Power distribution network load prediction method and system based on deep learning

Through deep learning methods, using technologies such as convolutional neural networks and fast Fourier transforms, the load data characteristics of distribution networks are extracted and fused, and the insufficient modeling of flexible resource influencing factors and insufficient extraction of multi-scale feature in the existing technology is solved. The problem improves the accuracy of load forecasting and the stability of power supply.

CN118940880APending Publication Date: 2024-11-12GUIZHOU POWER GRID CO LTD
View PDF 0 Cites 4 Cited by

Patent Information

Application Number
CN202410892246.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-04
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

The existing distribution network load prediction methods are difficult to fully utilize the influencing factors of flexible resources to model, and the multi-scale feature extraction is insufficient, resulting in low prediction accuracy, especially when processing long-term spans and multi-scale data, it is prone to overfitting or under-resolution. Fit.

Method used

Using a deep learning-based method, time series features are extracted through convolutional neural networks, combined with fast Fourier transform and two-dimensional convolutional networks, data preprocessing and feature extraction, slice division and adaptive fusion are constructed to build TimesNet network for loading. predict.

Benefits of technology

It improves the accuracy and stability of load prediction, ensures the accuracy and completeness of the data set, and improves the stability and reliability of the power supply.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118940880A_ABST
    Figure CN118940880A_ABST
Patent Text Reader

Abstract

The invention discloses a power distribution network load prediction method based on deep learning, and the method comprises the steps: obtaining regional load data and renewable energy power generation data, forming a data set, carrying out the preprocessing of the data set, and obtaining a first data set; extracting time sequence features in the first data set by using a convolutional neural network, converting the time sequence features into a data form of a deep learning model through an embedded layer, obtaining one-dimensional time sequence data, performing fast Fourier transform to obtain a frequency curve, extracting an amplitude value on the frequency curve, and calculating to obtain a corresponding period; selecting a corresponding period to slice and divide the one-dimensional time sequence data into a two-dimensional matrix; according to the method, a two-dimensional convolutional network is used for feature extraction, a two-dimensional matrix subjected to feature extraction is restored into a one-dimensional array, the one-dimensional array is subjected to adaptive fusion, first time sequence features are obtained and input into a full connection layer for weight calculation and linear transformation, a power distribution network load prediction result is obtained, and the stability and reliability of power supply are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network, and specifically to a distribution network load forecasting method and system based on deep learning. Background Art

[0002] As an important part of the power system, the reliability and efficiency of the distribution network directly affect the stable supply of electricity. With the access of new energy and the opening of the power market, traditional load forecasting methods can no longer meet the needs of modern distribution networks. The introduction of wind power and photovoltaics has provided a new way to optimize the operation of distribution networks, but it has also increased the complexity of load forecasting.

[0003] Traditional load forecasting methods, such as time series analysis and statistical regression analysis, often fail to fully utilize characteristic data related to distribution network load (such as weather data, user behavior data, distributed energy data, etc.), resulting in low prediction accuracy. In addition, these methods are prone to overfitting or underfitting when processing long-time span and multi-scale data, and it is difficult to accurately capture the dynamic characteristics of load changes. In order to solve these problems, deep learning methods have been introduced into the field of load forecasting in recent years. In particular, deep learning models based on the Transformer architecture have become a hot topic in load forecasting research due to their excellent performance in sequence data processing. Among them, the TimesNet network has great potential in distribution network load forecasting due to its unique advantages: strong long-time series modeling and multi-scale feature extraction capabilities, good robustness. Summary of the invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is: the present invention solves the limitations of existing distribution network load forecasting methods, insufficient modeling of influencing factors of flexible resources, and insufficient multi-scale feature extraction.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a distribution network load forecasting method based on deep learning, comprising:

[0008] Acquire regional load data and renewable energy power generation data, combine the regional load data and renewable energy power generation data into a data set, and pre-process the data set to obtain a first data set;

[0009] Using a convolutional neural network to extract time series features from the first data set, converting the time series features into a data form of a deep learning model through an embedding layer to obtain one-dimensional time series data;

[0010] Performing a fast Fourier transform on the one-dimensional time series data to obtain a frequency curve, extracting the amplitude value on the frequency curve, calculating the corresponding period according to the amplitude value, and selecting the corresponding period to slice and divide the one-dimensional time series data into a two-dimensional matrix;

[0011] Using a two-dimensional convolutional network to extract features from the two-dimensional matrix, restoring the two-dimensional matrix after feature extraction into a one-dimensional array, and adaptively fusing the one-dimensional array to obtain a first time series feature;

[0012] The first time series feature is input into the fully connected layer for weight calculation and linear transformation to obtain the distribution network load forecast result.

