A load prediction method and device fusing weather data and self-adapting a drift window

Abstract

This invention discloses a load forecasting method and apparatus that fuses weather data and employs an adaptive drift window, relating to the field of power system load forecasting technology. The method includes: acquiring historical load and weather data, constructing a dataset through preprocessing; calculating the load time-series similarity and weather correlation characteristics under different window lengths, fusing them to obtain a comprehensive similarity, selecting the window with the highest similarity as the optimal drift window, and extracting data from it; decomposing the load and weather sequences into trend, periodic, and noise components, and splicing them according to frequency to generate multiple types of fusion features; independently modeling each type of fusion feature using a corresponding prediction model to obtain branch prediction values ​​for each component; and adaptively weighting and fusing the branch prediction values ​​based on the prediction errors of each model on the validation set to obtain the final load forecast value. This invention fuses load time-series similarity and weather similarity, constructs a comprehensive similarity evaluation system, and dynamically filters suitable historical data to ensure the effectiveness of historical data for current forecasts.

Description

Technical Field

[0001] This invention relates to the field of power system load forecasting technology, and in particular to a load forecasting method and apparatus that integrates weather data and adaptive drift windows. Background Technology

[0002] Electricity load forecasting is a core technical support for ensuring the safe and stable operation of the power grid and optimizing the allocation of energy resources. With the increasing penetration rate of renewable energy, the diversification of user electricity consumption behavior, and the frequent occurrence of extreme weather, load sequences exhibit three major characteristics: strong time variability, multiple periods, and weather sensitivity.

[0003] In existing technologies, traditional methods often use fixed-length historical data windows (such as 720h or 1440h) to extract load features; existing technologies often directly concatenate weather data (temperature, rainfall, wind speed, etc.) as independent features into the load sequence; existing models (such as single LSTM and Transformer) use a unified architecture for modeling and are applied to power system load forecasting scenarios. However, existing technologies have the following drawbacks: (1) Poor time-series adaptability of fixed window: Fixed window cannot dynamically filter valid historical data, resulting in outdated data interfering with prediction accuracy.

[0004] (2) Weather data fusion is superficial: the dynamic nonlinear relationship between weather and load is not explored, and the correspondence between the frequency attributes of weather characteristics and load components cannot be distinguished.

[0005] (3) Insufficient modeling of multi-frequency load characteristics: The existing model adopts a unified architecture and cannot select an appropriate model for the characteristics of different frequency components.

[0006] Therefore, there is an urgent need for a load forecasting method that can dynamically adapt to load drift, deeply integrate weather data, and accurately model multi-frequency load characteristics. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a load forecasting method and apparatus that integrates weather data and adaptive drift windows.

[0008] To solve the above technical problems, the technical solution of the present invention is as follows: A load forecasting method that integrates weather data and adapts to a drift window includes: Historical load and weather characteristics of the area to be predicted are obtained, and the historical load and weather characteristics are preprocessed to construct a standardized input dataset, which is then divided into a training set and a validation set. Load time series similarity and weather correlation characteristics are calculated separately for different candidate historical window lengths, and the two are merged into a comprehensive similarity. The candidate window length with the highest comprehensive similarity is selected as the optimal drift window, and the load feature sequence and weather feature sequence within the optimal drift window are extracted. The variational mode decomposition algorithm is used to decompose the load feature sequence and the weather feature sequence into multiple frequency components, including at least trend components, periodic components and noise components. The frequency components of the load feature sequence and the frequency components of the weather feature sequence are then concatenated according to their frequency attributes to generate multi-class fusion features. For the fusion features of different categories, a matching prediction model is used to perform independent modeling and prediction to obtain the branch prediction value of each frequency component; The prediction error of each prediction model on the validation set is obtained, and the prediction values ​​of each branch are adaptively weighted and fused based on the prediction error to generate the final target load prediction value.

