Load forecasting and dynamic capacity control methods for dedicated transformer data acquisition terminals

Abstract

This invention discloses a load forecasting and dynamic capacity control method for a dedicated transformer data acquisition terminal, relating to the field of time series forecasting technology. The method includes: collecting electricity load data from the dedicated transformer data acquisition terminal, including historical and real-time electricity load time series; preprocessing the electricity load data to remove noise; using the historical electricity load time series as training samples to train a time series forecasting model, and saving the optimal model parameters as a pre-trained time series forecasting model; applying the pre-trained time series forecasting model to comprehensively analyze the real-time electricity load time series in both the time and frequency domains, and outputting the long-term electricity load forecasting results; and introducing a dynamic programming algorithm based on the electricity load forecasting results of the pre-trained time series forecasting model to control the capacity of the dedicated transformer data acquisition terminal in real time. This improves forecasting accuracy and efficiency, ensuring the stable operation of the power system.

Description

Technical Field

[0001] This application relates to the field of time series forecasting technology, and in particular to a method for load forecasting and dynamic capacity control of dedicated transformer data acquisition terminals. Background Technology

[0002] A dedicated transformer data acquisition terminal is a device used in power systems to collect, monitor, and control electricity consumption information from dedicated transformer users. Load forecasting by this terminal involves analyzing historical electricity consumption data to predict electricity load for a future period. Accurate load forecasting helps power departments rationally plan power generation, optimize grid operation, and reduce grid losses. It also provides crucial information for the operation of the electricity market, contributing to improved overall efficiency and reliability of the power system.

[0003] By analyzing load forecast data from dedicated transformer data acquisition terminals, the capacity of these transformers can be dynamically adjusted in real time to optimize the allocation of power resources and ensure the safe and stable operation of the power grid. Dynamic capacity control effectively improves the operating efficiency of the power grid and provides electricity users with more flexible electricity consumption options, promoting sustainable development. Combining load forecasting from dedicated transformer data acquisition terminals with dynamic capacity control is a crucial step in ensuring the stable operation of the power system.

[0004] Traditional load forecasting methods using dedicated transformer data acquisition terminals rely on statistical and regression analysis to uncover patterns and trends in historical load data to predict future loads. However, these methods require large amounts of data and have high data quality requirements, leading to low forecast accuracy. Existing time series analysis methods introduce mode decomposition for load forecasting, resulting in high computational complexity and applicability only to short-term forecasts with limited real-time performance. Traditional capacity control methods employ timed and manual control, operating based on pre-set rules and plans. This makes it difficult to make rapid and flexible adjustments and responses, and it cannot dynamically adjust based on real-time grid operating conditions, resulting in low capacity control efficiency. Summary of the Invention

[0005] To address the technical problems of the prior art, this application provides a method for load forecasting and dynamic capacity control of a dedicated transformer data acquisition terminal. A time series forecasting model based on a neural network is introduced to predict the electricity load time series data of the dedicated transformer data acquisition terminal in real time, outputting long-term electricity load data prediction results. Based on the prediction results, a dynamic programming algorithm is applied to determine the optimal capacity control strategy with the goal of minimizing cost. This improves the accuracy and efficiency of load forecasting and capacity control, ensuring the stable operation of the power system.

[0006] This application provides a method for load forecasting and dynamic capacity control of a dedicated transformer data acquisition terminal, including:

[0007] (1) Collect electricity load data from the dedicated transformer acquisition terminal, including historical electricity load time series and real-time electricity load time series;

[0008] (2) Preprocess the power load data of the dedicated transformer acquisition terminal to remove noise interference;

[0009] (3) Use the historical electricity load time series of the special transformer acquisition terminal as training samples to train the time series prediction model, and save the optimal model parameters as the pre-trained time series prediction model.

[0010] (4) Apply the pre-trained time series prediction model to comprehensively analyze the real-time electricity load time series from the time domain and frequency domain, and output the prediction results of long-term electricity load;

[0011] (5) Based on the power load prediction results of the pre-trained time series prediction model, a dynamic programming algorithm is introduced to control the capacity of the dedicated transformer acquisition terminal in real time.