[0013] As a preferred solution of the distribution network load forecasting method based on deep learning described in the present invention, the preprocessing of the data set includes:

[0014] Calculate the first quartile and the third quartile in the data set, and subtract the first quartile from the third quartile to get the interquartile range;

[0015] The normal value range is 1.5 times the interquartile range extending downward from the first quartile to obtain the lower limit, and 1.5 times the interquartile range extending upward from the third quartile to obtain the upper limit;

[0016] Values ​​outside the normal range are considered outliers and are removed from the data set;

[0017] For the missing parts in the data set, select the K value, calculate the Euclidean distance between each missing value sample and other samples, select the K samples closest to it, and use the weighted average of the K nearest neighbor samples to fill in the missing values.

[0018] As a preferred solution of the distribution network load forecasting method based on deep learning described in the present invention, the one-dimensional time series data includes:

[0019] The time series features of each column in the data set are converted into the data form processed by TimesBlock through the embedding layer. The data is transferred between each layer of TimesBlock through residual operation to obtain the input data form of the k-th layer TimesBlock, that is, one-dimensional time series data.

[0020] As a preferred solution of the distribution network load forecasting method based on deep learning described in the present invention, the fast Fourier transform includes:

[0021] Convert the one-dimensional time series data into a spectrum curve, extract the amplitude value corresponding to each positive frequency on the spectrum curve, select the first n frequency components of the average amplitude according to the setting of the hyperparameters, and find the corresponding period expression as:

[0022]

[0023] in, is one-dimensional time series data, FFT(〃) is fast Fourier transform, Amp(〃) is amplitude function, Avg(〃) is average function, {f1,...,f n} are n frequency components, {p1,...,p n} is the corresponding period, W is the average amplitude corresponding to the frequency component, and T is the data length.

[0024] As a preferred solution of the distribution network load forecasting method based on deep learning described in the present invention, wherein: selecting a corresponding period to slice and divide the one-dimensional time series data into a two-dimensional matrix includes:

[0025] According to the size of each period, the one-dimensional time series data is divided into multiple slices, each slice corresponds to the sequence data within a period;

[0026] Add 0 to the end of the one-dimensional time series data until the length of the one-dimensional time series can be divided by the period, and reorganize the one-dimensional time series to obtain a two-dimensional matrix.

[0027] As a preferred solution of the distribution network load forecasting method based on deep learning described in the present invention, wherein: using a two-dimensional convolutional network to extract features from the two-dimensional matrix, and restoring the two-dimensional matrix after feature extraction to a one-dimensional array includes;

[0028] The two-dimensional matrix is ​​used as input data, and the Inception convolutional network is used to perform convolution processing on the two-dimensional matrix to obtain the two-dimensional matrix after feature extraction;

[0029] The two-dimensional matrix after feature extraction is reshaped into a one-dimensional vector according to the number of rows and columns of the matrix, and the padding 0 added during the reshaping process is removed to obtain one-dimensional time series data, that is, a one-dimensional array.

[0030] As a preferred solution of the distribution network load forecasting method based on deep learning described in the present invention, wherein: adaptively fusing the one-dimensional array comprises:

[0031] Calculate the weight of each frequency component at k-1 steps and convert the weight into probability distribution through the Softmax function;

[0032] The weight of each frequency component is multiplied by the corresponding one-dimensional array, and the multiplication results are summed to obtain a weighted one-dimensional time array.

