small in size and highly compressible, which can save the limited local storage resources of the intelligent converged terminal in the distribution area. Moreover, binary byte read and write efficiency is high, which is suitable for the low computing power and low read and write speed of edge terminals, and helps to speed up the model preloading speed.

[0098] The verification field is a predefined checksum located at a specified field position in the binary model parameter file. It can be a hash value, checksum, checksum, or magic number, etc. For example, according to the preset binary field layout rules, the verification field at a specified offset position in the file can be located, the overall checksum of the current model file can be calculated in real time, and compared with the built-in verification field. If they match, the verification passes. Alternatively, hash digest verification can be used. The built-in hash value is used as the verification field when the model is published. Before the terminal loads the file, a hash operation is performed on the complete file, the digest information is compared, and file tampering is detected.

[0099] After the anti-tampering verification passes, data such as floating-point weights, integer configurations, and bias parameters can be read sequentially based on the predefined binary field order and data type. This data is then decoded and restored into a set of model parameters that the terminal can recognize, according to established rules. Parsing according to a fixed field order avoids parameter errors and dimension mismatches caused by out-of-order reading, ensuring accurate parameter parsing. Furthermore, structured binary parsing adapts to standardized model distribution, improving compatibility with multiple versions and types of candidate models.

[0100] Admission verification for candidate models can be performed by estimating the runtime memory usage or inference computation volume based on the number of model parameters, reading the remaining memory and available computing power of the terminal, comparing it with hardware thresholds, and determining whether the terminal can handle the candidate model. This achieves pre-screening of models at the hardware adaptation level, preventing the loading of models with excessive computational or memory requirements, and thus effectively avoiding terminal overload.

[0101] In some embodiments, the model parameter file includes: The Magic_Number field is a 4-byte unsigned integer used to identify the file type and perform preliminary verification of file validity. The version number field (Version) is used to record the iteration versions of the model parameter file; The Model Type field (Model_Type) is used to distinguish different network structures using encoding, and can include model types such as Multilayer Perceptron (MLP), Temporal Convolutional Network (TCN), and Long Short-Term Memory (LSTM); The input dimension (Input_Dim), hidden layer dimension (Hidden_Dim), and number of network layers (Num_Layers) are each represented by an integer field of corresponding length to solidify the basic structural information of the model, adapting to multi-channel input feature specifications such as three-phase load and imbalance. The weight data (Weight_Data) and bias data (Bias_Data) are stored sequentially in the form of floating-point arrays according to the network layer order, ensuring that parameters are parsed and loaded in an orderly manner. A 32-bit checksum field is set at the end of the file to generate a checksum value for all file content except itself. This enables integrity detection and anti-tampering verification of the model parameter file, thereby ensuring the security, reliability, and stable parsing of the model file during transmission, storage, and local loading.

[0102] The distribution area load forecasting method provided in this application uses a binary data structure to store model parameter files. Before the model is officially loaded, a triple mechanism is executed sequentially: binary file anti-tampering verification, ordered field parsing, and terminal resource access verification. The candidate model is only allowed to load and run when the file is complete and secure, the parameters are parsed normally, and the hardware resources are fully compatible. Through binary encrypted storage and built-in field verification mechanisms, the model parameter files can be effectively prevented from being tampered with or damaged, ensuring the reliable operation of load forecasting services. The unified binary field storage and fixed-order parsing can be adapted to the unified management of multiple types and versions of candidate models, and is suitable for model iteration upgrades and replacements. Compared with reducing the prediction error of a single model, it can effectively improve the accuracy of distribution area load forecasting.

[0103] In some embodiments of this application, model management is implemented in the form of micro-applications in the intelligent converged terminal of the distribution area; The micro-applications corresponding to model management are used for: In response to receiving a message in a preset format, the service type corresponding to the message is determined based on the interface identifier carried in the message; If the business type is model management, determine the corresponding operation type of the message under the model management business based on the operation identifier carried in the message, and execute the model operation corresponding to the operation type. The message in the preset format carries an interface identifier and an operation identifier; the interface identifier is used to indicate the business type corresponding to the message, and the operation identifier is used to indicate the operation type corresponding to the message under the business type.