[0012] Furthermore, processing the load forecasting data from the dedicated transformer acquisition terminal into a time series forecasting task involves long-term time series forecasting, which is a series-to-series forecasting problem. Let the electricity load data to be processed be... It is a length of The lookback window is a collection of elements of time variables, each representing a single point in time. Recording electricity load values, the time series forecasting task predicts the future step size as follows: time series .

[0013] Furthermore, the steps for preprocessing the power load data from the dedicated transformer acquisition terminal include:

[0014] Check for duplicate data points in the electricity load data to ensure that there is only a unique electricity load value at each point in time.

[0015] For missing values ​​in the electricity load data, a linear interpolation method is used to fill in the missing electricity load values ​​by calculating the average value of data points in the neighborhood;

[0016] The power load data is decomposed and reconstructed using the maximum overlap discrete wavelet transform to obtain the denoised power load time series.

[0017] Furthermore, the Discrete Wavelet Transform (DWT) is a multi-resolution analysis method that decomposes a time series into components of different frequencies. It decomposes and reconstructs the time series using wavelet basis functions. The Maximum Overlap Discrete Wavelet Transform (MEOT) is an improved version of the DWT. During the decomposition process, the wavelet function and the scaling function are convolved at each scale with maximum overlap, giving the DWT better localization properties and enabling it to better capture the local features of electricity load time series with complex dynamic characteristics while preserving important temporal information.

[0018] Furthermore, the power load data is denoised using the Maximum Overlap Discrete Wavelet Transform (MODWT), and the steps include:

[0019] The electricity load data was decomposed using MODWT to obtain approximation coefficients and detail coefficients at different scales.

[0020] Based on the characteristics of the electricity load data, a suitable threshold function is selected to perform threshold processing on the detail coefficients;

[0021] The power load data is obtained by reconstructing the inverse MODWT using the detail coefficients after thresholding.

[0022] Furthermore, the process of obtaining a pre-trained time series prediction model based on the historical electricity load time series of the dedicated transformer acquisition terminal includes:

[0023] The historical electricity load time series of the dedicated transformer acquisition terminal is used as the training sample and divided into training set, validation set and test set in a ratio of 8:1:1.

[0024] A time series forecasting model is constructed, which uses a time-frequency enhancement decomposition network to capture long-term latent patterns and time periodicity of time series from the time-frequency domain;

[0025] A time series prediction model was trained using a hybrid loss function, with the Adam optimizer used. An initial learning rate was set, and a cosine annealing strategy was employed to dynamically adjust the learning rate during the training process.

[0026] When the model's loss function reaches its minimum on the validation set, the model performance is evaluated using the test set, and the model parameters at this point are saved as a pre-trained time series prediction model.

[0027] Furthermore, the time series prediction model employs a time-frequency enhancement decomposition network, which processes the electricity load time series data from the dedicated transformer acquisition terminal through a multi-scale time-frequency enhancement encoder; the trend time-frequency module in the trend encoder learns the trend pattern of the time series by sharing the check time series, and the seasonal time-frequency module in the seasonal encoder learns the seasonal characteristics of the time series; finally, the future time series is predicted by fusing the feature representations of the trend encoder and the seasonal encoder.

[0028] Furthermore, the time series prediction model employs a hybrid loss mechanism to improve model robustness, combining L1 and L2 loss functions into a hybrid loss function. The expression is:

[0029]

[0030]

[0031] in, and These represent the actual and predicted values ​​of the time series, respectively. Represents the L1 loss function; Represents the L2 loss function; Used to control the weights of the L1 loss function and the L2 loss function. The L1 loss function is used to map the absolute value of the error between the true and predicted values ​​to the range [0,1] using the Tanh function, and serves as the control weight. The L1 loss function is particularly effective when the prediction error is large; the L2 loss function also contributes when the error is small.

[0032] Furthermore, based on the electricity load prediction results of the pre-trained time series prediction model, a dynamic programming algorithm is introduced to control the capacity of the dedicated transformer data acquisition terminal in real time. The steps include:

[0033] The objective of determining the capacity of the dedicated transformer data acquisition terminal is to dynamically adjust the capacity of the dedicated transformer data acquisition terminal according to changes in future electricity load, with the goal of minimizing the total cost.