[0033] In a second aspect, the present invention provides a system for distribution network load forecasting based on deep learning, comprising:

[0034] A preprocessing module, used for acquiring regional load data and renewable energy power generation data, combining the regional load data and renewable energy power generation data into a data set, and preprocessing the data set to obtain a first data set;

[0035] A data conversion module, used to extract time series features from the first data set using a convolutional neural network, and convert the time series features into a data form of a deep learning model through an embedding layer to obtain one-dimensional time series data;

[0036] A slicing module is used to perform a fast Fourier transform on the one-dimensional time series data to obtain a frequency curve, extract the amplitude value on the frequency curve, calculate the corresponding period according to the amplitude value, and select the corresponding period to slice and divide the one-dimensional time series data into a two-dimensional matrix;

[0037] A feature extraction module, used to extract features from the two-dimensional matrix using a two-dimensional convolutional network, restore the two-dimensional matrix after feature extraction to a one-dimensional array, and adaptively fuse the one-dimensional array to obtain a first time series feature;

[0038] The calculation module is used to input the first time series features into the fully connected layer for weight calculation and linear transformation to obtain the distribution network load prediction result.

[0039] In a third aspect, the present invention provides a computing device, comprising:

[0040] Memory and processor;

[0041] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the distribution network load forecasting method based on deep learning are implemented.

[0042] In a fourth aspect, the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the distribution network load forecasting method based on deep learning.

[0043] Beneficial effects of the invention: The distribution network load forecasting method based on deep learning proposed in the invention adopts the TimesNet network to construct a load forecasting model, combines the advantages of deep learning technology, improves the accuracy of load forecasting, and ensures the accuracy and completeness of the data set through data screening, outlier detection, missing value filling and data normalization processing, provides high-quality input data for the model, effectively manages the distribution network load, and improves the stability and reliability of power supply. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. Among them:

[0045] Figure 1 An overall flow chart of a distribution network load forecasting method based on deep learning provided by the present invention;

[0046] Figure 2 A wind power generation load prediction result diagram of a distribution network load prediction method based on deep learning provided by the present invention;

[0047] Figure 3 A photovoltaic power generation load prediction result diagram of a distribution network load prediction method based on deep learning provided by the present invention;

[0048] Figure 4 A regional load forecasting result diagram of a distribution network load forecasting method based on deep learning provided by the present invention;

[0049] Figure 5 A distribution network load forecasting result diagram considering wind power and photovoltaic access for a distribution network load forecasting method based on deep learning provided by the present invention. DETAILED DESCRIPTION

[0050] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0051] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0052] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0053] The present invention is described in detail with reference to schematic diagrams. When describing the embodiments of the present invention, for the sake of convenience, the cross-sectional diagrams showing the device structure will not be partially enlarged according to the general scale, and the schematic diagrams are only examples, which should not limit the scope of protection of the present invention. In addition, in actual production, the three-dimensional dimensions of length, width and depth should be included.

[0054] At the same time, in the description of the present invention, it should be noted that the directions or positional relationships indicated by the terms "upper, lower, inner and outer" are based on the directions or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore cannot be understood as limiting the present invention. In addition, the terms "first, second or third" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0055] In the present invention, unless otherwise clearly specified and limited, the terms "install, connect, connect" should be understood in a broad sense, for example: it can be a fixed connection, a detachable connection or an integral connection; it can also be a mechanical connection, an electrical connection or a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0056] Example 1

[0057] Reference Figure 1 , is an embodiment of the present invention, and provides a distribution network load forecasting method based on deep learning, comprising:

[0058] S100: Acquire regional load data and renewable energy generation data, combine the regional load data and renewable energy generation data into a data set, and pre-process the data set to obtain a first data set;

[0059] In the application embodiment, the first quartile and the third quartile in the data set are calculated, and the interquartile range is obtained by subtracting the first quartile from the third quartile;

[0060] The normal value range is 1.5 times the interquartile range extending downward from the first quartile to obtain the lower limit, and 1.5 times the interquartile range extending upward from the third quartile to obtain the upper limit;

[0061] Values ​​outside the normal range are considered outliers and are removed from the data set;

[0062] For the missing parts in the data set, select the K value, calculate the Euclidean distance between each missing value sample and other samples, select the K samples closest to it, and use the weighted average of the K nearest neighbor samples to fill in the missing values.

