Loss reduction method and device for adjusting phase of single-phase user node, equipment and medium

Abstract

The invention relates to a loss reduction method and device for adjusting the phase of a single-phase user node, equipment and a medium. The method comprises the following steps: acquiring an initial source load power data feature, a meteorological data feature and a time data feature of a transformer area, obtaining a data interpolation result according to the initial source load power data feature and a data interpolation model, and updating the initial source load power data feature into a source load power data feature by using the data interpolation result, and inputting the source load power data features, the meteorological data features and the time data features into a prediction model to obtain dynamic and static data features, obtaining prediction data corresponding to the transformer area according to the dynamic and static data features, and making a regulation and control scheme according to the prediction data to adjust the phase of the single-phase user node of the transformer area. Data interpolation is carried out, the data availability is improved, data features are extracted through a parallel architecture, the accuracy of prediction data is improved, meanwhile, phase adjustment is carried out, and the operation stability of a power grid is guaranteed.

Description

Technical Field

[0001] The present application relates to the technical field of electric power energy, and particularly relates to a loss reduction method, device, computer device, computer-readable storage medium, and computer program product for adjusting the phase of single-phase user nodes. Background Art

[0002] With the rapid development of new energy and the increasing complexity of power system operation, short-term forecasting of grid load and photovoltaic power generation has gradually become an important research direction for ensuring grid stability and improving energy utilization efficiency. Specifically, it is to use the forecasting results of grid load and photovoltaic power generation to reduce losses by adjusting the phase of single-phase user nodes in the power grid.

[0003] However, the current loss reduction method for adjusting the phase of single-phase user nodes has the problem of low accuracy in phase adjustment. Summary of the Invention

[0004] Based on this, it is necessary to provide a loss reduction method, device, computer device, computer-readable storage medium, and computer program product for adjusting the phase of single-phase user nodes, which can improve the accuracy of phase adjustment, in view of the above technical problems.

[0005] In a first aspect, the present application provides a loss reduction method for adjusting the phase of single-phase user nodes, including:

[0006] Obtain the source-load power data, meteorological information, and time information of the distribution area, and obtain the initial source-load power data characteristics corresponding to the source-load power data, as well as the meteorological data characteristics and time data characteristics corresponding to the meteorological information and time information respectively;

[0007] According to the initial source-load power data characteristics and a pre-constructed data interpolation model based on the self-attention mechanism, obtain a data interpolation result, and use the data interpolation result to update the initial source-load power data characteristics to source-load power data characteristics;

[0008] Input the source-load power data characteristics, meteorological data characteristics, and time data characteristics into a pre-constructed source-load power data prediction model, obtain static data characteristics through the convolutional neural network unit in the source-load power data prediction model, and obtain dynamic data characteristics through the gated recurrent unit in the source-load power data prediction model, and obtain the predicted source-load power data corresponding to the distribution area according to the static data characteristics and dynamic data characteristics; the static data characteristics are used to characterize data characteristics with stability and repeatability, and the dynamic data characteristics are used to describe data characteristics that reflect temporal dependence and change trends;

[0009] Formulate a control plan based on the predicted source-load power data, and adjust the phase of the single-phase user nodes in the distribution area to achieve line loss reduction.

[0010] In one embodiment, the data imputation model includes a first diagonal masked self-attention module, a second diagonal masked self-attention module, and a dynamic weighted combination module;

[0011] According to the initial source-load power data features and the pre-constructed data imputation model based on the self-attention mechanism, the data imputation result is obtained, including:

[0012] Obtain the missing mask matrix pre-constructed according to the source-load power data, and perform mask processing on the initial source-load power data features to obtain the primary source-load power data features;

[0013] Concatenate the missing mask matrix and the primary source-load power data features to obtain the secondary source-load power data features;

[0014] Input the secondary source-load power data features into the first diagonal masked self-attention module to obtain the first-stage imputation result;

[0015] Use the first-stage imputation result to update the primary source-load power data features, and concatenate the updated primary source-load power data features and the missing mask matrix to obtain the advanced source-load power data features;

[0016] Input the advanced source-load power data features into the second diagonal masked self-attention module to obtain the second-stage imputation result;

[0017] Input the missing mask matrix, the first-stage imputation result, and the second-stage imputation result into the dynamic weighted combination module to obtain the data imputation result.

[0018] In one embodiment, the first diagonal masked self-attention module includes a first feature projection unit, a first feature enhancement unit, and a first feature restoration unit; the first feature enhancement unit is constructed based on the multi-head self-attention mechanism;

[0019] Input the secondary source-load power data features into the first diagonal masked self-attention module to obtain the first-stage imputation result, including:

[0020] Input the secondary source-load power data features into the first feature projection unit to obtain the secondary source-load power data features in the target dimension;

[0021] Input the secondary source-load power data features in the target dimension into the first feature enhancement unit to obtain the secondary source-load power data features with enhanced features;

[0022] Input the secondary source-load power data features with enhanced features into the first feature restoration unit to obtain the first-stage imputation result.

[0023] In one embodiment, the second diagonal masked self-attention module includes a second feature projection unit, a second feature enhancement unit, and a second feature restoration unit; the second feature enhancement unit is constructed based on the multi-head self-attention mechanism;

[0024] Input the high-level source-load power data features into the second diagonal masked self-attention module to obtain the second-stage interpolation result, including:

[0025] Input the high-level source-load power data features into the second feature projection unit to obtain the high-level source-load power data features of the target dimension;

[0026] Input the high-level source-load power data features of the target dimension into the second feature enhancement unit to obtain the high-level source-load power data features after feature enhancement;

[0027] Input the high-level source-load power features after feature enhancement into the second feature restoration unit to obtain the second-stage interpolation result.