[0034] A dynamic programming model is established, defining the capacity of the dedicated transformer data acquisition terminal at each future time point as a state variable, the capacity adjustment amount at each future time point as a decision variable, and the objective function as minimizing the total cost, including equipment configuration cost, operation and maintenance cost, and energy consumption cost. Constraints are determined, including capacity constraints, load demand constraints, and capacity adjustment constraints.

[0035] Divide the future time series into multiple stages, each stage corresponding to a point in time; update the state variables based on the decision variables, and establish a dynamic programming recursive formula.

[0036] Starting from the initial time point, the optimal cost and decision are calculated based on the dynamic programming recursive formula. The optimal capacity adjustment strategy for each time point is gradually extracted, and the capacity of the dedicated transformer acquisition terminal is adjusted according to the optimal strategy.

[0037] Based on operational data and feedback, the parameters and constraints of the dynamic programming model are continuously optimized; electricity load forecast data is updated regularly to ensure the accuracy and adaptability of the dynamic programming algorithm.

[0038] The present invention discloses the following technical effects:

[0039] This invention proposes a load forecasting and dynamic capacity control method for dedicated transformer data acquisition terminals. By collecting historical and real-time electricity load time series from the dedicated transformer data acquisition terminals, a time series forecasting model is established to achieve real-time forecasting of electricity load data. Based on the forecast results, a dynamic programming algorithm is introduced to control the capacity of the dedicated transformer data acquisition terminals in real time. Furthermore, during the preprocessing of the electricity load data from the dedicated transformer data acquisition terminals, maximally overlapping discrete wavelet transform is used to denoise the data. For electricity load data with complex dynamic characteristics and containing locally important information, this method can preserve complete time series information and fully capture local features. The time series forecasting model is constructed based on a time-frequency enhancement decomposition network, comprehensively analyzing the time series characteristics of electricity load from the time and frequency domains. Multi-scale trend and seasonal encoders are applied to capture long-term basic patterns and time periodicity from the time and frequency domains in the time series, improving the forecasting accuracy of electricity load data. This invention employs a dynamic programming algorithm, establishing an optimization objective and using the load forecast results as constraints to find the optimal capacity control strategy to achieve the objective. Dynamically adjusting the capacity of the dedicated transformer data acquisition terminals allows for personalized design of capacity control strategies, improving control flexibility. This invention combines load forecasting and capacity control to comprehensively analyze the time series of electricity load of the dedicated transformer data acquisition terminal, ensuring forecast accuracy and effectively promoting the stable operation of the dedicated transformer data acquisition terminal. Attached Figure Description

[0040] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. Flowcharts are used in this application to illustrate the operations performed according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed.

[0041] Figure 1 This is a flowchart illustrating the load prediction and dynamic capacity control method for a dedicated transformer data acquisition terminal provided in this application embodiment.

[0042] Figure 2 This is a schematic diagram of the time series prediction model structure based on the time-frequency enhancement decomposition network provided in the embodiments of this application.

[0043] Figure 3 A schematic diagram of the structure of the seasonal encoder and trend encoder in the time-frequency enhancement decomposition network provided in the application embodiment. Detailed Implementation

[0044] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application.

[0045] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of this application will be provided in conjunction with the accompanying drawings. The described embodiments should not be considered as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0046] In the following description, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those explicitly listed, but may include other steps or modules not explicitly listed or inherent to such processes, methods, products, or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only.

[0047] Example 1: This application provides a method for load forecasting and dynamic capacity control of a dedicated transformer data acquisition terminal, such as... Figure 1 As shown, the method includes:

[0048] Step S10: Collect the power load data from the dedicated transformer acquisition terminal, including historical power load time series and real-time power load time series.

[0049] In this embodiment, the dedicated transformer data acquisition terminal has a local storage function, storing a large amount of historical electricity load data in the terminal's memory. The method for acquiring the data is as follows:

[0050] Connect the dedicated transformer data acquisition terminal through the communication interface and log in to the terminal's management system; find the data storage module in the management system, select the required historical electricity load data time period; transmit the data to the computer device to construct the time series of historical electricity load data.

[0051] Real-time power load data from the dedicated transformer acquisition terminal can be directly obtained from the terminal. By observing the real-time data on the terminal display screen, the sampling period and frequency can be determined. Power load data over a period of time can be collected to form a real-time power load time series, with a corresponding power load value at each time point.