[0063] Exemplarily, obtaining regional load data and renewable energy generation data includes collecting relevant characteristic factors affecting wind power generation, including wind speed, wind direction, temperature and air pressure, so as to characterize wind power generation as much as possible. Wind speed and wind direction directly affect the power generation efficiency of wind turbines, while temperature and air pressure affect air density, which in turn directly affects the kinetic energy content of wind energy and the efficiency of wind energy conversion. At the same time, data related to photovoltaic power generation are collected, including solar irradiance, total cloud cover, temperature and air pressure, to ensure that the environmental conditions for photovoltaic power generation are fully considered. Solar irradiance and total cloud cover directly affect the power generation capacity of photovoltaic panels, while temperature and air pressure affect the overall performance of photovoltaic systems. In addition, regional load-related data are collected, covering historical load data and user electricity consumption behavior, providing background information on load changes in regional distribution networks. Historical load data helps to understand past electricity consumption patterns and load peak and valley changes, while user electricity consumption behavior can reflect the periodicity and suddenness of electricity demand, thereby more accurately predicting future load changes.

[0064] For the acquired data, through correlation analysis and calculation, the features with significant influence on wind energy, photovoltaic power and regional load are screened out respectively to form a data set, ensuring that the model only uses the most relevant data, reducing data redundancy and unnecessary computing power consumption; the data set is cleaned, and the interquartile range (IQR) method is used to detect and remove outliers in the data set, including calculating the first quartile (Q1) and the third quartile (Q3) of the data set, and calculating the interquartile range (IQR): IQR = Q3-Q1; defining the normal value range: Q = [Q1-1.5*IQR, Q3+1.5*IQR];

[0065] Use K-nearest neighbor interpolation (KNN) to fill in the missing parts of the data set to ensure the accuracy and completeness of the data set. Select the K value, which can be 5. Calculate the Euclidean distance between each missing value sample and other samples as follows:

[0066]

[0067] Where d(x,y) is the Euclidean distance between sample x and sample y;

[0068] Filling missing values ​​with the weighted average of K nearest neighbor samples is expressed as:

[0069]

[0070] in, is the missing value after filling, x i is the value of the i-th nearest neighbor sample, and K is the number of selected neighboring samples.

[0071] The data in the wind energy, photovoltaic and regional load datasets are normalized using the minimum-maximum normalization method, and the complex nonlinear time series features in the datasets are extracted using a convolutional neural network (CNN).

[0072] Specifically, the methods for processing missing values ​​include mean / median filling, linear interpolation, and K nearest neighbor interpolation. The methods for detecting outliers include standard deviation method, model-based method, and interquartile range method. The present invention prefers interquartile range method and K nearest neighbor interpolation method.

[0073] It should be noted that there will be outliers in the distribution network load data and renewable energy generation data, which will interfere with the training of the deep learning model and affect the prediction accuracy. Using the interquartile range method to remove outliers can effectively identify and remove abnormal data points to ensure the stability and consistency of the data. Compared with other outlier detection methods, the interquartile range method has greater data fluctuations for distribution network load data. The interquartile range method can flexibly respond to these fluctuations and ensure the robustness of data preprocessing;

[0074] In the actual data collection process, some data will be missing. If not processed, these missing values ​​will also affect the training of the model. Compared with other missing value processing methods, the K nearest neighbor method (KNN) can use the existing sample information in the data set to fill the missing values ​​by calculating the similarity, maintain the integrity and continuity of the data, and ensure that the filled data is reasonable and accurate.

[0075] Removing outliers by the IQR method can reduce noise and redundancy in the data set and improve the training efficiency and accuracy of the model. KNN fills in missing values, ensures the integrity of the data set, and avoids information loss caused by missing data. The interquartile range method and K nearest neighbor interpolation method have obvious advantages in data preprocessing. They can improve data quality, reduce the consumption of computing resources, and provide a more accurate and reliable data basis for subsequent model training and prediction.

[0076] S200: extracting time series features from the first data set using a convolutional neural network, and converting the time series features into a data form of a deep learning model through an embedding layer to obtain one-dimensional time series data;

[0077] In an embodiment of the present application, the time series features of each column in the data set are converted into a data format processed by TimesBlock through an embedding layer, and data is transferred between each layer of TimesBlock through a residual operation to obtain the input data format of the kth layer of TimesBlock, that is, one-dimensional time series data.

[0078] Exemplarily, each column of time series data in the wind energy, photovoltaic and regional load data sets is transformed into: Among them, Z 1D is each column of time series data in the wind energy, photovoltaic and regional load data sets, and F(·) is the 1D Input is processed into the embedding layer. is the input data of the first layer TimeBlock in the model. The input data of the kth layer TimeBlock module is expressed as:

[0079]

[0080] in, It is the input data form of the k-1th layer TimeBlock module, and TimesBlock(〃) is the conversion form of the TimeBlock module.