[0009] As a preferred embodiment of the load forecasting method that integrates weather data and adaptive drift windows according to the present invention, the data preprocessing of the historical load characteristics and the weather characteristics includes: Calculate the mean of the original data and standard deviation If the eigenvalue satisfy If the value is not found in the j-th dimension of the i-th sample, it is considered an outlier. For the identified outlier, the mean of adjacent time steps is used for replacement and correction: Let the outlier be located in the j-th dimension of the i-th sample. If this sample is neither the first nor the last sample, the replacement formula is: If the sample is the first or last sample, then the feature value of the adjacent available time is used to replace it; Z-score standardization is used to standardize the corrected data. The standardization formula is as follows: ,in, These are the standardized eigenvalues. Let be the mean of the j-th feature. Let be the standard deviation of the j-th feature.

[0010] As a preferred embodiment of the load forecasting method that integrates weather data and adaptively shifts windows as described in this invention, the method for calculating the load time series similarity is as follows: constructing a distance matrix D between the current window load sequence and the historical window load sequence, and initializing... , , According to the cumulative distance formula Calculate the matrix elements, where The final dynamic time-normalized distance is: (Euclidean distance) The distance is converted into a similarity between 0 and 1 using the following formula: ; The calculation method for the weather correlation characteristics is as follows: Min-max normalization is applied to the current window weather sequence and the historical window weather sequence to eliminate dimensional differences. After normalization, the absolute difference between the two normalized sequences is calculated time-by-time, and a resolution coefficient is defined. The correlation coefficient is: ,in, Let be the absolute difference at time t, and let be the average value of the correlation coefficients. .

[0011] As a preferred embodiment of the load forecasting method that fuses weather data and adapts to drift windows as described in this invention, wherein fusing the two data into a comprehensive similarity score includes: The load time series similarity and the weather-related characteristics Integrate into comprehensive similarity The formula is: , where α is the load time series similarity weight, β is the weather association characteristic weight, and α > β.

[0012] As a preferred embodiment of the load forecasting method that fuses weather data and adapts to a drift window as described in this invention, the method employs a variational mode decomposition algorithm to decompose the load feature sequence and the weather feature sequence into multiple frequency components, including at least trend components, periodic components, and noise components. The frequency components of the load feature sequence are then concatenated with the frequency components of the weather feature sequence according to their frequency attributes to generate multi-category fused features, including: Variational mode decomposition algorithm is used for single-dimensional sequences Decomposed into K=3 intrinsic mode functions ,satisfy: ,in, The center frequencies of each intrinsic mode function are... It's a countdown to the end of the day. This is a convolution operation; The variational mode decomposition algorithm is applied dimension-by-dimensionally to the load feature matrix and the weather feature matrix to obtain the trend terms of the load features. Periodic terms Noise item and trend items of weather characteristics Periodic terms Noise item ; The same frequency components of load characteristics and weather characteristics are spliced ​​together to form a fused feature: ,in, , , These are trend fusion features, periodic fusion features, and noise fusion features, respectively.

[0013] As a preferred embodiment of the load forecasting method that fuses weather data and adapts to a drift window as described in this invention, the step of independently modeling and forecasting different categories of the fused features using matching prediction models to obtain branch prediction values ​​for each frequency component includes: Regarding the aforementioned trend fusion features The multilayer perceptron is used as the prediction model, and its forward propagation formula is: ,in, , Here is the weight matrix, and b1 and b2 are the bias terms. For activation function, To output the predicted value of the trend term; Regarding the aforementioned periodic fusion features A bidirectional gated cyclic unit is used as the prediction model, and the output is the predicted value of the periodic term. Its hidden state update formula is: ,in, Let be the input features at time t. Let be the hidden state at time t. It is the sigmoid activation function. For element-wise multiplication, the bidirectional structure includes both forward and backward GRU computations. Regarding the noise fusion characteristics Extreme gradient boosting trees are used as the prediction model, and the output is the predicted value of the noise term. Its objective function is: ,in, For loss function, Here, T is the number of leaf nodes in the tree, and T is the regularization term. The weight of the leaf node. and For regularization parameters.

[0014] As a preferred embodiment of the load forecasting method for fusing weather data and adaptively shifting windows according to the present invention, the step of obtaining the prediction error of each prediction model on the validation set, and adaptively weighting and fusing the branch prediction values ​​based on the prediction error to generate the final target load forecast value includes: The prediction error is the mean absolute error of each prediction model on the validation set. The weights of the adaptive weighted fusion for ,in, ; The final target load forecast after merging for: .