[0028] In an exemplary embodiment, the second-stage interpolation result carries the weight information of the second feature enhancement unit;

[0029] Input the missing mask matrix, the first-stage interpolation result, and the second-stage interpolation result into the dynamic weighted combination module to obtain the data interpolation result, including:

[0030] Obtain the weight factor according to the weight information and the missing mask matrix;

[0031] Use the weight factor to perform weighted summation on the first-stage interpolation result and the second-stage interpolation result to obtain the data interpolation result.

[0032] In one embodiment, the initial source-load power data features include source-load trend features, source-load seasonal features, and average daily source-load features;

[0033] Obtain the initial source-load power data features corresponding to the source-load power data, including:

[0034] Generate a corresponding time series based on the initial source-load power data, and obtain the average daily source-load features according to the time series; the time series includes a source-load trend component, a source-load seasonal component, and a source-load residual component;

[0035] Subtract the preset source-load trend component from the time series to obtain the detrended time series;

[0036] Decompose the detrended time series according to a preset period to obtain multiple time subsequences, and obtain a new source-load seasonal component according to the time subsequences;

[0037] Subtract the new source-load seasonal component from the detrended time series to obtain the deseasonalized time series;

[0038] Obtain the source-load trend characteristics and source-load seasonal characteristics based on the deseasonalized time series and the new source-load seasonal components.

[0039] In one embodiment, obtaining the source-load trend characteristics and source-load seasonal characteristics based on the deseasonalized time series and the new source-load seasonal components includes:

[0040] Perform locally weighted regression filtering on the deseasonalized time series, and extract a new source-load trend component;

[0041] If the similarity between the new source-load trend component and the source-load trend component is greater than or equal to a preset first similarity, and the new source-load seasonal component and the source-load seasonal component are greater than or equal to a preset second similarity, then determine the new source-load trend component as the source-load trend characteristic and the new source-load seasonal component as the source-load seasonal characteristic.

[0042] In a second aspect, the present application also provides a loss reduction device for adjusting the phase of a single-phase user node, including:

[0043] An acquisition module, configured to acquire the source-load power data, meteorological information, and time information of the substation area, and acquire the initial source-load power data characteristics corresponding to the source-load power data, as well as the meteorological data characteristics and time data characteristics corresponding to the meteorological information and time information respectively;

[0044] An interpolation module, configured to obtain an interpolation result according to the initial source-load power data characteristics and a pre-constructed data interpolation model based on the self-attention mechanism, and update the initial source-load power data characteristics to source-load power data characteristics by using the interpolation result;

[0045] A prediction module, configured to input the source-load power data characteristics, meteorological data characteristics, and time data characteristics into a pre-constructed source-load power data prediction model, obtain static data characteristics through the convolutional neural network unit in the source-load power data prediction model, and obtain dynamic data characteristics through the gated recurrent unit in the source-load power data prediction model, and obtain the predicted source-load power data corresponding to the substation area according to the static data characteristics and the dynamic data characteristics; the static data characteristics are used to characterize data characteristics with stability and repeatability, and the dynamic data characteristics are used to describe data characteristics reflecting time series dependence and change trends;

[0046] An adjustment module, configured to formulate a regulation plan based on the predicted source-load power data, and adjust the phase of the single-phase user node in the substation area to achieve line loss reduction.

[0047] In a third aspect, the present application also provides a computer device, including a memory and a processor, where the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0048] Obtain the source-load power data, meteorological information, and time information of the substation area, and obtain the initial source-load power data characteristics corresponding to the source-load power data, as well as the meteorological data characteristics and time data characteristics corresponding to the meteorological information and time information respectively;

[0049] According to the initial source-load power data characteristics and a pre-constructed data interpolation model based on the self-attention mechanism, obtain the data interpolation result, and use the data interpolation result to update the initial source-load power data characteristics to source-load power data characteristics;

[0050] Input the source-load power data characteristics, meteorological data characteristics, and time data characteristics into a pre-constructed source-load power data prediction model. Obtain static data characteristics through the convolutional neural network unit in the source-load power data prediction model, and obtain dynamic data characteristics through the gated recurrent unit in the source-load power data prediction model. And according to the static data characteristics and dynamic data characteristics, obtain the predicted source-load power data corresponding to the substation area; Static data characteristics are used to characterize data characteristics with stability and repeatability, and dynamic data characteristics are used to describe data characteristics that reflect time series dependence and change trends;

[0051] Formulate a regulation plan based on the predicted source-load power data, and adjust the phases of single-phase user nodes in the substation area to achieve line loss reduction.

[0052] In a fourth aspect, the present application also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0053] Obtain the source-load power data, meteorological information, and time information of the substation area, and obtain the initial source-load power data characteristics corresponding to the source-load power data, as well as the meteorological data characteristics and time data characteristics corresponding to the meteorological information and time information respectively;

[0054] According to the initial source-load power data characteristics and a pre-constructed data interpolation model based on the self-attention mechanism, obtain the data interpolation result, and use the data interpolation result to update the initial source-load power data characteristics to source-load power data characteristics;

[0055] Input the source-load power data characteristics, meteorological data characteristics, and time data characteristics into a pre-constructed source-load power data prediction model. Obtain static data characteristics through the convolutional neural network unit in the source-load power data prediction model, and obtain dynamic data characteristics through the gated recurrent unit in the source-load power data prediction model. And according to the static data characteristics and dynamic data characteristics, obtain the predicted source-load power data corresponding to the substation area; Static data characteristics are used to characterize data characteristics with stability and repeatability, and dynamic data characteristics are used to describe data characteristics that reflect time series dependence and change trends;

[0056] Based on the predicted source-load power data, a regulation plan is formulated to adjust the phases of single-phase user nodes in the substation area to achieve line loss reduction.