[0081] It should be noted that TimesBlock, as a deep learning structure, can capture complex patterns and long-term dependencies in time series data. Through the embedding layer, the original data can be converted into a richer feature representation. Residual connections help alleviate the gradient vanishing problem in deep learning models, allowing deep networks to be trained more efficiently. In time series models, this helps maintain the transmission of long-term dependency information. Through the combination of embedding layers and TimesBlock, the model is able to learn more abstract feature representations, which helps improve the model's generalization ability for unseen data.

[0082] S300: Perform fast Fourier transform on the one-dimensional time series data to obtain a frequency curve, extract the amplitude value on the frequency curve, calculate the corresponding period according to the amplitude value, and select the corresponding period to slice and divide the one-dimensional time series data into a two-dimensional matrix;

[0083] In the embodiment of the present application, the one-dimensional time series data is converted into a spectrum curve, the amplitude value corresponding to each positive frequency on the spectrum curve is extracted, and according to the setting of the hyperparameters, the first n frequency components of the average amplitude are selected, and the corresponding period is expressed as:

[0084]

[0085] in, is one-dimensional time series data, FFT(〃) is fast Fourier transform, Amp(〃) is amplitude function, Avg(〃) is average function, {f1,...,f n} are n frequency components, {p1,...,p n} is the corresponding period, W is the average amplitude corresponding to the frequency component, and T is the data length.

[0086] It should be noted that FFT can convert time series data from the time domain to the frequency domain, making the periodic features in the data more obvious, which helps to identify and analyze the periodic components in the data; by selecting frequency components with larger amplitudes, the features that contribute most to the data pattern can be extracted, thereby simplifying the model and possibly improving prediction accuracy; by selecting only n frequency components with larger average amplitudes, the data can be compressed and the amount of computational effort for subsequent processing can be reduced; applying FFT in the TimesBlock module can effectively process and analyze time series data, providing strong support for subsequent feature learning and model training.

[0087] In the embodiment of the present application, the one-dimensional time series data is divided into multiple slices according to the size of each period, and each slice corresponds to the sequence data within one period;

[0088] Add 0 to the end of the one-dimensional time series data until the length of the one-dimensional time series can be divided by the period, and reorganize the one-dimensional time series to obtain a two-dimensional matrix.

[0089] Specifically, based on the selected n cycles, The slices are divided into n groups of two-dimensional matrices, each of which has a dimension of (p i ,f i );

[0090]

[0091] in, According to the size of the i-th cycle The two-dimensional matrix of slice partitioning, i∈{1,…,n}; Padding(〃) is the The end of the sequence is p-padded with 0 so that the sequence length can be i Divide it by integer to get a set of two-dimensional matrices The corresponding period is p i Two-dimensional time series changes;

[0092] S400: extracting features from the two-dimensional matrix using a two-dimensional convolutional network, restoring the two-dimensional matrix after feature extraction to a one-dimensional array, and adaptively fusing the one-dimensional array to obtain a first time series feature;

[0093] In the embodiment of the present application, a two-dimensional matrix is ​​used as input data, and the two-dimensional matrix is ​​convolved using the Inception convolutional network to obtain a two-dimensional matrix after feature extraction;

[0094] Reshape the two-dimensional matrix after feature extraction into a one-dimensional vector according to the number of rows and columns of the matrix, remove the padding 0 added during the reshaping process, and obtain one-dimensional time series data, that is, a one-dimensional array;

[0095] Specifically, a two-dimensional convolutional network is used to train n two-dimensional matrices Feature extraction is expressed as: in, According to the i-th cycle The two-dimensional matrix after feature extraction, f(·) is a two-dimensional matrix Perform convolution processing; the convolution network can use Inception convolution network, ResNet network and Xception network, and the present invention preferably uses Inception convolution network.

[0096] It should be noted that compared to other convolutional networks, the Inception network can simultaneously capture the features of images at different scales by applying convolution kernels of different sizes (1x1, 3x3, 5x5) and maximum pooling in parallel, allowing the network to understand the image content more comprehensively. The 1x1 convolution kernel in the Inception module is used to reduce the dimension of the feature map before performing large-scale convolution operations, thereby reducing the amount of calculation and preventing overfitting. Multiple convolution kernels of different sizes are used to capture feature information of different scales. This multi-scale convolution operation is conducive to extracting rich feature information.

