Large-scale distribution transformer load prediction method

Abstract

The invention discloses a large-scale distribution transformer load prediction method, and relates to the technical field of power system load prediction, and the method comprises the following steps: A, carrying out the integrity verification of input historical load data of a distribution transformer, and screening out a complete data set meeting the preset day number requirement; b, grouping the verified historical load data according to distribution transformers and arranging the historical load data according to a time sequence; and C, establishing a TimeFM model, inputting the historical load data into the TimeFM model according to a time sequence, and predicting the load of the distribution transformer. The defects in the prior art can be overcome, and the accuracy of power grid load prediction is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system load forecasting, and particularly relates to a large-scale distribution transformer load forecasting method. BACKGROUND

[0002] With the rapid development of modern society, the stability, safety and efficiency of the power system as the infrastructure supporting the operation of society are particularly important. As a key node in the power grid, the distribution transformer (referred to as "distribution transformer") is responsible for converting high-voltage power into low-voltage power and distributing it to end users. Its operating state is directly related to the quality and reliability of power supply. In the power technology frontier regions such as Europe and the United States, artificial intelligence algorithms are deeply integrated into distribution transformer load forecasting work. Many power research institutions in the United States use long short-term memory networks (LSTM) and its derivative gated recurrent units (GRU) in deep learning to carry out prediction research. Through the LSTM model, the load data for several years is trained. Due to the different use of air conditioners and heating equipment in different seasons, the load changes significantly. The LSTM model can effectively capture these complex time series characteristics and accurately predict the load difference between weekdays and weekends and different seasons, providing accurate basis for power dispatching and equipment maintenance. However, facing a large number of distribution transformer points, since LSTM contains multiple gating mechanisms (input gate, forget gate and output gate), the number of its parameters is much larger than that of ordinary recurrent neural networks (RNN), resulting in increased computational overhead, and its model training will be very slow, only suitable for small-scale distribution transformer load forecasting in some areas.

[0003] In China, machine learning and deep learning techniques have achieved fruitful results in distribution transformer load forecasting. Many universities and research institutions actively carry out related research. For example, a research team from Tsinghua University used support vector machines (SVM) to predict the load of a residential area distribution transformer in Beijing. They conducted in-depth analysis of the historical load data of the area for many years, combined with local meteorological factors (temperature, humidity, rainfall, etc.), through feature selection and data preprocessing, removed noise and redundant information, and extracted feature variables that had a greater impact on load. Using the powerful nonlinear classification ability of SVM, a prediction model was established. Experimental results show that the average absolute error of the model in the load prediction of the residential area is reduced by about 12%, effectively improving the prediction accuracy. However, SVM technology has some problems in practical application. First, SVM is less efficient when dealing with large-scale data. When the number of users in the area where the distribution transformer is located increases significantly and the electricity consumption data grows rapidly, the training time of the model will be significantly extended, making it difficult to meet the demand for real-time prediction. Second, SVM is sensitive to the selection of kernel functions and parameter settings. Different kernel functions and parameter combinations can result in significant differences in model performance. When faced with complex and variable distribution transformer load scenarios, a lot of time is spent on parameter tuning, and it is difficult to find the optimal combination, affecting the stability and generalization ability of the model. In addition, SVM is prone to overfitting when dealing with high-dimensional sparse data. Although it performs well in the specific scenario of a residential area in Beijing, the prediction accuracy will decrease significantly when applied to other areas with different electricity consumption patterns.

[0004] TimesFM, a time series prediction model based on the Transformer structure that has emerged in recent years, has shown unique advantages in handling multivariate and long-period time series data due to its strong ability to capture long sequence dependencies. However, applying it to large-scale distribution transformer load forecasting scenarios still requires overcoming multiple technical bottlenecks: how to adapt the model to the individual load characteristics of a large number of transformers to avoid precision loss caused by "one-size-fits-all" prediction; how to efficiently handle missing data and integrate data completion and prediction without sacrificing computational efficiency; how to optimize the model architecture to reduce resource consumption during large-scale parallel computing and ensure real-time prediction performance even when a large amount of data is incoming - solving these problems is key to promoting the application of TimesFM in intelligent power grid load forecasting. SUMMARY

[0005] The technical problem to be solved by the present application is to provide a large-scale distribution transformer load forecasting method that can overcome the shortcomings of the prior art and improve the accuracy of power grid load forecasting.

[0006] To solve the above technical problems, the technical solutions adopted by the present application are as follows.

[0007] A method for large-scale distribution transformer load forecasting includes the following steps:

[0008] A. Verify the integrity of the input historical load data of the distribution transformer and select the complete dataset that meets the preset number of days requirement;

[0009] B. Group the verified historical load data by distribution transformer and arrange them in chronological order;

[0010] C. Establish a TimesFM model, input historical load data into the TimesFM model according to time sequence, and predict the load of the distribution transformer.