[0015] The present invention also provides a load forecasting device that fuses weather data and adapts to a drift window, comprising: The data preprocessing module is used to obtain historical load characteristics and weather characteristics of the area to be predicted, perform data preprocessing on the historical load characteristics and weather characteristics, construct a standardized input dataset, and divide the standardized input dataset into a training set and a validation set. The drift window determination module is used to calculate the load time series similarity and weather correlation characteristics under different candidate historical window lengths, and merge the two into a comprehensive similarity. The candidate window length with the highest comprehensive similarity is selected as the optimal drift window, and the load feature sequence and weather feature sequence within the optimal drift window are extracted. The feature decomposition and fusion module is used to decompose the load feature sequence and the weather feature sequence into multiple frequency components, including at least trend components, periodic components and noise components, respectively, using a variational mode decomposition algorithm, and to concatenate the frequency components of the load feature sequence with the frequency components of the weather feature sequence according to the frequency attributes to generate multi-class fusion features. The branch prediction module is used to independently model and predict the fusion features of different categories using matching prediction models to obtain the branch prediction values ​​of each frequency component. The prediction result integration module is used to obtain the prediction error of each prediction model on the validation set, and to perform adaptive weighted fusion of the prediction values ​​of each branch based on the prediction error to generate the final target load prediction value.

[0016] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in any of the above-described load forecasting methods that fuse weather data and adaptive drift windows.

[0017] The present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that: when the program is executed by a processor, it implements the method described in any of the above-mentioned load forecasting methods that fuse weather data and adaptively shift windows.

[0018] The beneficial effects of this invention are: (1) This invention utilizes a dynamic generation mechanism with a drift window to integrate load temporal similarity and weather similarity, constructing a comprehensive similarity evaluation system. It dynamically filters historical data that is highly compatible with the current scenario, and the window length can be adaptively adjusted within the range of 240h to 1440h. Compared with the traditional fixed window method, in load drift scenarios (such as changes in cooling / heating load characteristics during seasonal transitions), the prediction error is reduced by 15% to 20%, and the temporal adaptability is improved by more than 30%.

[0019] (2) This invention employs a variational mode decomposition algorithm to decompose both weather and load sequences into three types of frequency components: trend, period, and noise, establishing a one-to-one correspondence between "weather frequency components and load frequency components." Compared to the traditional direct splicing method, in extreme weather scenarios (such as high temperatures ≥35℃ and daily rainfall ≥50mm), the prediction accuracy is improved by 20%~25%, and the utilization rate of weather features is improved by 40%.

[0020] (3) This invention designs a multi-branch adaptive modeling architecture, selecting an adaptive model for the "load-weather" fusion features of different frequencies: Multilayer Perceptron (MLP) is used to adapt to trend terms (linear features), Bidirectional Gated Recurrent Unit (BiGRU) is used to adapt to periodic terms (time-dependent features), and Extreme Gradient Boosting Tree (XGBoost) is used to adapt to noise terms (nonlinear burst features). Under diverse scenarios such as weekdays / weekends, peak and off-peak periods, and extreme weather, the mean absolute error (MAE) is stable within 5%, while the MAE of traditional single models often exceeds 10% under extreme weather conditions.

[0021] (4) Based on conventional load data and meteorological data, this invention requires no additional hardware deployment and the prediction time for a single sample is less than 100ms. At the same time, based on the adaptive weighted fusion strategy of validation set error, the weights are dynamically allocated and normalized, and the optimal fusion can be achieved without manual parameter tuning, which significantly reduces the threshold for engineering applications. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 A flowchart illustrating the load forecasting method that integrates weather data and adaptive drift windows provided by this invention; Figure 2 A schematic diagram illustrating the specific process of the load forecasting method that integrates weather data and adaptive drift windows provided by the present invention; Figure 3The network structure diagram of the load prediction model that integrates weather data and adaptive drift window provided by the present invention; Figure 4 A schematic diagram of the load forecasting device that integrates weather data and adaptive drift window provided by the present invention; Figure 5 A schematic diagram of the computer device provided by the present invention. Detailed Implementation

[0024] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0025] See Figure 1 This embodiment provides a load forecasting method that integrates weather data and an adaptive drift window. The method specifically includes the following steps: Step S1: Data preprocessing.