[0097] Specifically, for the two-dimensional matrix that completes feature extraction, the two-dimensional matrix is ​​restored to one dimension as follows:

[0098]

[0099] in, The dimension is (1,(p i * f i )), Trunc(〃) is to remove the 0 added by Padding(〃).

[0100] In the embodiment of the present application, the weight of each frequency component at the k-1 step is calculated, and the weight is converted into a probability distribution through the Softmax function;

[0101] Multiply the weight of each frequency component by the corresponding one-dimensional array, and sum the multiplication results to obtain a weighted one-dimensional time array;

[0102] In the embodiment of the present application, the obtained one-dimensional array The weighted summation of the average amplitude of the corresponding frequency is expressed as:

[0103]

[0104] in, is the frequency component {f1,f 2,…, f n} is the weight of the matrix, and Softmax(〃) is the activation function.

[0105] Specifically, TimesNet extracts two-dimensional time series changes in multiple cycles and then performs adaptive fusion time series change modeling. The final time series features are flattened and input into the fully connected layer. The fully connected layer performs weight calculation and linear transformation on the flattened features to generate load forecasting results. In addition to the TimesNet model, the Informer model and LSTM model can also be used.

[0106] It should be noted that the TimesNet model is based on the Transformer architecture and is good at processing long-span data. It can capture long-term dependencies in time series, that is, it can more accurately capture the time series characteristics in the data. The key indicator that affects the accuracy of load forecasting is whether the time series characteristics in time series data can be accurately extracted. This makes it have higher prediction accuracy and robustness in distribution network load forecasting, thereby better meeting the prediction needs of modern distribution networks.

[0107] Compared with other deep learning models, TimesNet is preferred to build a load forecasting model. By converting one-dimensional time series data into two-dimensional tensors, TimesNet can simultaneously capture the intra-cycle changes and inter-cycle changes of time series data, thereby more accurately modeling time series data with multi-periodicity. TimesNet has demonstrated performance that surpasses other models in multiple time series tasks, including short-term and long-term prediction, classification, anomaly detection, and missing value processing.

[0108] The above is a schematic scheme of a distribution network load prediction method based on deep learning in this embodiment. It should be noted that the technical scheme of the distribution network load prediction device based on deep learning and the technical scheme of the distribution network load prediction method based on deep learning belong to the same concept. The details of the technical scheme of the distribution network load prediction device based on deep learning in this embodiment that are not described in detail can be found in the description of the technical scheme of the distribution network load prediction method based on deep learning.

[0109] The device for distribution network load prediction based on deep learning in this embodiment includes:

[0110] A preprocessing module, used for acquiring regional load data and renewable energy power generation data, combining the regional load data and renewable energy power generation data into a data set, and preprocessing the data set to obtain a first data set;

[0111] A data conversion module, used to extract time series features from the first data set using a convolutional neural network, and convert the time series features into a data form of a deep learning model through an embedding layer to obtain one-dimensional time series data;

[0112] A slicing module is used to perform a fast Fourier transform on the one-dimensional time series data to obtain a frequency curve, extract the amplitude value on the frequency curve, calculate the corresponding period according to the amplitude value, and select the corresponding period to slice and divide the one-dimensional time series data into a two-dimensional matrix;

[0113] A feature extraction module, used to extract features from the two-dimensional matrix using a two-dimensional convolutional network, restore the two-dimensional matrix after feature extraction to a one-dimensional array, and adaptively fuse the one-dimensional array to obtain a first time series feature;

[0114] The calculation module is used to input the first time series features into the fully connected layer for weight calculation and linear transformation to obtain the distribution network load prediction result.

[0115] This embodiment also provides a computing device, which is applicable to distribution network load prediction based on deep learning, including:

[0116] Memory and processor; the memory is used to store computer executable instructions, and the processor is used to execute computer executable instructions to implement the distribution network load forecasting method based on deep learning as proposed in the above embodiment.

[0117] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, the method for load forecasting of a distribution network based on deep learning as proposed in the above embodiment is implemented.