[0011] As a preferred option, in step A, a minimum consecutive days threshold for historical load data is set to filter distribution transformers that meet the requirements and exclude distribution transformer areas with incomplete data.

[0012] As a preferred approach, a nonlinear fit is performed on the selected complete dataset, and the time points corresponding to data points whose distance from the fitted curve exceeds a set threshold are marked.

[0013] As a preferred option, the TimesFM model includes an input layer, a Transformer layer, and an output layer arranged in series.

[0014] Preferably, the input layer includes a cascaded activation function module and a convolution module. The activation function module enhances the features of the fitted curve of the input, X. ’ =ReLU(W·X+θ), where X is the input fitted curve, ReLU() is the activation function, W is the weight matrix, θ is the bias of the weight matrix, and X... ’ The output of the activation function module; the convolution module performs a one-dimensional convolution operation on the input data to increase the data dimensionality, X ’’ =Conv(X ’ ), Conv() is a one-dimensional convolution kernel.

[0015] Preferably, the Transformer layer includes a self-attention mechanism module, a residual connection module, and a feedforward neural network module configured sequentially. The self-attention mechanism module obtains the query vector Q, key vector K, and value vector V through linear transformation and calculates the temporal dependencies of the input data. , As a dimensionality scaling factor, time-series data and time-series dependencies are input into the feedforward neural network module through the residual connection module. The feedforward neural network module enhances the local features of each data point according to the time sequence.

[0016] Preferably, the hidden layer of the feedforward neural network module consists of dynamic hidden layers and static hidden layers. The static hidden layer remains online. When enhancing the local features of the data at the labeled time point, some of the dynamic hidden layers are connected to the static hidden layers. The feature extraction depth for enhancing the data at the labeled time point is increased by increasing the number of hidden layers. The number of dynamic hidden layers increases as the deviation between the data at the labeled time point and the fitted curve increases.

[0017] As a preferred approach, the output layer generates prediction data through normalization processing.

[0018] The beneficial effects of adopting the above technical solution are as follows: This invention improves prediction accuracy by introducing data integrity verification and by improving the TimesFM model. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of a specific embodiment of the present invention. Detailed Implementation

[0020] In the following description of the embodiments, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0021] Reference Figure 1 The overall process of this invention is as follows:

[0022] A. Verify the integrity of the input historical load data of the distribution transformer and select the complete dataset that meets the preset number of days.

[0023] B. Group the verified historical load data by distribution transformer and arrange them in chronological order.

[0024] C. Establish a TimesFM model, input historical load data into the TimesFM model according to time sequence, and predict the load of the distribution transformer.

[0025] In step A, a minimum threshold for the number of consecutive days of historical load data is set to filter distribution transformers that meet the requirements, excluding distribution transformer areas with incomplete data. Specifically, the data includes the device identifier (dev_id), date (data_date), and load values ​​at multiple time points (e.g., 96 time points / day). It is verified whether each device has data within a consecutive 7 (or 10) day period; if a device has incomplete data, that device is excluded. This completely solves the problem of "poor data quality leading to low prediction accuracy."

[0026] In step B, for each device, the load data from multiple consecutive days is concatenated into a long sequence in chronological order (e.g., 960 time points over 10 days). Specifically, non-numeric columns (such as `dev_id`, `data_date`, etc.) are excluded, and the load values ​​from the 96 time points each day are concatenated row-wise to form a one-dimensional long sequence. Then, the long sequence array is divided into multiple batches according to batch size (e.g., 400), with each batch containing long sequences from multiple devices.

[0027] Finally, a nonlinear fit is performed on the selected complete dataset, and the time points corresponding to data points whose distance from the fitted curve exceeds a set threshold are marked. The processed data is then input into the TimesFM model.

[0028] To improve prediction efficiency and accuracy, this invention modifies the structure of the TimesFM model. The TimesFM model includes a cascaded input layer, a Transformer layer, and an output layer. First, the input layer is designed as a cascaded combination of an activation function module and a convolutional module. The activation function module enhances the features of the fitted curve of the input, X... ’ =ReLU(W·X+θ), where X is the input fitted curve, ReLU() is the activation function, W is the weight matrix, θ is the bias of the weight matrix, and X... ’ The output of the activation function module; the convolution module performs a one-dimensional convolution operation on the input data to increase the data dimensionality, X ’’ =Conv(X ’ Conv() is a one-dimensional convolution kernel. The Transformer layer includes a self-attention mechanism module, a residual connection module, and a feedforward neural network module configured sequentially. The self-attention mechanism module obtains the query vector Q, key vector K, and value vector V through linear transformation, and calculates the temporal dependencies of the input data. , As a dimensionality scaling factor, temporal data and temporal dependencies are input into the feedforward neural network module via a residual connection module. The feedforward neural network module enhances the local features of each data point according to the temporal order. We improved upon the traditional feedforward neural network by designing its hidden layers to consist of dynamic and static hidden layers. The static hidden layer remains continuous. When enhancing the local features of the data corresponding to the labeled time point, part of the dynamic hidden layer is connected to the static hidden layer. By increasing the number of hidden layers, the feature extraction depth for enhancing the data corresponding to the labeled time point is increased. The number of dynamic hidden layers increases as the deviation between the data corresponding to the labeled time point and the fitted curve increases. Finally, the output layer undergoes normalization processing to generate predicted data.