[0026] Specifically, load data is collected from real-time monitoring and historical records from the power grid dispatching platform, including 12-dimensional load correlation characteristics: historical load sequence (power factor, transformer utilization rate, hourly load, hourly maintenance plan, current, power, failure rate, reactive power compensation value, load growth rate, concurrent load, and load structure proportion). Hourly weather data is collected from the meteorological monitoring platform: temperature, rainfall, wind speed, and irradiance, totaling 4-dimensional weather characteristics.

[0027] Let the load data matrix be (N is the number of samples, 12 is the load characteristic dimension), the weather data matrix is ​​as follows: (4 represents the weather data dimension), which is then combined into a feature matrix X by column concatenation: .

[0028] Meanwhile, the target load value vector It is generated by a nonlinear combination of load data, and the formula is: ,in, This represents the j-th dimension of the load feature of the i-th sample. This represents the k-th dimension weather feature of the i-th sample. This is the random noise term (simulating uncertainties in actual operating conditions).

[0029] For outliers in the sequence, use Criterion identification: Calculate the mean of the raw data and standard deviation If the eigenvalue satisfy If the value is not found in the j-th dimension of the i-th sample, it is considered an outlier. The outlier is corrected using the average of adjacent time steps: If the outlier is located in the j-th dimension of the i-th sample, and is neither the first nor the last sample, then the replacement formula is: If it is the first or last position, the feature value of the adjacent available time is used for replacement.

[0030] To eliminate dimensional differences, Z-score standardization is used to process the cleaned features. The standardization formula is as follows: ,in, These are the standardized eigenvalues. and , where are the mean and standard deviation of the j-th feature, respectively. After standardization, the data has a mean of 0 and a standard deviation of 1. The dataset is divided into training and validation sets in an 8:2 ratio.

[0031] Step S2: Dynamically generate the drift window.

[0032] The dynamic generation of the drift window is designed to account for the multi-timescale variations in cover load and weather. The window length is set to range from 240 to 1440 hours (10 to 60 days), with a step size of 24 hours (1 day), ensuring that the adjustment granularity of the window length matches the diurnal cycle characteristics. Let the window length variable be... ,but Load data exhibits long-term temporal dependencies (such as daily peaks and troughs, and weekly cycles). The DTW algorithm is used to calculate the similarity of load sequences in different windows, which can effectively overcome the sensitivity of traditional Euclidean distance to temporal misalignment.

[0033] For the current prediction time index Extract the current window load sequence and historical window load sequence Load similarity is calculated using Dynamic Time Warping (DTW): [Construction] Given a distance matrix D, initialize D(0,0)=0, D(i,0)=∞, D(0,j)=∞, and use the cumulative distance formula: Calculate the matrix elements, where The final dynamic time-normalized distance is: (Euclidean distance) The smaller the DTW distance, the higher the sequence similarity. The distance can be converted into a similarity interval of 0-1 using the following formula: .

[0034] For weather data, there is a strong nonlinear correlation with load. The GRA (Grey Weighted Relationship Analysis) algorithm can quantify the degree of correlation between two sets of weather sequences, avoiding the limitations of comparing a single feature. Specifically, it uses grey relational analysis (GRA) to calculate weather similarity: min-max normalization is performed on the current window weather sequence and the historical window weather sequence to eliminate dimensional differences. .

[0035] After normalization, the absolute difference between the two normalized sequences is calculated time-by-time. Determine Δmin and Δmax. The correlation coefficient is: The resolution coefficient The value is 0.5. The weather correlation characteristic is the average of the correlation coefficients over all time periods: .

[0036] Since the task focuses on load forecasting, the weight of load time-series characteristics in the forecast is higher than that of weather correlation characteristics. The load similarity weight is set as α, and the weather similarity weight is set as β. The comprehensive similarity formula is: In this embodiment, α=0.6 and β=0.4 (α>β).