[0057] Fifthly, the present application also provides a computer program product, including a computer program, which when executed by a processor implements the following steps:

[0058] Obtain the source-load power data, meteorological information, and time information of the substation area, and obtain the initial source-load power data characteristics corresponding to the source-load power data, as well as the meteorological data characteristics and time data characteristics corresponding to the meteorological information and time information respectively;

[0059] According to the initial source-load power data characteristics and the pre-constructed data interpolation model based on the self-attention mechanism, obtain the data interpolation result, and use the data interpolation result to update the initial source-load power data characteristics to source-load power data characteristics;

[0060] Input the source-load power data characteristics, meteorological data characteristics, and time data characteristics into the pre-constructed source-load power data prediction model. Obtain static data characteristics through the convolutional neural network unit in the source-load power data prediction model, and obtain dynamic data characteristics through the gated recurrent unit in the source-load power data prediction model. And according to the static data characteristics and dynamic data characteristics, obtain the predicted source-load power data corresponding to the substation area; Static data characteristics are used to characterize data characteristics with stability and repeatability, and dynamic data characteristics are used to describe data characteristics that reflect temporal dependence and change trends;

[0061] Based on the predicted source-load power data, a regulation plan is formulated to adjust the phases of single-phase user nodes in the substation area to achieve line loss reduction.

[0062] The above loss reduction method, device, computer equipment, computer-readable storage medium and computer program product for adjusting the phase of single-phase user nodes obtain the source-load power data, meteorological information and time information of the power distribution area, and obtain the initial source-load power data characteristics corresponding to the source-load power data, as well as the meteorological data characteristics and time data characteristics corresponding to the meteorological information and time information. According to the initial source-load power data characteristics and a pre-constructed data interpolation model based on the self-attention mechanism, a data interpolation result is obtained, and the initial source-load power data characteristics are updated to source-load power data characteristics by using the data interpolation result. The source-load power data characteristics, meteorological data characteristics and time data characteristics are input into a pre-constructed source-load power data prediction model. Static data characteristics are obtained through the convolutional neural network in the source-load power data prediction model, and dynamic data characteristics are obtained through the gated recurrent unit in the source-load power data prediction model. According to the static data characteristics and dynamic data characteristics, the predicted source-load power data corresponding to the power distribution area is obtained. Among them, the static data characteristics are used to represent data characteristics with stability and repeatability, and the dynamic data characteristics are used to describe data characteristics reflecting time series dependence and change trends. A regulation plan is formulated based on the predicted source-load power data to adjust the phase of the single-phase user nodes in the power distribution area. Obtaining the initial source-load power data characteristics, meteorological data characteristics and time data characteristics of the power distribution area, and performing data interpolation on the initial source-load power data characteristics to improve the data availability of the initial source-load power data characteristics. In addition, inputting each characteristic into a pre-constructed prediction model, using the parallel architecture in the model to extract dynamic and static data characteristics respectively, splicing the dynamic and static data characteristics, and then obtaining the predicted source-load power data corresponding to the large area based on this can fully mine the local characteristics and long-term time series dependence relationships in the data, improve the accuracy of the predicted source-load power data, and finally use the predicted source-load power data to formulate a regulation plan to adjust the phase of the single-phase user nodes in the power distribution area to ensure the operation stability of the power grid. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following will briefly introduce the drawings required for use in the description of the embodiments of the present application or related technologies. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other related drawings can be obtained without creative efforts based on these drawings.

[0064] Figure 1 It is an application environment diagram of the loss reduction method for adjusting the phase of single-phase user nodes in an embodiment;

[0065] Figure 2 It is a schematic flowchart of the loss reduction method for adjusting the phase of single-phase user nodes in an embodiment;

[0066] Figure 3Schematic flow chart of the method for obtaining the initial source-load power data characteristics in an embodiment;

[0067] Figure 4 Schematic diagram of the joint optimization training method in another embodiment;

[0068] Figure 5 Schematic flow chart of data interpolation in an embodiment;

[0069] Figure 6 Schematic flow chart of the diagonal mask attention mechanism in an embodiment;

[0070] Figure 7 The first diagonal mask self-attention processing flow in another embodiment;

[0071] Figure 8 The second diagonal mask self-attention processing flow in an embodiment;

[0072] Figure 9 Structural diagram of the convolutional neural network in an embodiment;

[0073] Figure 10 Gated recurrent unit diagram in another embodiment;

[0074] Figure 11 Structural block diagram of the loss reduction device for adjusting the phase of single-phase user nodes in an embodiment;

[0075] Figure 12 Internal structural diagram of a computer device in an embodiment. Detailed implementation manners

[0076] In order to make the objectives, technical solutions and advantages of the present application clearer and more understandable, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0077] With the rapid development of new energy and the increasing complexity of power system operation, the short-term prediction of power grid load and photovoltaic power generation has gradually become an important research direction for ensuring power grid stability and improving energy utilization efficiency. However, existing methods have significant limitations in dealing with multi-source heterogeneous data, coping with data missing, and capturing complex non-linear features, specifically manifested as follows: ① Data missing problem: Due to reasons such as sensor failures and communication interruptions, time series data often has missing values. Traditional interpolation methods have limited accuracy and are prone to introducing noise or losing data correlation. ② Insufficient feature fusion of multi-source heterogeneous data: The features of power load and photovoltaic power generation are affected by various factors, including historical load data, meteorological conditions, holidays, etc. Existing models are difficult to efficiently fuse multi-source dynamic and static data. ③ Limited prediction accuracy and generalization ability: Complex non-linear features and multi-variable correlations pose challenges to the learning ability of the model. Traditional single-architecture models have deficiencies in capturing spatio-temporal characteristics. In addition, traditional three-phase unbalance regulation methods mainly rely on manual operations or simple optimization algorithms, lacking the ability of real-time dynamic adjustment and unable to make rapid responses under the conditions of complex power load and photovoltaic power generation fluctuations, resulting in low power grid efficiency and high line losses. How to effectively regulate based on the short-term prediction results of load and photovoltaic power generation to reduce line losses and improve power grid operation efficiency has also become an important issue faced by the current power field.