[0052] Step S20: Preprocess the power load data of the dedicated transformer acquisition terminal to remove noise interference.

[0053] In this embodiment, the steps for preprocessing the electricity load time series data of the dedicated transformer acquisition terminal include:

[0054] Check for duplicate data points in the electricity load data to ensure that there is only one unique electricity load value at each time point:

[0055] For a large amount of electricity load data, data processing tools are used to check for duplicate data records. This is achieved by calling the `duplicated` function in the pandas library of Python through a computer program.

[0056] For missing values ​​in the electricity load data, a linear interpolation method is used to fill in the missing electricity load values ​​by calculating the average value of data points in the neighborhood:

[0057] To quickly identify missing data in electricity load data, a computer program calls the `isnull()` and `loc()` functions from the pandas library in Python to check for missing data and locate its position. The missing data is then filled in using a linear interpolation formula.

[0058]

[0059] in, This represents the calculated filler electrical load value; and These are the timestamps of the known data points on either side of the missing data; These are timestamps of the missing data; and These are the power load values ​​of the known data points on both sides of the missing data.

[0060] The electricity load data is decomposed and reconstructed using the Maximum Overlap Discrete Wavelet Transform (MODWT) to obtain denoised electricity load data. The steps include:

[0061] The electricity load data is decomposed using MODWT to obtain approximation coefficients and detail coefficients at different scales. For electricity load data, For time indexing, and These are the wavelet function and the scaling function, i.e., the high-pass filter and the low-pass filter, respectively; the decomposition process of MODWT is expressed as:

[0062]

[0063]

[0064] in, and These represent the detail coefficients and the approximation coefficients, respectively. This is the scaling factor, representing the index of the decomposition level. The shift coefficient represents the filter index;

[0065] Based on the characteristics of the electricity load data, a suitable threshold function is selected to perform thresholding on the detail coefficients. Let the threshold be... The detail coefficients below a threshold are set to zero, while the detail coefficients above a threshold are shrunk, thereby removing noise components from the time series.

[0066]

[0067] in, Indicates the detail factor after shrinkage; Represents a symbolic function; For maximum value operation; threshold The calculation formula is as follows:

[0068]

[0069]

[0070] in, This represents the scale factor of the first layer in the MODWT decomposition process. The noise level is represented by an assessment of the median absolute deviation. It is the length of the time series. This indicates a median-taking operation;

[0071] The inverse MODWT reconstruction is performed using the detail coefficients after thresholding to obtain the denoised power load time series. The MODWT reconstruction process is represented as follows:

[0072]

[0073] in, It is the total number of decomposition levels. It is the order of the wavelet transform. The reconstructed time series does not contain noise components.

[0074] Step S30: Use the historical electricity load time series of the dedicated transformer acquisition terminal as training samples to train the time series prediction model, and save the optimal model parameters as the pre-trained time series prediction model.

[0075] In this embodiment, the process of obtaining a pre-trained time series prediction model includes:

[0076] The historical electricity load time series of the dedicated transformer acquisition terminal is used as the training sample and divided into training set, validation set and test set in a ratio of 8:1:1.

[0077] A time series forecasting model is constructed, which uses a time-frequency enhancement decomposition network to capture long-term latent patterns and time periodicity of time series from the time-frequency domain;

[0078] A time series prediction model was trained using a hybrid loss function with the Adam optimizer as the optimizer. An initial learning rate was set, and a cosine annealing strategy was used to dynamically adjust the learning rate during training.

[0079] Hybrid loss function Combining the L1 and L2 loss functions, the expression is:

[0080]

[0081]

[0082] in, and These represent the actual and predicted values ​​of the time series, respectively. Represents the L1 loss function; Represents the L2 loss function; Used to control the weights of the L1 loss function and the L2 loss function. Based on the absolute value of the error between the true value and the predicted value, and mapped to the range [0,1] by the Tanh function, it is used as a control weight;

[0083] Cosine annealing causes the learning rate to change according to the shape of a cosine function during training, helping the model converge quickly in the early stages of training and avoiding getting trapped in local optima in the later stages. The formula for updating the learning rate is expressed as:

[0084]

[0085] in, This is the current learning rate. and These represent the maximum and minimum values ​​of the learning rate, respectively. This indicates the number of steps in the current training cycle. It is the total number of steps in a training cycle.