[0037] For each candidate window length Calculate their overall similarity , choose to The maximum window length is used as the optimal drift window length. .

[0038] Determine the optimal window length Then, extract the complete data from the window, including: Window time range: ; Load characteristics: ; Weather characteristics: ; Target load: ; This data is then encapsulated into a dictionary format: .

[0039] Step S3: Multi-frequency feature decomposition and fusion.

[0040] The optimal adaptation window data for the output (including load characteristics) Weather characteristics The variational mode decomposition (VMD) algorithm is used to extract multi-scale frequency components, providing feature support for subsequent multi-branch modeling.

[0041] First, the input data consists of structured data within the optimal window, where: The load characteristic matrix is ​​as follows: ( (where 12 represents the optimal window length and the load characteristic dimension). The weather feature matrix is ​​as follows: (4 represents weather dimension features).

[0042] The decomposition goal is to break down the two types of features into three components based on frequency characteristics: trend components (low frequency, long-term changes), periodic components (medium frequency, medium-term cycles), and noise components (high frequency, short-term fluctuations), and to establish a one-to-one correspondence between "weather frequency components and load frequency components".

[0043] Variational Mode Decomposition (VMD) algorithm decomposes signals by iteratively optimizing a variational model for each dimension of the sequence. (Such as time series data of a certain load characteristic or weather characteristic), which is decomposed into three intrinsic mode functions (IMF) by constructing a variational problem. It satisfies the following constraints: ,in, The center frequencies of each intrinsic mode function are... It's a countdown to the end of the day. This is a convolution operation. The constraints ensure that the sum of the IMFs after decomposition equals the original sequence.

[0044] Therefore, for 12-dimensional load characteristics and Four-dimensional weather features, applying VMD (parameter: K=3, bandwidth control coefficient) dimension by dimension. Convergence tolerance This process yields three frequency components for each feature. Then, the trend, periodic, and noise components of the 12-dimensional feature are integrated into... , , Similarly, the three components of weather characteristics can be obtained. , , Based on the correspondence between "trend items - trend items", "periodic items - periodic items", and "noise items - noise items", the same frequency components of load and weather are spliced ​​together to form a fused feature: .

[0045] The final output consists of three types of fused features. , , The data flowed into the MLP, BiGRU, and XGBoost branches respectively for adaptation modeling.

[0046] Step S4: Multi-branch fusion modeling.

[0047] For the three types of fused features (trend, period, and noise) output by the multi-frequency feature decomposition module, different models are used for adaptation modeling to give full play to the advantages of each model in processing the corresponding frequency components, and finally output the prediction results of each branch.

[0048] First, the three types of fusion features in the above-mentioned multi-frequency feature decomposition module are: trend term fusion features. (12-dimensional load + 4-dimensional weather), periodic term fusion features Noise Term Fusion Features The module assigns different prediction branches to the three signals, adapting them to three types of frequency components. The three branches are: an MLP (Multilayer Perceptron) branch to adapt to the linear characteristics of the trend term; a BiGRU branch to adapt to the time-dependent characteristics of the periodic term; and an XGBoost branch to adapt to the nonlinear burst characteristics of the noise term. Specifically: A Multilayer Perceptron (MLP) fits a linear relationship between the trend term through fully connected layers. The model structure is "input layer → hidden layer → output layer". Let the input be... Output the predicted value of the trend term. Its forward propagation formula is: ,in, , Here is the weight matrix, and b1 and b2 are the bias terms. This is the activation function.

[0049] The Bidirectional Gated Recurrent Unit (BiGRU) captures the temporal dependency of periodic terms through a bidirectional recurrent structure. The model structure is "input layer → BiGRU layer → fully connected output layer". Let the input be... The output is the predicted value of the periodic term. Its hidden state update formula is: ,in, Let be the input features at time t. Let be the hidden state at time t. It is the sigmoid activation function. For element-wise multiplication, the bidirectional structure includes both forward and backward GRU computations.