[0078] The loss reduction method for adjusting the phase of single-phase user nodes provided by the embodiments of this application can be applied to an application environment such as Figure 1 shown. Among them, the server 102 communicates with the substation area through the network. The data storage system can store the data that the server 102 needs to process. The data storage system can be integrated on the server 102, or placed in the cloud or other network servers.

[0079] The server 102 obtains the source-load power data of the power grid area, obtains the meteorological information and time information of the power grid area from the data storage system, and obtains the initial source-load power data features corresponding to the source-load power data, as well as the meteorological data features and time data features corresponding to the meteorological information and time information. According to the initial source-load power data features and the pre-constructed data interpolation model based on the self-attention mechanism, the data interpolation result is obtained, and the initial source-load power data features are changed to the source-load power data features by using the data interpolation result. The source-load power data features, meteorological data features, and time data features are input into the pre-constructed source-load power data prediction model. The static data features are obtained through the convolutional neural network unit in the source-load power data prediction model, and the dynamic data features are obtained through the gated recurrent unit in the source-load power data prediction model. According to the static data features and dynamic data features, the predicted source-load power data corresponding to the power grid area is obtained, and a regulation plan is formulated based on the predicted source-load power data to adjust the phase of the single-phase user nodes in the power grid area. Among them, the server 102 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0080] In an exemplary embodiment, as Figure 2 shown, a loss reduction method for adjusting the phase of single-phase user nodes is provided. Taking this method applied to Figure 1 the server 102 in it as an example for explanation, it includes the following steps S201 to S204. Among them:

[0081] Step S201, obtain the source-load power data, meteorological information, and time information of the power grid area, and obtain the initial source-load power data features corresponding to the source-load power data, as well as the meteorological data features and time data features corresponding to the meteorological information and time information respectively.

[0082] Among them, the power grid area can be understood as the area covered by the electrical signal; the source-load power data can be understood as the data related to the power generation source (source) and load demand (load) in the power system, such as power load power, photovoltaic power generation power data. The meteorological information can be understood as the climate information inside the power grid area, such as temperature, humidity, wind direction, wind speed, etc.

[0083] Exemplarily, the server 102 uses devices such as electric meters to collect the source-load power data of the power grid area, including power load power, photovoltaic power generation power data, etc., through the weather website API (Application Programming Interface) interface, including temperature, humidity, wind direction, wind speed, etc., and at the same time obtains the time information, including time stamps, holidays, etc., and performs the following data preprocessing on the above data:

[0084] Data cleaning and integrity verification:

[0085] Outlier Detection and Handling:

[0086] For features that cannot be negative (such as source-load power data, wind speed, etc.), if negative values appear, they are marked as outliers. Then, the box plot method is used to statistically analyze the distribution of all time series data to detect outliers. After processing, all outliers are marked as missing values.

[0087] Duplicate and Redundancy Check:

[0088] Check for duplicate timestamp records, and for duplicate data, delete redundant samples.

[0089] Integrity Check:

[0090] Calculate the missing ratio of all features. For features with a large missing ratio, based on their importance, choose whether to exclude them. Key features include source-load power data, temperature, etc., and it is necessary to ensure that there are no large-scale missing values.

[0091] Missing Mask and Data Output:

[0092] Construct a missing mask matrix, output the processed data, and save it in a unified format, including the complete time series and the missing mask.

[0093] Time Series Alignment:

[0094] Based on the acquisition frequency of source-load power data, align all data to a unified time step (the acquisition frequency of source-load power data is A minutes, so the time step of this application is set to A minutes). Make the dynamic features match the static features.

[0095] Then, use the STL (Seasonal and Trend decomposition using Loess) decomposition method to decompose the source-load power data for time series decomposition, and obtain the initial source-load power data features, including the average daily load / photovoltaic, trend, seasonality, etc. features of the user load power and photovoltaic power generation. First, classify and code different holiday information, and at the same time add time-related features, including minute, hour, day, week, month, season, year, weekday, etc. markers, and extract and quantify meteorological features such as temperature, humidity, weather type, wind speed, etc. Calculate the correlation between the features and the target variable using the Pearson correlation coefficient, and use the Granger Causality Test to verify the causal relationship between the input features and the prediction target. Retain the features with strong correlation as the model input, and perform standardization processing on the retained initial source-load power data features, meteorological data features, and time data features to normalize the features to the range of [0, 1].

[0096] Based on the above embodiments, by obtaining multi-source data within the power distribution area and performing data preprocessing on the obtained multi-source data, the availability and effectiveness of the data are improved, laying a data foundation for subsequent prediction of source-load power data greater than; and corresponding feature extraction is performed on the cleaned multi-source data to explore the internal relationships of the multi-source data and improve the prediction accuracy of the model.

[0097] Step S202: According to the initial source-load power data features and a pre-constructed data imputation model based on the self-attention mechanism, obtain a data imputation result, and use the data imputation result to update the initial source-load power data features to source-load power data features.

[0098] Among them, the data imputation model based on the self-attention mechanism can be understood as a model that uses the self-attention mechanism to explicitly capture time dependence and feature correlation, and at the same time aggregates dynamic weights to optimize the imputation result to achieve efficient imputation of complex multivariate time series.

[0099] Optionally, the server 102 obtains a data imputation result through the data imputation model according to the initial source-load power data features and a pre-constructed data imputation model based on the self-attention mechanism, and uses the data imputation result to correspondingly modify the data features in the initial source-load power features to obtain source-load power data features. The diagonal masked self-attention module is used to capture time dependence and the correlation between features, and the imputation accuracy is improved through a joint optimization training strategy, ensuring the integrity of the input data.

[0100] Step S203: Input the source-load power data features, meteorological data features, and time data features into a pre-constructed source-load power data prediction model. Obtain static data features through the convolutional neural network unit in the source-load power data prediction model, and obtain dynamic data features through the gated recurrent unit in the source-load power data prediction model. And based on the static data features and dynamic data features, obtain the predicted source-load power data corresponding to the power distribution area; the static data features are used to characterize data features with stability and repeatability, and the dynamic data features are used to describe data features that reflect time series dependence and change trends.