[0104] A micro application refers to software that runs within a terminal, conforms to edge computing architecture, can be rapidly developed, freely expanded, and meets the needs of power distribution and new services. In other words, the intelligent converged terminal for distribution areas adopts a micro application architecture, encapsulating all model management logic (message parsing, operation scheduling, model control, etc.) into independent micro applications. These micro applications are completely isolated from other terminal services, independently occupy system resources, and can be started, upgraded, or uninstalled separately.

[0105] Pre-formatted messages refer to the standardized data message format agreed upon in advance when the micro-application corresponding to the model management interacts with external entities (such as servers or other micro-applications running on the smart converged terminal in the distribution area). This format can be used to ensure the consistency of message transmission and parsing, and avoid business anomalies caused by format confusion.

[0106] Specifically, the micro-applications corresponding to model management can monitor the message receiving port in real time. When an external message is received, the preset format identifier in the message header is parsed first. After confirming that the message format is valid, the interface identifier in a fixed position in the message is extracted. The interface identifier in the terminal is compared with the business type mapping table to quickly determine the business type to which the message belongs. Once the message service type is determined to be model management, the operation identifier in the message is extracted, the operation identifier is compared with the model operation mapping table, the specific operation type is determined (such as model loading, parameter update, model unloading, version rollback), and the corresponding operation interface built into the micro-application is called to execute the model management action.

[0107] Model management encompasses various model-related tasks, including model loading, model unloading, model startup, model stopping, incremental parameter training, full parameter training, and prediction. In some embodiments, the micro-application corresponding to model management is specifically used to perform at least one of the following: Determine the model configuration information, and based on the model configuration information, determine at least one candidate model for load forecasting of distribution radio areas; Load each candidate model according to the model parameter file with the preset data structure of each candidate model; Based on the historical load data of the distribution substation, each candidate model is warm-started to ensure that each candidate model meets the load forecasting execution conditions. Load forecasting was performed using each candidate model, and the load forecasting results corresponding to each candidate model were obtained. Based on the load forecast results of each candidate model, the load forecast results for the distribution area are determined. The primary model is trained when the training trigger conditions set according to the preset time and / or preset data volume are met. The model parameter file is updated based on the trained model parameters, and the trained primary model is loaded based on the updated model parameter file; wherein, during the training process of the primary model: According to the preset iteration data volume rules, the model parameters during the training process are stored in the intelligent fusion terminal of the distribution area to generate checkpoint files. The checkpoint files are used to resume training after a break in the training process. If the computing power of the intelligent fusion terminal in the distribution area does not meet the training requirements, the load data of the distribution area will be uploaded to the server in a preset format for collaborative training, so as to load the trained main model according to the model parameter file issued by the server. If the load rate of the intelligent converged terminal in the distribution area exceeds the overload limit and / or the operating temperature exceeds the overheat limit, a sleep delay will be inserted during the training process.

[0108] In some embodiments, the preset format is the message format corresponding to the Message Queuing Telemetry Transport (MQTT) protocol; the message includes a header format identifier, an interface identifier, an operation identifier, message content, and a checksum, wherein the interface identifier and the operation identifier are mandatory fields. The interface identifier is used to distinguish the service type, and the operation identifier is used to distinguish the specific operation under the same service type. A specific example is given below: Taking the main station (server) sending control commands to the intelligent converged terminal in the distribution area via the MQTT protocol as an example, different function types are distinguished by Interface Identifier (IID) and Operation Identifier (IOP) encoding. The messages carry business parameters in JavaScript Object Notation (JSON) format, realizing remote management and control of the entire lifecycle of the model. The specifications of various interaction commands are as follows: For model management services, the unified configuration interface identifier (IID) is 0x4001, and specific operation types are distinguished by different operation identifiers (IOP). Specifically, operation identifier IOP=0x01 is a model loading instruction, carrying key information such as the model number, file download address, and verification value, used to instruct the micro-application to pull and load the target model (the corresponding model operation is model loading); operation identifier IOP=0x02 is a model unloading instruction, used to instruct the micro-application to unload a specified candidate model based on the model number (the corresponding model operation is model unloading); and operation identifier IOP=0x03 is a parameter configuration instruction, used to instruct the micro-application to configure model input length, evaluation threshold, etc. The system includes operational parameters (corresponding model operations are parameter configuration); operation identifiers IOP=0x04 and IOP=0x05 correspond to model start and model stop commands, respectively, enabling remote start and stop control of the prediction task (corresponding model operations are model start and model stop); operation identifier IOP=0x06 is an incremental training command, used to instruct the micro-application to configure hyperparameters such as iteration rounds and learning rate for minor iterative optimization of the main model (corresponding model operation is parameter incremental training); operation identifier IOP=0x07 is a full training command, used to instruct the micro-application to perform large-scale, multi-round complete training (corresponding model operation is parameter full training). Additionally, the load forecasting service adopts a two-way message interaction mode. Operation identifier IOP=0x08 is the prediction execution command. The main station sends parameters such as the prediction step size to trigger terminal inference calculations. After the terminal completes the prediction, it sends back multi-dimensional load forecast results via uplink messages (corresponding model operation is execution prediction), achieving a closed-loop interaction of command triggering, edge inference, and result reporting. Through unified overall interaction specifications and command division, it can support remote control of multiple models and collaborative operation of services.