[0050] Extreme Gradient Boosting (XGBoost) fits the nonlinear relationship of the noise term by ensembling multiple decision trees. The model objective function is: ,in, For loss function, Here, T is the number of leaf nodes in the tree, and T is the regularization term. The weight of the leaf node. and This is the regular expression parameter. The input is... The output is the predicted value of the noise term. .

[0051] By leveraging the synergy of multiple models, the linear fitting capability of MLP, the time-series modeling capability of BiGRU, and the nonlinear fitting capability of XGBoost are fully utilized to provide accurate prediction outputs for each frequency component.

[0052] Step S5: Integrate prediction results.

[0053] This step performs weighted fusion of the three types of prediction results (trend, period, and noise predictions) output by the multi-branch fusion modeling module to generate the final load prediction. The fusion weights are then dynamically optimized through error feedback to improve prediction accuracy.

[0054] The output of each item in step S4 ( , , The integration goal is to merge the three types of forecasts into a final load forecast result through dynamic weight allocation. This minimizes the error (such as root mean square error) in the fusion result. A dynamic weighted fusion strategy is adopted, where the weights are adaptively adjusted based on the historical prediction errors of each branch (the smaller the error, the higher the weight): Let the k-th branch... The historical average absolute error is Then its weight for: ,in, This ensures that the weights are normalized. Therefore, the final target load forecast after fusion... for: Output The forecast results are used for power grid dispatching or power trading decisions.

[0055] See Figure 4 This application also provides a load forecasting device that fuses weather data and adapts to a drift window. The device includes: a data preprocessing module, a drift window determination module, a feature decomposition and fusion module, a branch forecasting module, and a forecast result integration module.

[0056] Specifically, the data preprocessing module is used to obtain historical load characteristics and weather characteristics of the area to be predicted, and to preprocess the historical load characteristics and weather characteristics to construct a standardized input dataset, and then divide the standardized input dataset into a training set and a validation set.

[0057] The drift window determination module is used to calculate the load time series similarity and weather correlation characteristics under different candidate historical window lengths, and integrate the two into a comprehensive similarity. The candidate window length with the highest comprehensive similarity is selected as the optimal drift window, and the load feature sequence and weather feature sequence within the optimal drift window are extracted.

[0058] The feature decomposition and fusion module is used to decompose the load feature sequence and the weather feature sequence into multiple frequency components, including at least trend components, periodic components and noise components, using the variational mode decomposition algorithm. The frequency components of the load feature sequence are then concatenated with the frequency components of the weather feature sequence according to their frequency attributes to generate multi-class fused features.

[0059] The branch prediction module is used to independently model and predict different categories of fusion features using matching prediction models, thereby obtaining the branch prediction values ​​for each frequency component.

[0060] The prediction results integration module is used to obtain the prediction error of each prediction model on the validation set, and to adaptively weight and fuse the prediction values ​​of each branch based on the prediction error to generate the final target load prediction value.

[0061] See Figure 5 This embodiment also provides a computer device, the components of which may include, but are not limited to: one or more processors or processing units, system memory, and buses connecting different system components (including system memory and processing units).

[0062] A bus refers to one or more of several bus architectures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration port, a processor, or a local bus using any of the various bus architectures. Examples of these architectures include, but are not limited to, the Industry Standard Architecture (ISA) bus, the Micro Channel Architecture (MAC) bus, the Enhanced ISA bus, the Video Electronics Standards Association (VESA) local bus, and the Peripheral Component Interconnect (PCI) bus.

[0063] Computer systems / servers typically include a variety of computer system-readable media. These media can be any available media that can be accessed by the computer system / server, including volatile and non-volatile media, and removable and non-removable media.

[0064] System memory may include computer system readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory. The computer device may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, the storage system may be used to read and write non-removable, non-volatile magnetic media. Disk drives for reading and writing to removable non-volatile disks (e.g., "floppy disks") and optical disc drives for reading and writing to removable non-volatile optical discs (e.g., CD-ROMs, DVD-ROMs, or other optical media) may be provided. In these cases, each drive may be connected to a bus via one or more data media interfaces. The memory may include at least one program product having a set (e.g., at least one) of program modules configured to perform the functions of the embodiments of the present invention.