[0101] Among them, the static data features can be understood as data features used to characterize data with stability and repeatability. In this application, they mainly include meteorological data features and time data features. The dynamic data features can be understood as data features used to describe data that reflects time series dependence and change trends. In this application, they mainly include source-load power data features, such as power load power and photovoltaic power generation power, etc.

[0102] Exemplarily, the server 102 inputs the source-load power data features, meteorological data features, and time data features into a pre-constructed source-load power data prediction model, and obtains static data features through the convolutional neural network unit in the source-load power data prediction model:

[0103] Since the convolutional layer can effectively capture static features such as periodic changes in the time series and extract the spatial local dependence relationship of the data, this application uses a convolutional neural network to process static features. Specifically, the convolutional neural network includes two convolutional layers, two max pooling layers, and one fully connected layer.

[0104] Its output can be expressed as follows:

[0105]

[0106] where is the fully connected layer, is the activation function, is the fully connected layer, , are the weight parameters, , are the bias terms.

[0107] and obtains dynamic data features through the gated recurrent unit in the source-load power data prediction model:

[0108] The multi-layer gated recurrent unit can use the gating mechanism (update gate and reset gate) to control the transmission of information flow, update or forget past information according to the information at the current time step, so as to capture the dynamic features and temporal dependencies of the time series. The gated recurrent unit mainly consists of a forget gate, an update gate, and a hidden state.

[0109] Among them, represents the input information at the current moment, represents the hidden state at the previous moment, represents the hidden state (new weight matrix) passed to the next moment, represents the hidden candidate state, represents the fractional forget gate, represents the update gate, represents the activation function, represents the hyperbolic tangent function, specifically:

[0110]

[0111]

[0112]

[0113]

[0114] Among them, represents element-wise multiplication. By using a multi-layer gated recurrent unit to obtain temporal features, in this application, the hidden layer of the multi-layer gated recurrent unit is set to two layers, and the extracted features are mapped through a fully connected layer and converted into the target dimension.

[0115] Model output:

[0116] The output results of the convolutional neural network unit and the gated recurrent unit are concatenated together, and the final model prediction result is output by using a fully connected layer.

[0117] Based on the foregoing embodiments, by designing the parallel architecture of the above-mentioned convolutional neural network source and the gated recurrent unit, the spatial features and temporal dependencies of multi-source data can be fully mined, the local features and long-term dependencies of multi-source data can be extracted, their internal relationships can be mined, and the prediction accuracy can be improved.

[0118] Step S204, formulate a regulation plan based on the predicted source-load power data, and adjust the phase of the single-phase user nodes in the substation area to achieve line loss reduction.

[0119] Optionally, the server 102 formulates a regulation plan based on the predicted source-load power data and adjusts the phase of the single-phase user nodes in the substation area, including day-ahead scheduling and ultra-short-term scheduling, where:

[0120] Day-ahead scheduling is to make a prediction the day before (a 24-hour prediction for the future), generate a preliminary regulation plan according to the prediction result, then obtain the real-time source-load power data of the substation area on the same day, calculate reverse overload and the like, and perform a rolling prediction (a 3-hour prediction for the future) to fine-tune the day-ahead regulation plan, and adjust the phase of the single-phase user nodes in the substation area according to the fine-tuned regulation plan.

[0121] Ultra-short-term scheduling refers to directly judging whether there is reverse overload in the substation area in the future period according to the predicted source-load power data. If so, the phase of the single-phase user nodes in the substation area is adjusted according to the corresponding regulation plan (this method is based on the premise of a high prediction result accuracy).

[0122] The current node phase of the user is known, and the adjustable nodes are also known. What needs to be done is to generate the phase of the adjustable nodes according to the regulation algorithm when the unbalance degree or line loss is relatively large to reduce the unbalance degree or line loss. Determine which adjustable points need to be changed according to the regulation plan. Combine with the same communication method to send a command to the phase changer that controls that node to make it change phase. Based on the short-term load and photovoltaic power generation prediction results, implement three-phase unbalance regulation, adjust the grid load, reduce line loss, and improve the grid operation efficiency.

[0123] In the above-mentioned loss reduction method for adjusting the phase of a single-phase user node, the source load power data, meteorological information and time information of the substation are obtained, and the initial source load power data features corresponding to the source load power data, as well as the meteorological data features and time data features corresponding to the meteorological information and time information are obtained. According to the initial source load power data features and a pre-built data interpolation model based on a self-attention mechanism, a data interpolation result is obtained, and the initial source load power data features are updated to source load power data features using the data interpolation result. The source load power data features, meteorological data features and time data features are input into a pre-built source load power data prediction model, static data features are obtained through a convolutional neural network in the source load power data prediction model, and dynamic data features are obtained through a gated recurrent unit in the source load power data prediction model. According to the static data features and the dynamic data features, the predicted source load power data corresponding to the substation is obtained, wherein the static data features are used to characterize data features with stability and repeatability, and the dynamic data features are used to describe data features reflecting time series dependence and change trends. A control plan is formulated based on the predicted source load power data to adjust the phase of the single-phase user node in the substation. The initial source-load power data characteristics, meteorological data characteristics and time data characteristics of the substation are obtained, and data interpolation is performed on the initial source-load power data characteristics to improve the data availability of the initial source-load power data characteristics. In addition, each characteristic is input into a pre-built prediction model, and the parallel architecture in the model is used to extract dynamic and static data characteristics respectively, and the dynamic and static data characteristics are spliced to obtain the corresponding predicted source-load power data of the large area based on this. This can fully explore the local characteristics and long-term temporal dependencies in the data, improve the accuracy of the predicted source-load power data, and finally use the predicted source-load power data to formulate a control plan, adjust the phase of the single-phase user node in the substation, and ensure the operational stability of the power grid.