[0086] When the model's loss function reaches its minimum on the validation set, the model performance is evaluated using the test set, and the model parameters at this point are saved as a pre-trained time series prediction model.

[0087] Step S40: Apply the pre-trained time series prediction model to comprehensively analyze the real-time electricity load time series from the time domain and frequency domain, and output the prediction results of long-term electricity load.

[0088] In this embodiment, the process of analyzing the electricity load time series using the pre-trained time series prediction model mainly consists of three stages:

[0089] In the first stage, the time series of the input power load of the dedicated transformer acquisition terminal is normalized, and the time series is further divided into trend time series and seasonal time series.

[0090] In the second stage, the multi-scale time-frequency enhancement encoder uses multiple sliding windows of different sizes. The trend decoder and seasonal encoder extract time-series features from the trend time series and seasonal time series, respectively, for windows of different sizes.

[0091] In the third stage, the outputs of the trend encoder and the seasonal encoder are added together, and after linear mapping and inverse normalization, the predicted future electricity load time series is output.

[0092] Step S50: Based on the electricity load prediction results of the pre-trained time series prediction model, a dynamic programming algorithm is introduced to control the capacity of the dedicated transformer acquisition terminal in real time.

[0093] In this embodiment, the steps for using a dynamic programming algorithm to control the capacity of the dedicated transformer data acquisition terminal in real time include:

[0094] The objective of determining the capacity of the dedicated transformer data acquisition terminal is to dynamically adjust the capacity of the dedicated transformer data acquisition terminal according to changes in future electricity load, with the goal of minimizing the total cost.

[0095] A dynamic programming model is established, defining the capacity of the dedicated transformer data acquisition terminal at each future time point as a state variable, and the capacity adjustment amount at each future time point as a decision variable. The objective function is to minimize the total cost, including equipment configuration cost, operation and maintenance cost, and energy consumption cost. Constraints are determined, including capacity constraints, load demand constraints, and capacity adjustment constraints.

[0096] Let the state variable be... For time points The capacity of the dedicated transformer data acquisition terminal;

[0097] Let decision variables be defined. For time points The capacity adjustment amount, i.e., the increase or decrease in capacity;

[0098] The objective function is expressed by the formula: ,in, To perform the minimum value operation, , and These are the equipment configuration costs, operation and maintenance costs, and energy consumption costs for a specific capacity.

[0099] Obtain the maximum capacity from the technical manual of the dedicated transformer data acquisition terminal equipment. , minimum capacity and capacity adjustment step size The capacity constraint is: ;

[0100] The load demand constraints are: ,in For the predicted time point The electrical load value;

[0101] The capacity adjustment constraints are: .

[0102] Divide the future time series into multiple stages, each corresponding to a specific time point; update the state variables based on the decision variables: ,in For the updated state variables;

[0103] Establish a dynamic programming recurrence formula: starting from the last time point Begin by recursively calculating the optimal cost and optimal decision at each time point; for each time point... Calculate all possible capacity adjustment amounts Calculate the total cost and select the minimum value:

[0104]

[0105] in, and They represent time. and And the capacity is and The minimum total cost at which the minimum cost is achieved is retained, along with the capacity adjustment amount at which the minimum cost is achieved. ;

[0106] Starting from the initial time point, the optimal cost is calculated using the dynamic programming recursive formula. and decision Extract each time point step by step The optimal capacity adjustment strategy is used to adjust the capacity of the dedicated transformer data acquisition terminal according to the optimal strategy.

[0107] Based on operational data and feedback, the parameters and constraints of the dynamic programming model are continuously optimized; electricity load forecast data is updated regularly to ensure the accuracy and adaptability of the dynamic programming algorithm.

[0108] Example 2: This application provides a method for load forecasting and dynamic capacity control of a dedicated transformer data acquisition terminal. The detailed structure of the time series forecasting model is as follows: Figure 2 As shown:

[0109] In this embodiment, the time series prediction model is a Time-Frequency Augmentation Decomposition Network (TFDNet), which consists of three stages. In the first stage, the input time series... Normalization and decomposition are performed. A reversible instance normalization operation is used to process the input time series, further decomposing the time series into seasonal components. and trend components To capture different potential patterns in the seasonal and trend sections:

[0110]

[0111]

[0112] in, This indicates that the moving average operation is used to smooth out periodic hidden variables. This is a padding operation that ensures the sequence length remains unchanged.