[0118] The storage medium proposed in this embodiment and the method for implementing distribution network load forecasting based on deep learning proposed in the above embodiment belong to the same inventive concept. The technical details not fully described in this embodiment can be referred to the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0119] Through the above description of the implementation method, the technicians in the relevant field can clearly understand that the present invention can be implemented by means of software and necessary general hardware, and of course it can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ReadOnly, Memory, ROM), random access memory (RandomAccessMemory, RAM), flash memory (FLASH), hard disk or optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform the methods of various embodiments of the present invention.

[0120] Example 2

[0121] Reference Figure 2-5 As shown, an embodiment of the present invention provides a distribution network load forecasting method based on deep learning to verify and illustrate the technical effects adopted in this method.

[0122] Taking the distribution network system of a coastal city as an example, the system comprehensively utilizes wind energy and photovoltaic power generation to meet the electricity needs of local residents and industries.

[0123] The voltage level of the distribution network system is 10kV, and there are 10 wind turbines and one photovoltaic power station installed in the coverage area. The rated power of each wind turbine is 500kW, with a total installed capacity of 5000kW, while the total installed capacity of the photovoltaic power station is 3000kW;

[0124] The system collects environmental data such as wind speed, wind direction, solar irradiance, total cloud cover, temperature and air pressure, and combines historical load data and user electricity consumption behavior data to build a high-precision load forecasting model, thereby achieving accurate prediction and effective management of distribution network load.

[0125] like Figure 2 , Figure 3 and Figure 4 They are wind power generation load, photovoltaic power generation load and regional load forecast results. Overall, the predicted values ​​(circular marks) and actual values ​​(diamond marks) have a good fit;

[0126] from Figure 2 It can be seen that the wind power load power drops rapidly from about 350kW at 1 hour to about 150kW at 5 hours, then fluctuates significantly at 10 hours and 15 hours, and finally approaches zero after 20 hours. The changing trend within 1-24 hours is accurately captured by the model;

[0127] like Figure 3 As shown in the figure, following the natural law of sunrise and sunset, the photovoltaic power generation gradually increases in the morning, reaches a peak of about 700kW at noon, and then gradually decreases, approaching zero in the evening. The model accurately predicts this trend;

[0128] like Figure 4 As shown in the figure, the regional load gradually increased from about 1000kW in the early morning to 1800kW around 6 hours, and then maintained a high level with small fluctuations thereafter, and finally dropped rapidly after 20 hours. The predicted value and the actual value are basically consistent in the overall trend and detailed changes, showing the high accuracy of the model in regional load forecasting.

[0129] Figure 5 The result of the regional load forecast for the distribution network with flexible resource access is shown in the figure. Since the access of wind energy and photovoltaic power significantly affects the original regional load curve in the distribution network, photovoltaic power generation provides power supplement for the distribution network during the day, significantly reducing the load peak-valley difference; while wind power generation provides power support for the distribution network at different time periods throughout the day. Overall, the access of these two renewable energy sources has effectively improved the stability and reliability of power supply.

[0130] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A distribution network load forecasting method based on deep learning, characterized in that: include: Acquire regional load data and renewable energy power generation data, combine the regional load data and renewable energy power generation data into a data set, and pre-process the data set to obtain a first data set; Extracting time series features from the first data set using a convolutional neural network, converting the time series features into a data form of a deep learning model through an embedding layer, and obtaining one-dimensional time series data; Performing a fast Fourier transform on the one-dimensional time series data to obtain a frequency curve, extracting the amplitude value on the frequency curve, calculating the corresponding period according to the amplitude value, and selecting the corresponding period to slice and divide the one-dimensional time series data into a two-dimensional matrix; Using a two-dimensional convolutional network to extract features from the two-dimensional matrix, restoring the two-dimensional matrix after feature extraction into a one-dimensional array, and adaptively fusing the one-dimensional array to obtain a first time series feature; The first time series features are input into the fully connected layer for weight calculation and linear transformation to obtain the distribution network load forecast result.

2. The distribution network load forecasting method based on deep learning as claimed in claim 1, characterized in that: Preprocessing the data set includes: Calculate the first quartile and the third quartile in the data set, and subtract the first quartile from the third quartile to get the interquartile range; The normal value range is 1.5 times the interquartile range extending downward from the first quartile to obtain the lower limit, and 1.5 times the interquartile range extending upward from the third quartile to obtain the upper limit; Values ​​outside the normal range are considered outliers and are removed from the data set; For the missing parts in the data set, select the K value, calculate the Euclidean distance between each missing value sample and other samples, select the K samples closest to it, and use the weighted average of the K nearest neighbor samples to fill in the missing values.