[0124] In one embodiment, the initial source load power data features include source load trend features, source load seasonal features, and average daily source load features; the initial source load power data features corresponding to the acquired source load power data include:

[0125] A corresponding time series is generated based on the initial source-load power data, and the average daily source-load characteristics are obtained according to the time series; the time series includes a source-load trend component, a source-load seasonal component and a source-load residual component; a pre-set source-load trend component is subtracted from the time series to obtain a detrended time series; the detrended time series is decomposed according to a preset period to obtain multiple time subseries, and a new source-load seasonal component is obtained according to the time subsequences; the new source-load seasonal component is subtracted from the detrended time series to obtain a deseasonalized time series; and the source-load trend characteristics and source-load seasonal characteristics are obtained based on the deseasonalized time series and the new source-load seasonal component.

[0126] Exemplarily, extraction of source-load power data features:

[0127] Use the STL decomposition method to perform time series decomposition on the source-load power data to obtain features such as average daily load / photovoltaic, trend, and seasonality of the user load power and photovoltaic power generation. The STL feature decomposition process is as Figure 3 shown:

[0128] For the user load power, if you need to obtain average daily load, trend, and seasonality features, the steps are as follows:

[0129] Specifically, as in step S201, organize the original load data into a time series format with a step size of A minutes. The time series can be expressed as follows:

[0130]

[0131] where, is the load trend component, is the load seasonal component, is the load residual component. Let the initial load trend component .

[0132] (1) Detrending. Let the original time series subtract the trend component (with an initial value of 0) to obtain the detrended time series ( ). Thus, the influence of the long-term trend is removed to facilitate the extraction of seasonal and other components.

[0133] (2) Subsequence smoothing. Decompose the time series into each periodic segment according to the period. Set a period of one day (96 points), decompose the sequence into subsequences of several days. And perform locally weighted regression (LOESS, Locally Estimated Scatterplot Smoothing) filtering on each subsequence, and expand points before and after the sliding window to generate a smooth periodic component .

[0134] (3) Obtaining the seasonal sequence. Perform 3-time moving average on the subsequence smoothing result to further smooth the sequence and remove randomness. At the same time, use locally weighted regression filtering to obtain the complete smooth seasonal sequence .

[0135] (4) Removal of seasonality and trend. Subtract the seasonal sequence from the detrended time series , and then use locally weighted regression filtering to extract the new trend component , the residual part is the unexplained random fluctuation.

[0136] Based on the deseasonalized time series and the new load seasonal component, obtain the load trend feature and the load seasonal feature.

[0137] The above method extracts features from the source-load power data by using the STL decomposition method, mines the internal relationship of multi-source data, enhances the data expression ability of the extracted features, and thus improves the prediction accuracy of the source-load power data.

[0138] In one embodiment, based on the deseasonalized time series and the new source-load seasonal component, obtain the source-load trend feature and the source-load seasonal feature, including: performing locally weighted regression filtering on the deseasonalized time series to extract a new source-load trend component; if the similarity between the new source-load trend component and the source-load trend component is greater than or equal to a preset first similarity, and the similarity between the new source-load seasonal component and the source-load seasonal component is greater than or equal to a preset second similarity, then determine the new source-load trend component as the source-load trend feature and determine the new source-load seasonal component as the source-load seasonal feature.

[0139] Optionally, the server 102 performs locally weighted regression filtering on the deseasonalized time series to extract a new source-load trend component. If the similarity between the new source-load trend component and the source-load trend component is greater than or equal to a preset first similarity, and the similarity between the new source-load seasonal component and the source-load seasonal component is greater than or equal to a preset second similarity, it indicates that the STL decomposition process has converged. At this time, determine the new source-load trend component as the source-load trend feature , and determine the new source-load seasonal component as the source-load seasonal feature , the average daily source-load feature is the mean of 96 points per day, and the calculation method is as follows:

[0140]

[0141] Based on the foregoing embodiments, by extracting the new source-load trend feature and the source-load seasonal feature, the dynamic changes of the time series can be more accurately reflected, and thus the prediction accuracy of the model is improved.

[0142] In an exemplary embodiment, the data imputation model includes a first diagonal masked self-attention module, a second diagonal masked self-attention module, and a dynamic weighted combination module; based on the initial source-load power data feature and the pre-constructed data imputation model based on the self-attention mechanism, obtain the data imputation result, including:

[0143] Obtain the missing mask matrix pre-constructed according to the source-load power data, and perform mask processing on the initial source-load power data features to obtain the primary source-load power data features; concatenate the missing mask matrix and the primary source-load power data features to obtain the secondary source-load power data features; input the secondary source-load power data features into the first diagonal mask self-attention module to obtain the first-stage interpolation result; use the first-stage interpolation result to update the primary source-load power data features, and concatenate the updated primary source-load power data features and the missing mask matrix to obtain the advanced source-load power data features; input the advanced source-load power data features into the second diagonal mask self-attention module to obtain the second-stage interpolation result; input the missing mask matrix, the first-stage interpolation result, and the second-stage interpolation result into the dynamic weighted combination module to obtain the data interpolation result.

[0144] Exemplarily, subsequently, generate the corresponding missing mask matrix according to the load trend feature and the load seasonal feature:

[0145] Missing mask matrix preparation:

[0146] Input the load trend feature and the load seasonal feature into a multivariate time series matrix, denoted as , , where represents the input feature dimension, represents the time step, represents the -th step input feature, , where each value may be missing. To represent the missing variables in the input time series matrix , generate the missing mask matrix according to the aforementioned data missing situation, and denote it as , , where the observed value is 1 and the missing value is 0, which can be expressed as follows:

[0147]

[0148] Joint optimization training method:

[0149] Design a joint optimization training method based on interpolation and reconstruction, which specifically includes two parts: the mask interpolation task and the observation reconstruction task, as Figure 4 shown.