[0109] The distribution area load forecasting method provided in this application adopts a micro-application architecture to encapsulate model management functions. The model management micro-application listens to messages in a preset format, quickly locates model management services based on interface identifiers, determines specific model operations through operation identifiers, and finally executes the corresponding model operations. Through the collaborative design of preset format messages, interface identifiers, and operation identifiers, standardized scheduling of model operations is achieved, enabling automated execution of model management operations and improving the efficiency and reliability of model management in the intelligent converged terminal of the distribution area. The modular design of the micro-application and the configurability of the identifier encoding facilitate the addition of model management functions or expansion of service types in the future, and are applicable to intelligent converged terminals of different models and configurations. It eliminates the need for recompilation and re-release for each model upgrade or replacement, reducing maintenance costs.

[0110] Figure 2 This is a second schematic flowchart of a distribution area load forecasting method provided in some embodiments of this application. For example... Figure 2 As shown, the distribution area load forecasting method also includes: The intelligent converged terminal in the distribution area powers on, starts up, completes system initialization, and enters the model self-management process; The intelligent converged terminal in the distribution area determines the model configuration information and obtains the operating parameters, including the operating mode (Run_Mode), the primary model identifier, and the list of backup models. When the running mode is single-model independent running mode (Run_Mode = single): Load the primary model corresponding to the primary model identifier; input the preprocessed multi-channel input data into the primary model and perform independent inference calculations; directly output the load forecast results of the primary model and complete the data reporting; When the running mode is multi-model parallel running mode (Run_Mode = parallel): The system synchronously loads all candidate models in the primary and backup model lists; each loaded model shares the same set of input data and performs load forecasting inference synchronously or in a time-sharing manner; based on the forecast results of each model and the actual load data, it synchronously calculates the accuracy scores of MAE, RMSE, and MAPE to complete the model scoring; it automatically selects the optimal model based on the scoring results, outputs the load forecast results and completes data reporting, and provides a basis for model self-training and primary / backup switching.

[0111] The distribution area load forecasting method provided in this application can be executed by a distribution area load forecasting device. This application uses the distribution area load forecasting device executing the distribution area load forecasting method as an example to illustrate the distribution area load forecasting device provided in this application.

[0112] Figure 3This is a schematic diagram of the structure of a distribution substation load forecasting device provided in some embodiments of this application. For example... Figure 3 As shown, the distribution area load forecasting device 300 includes: The determination module 301 is used to determine model configuration information and determine at least one candidate model for load forecasting of the distribution area based on the model configuration information. Loading module 302 is used to load each candidate model according to the model parameter file of the preset data structure of each candidate model; The startup module 303 is used to perform a hot start on each candidate model based on the historical load data of the distribution substation, so that each candidate model meets the load forecasting execution conditions. Prediction module 304 is used to perform load prediction using each candidate model and obtain the load prediction results corresponding to each candidate model; The prediction module 304 is also used to determine the load prediction result of the distribution substation based on the load prediction results of each candidate model.

[0113] In some embodiments, the startup module 303 is used for: Obtain historical load data for each candidate model; For each candidate model, the corresponding historical load data is input into the loaded candidate model, and model inference is performed to ensure that the candidate model meets the load forecasting execution conditions.