3. The distribution network load forecasting method based on deep learning as claimed in claim 1 or 2, characterized in that: One-dimensional time series data includes: The time series features of each column in the data set are converted into the data form processed by TimesBlock through the embedding layer. The data is transferred between each layer of TimesBlock through residual operation to obtain the input data form of the k-th layer TimesBlock, that is, one-dimensional time series data.

4. The distribution network load forecasting method based on deep learning as claimed in claim 3, characterized in that: Fast Fourier Transform includes: Convert the one-dimensional time series data into a spectrum curve, extract the amplitude value corresponding to each positive frequency on the spectrum curve, select the first n frequency components of the average amplitude according to the setting of the hyperparameters, and find the corresponding period expression as: in, is one-dimensional time series data, FFT(·) is fast Fourier transform, Amp(·) is amplitude function, Avg(·) is average function, {f1,...,f n } are n frequency components, {p1,...,p n } is the corresponding period, W is the average amplitude corresponding to the frequency component, and T is the data length.

5. The distribution network load forecasting method based on deep learning as claimed in claim 4, characterized in that: Select the corresponding period to slice and divide the one-dimensional time series data into a two-dimensional matrix including: According to the size of each period, the one-dimensional time series data is divided into multiple slices, each of which corresponds to the sequence data within a period; Add 0 to the end of the one-dimensional time series data until the length of the one-dimensional time series can be divided by the period, and reorganize the one-dimensional time series to obtain a two-dimensional matrix.

6. The distribution network load forecasting method based on deep learning as claimed in claim 5, characterized in that: Using a two-dimensional convolutional network to extract features from the two-dimensional matrix, and restoring the two-dimensional matrix after feature extraction to a one-dimensional array includes: The two-dimensional matrix is ​​used as input data, and the Inception convolutional network is used to perform convolution processing on the two-dimensional matrix to obtain the two-dimensional matrix after feature extraction; The two-dimensional matrix after feature extraction is reshaped into a one-dimensional vector according to the number of rows and columns of the matrix, and the padding 0 added during the reshaping process is removed to obtain one-dimensional time series data, that is, a one-dimensional array.

7. The distribution network load forecasting method based on deep learning as claimed in claim 6, characterized in that: Adaptively fusing the one-dimensional array comprises: Calculate the weight of each frequency component at k-1 steps and convert the weight into probability distribution through the Softmax function; The weight of each frequency component is multiplied by the corresponding one-dimensional array, and the multiplication results are summed to obtain a weighted one-dimensional time array.

8. A system for distribution network load forecasting method based on deep learning, characterized in that: include: A preprocessing module, used for acquiring regional load data and renewable energy power generation data, combining the regional load data and renewable energy power generation data into a data set, and preprocessing the data set to obtain a first data set; A data conversion module, used to extract time series features from the first data set using a convolutional neural network, and convert the time series features into a data form of a deep learning model through an embedding layer to obtain one-dimensional time series data; A slicing module is used to perform a fast Fourier transform on the one-dimensional time series data to obtain a frequency curve, extract the amplitude value on the frequency curve, calculate the corresponding period according to the amplitude value, and select the corresponding period to slice and divide the one-dimensional time series data into a two-dimensional matrix; A feature extraction module, used to extract features from the two-dimensional matrix using a two-dimensional convolutional network, restore the two-dimensional matrix after feature extraction to a one-dimensional array, and adaptively fuse the one-dimensional array to obtain a first time series feature; The calculation module is used to input the first time series features into the fully connected layer for weight calculation and linear transformation to obtain the distribution network load prediction result.

9. An electronic device, characterized in that: The device comprises: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Citation Information

Cited By

  • Multi-source heterogeneous data fusion and processing method based on localized LLM model

    CN119357880A

  • Power consumption anomaly detection method based on learnable time sequence diagram and periodic convolution fusion

    CN122528002A

  • Electricity consumption anomaly detection method based on fusion of learnable temporal graph and periodic convolution

    CN122528002B

  • Deep learning-based power distribution network load prediction method and system

    WO2026007318A1