[0065] A program / utility having a set (at least one) of program modules can be stored, for example, in memory. Such program modules include—but are not limited to—an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment. The program modules typically perform the functions and / or methods described in the embodiments of this invention.

[0066] Computer devices can also communicate with one or more external devices (such as keyboards, pointing devices, monitors, etc.). This communication can be done through input / output (I / O) interfaces. Furthermore, computer devices can communicate with one or more networks (such as local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) via network adapters.

[0067] The processing unit executes the functions and / or methods described in the embodiments of the present invention by running programs stored in the system memory.

[0068] The aforementioned computer program can be stored in a computer storage medium, that is, the computer storage medium is encoded with a computer program, which, when executed by one or more computers, causes one or more computers to perform the method flow and / or device operation shown in the above embodiments of the present invention.

[0069] With the development of time and technology, the meaning of "medium" has become increasingly broad. The dissemination of computer programs is no longer limited to tangible media; they can also be downloaded directly from the network. Any combination of one or more computer-readable media can be used. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. Computer-readable storage media can be, for example,—but not limited to—electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0070] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including—but not limited to—electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of transmitting, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0071] In addition to the above embodiments, the present invention may have other implementation methods; all technical solutions formed by equivalent substitution or equivalent transformation fall within the protection scope claimed by the present invention.

Claims

1. A load forecasting method that integrates weather data and adapts to a drift window, characterized in that: include: Historical load and weather characteristics of the area to be predicted are obtained, and the historical load and weather characteristics are preprocessed to construct a standardized input dataset, which is then divided into a training set and a validation set. Load time series similarity and weather correlation characteristics are calculated separately for different candidate historical window lengths, and the two are merged into a comprehensive similarity. The candidate window length with the highest comprehensive similarity is selected as the optimal drift window, and the load feature sequence and weather feature sequence within the optimal drift window are extracted. The variational mode decomposition algorithm is used to decompose the load feature sequence and the weather feature sequence into multiple frequency components, including at least trend components, periodic components and noise components. The frequency components of the load feature sequence and the frequency components of the weather feature sequence are then concatenated according to their frequency attributes to generate multi-class fusion features. For the fusion features of different categories, a matching prediction model is used to perform independent modeling and prediction to obtain the branch prediction value of each frequency component; The prediction error of each prediction model on the validation set is obtained, and the prediction values ​​of each branch are adaptively weighted and fused based on the prediction error to generate the final target load prediction value.

2. The load forecasting method based on fusion of weather data and adaptive drift window as described in claim 1, characterized in that: Data preprocessing for the historical load characteristics and the weather characteristics includes: Calculate the mean of the original data and standard deviation If the eigenvalue satisfy If the value is not found in the j-th dimension of the i-th sample, it is considered an outlier. For the identified outlier, the mean of adjacent time steps is used for replacement and correction: Let the outlier be located in the j-th dimension of the i-th sample. If this sample is neither the first nor the last sample, the replacement formula is: If the sample is the first or last sample, then the feature value of the adjacent available time is used to replace it; Z-score standardization is used to standardize the corrected data. The standardization formula is as follows: ,in, These are the standardized eigenvalues. Let be the mean of the j-th feature. Let be the standard deviation of the j-th feature.

3. The load forecasting method based on fusion of weather data and adaptive drift window as described in claim 1, characterized in that: The method for calculating the load time series similarity is as follows: construct the distance matrix D between the current window load sequence and the historical window load sequences, and initialize... , , According to the cumulative distance formula Calculate the matrix elements, where The final dynamic time-normalized distance is: (Euclidean distance) The distance is converted into a similarity between 0 and 1 using the following formula: ; The calculation method for the weather correlation characteristics is as follows: Min-max normalization is applied to the current window weather sequence and the historical window weather sequence to eliminate dimensional differences. After normalization, the absolute difference between the two normalized sequences is calculated time-by-time, and a resolution coefficient is defined. The correlation coefficient is: ,in, Let be the absolute difference at time t, and let be the average value of the correlation coefficients. .