[0114] In some embodiments, historical load data is stored in the smart converged terminal of the distribution area using a variable time granularity strategy; the variable time granularity strategy includes: The sampling frequency of storing historical load data is adjusted in stages based on the time elapsed between the historical load data and the current time. The time elapsed is negatively correlated with the sampling frequency.

[0115] In some embodiments, a short-time buffer and a long-time buffer are preset; the short-time buffer is used to store historical load data within a first time period from the current time at a first sampling frequency, and the long-time buffer is used to store historical load data exceeding the first time period from the current time at a second sampling frequency; the first sampling frequency is higher than the second sampling frequency; Obtain historical load data for each candidate model, including: For each candidate model, historical load data corresponding to the time window length is obtained from the short-time buffer and / or long-time buffer, based on the time window length corresponding to the candidate model. In cases where the time window length exceeds the first duration, historical load data exceeding the first duration is retrieved from the long-term buffer, and missing value imputation is performed to make the historical load data meet the input requirements of the candidate model.

[0116] In some embodiments, the input data for performing load forecasting using each candidate model includes three-phase load data of the distribution substation; In this process, the load data of each phase in the three-phase load data is processed and used as an independent channel input for the candidate model. The imbalance between the load data of each phase is calculated and used as a new channel input for the candidate model, and is input into the candidate model in conjunction with the load data of each phase.

[0117] In some embodiments, the prediction module 304 is used for: If the load rate of the intelligent converged terminal in the distribution area exceeds the overload limit and / or the operating temperature exceeds the overheat limit, a sleep delay will be inserted during the load prediction process. If the memory usage of the intelligent converged terminal in the distribution area exceeds the usage limit and / or the load prediction time exceeds the time limit, a target candidate model with a lower load prediction calculation amount than the primary model will be selected from the backup models as the new primary model.

[0118] In some embodiments, the candidate models include a primary model and a backup model; Obtain the load forecast results of each candidate model, and calculate the root mean square error, mean absolute error, and mean absolute percentage error for each candidate model by combining the actual load data of the distribution area. The root mean square error, mean absolute error, and mean absolute percentage error are fused and calculated according to the preset weighting rules to obtain the score results of each candidate model. The primary and backup models among the candidate models are determined based on the scoring results.

[0119] In some embodiments, the distribution area load forecasting device 300 further includes a training module, which is used for: The primary model is trained when the training trigger conditions set according to the preset time and / or preset data volume are met. Update the model parameter file based on the model parameters obtained from training, and load the trained main model based on the updated model parameter file; During the training process of the primary model: According to the preset iteration data volume rules, the model parameters during the training process are stored in the intelligent fusion terminal of the distribution area to generate checkpoint files. The checkpoint files are used to resume training after a break in the training process. If the computing power of the intelligent fusion terminal in the distribution area does not meet the training requirements, the load data of the distribution area will be uploaded to the server in a preset format for collaborative training, so as to load the trained main model according to the model parameter file issued by the server. If the load rate of the intelligent converged terminal in the distribution area exceeds the overload limit and / or the operating temperature exceeds the overheat limit, a sleep delay will be inserted during the training process.

[0120] In some embodiments, the preset data structure is a binary data structure; The loading module 302 is used to: determine the verification fields in the model parameter file according to the field order of the binary data structure for each candidate model, and perform anti-tampering verification on the model parameter file based on the verification fields; After the anti-tampering verification is passed, the model parameter file is parsed according to the field order to obtain the model parameters of the candidate model; Based on the model parameters, and combined with the storage and computing power information of the intelligent converged terminal in the distribution area, the candidate models are subject to admission verification. After the admission verification is passed, the candidate model is loaded according to the model parameter file.

[0121] In some embodiments, model management is implemented in the form of micro-applications in the intelligent converged terminal of the distribution area; The micro-applications corresponding to model management are used for: In response to receiving a message in a preset format, the service type corresponding to the message is determined based on the interface identifier carried in the message; If the business type is model management, determine the corresponding operation type of the message under the model management business based on the operation identifier carried in the message, and execute the model operation corresponding to the operation type. The message in the preset format carries an interface identifier and an operation identifier; the interface identifier is used to indicate the business type corresponding to the message, and the operation identifier is used to indicate the operation type corresponding to the message under the business type.