[0150] First, randomly set a part of the observed values of the input data to the "missing" state to obtain the masked input time series , the corresponding missing mask vector , and denote the manually marked mask as to distinguish the artificial mask from the original missing values. and are expressed as follows:

[0151] ,

[0152] Input and into the input model, and the imputation result of the model output is denoted as .

[0153] The masked imputation task is used to predict the randomly masked observed values. By training the model to learn how to fill in the missing values, the imputation ability of the model for real missing values is improved. Compare the imputation result of the artificially marked missing values with the real values, calculate the imputation error, and use the mean absolute error (MAE, Mean Absolute Error) as the loss function to calculate the masked imputation task loss :

[0154]

[0155]

[0156] Among them, represents the input feature dimension, represents the time step, represents element-wise multiplication, represents the imputation result, represents the original input data, represents the artificially marked mask.

[0157] The observation reconstruction task is used to reconstruct the original observed values to ensure that the model can learn the distribution characteristics of the observed data. Similarly, the mean absolute error is used as the loss function to calculate the masked imputation task loss :

[0158]

[0159] The data imputation model based on the self-attention mechanism is as Figure 5 shown, mainly including the first diagonal masked self-attention module (DMSA_1, Diagonal Masked Self-Attention Module 1), the second diagonal masked self-attention module (DMSA_2, Diagonal Masked Self-Attention Module 2), and the dynamic weighted combination module, where:

[0160] Based on the above embodiments, the self-attention mechanism is used to explicitly capture temporal dependencies and feature correlations, and at the same time, the dynamic weight adjustment module is combined to optimize the imputation results, realizing the efficient imputation of complex multivariate time series, and using the joint training method to train the data imputation process, which can improve the accuracy of data imputation, and further improve the effectiveness and accuracy of the imputation results.

[0161] In one embodiment, the first diagonal masked self-attention module includes a first feature projection unit, a first feature enhancement unit, and a first feature restoration unit; the first feature enhancement unit is constructed based on the multi-head self-attention mechanism; the secondary source-load power data features are input into the first diagonal masked self-attention module to obtain the first-stage imputation results, including:

[0162] Input the secondary source-load power data features into the first feature projection unit to obtain the secondary source-load power data features in the target dimension; input the secondary source-load power data features in the target dimension into the first feature enhancement unit to obtain the feature-enhanced secondary source-load power data features; input the feature-enhanced secondary source-load power data features into the first feature restoration unit to obtain the first-stage imputation results.

[0163] Optionally, the first diagonal masked self-attention module DMSA_1:

[0164] This module is used to extract the basic features of the time series and generate preliminary imputation results, and can input the time series feature matrix and the missing mask .

[0165] ① First, and are concatenated as the input, and the result is denoted as , , then is projected to dimensions by using a linear mapping, and position encoding information is added, which can be expressed as follows:

[0166]

[0167] where represents the feature concatenation operation, and represent the weight and bias parameters respectively, , , represents the position encoding matrix (the foregoing is the first feature projection unit).

[0168] ② Use the multi-head self-attention mechanism to capture temporal dependencies and feature interaction relationships, and use the feed-forward network to perform non-linear transformation on the attention results to enhance the feature expression ability and generate the first-stage imputation results.

[0169]

[0170] Among them, represents the diagonal masked multi-head attention mechanism, represents the feed-forward network, represents stacking layers. Its specific calculation formula is as follows:

[0171]

[0172] represents the diagonal masking operation. For the input feature , the diagonal elements are set to .

[0173]

[0174] represents the diagonal masked self-attention mechanism, represents the query vector, represents the key vector, represents the value vector, is the normalization function, represents the dimension. Specifically, for a time series with a data length of 5, the calculation process of the diagonal masked multi-head attention mechanism is as Figure 6 shown.

[0175]

[0176]

[0177] represents the diagonal masked multi-head attention mechanism, represents the feature concatenation operation, represents the th head, represents the number of heads, represents the output layer weight matrix, , , respectively represent , , 's projection matrix (the above is the first feature enhancement unit).

[0178] The feed-forward network includes two layers of fully connected networks and the ReLU (Rectified Linear Unit) activation function, and can be expressed as follows:

[0179]

[0180] wherein and are weight parameters, and are bias parameters.

[0181] The first-stage interpolation process is as shown in Figure 7 The first-stage interpolation result can be expressed as follows: , are weight parameters, are bias parameters, and then the missing points in are replaced into the original data to obtain the complete feature vector , , represents element-wise multiplication, represents the first-stage interpolation result, represents the feature matrix, represents the missing mask matrix.

[0182] Based on the foregoing embodiments, through the self-attention mechanism, the module can better capture temporal dependencies and feature interactions, enhance feature representation, and the interpolation result of the first stage can effectively handle missing data, improving the adaptability of the model to incomplete time series, laying a data foundation for further interpolation in the second stage.

[0183] In one embodiment, the second diagonal mask self-attention module includes a second feature projection unit, a second feature enhancement unit, and a second feature restoration unit; the second feature enhancement unit is constructed based on the multi-head self-attention mechanism; the high-level source-load power data features are input into the second diagonal mask self-attention module to obtain the second-stage interpolation result, including:

[0184] The high-level source-load power data features are input into the second feature projection unit to obtain the high-level source-load power data features in the target dimension; the high-level source-load power data features in the target dimension are input into the second feature enhancement unit to obtain the enhanced high-level source-load power data features; the enhanced high-level source-load power features are input into the second feature restoration unit to obtain the second-stage interpolation result.

[0185] Exemplarily, the second diagonal mask self-attention module DMSA_2:

[0186] The main steps are the same as those of the first diagonal mask self-attention block, and its input includes the output of the first diagonal mask self-attention block , the mask matrix , and the specific calculation process is as follows:

[0187]

[0188]

[0189]

[0190] For the specific parameter description, refer to the first diagonal mask self-attention module. The specific steps are as Figure 8 shown.