4. The load forecasting method based on fusion of weather data and adaptive drift window as described in claim 3, characterized in that: The fusion of the two into a comprehensive similarity includes: The load time series similarity and the weather-related characteristics Integrate into comprehensive similarity The formula is: , where α is the load time series similarity weight, β is the weather association characteristic weight, and α > β.

5. The load forecasting method based on fusion of weather data and adaptive drift window as described in claim 1, characterized in that: The variational mode decomposition algorithm is used to decompose the load feature sequence and the weather feature sequence into multiple frequency components, including at least trend components, periodic components, and noise components. The frequency components of the load feature sequence and the frequency components of the weather feature sequence are then concatenated according to their frequency attributes to generate multi-class fusion features, including: Variational mode decomposition algorithm is used for single-dimensional sequences Decomposed into K=3 intrinsic mode functions ,satisfy: ,in, The center frequencies of each intrinsic mode function are... It's a countdown to the end of the day. This is a convolution operation; The variational mode decomposition algorithm is applied dimension-by-dimensionally to the load feature matrix and the weather feature matrix to obtain the trend terms of the load features. Periodic terms Noise item and trend items of weather characteristics Periodic terms Noise item ; The same frequency components of load characteristics and weather characteristics are spliced ​​together to form a fused feature: ,in, , , These are trend fusion features, periodic fusion features, and noise fusion features, respectively.

6. The load forecasting method based on fusion of weather data and adaptive drift window as described in claim 5, characterized in that: The fusion features of different categories are independently modeled and predicted using matching prediction models to obtain branch prediction values ​​for each frequency component, including: Regarding the aforementioned trend fusion features The multilayer perceptron is used as the prediction model, and its forward propagation formula is: ,in, , Here is the weight matrix, and b1 and b2 are the bias terms. For activation function, To output the predicted value of the trend term; Regarding the aforementioned periodic fusion features A bidirectional gated cyclic unit is used as the prediction model, and the output is the predicted value of the periodic term. Its hidden state update formula is: ,in, Let be the input features at time t. Let be the hidden state at time t. It is the sigmoid activation function. For element-wise multiplication, the bidirectional structure includes both forward and backward GRU computations. Regarding the noise fusion characteristics Extreme gradient boosting trees are used as the prediction model, and the output is the predicted value of the noise term. Its objective function is: ,in, For loss function, Here, T is the number of leaf nodes in the tree, and T is the regularization term. The weight of the leaf node. and For regularization parameters.

7. The load forecasting method based on fusion of weather data and adaptive drift window as described in claim 6, characterized in that: The step of obtaining the prediction error of each prediction model on the validation set, and adaptively weighting and fusing the prediction values ​​of each branch based on the prediction error to generate the final target load prediction value includes: The prediction error is the mean absolute error of each prediction model on the validation set. The weights of the adaptive weighted fusion for ,in, ; The final target load forecast after merging for: .

8. A load forecasting device that integrates weather data and adapts to a drift window, characterized in that: include: The data preprocessing module is used to obtain historical load characteristics and weather characteristics of the area to be predicted, perform data preprocessing on the historical load characteristics and weather characteristics, construct a standardized input dataset, and divide the standardized input dataset into a training set and a validation set. The drift window determination module is used to calculate the load time series similarity and weather correlation characteristics under different candidate historical window lengths, and merge the two into a comprehensive similarity. The candidate window length with the highest comprehensive similarity is selected as the optimal drift window, and the load feature sequence and weather feature sequence within the optimal drift window are extracted. The feature decomposition and fusion module is used to decompose the load feature sequence and the weather feature sequence into multiple frequency components, including at least trend components, periodic components and noise components, respectively, using a variational mode decomposition algorithm, and to concatenate the frequency components of the load feature sequence with the frequency components of the weather feature sequence according to the frequency attributes to generate multi-class fusion features. The branch prediction module is used to independently model and predict the fusion features of different categories using matching prediction models to obtain the branch prediction values ​​of each frequency component. The prediction result integration module is used to obtain the prediction error of each prediction model on the validation set, and to perform adaptive weighted fusion of the prediction values ​​of each branch based on the prediction error to generate the final target load prediction value.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it implements the method as described in any one of claims 1 to 7.

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

Images