[0122] The distribution area load forecasting device in this application embodiment can be an electronic device or a component of an electronic device, such as an integrated circuit or a chip. The electronic device can be a terminal or other devices besides a terminal. For example, the electronic device can be a mobile phone, tablet computer, laptop computer, PDA, in-vehicle electronic device, mobile internet device (MID), augmented reality (AR) / virtual reality (VR) device, robot, wearable device, ultra-mobile personal computer (UMPC), netbook, or personal digital assistant (PDA), etc. It can also be a server, network attached storage (NAS), personal computer (PC), television (TV), ATM, or self-service machine, etc. This application embodiment does not specifically limit the specific devices.

[0123] The distribution area load forecasting device in this application embodiment can be a device with an operating system. This operating system can be a Microsoft (Windows) operating system, an Android operating system, an iOS operating system, or other possible operating systems; this application embodiment does not specifically limit it.

[0124] The distribution area load forecasting device provided in this application embodiment can realize all the processes implemented in the above-described distribution area load forecasting method embodiment and achieve the same technical effect. To avoid repetition, it will not be described again here.

[0125] Figure 4 These are schematic diagrams of the structure of an electronic device provided in some embodiments of this application. In some embodiments, such as Figure 4 As shown, this application embodiment also provides an electronic device 400, including a processor 401, a memory 402, and a computer program stored in the memory 402 and executable on the processor 401. When the program is executed by the processor 401, it implements the various processes of the above-described distribution area load forecasting method embodiment and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0126] It should be noted that the electronic devices in the embodiments of this application include the aforementioned mobile electronic devices and non-mobile electronic devices.

[0127] This application also provides a non-transitory computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the various processes of the above-described distribution area load forecasting method embodiments and achieves the same technical effect. To avoid repetition, it will not be described again here.

[0128] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0129] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described distribution area load forecasting method.

[0130] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0131] This application also provides a chip, which includes a processor and a communication interface. The communication interface and the processor are coupled. The processor is used to run programs or instructions to implement the various processes of the above-described distribution area load forecasting method embodiments and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0132] It should be understood that the chip mentioned in the embodiments of this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.

[0133] It should be noted that, in this document, 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 a process, method, article, or apparatus. Without further limitations, 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. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0134] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the related technology, can be embodied in the form of a computer software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of this application.

[0135] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

[0136] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0137] Although embodiments of this application have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for predicting load in a distribution area, characterized in that, A smart converged terminal for distribution radio areas, the method comprising: Determine the model configuration information, and determine at least one candidate model for load forecasting of the distribution area based on the model configuration information; Load each candidate model according to the model parameter file of the preset data structure of each candidate model; Based on the historical load data of the distribution area, each candidate model is warm-started to ensure that each candidate model meets the load forecasting execution conditions. Load forecasting is performed using each of the candidate models, and the load forecasting results corresponding to each candidate model are obtained; Based on the load forecast results of each candidate model, the load forecast result of the distribution radio area is determined.

2. The method according to claim 1, characterized in that, The step of performing a warm start on each candidate model based on the historical load data of the distribution substation, so that each candidate model meets the load forecasting execution conditions, includes: Obtain the historical load data corresponding to each of the candidate models; For each candidate model, the corresponding historical load data is input into the loaded candidate model, and model inference is performed to make the candidate model meet the load prediction execution conditions.

3. The method according to claim 1 or 2, characterized in that, The historical load data is stored in the intelligent converged terminal of the distribution area using a variable time granularity strategy. The variable time granularity strategy includes: The sampling frequency for storing the historical load data is adjusted in stages based on the time elapsed between the historical load data and the current time, wherein the time elapsed is negatively correlated with the sampling frequency.

4. The method according to claim 2, characterized in that, The system includes a short-time buffer and a long-time buffer. The short-time buffer is used to store historical load data within a first time period from the current time at a first sampling frequency, and the long-time buffer is used to store historical load data exceeding the first time period from the current time at a second sampling frequency. The first sampling frequency is higher than the second sampling frequency; The step of obtaining the historical load data corresponding to each of the candidate models includes: For each candidate model, historical load data corresponding to the time window length is obtained from the short-time buffer and / or the long-time buffer, based on the time window length corresponding to the candidate model. In cases where the time window length exceeds the first duration, historical load data exceeding the first duration is obtained from the long-term buffer, and missing value imputation is performed to make the historical load data meet the input requirements of the candidate model.