[0029] In addition, to address the issue that the TimesFM model cannot effectively identify periods of heavy overload, this invention identifies heavy overload states as follows.

[0030] Load factor = Real-time load (kVA) / Rated capacity (kVA) × 100%. Let: a load factor ≤ 80% be considered normal; a load factor between 80% and 100% be considered full load; a load factor exceeding 100% be considered overload. If a distribution transformer remains under heavy load or overload for two consecutive hours, it is considered to have experienced a severe overload.

[0031] The three phases (a, b, and c) and the total active load are independently assessed to identify the periods when each load exceeds the threshold. Simultaneously, attention is paid to whether the total load exceeds the limit and whether the three-phase load distribution is balanced. The core evaluation focuses on the matching relationship between the actual load of each phase (A, B, and C) and the apparent capacity of a single phase. For example, with the same 100kVA transformer, the total load is 80kVA (total load factor 80%), but phase A load is 60kVA (apparent capacity of approximately 33kVA per phase, load factor 180%), while phases B and C each have 10kVA. This is judged as "phase A overload," improving the practicality of heavy overload prediction.

[0032] Traditional predictive modeling is only suitable for distribution transformers with a single trend. When faced with the integration of other distributed energy sources such as distributed photovoltaics, traditional modeling methods struggle to identify the type of distribution transformer, easily leading to inaccurate predictions. This patented prediction method effectively solves the complex problem of predicting the trends of various distribution transformers.

[0033] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0034] The present invention can implement all or part of the processes in the methods of the above embodiments, or it can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals.

[0035] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for large-scale distribution transformer load forecasting, characterized in that... Includes the following steps: A. Verify the integrity of the input historical load data of the distribution transformer and select the complete dataset that meets the preset number of days requirement; B. Group the verified historical load data by distribution transformer and arrange them in chronological order; C. Establish a TimesFM model, input historical load data into the TimesFM model according to time sequence, and predict the load of the distribution transformer.

2. The method for large-scale distribution transformer load forecasting according to claim 1, characterized in that: In step A, a minimum consecutive days threshold for historical load data is set, and distribution transformers that meet the requirements are screened out, while distribution transformer areas with incomplete data are excluded.

3. The method for large-scale distribution transformer load forecasting according to claim 2, characterized in that: The selected complete dataset is subjected to nonlinear fitting, and the time points corresponding to data points whose distance from the fitted curve exceeds a set threshold are marked.

4. The method for large-scale distribution transformer load forecasting according to claim 3, characterized in that: The TimesFM model consists of a serially configured input layer, Transformer layer, and output layer.

5. The method for large-scale distribution transformer load forecasting according to claim 4, characterized in that: The input layer includes a concatenated activation function module and a convolutional module. The activation function module enhances the features of the fitted curve of the input. ’ =ReLU(W·X+θ), where X is the input fitted curve, ReLU() is the activation function, W is the weight matrix, θ is the bias of the weight matrix, and X... ’ The output of the activation function module; the convolution module performs a one-dimensional convolution operation on the input data to increase the data dimensionality, X ’’ =Conv(X ’ ), Conv() is a one-dimensional convolution kernel.

6. The method for large-scale distribution transformer load forecasting according to claim 5, characterized in that: The Transformer layer includes a sequentially configured self-attention mechanism module, a residual connection module, and a feedforward neural network module. The self-attention mechanism module obtains the query vector Q, key vector K, and value vector V through linear transformations and calculates the temporal dependencies of the input data. , As a dimensionality scaling factor, time-series data and time-series dependencies are input into the feedforward neural network module through the residual connection module. The feedforward neural network module enhances the local features of each data point according to the time sequence.

7. The method for large-scale distribution transformer load forecasting according to claim 6, characterized in that: The hidden layers of the feedforward neural network module consist of dynamic hidden layers and static hidden layers. The static hidden layers remain online. When enhancing the local features of the data at the labeled time point, some dynamic hidden layers are connected to the static hidden layers. The feature extraction depth for enhancing the data at the labeled time point is increased by increasing the number of hidden layers. The number of dynamic hidden layers increases as the deviation between the data at the labeled time point and the fitted curve increases.

8. The method for large-scale distribution transformer load forecasting according to claim 7, characterized in that: The output layer generates prediction data through normalization.

Images