[0113] In the second stage, the multi-scale time-frequency encoder includes a seasonal encoder and a trend encoder, which capture latent time-domain and frequency-domain patterns in seasonal and trend components, respectively. The feature extraction process of the multi-scale time-frequency encoder is expressed by the following formula:

[0114]

[0115]

[0116] in, and These represent the output features of the multi-scale seasonal encoder and the trend encoder, respectively. This represents a linear layer, used to adjust the size of the output features, converting the output feature dimensions concatenated from encoders of different scales into the original feature dimensions; Indicates different sliding window sizes; and These represent the seasonal encoder and the trend encoder, respectively.

[0117] The detailed structure of seasonal encoders and trend encoders is as follows: Figure 3 As shown, the encoder's processing procedure is represented as follows:

[0118]

[0119]

[0120]

[0121]

[0122] in, The input sequence for the encoder, After short-time Fourier transform The generated spectrum; It is a time-frequency module The output characteristics; yes The sum of frequency and feedforward layer output, It is a frequency feedforward layer. It is the final output feature of the encoder. This represents the inverse short-time Fourier transform.

[0123] The Short-Time Fourier Transform (SFT) can transform a time series from the time domain to the time-frequency domain. This method divides the input time series into overlapping frames and performs a Discrete Fourier Transform on each frame. This process is represented as follows:

[0124]

[0125] Among them, window functions The value is 1 throughout the entire window area, and the window size is [value missing]. , It is a time index. Indicates displacement. It is the step size of the moving window. Indicates frequency, For frequency indexing. The size is ,in The magnitudes of the components of the Fourier transform conform to conjugate symmetry. Indicates the number of time frames, and , This refers to the length of the time series. The multi-scale strategy reflects the sliding window size in the short-time Fourier transform. different.

[0126] The time-frequency module is used internally by the encoder to capture latent patterns in the time-frequency domain, applying kernel operations, i.e., convolution operations, to the same frequency across different time frames. Each channel's features in each frequency band have their own convolutional kernel weight matrix. Based on the inherent characteristics of different sequence components, the time-frequency modules within the seasonal encoder and trend encoder employ different weighting strategies, either sharing a weight matrix or having different values ​​for the weight matrix in each channel. The frequency-feedforward layer is a fully connected layer with a Tanh activation function, used to accumulate time-frequency information along the frequency dimension within a single time step.

[0127] The final stage predicts the time series for the next T time points based on the fusion encoder representation, including linear mapping and inverse normalization operations.

[0128] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application. In some cases, the actions or steps described in this application can be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results.

Claims

1. A method for load forecasting and dynamic capacity control of a dedicated transformer data acquisition terminal, characterized in that, The method includes: (1) Collect electricity load data from the dedicated transformer acquisition terminal, including historical electricity load time series and real-time electricity load time series; Specifically, the load forecasting data from the dedicated transformer acquisition terminal is processed into a time series forecasting task. Let the electricity load data to be processed be... It is a length of The lookback window is a collection of elements of time variables, each representing a single point in time. Recording electricity load values, the time series forecasting task predicts the future step size as follows: time series ; (2) Preprocess the power load data of the dedicated transformer acquisition terminal to remove noise interference; The steps for preprocessing the power load data from the dedicated transformer acquisition terminal include: Check for duplicate data points in the electricity load data to ensure that there is only a unique electricity load value at each point in time. For missing values ​​in the electricity load data, a linear interpolation method is used to fill in the missing electricity load values ​​by calculating the average value of data points in the neighborhood; The electricity load data is decomposed and reconstructed using the maximum overlap discrete wavelet transform to obtain the denoised electricity load time series: The electricity load data is decomposed by maximum overlap discrete wavelet transform to obtain approximation coefficients and detail coefficients at different scales. According to the characteristics of the electricity load data, an appropriate threshold function is selected to perform threshold processing on the detail coefficients. The thresholded detail coefficients are then used to perform inverse maximum overlap discrete wavelet transform reconstruction to obtain the denoised electricity load data. The decomposition process of the maximally overlapping discrete wavelet transform is expressed as follows: in, For electricity load data, For time indexing, and These are wavelet functions and scaling functions, respectively. and These represent the detail coefficients and the approximation coefficients, respectively. is the scaling factor, representing the index of the decomposition level; The translation coefficients represent the filter indexes; the reconstruction process of the maximum overlap discrete wavelet transform is expressed as follows: in, It is the total number of decomposition levels. It is the order of the wavelet transform. These are the detail coefficients after thresholding. (3) Use the historical electricity load time series of the special transformer acquisition terminal as training samples to train the time series prediction model, and save the optimal model parameters as the pre-trained time series prediction model. The process of obtaining a pre-trained time series prediction model includes: The historical electricity load time series of the dedicated transformer acquisition terminal is used as the training sample and divided into training set, validation set and test set in a ratio of 8:1:

1. A time series prediction model is constructed, which uses a time-frequency enhancement decomposition network to capture the long-term potential patterns and time periodicity of the time series from the time-frequency domain. The time series prediction model uses a time-frequency enhancement decomposition network to process the electricity load time series data of the dedicated transformer acquisition terminal through a multi-scale time-frequency enhancement encoder. The trend time-frequency module in the trend encoder learns the trend pattern of the time series by sharing the core, and the seasonal time-frequency module in the seasonal encoder learns the seasonal characteristics of the time series. A time series prediction model was trained using a hybrid loss function, with the Adam optimizer employed. An initial learning rate was set, and a cosine annealing strategy was used to dynamically adjust the learning rate during training. The hybrid loss function... The L1 loss function and the L2 loss function are weighted and summed, and the expression is: in, and These represent the actual and predicted values ​​of the time series, respectively. Represents the L1 loss function; Represents the L2 loss function; Used to control the weights of the L1 loss function and the L2 loss function. Based on the absolute value of the error between the true value and the predicted value, and mapped to the range [0,1] by the Tanh function, it is used as a control weight; When the model's loss function reaches its minimum on the validation set, the model performance is evaluated using the test set, and the model parameters at this point are saved as a pre-trained time series prediction model. (4) Apply the pre-trained time series prediction model to comprehensively analyze the real-time power load time series from the time domain and frequency domain, and output the prediction results of long-term power load; The process of analyzing electricity load time series using a pre-trained time series prediction model mainly consists of three stages: In the first stage, the time series of the input power load of the dedicated transformer acquisition terminal is normalized, and the time series is further divided into trend time series and seasonal time series. In the second stage, the multi-scale time-frequency enhancement encoder uses multiple sliding windows of different sizes. The trend decoder and seasonal encoder extract time-series features from the trend time series and seasonal time series, respectively, for windows of different sizes. In the third stage, the outputs of the trend encoder and the seasonal encoder are added together, and after linear mapping and inverse normalization, the predicted future electricity load time series is output. (5) Based on the power load prediction results of the pre-trained time series prediction model, a dynamic programming algorithm is introduced to control the capacity of the dedicated transformer acquisition terminal in real time.

2. The load forecasting and dynamic capacity control method for a dedicated transformer data acquisition terminal as described in claim 1, characterized in that, Step (5), which uses a dynamic programming algorithm to control the capacity of the dedicated transformer data acquisition terminal in real time, includes the following steps: The objective of determining the capacity of the dedicated transformer data acquisition terminal is to dynamically adjust the capacity of the dedicated transformer data acquisition terminal according to changes in future electricity load, with the goal of minimizing the total cost. A dynamic programming model is established, defining the capacity of the dedicated transformer data acquisition terminal at each future time point as a state variable, the capacity adjustment amount at each future time point as a decision variable, the objective function as minimizing the total cost, and determining the constraints, including capacity constraints, load demand constraints, and capacity adjustment constraints. Divide the future time series into multiple stages, each stage corresponding to a point in time; update the state variables based on the decision variables, and establish a dynamic programming recursive formula. Starting from the initial time point, the optimal cost and decision are calculated based on the dynamic programming recursive formula. The optimal capacity adjustment strategy for each time point is gradually extracted, and the capacity of the dedicated transformer acquisition terminal is adjusted according to the optimal strategy. Based on operational data and feedback, the parameters and constraints of the dynamic programming model are continuously optimized; electricity load forecast data is updated regularly to ensure the accuracy and adaptability of the dynamic programming algorithm.

Images