5. The method according to claim 1, characterized in that, The input data for performing load forecasting using each of the candidate models includes the three-phase load data of the distribution substation. In this process, the load data of each phase in the three-phase load data is used as an independent channel input for the candidate model after data processing. The imbalance between the load data of each phase is calculated and used as a new channel input for the candidate model, and is input into the candidate model in conjunction with the load data of each phase.

6. The method according to claim 1, characterized in that, The candidate models include a primary model and a backup model; the primary model and the backup model are determined through the following steps: Obtain the load prediction results of each candidate model, and calculate the root mean square error, mean absolute error, and mean absolute percentage error corresponding to each candidate model in combination with the actual load data of the distribution area. The root mean square error, mean absolute error, and mean absolute percentage error are fused and calculated according to a preset weighting rule to obtain the score results of each candidate model. The primary and backup models among the candidate models are determined based on the scoring results.

7. The method according to claim 6, characterized in that, When using the primary model to perform load forecasting, the method further includes: If the load rate of the intelligent converged terminal in the distribution area exceeds the overload limit and / or the operating temperature exceeds the overheat limit, a sleep delay is inserted during the load prediction process. If the memory usage of the intelligent converged terminal in the distribution area exceeds the usage limit and / or the load prediction time exceeds the time limit, a target candidate model with a lower load prediction calculation amount than the primary model is selected as the new primary model from the backup models.

8. The method according to claim 6 or 7, characterized in that, The method further includes: The primary model is trained when the training trigger conditions set according to the preset time and / or preset data volume are met. The model parameter file is updated based on the model parameters obtained from training, and the trained main model is loaded based on the updated model parameter file; During the training process of the primary model: According to the preset iterative data volume rules, the model parameters during the training process are stored in the intelligent fusion terminal of the transformer area to generate a checkpoint file. The checkpoint file is used to realize the resumption of training after the interruption during the training process. If the computing power of the intelligent fusion terminal of the distribution area does not meet the training requirements, the load data of the distribution area is uploaded to the server in a preset format for collaborative training, so as to load the trained main model according to the model parameter file issued by the server. If the load rate of the intelligent fusion terminal in the distribution area exceeds the overload limit and / or the operating temperature exceeds the overheat limit, a sleep delay is inserted during the training process.

9. The method according to claim 1, characterized in that, The preset data structure is a binary data structure; The step of loading each candidate model according to the model parameter file with a preset data structure includes: For each candidate model, the verification field in the model parameter file is determined according to the field order of the binary data structure, and the model parameter file is verified against tampering based on the verification field. After the anti-tampering verification is passed, the model parameter file is parsed according to the field order to obtain the model parameters of the candidate model; Based on the model parameters, and combined with the storage and computing power information of the intelligent converged terminal in the distribution area, the candidate models are subject to admission verification. After the admission verification is passed, the candidate model is loaded according to the model parameter file.

10. The method according to claim 1, characterized in that, The intelligent converged terminal for the transformer area implements model management in the form of micro-applications; The micro-application corresponding to the model management is used for: In response to receiving a message in a preset format, the service type corresponding to the message is determined based on the interface identifier carried in the message; If the business type is model management, determine the operation type corresponding to the message under the model management business based on the operation identifier carried by the message, and execute the model operation corresponding to the operation type. The message in the preset format carries an interface identifier and an operation identifier; the interface identifier is used to indicate the service type corresponding to the message, and the operation identifier is used to indicate the operation type corresponding to the message under the service type.

11. A load forecasting device for a distribution area, characterized in that, include: A determination module is used to determine model configuration information and, based on the model configuration information, determine at least one candidate model for load forecasting of the distribution area; The loading module is used to load each candidate model according to the model parameter file of the preset data structure of each candidate model; The startup module is used to perform a hot start on each of the candidate models based on the historical load data of the distribution area, so that each of the candidate models meets the load forecasting execution conditions. The prediction module is used to perform load prediction using each of the candidate models to obtain the load prediction results corresponding to each candidate model. The prediction module is further configured to determine the load prediction result of the distribution radio area based on the load prediction results of each candidate model.

12. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the distribution area load forecasting method as described in any one of claims 1 to 10.

13. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the distribution area load forecasting method as described in any one of claims 1 to 10.

Images