[0191] According to the foregoing embodiments, through the self-attention mechanism, the module can better capture temporal dependencies and feature interactions, enhance feature representation. The imputation results in the second stage can effectively handle missing data, improve the adaptability of the model to incomplete time series, and thus improve the accuracy of model prediction.

[0192] In an exemplary embodiment, the imputation result in the second stage carries the weight information of the second feature enhancement unit; the missing mask matrix, the imputation result in the first stage, and the imputation result in the second stage are input into the dynamic weighted combination module to obtain the data imputation result, including: obtaining a weight factor according to the weight information and the missing mask matrix; using the weight factor to perform weighted summation on the imputation result in the first stage and the imputation result in the second stage to obtain the data imputation result.

[0193] Optionally, first extract the weight matrix from the multi-head attention of DMSA_2 , calculate the multi-head average attention weight , as shown in the following formula:

[0194]

[0195] represents the number of heads, represents the th weight matrix of the head.

[0196] Then calculate the weight factor according to the multi-head average attention weight :

[0197]

[0198] represents the activation function, represents the missing mask matrix, represents the weight parameter, represents the bias parameter.

[0199] Dynamically fuse the imputation results of DMSA_1 and DMSA_2 to generate a weighted combination imputation result:

[0200]

[0201] Then Replace the missing points in the data with the original data to obtain the final output result :

[0202]

[0203] Based on the above embodiments, use to replace the data at the corresponding positions in the characteristics of the initial source-load power data, so as to realize data interpolation. By combining the interpolation results of DMSA_1 and DMSA_2, the advantages of the two modules can be fully utilized, thereby improving the accuracy and reliability of interpolation. The complete interpolated data can significantly improve the performance of subsequent models (such as prediction models).

[0204] In an exemplary embodiment, a specific implementation manner of a loss reduction method for adjusting the phase of single-phase user nodes is provided (the following data are all specific examples, and it is not limited that the present application can only be implemented in this case), wherein:

[0205] I. Data collection and preprocessing:

[0206] 1.1 Data acquisition:

[0207] Use devices such as electric meters to collect source-load power data, including power load power and photovoltaic power generation power data. Obtain meteorological information, including temperature, humidity, wind direction, wind speed, etc., through the weather website API interface. At the same time, obtain time information, including timestamps, holidays, etc. Among them, the source-load power data belongs to dynamic characteristics, and other data belongs to static characteristics.

[0208] 1.2 Data cleaning and integrity verification:

[0209] 1.2.1 Outlier detection and processing:

[0210] For features that cannot be negative values (such as source-load power data, wind speed, etc.), if negative values appear, mark them as outliers. Then, use the box plot method to statistically analyze the distribution of all time series data to detect outliers. After processing, mark all outliers as missing values.

[0211] 1.2.2 Duplicate and redundancy check:

[0212] Check for duplicate timestamp records. For duplicate data, delete redundant samples.

[0213] 1.2.3 Integrity check:

[0214] Calculate the missing ratio of all features. For features with a large missing ratio, select whether to eliminate them based on their importance (according to the feature selection results in 1.4.4). Key features include source-load power data, temperature, etc., and it is necessary to ensure that there are no large-area missing values.

[0215] 1.2.4 Missing Mask and Data Output:

[0216] Construct a missing mask matrix, output the processed data, and save it in a unified format, including the complete time series and the missing mask.

[0217] 1.3 Time Series Alignment:

[0218] Based on the source-load power data acquisition frequency, align all data to a unified time step (the source-load power data acquisition frequency is 15 minutes, so the time step of this application is set to 15 minutes). Make the dynamic features match the static features.

[0219] 1.4 Feature Engineering:

[0220] 1.4.1 Feature Extraction of Source-Load Power Data:

[0221] Use the STL decomposition method to decompose the source-load power data in time series to obtain features such as the average daily load / photovoltaic, trend, and seasonality of the user load power and photovoltaic power generation. The STL feature decomposition process is as Figure 3 shown:

[0222] For the user load power, if you need to obtain the average daily load, trend, and seasonal features, the steps are as follows:

[0223] Specifically, as in step 1.3, organize the original load data into a time series format with a time step of 15 minutes. The time series can be expressed as follows:

[0224]

[0225] where, is the load trend component, is the load seasonal component, is the load residual component. Let the initial load trend component .

[0226] (1) Detrending. Let the original time series subtract the trend component (with an initial value of 0) to obtain the detrended time series ( ). Thus, the influence of the long-term trend is removed to facilitate the extraction of seasonal and other components.

[0227] (2) Subsequence Smoothing. Decompose each period segment of the time series according to the period. Set a period of one day (96 points), decompose the sequence into subsequences of several days. And perform locally weighted regression (LOESS) filtering on each subsequence, and expand according to the sliding window before and after Generate smooth periodic components at several points .

[0228] (3)Seasonal sequence acquisition. Perform a 3-point moving average on the smoothed result of the subsequence to further smooth the sequence and remove randomness. At the same time, use locally weighted regression filtering to obtain a complete smoothed seasonal sequence .

[0229] (4)Seasonality and trend removal. Subtract the seasonal sequence from the detrended time series , and then use partial weighted regression filtering to extract a new trend component . The residual part is the unexplained random fluctuation.

[0230] According to the above process, continuously optimize the trend and seasonal components until convergence.

[0231] Finally, the trend feature , seasonal feature can be obtained. The average daily load feature is the mean of 96 points per day, and the calculation method is as follows:

[0232]

[0233] 1.4.2 Time data feature extraction:

[0234] First, classify and code different holiday information, and at the same time add time-related features, including markers for minutes, hours, days, weeks, months, seasons, years, weekdays, etc.

[0235] 1.4.3 Meteorological data feature extraction:

[0236] Extract meteorological features such as temperature, humidity, weather type, wind speed, etc. and quantify them.

[0237] 1.4.4 Correlation analysis:

[0238] Use the Pearson correlation coefficient to calculate the correlation between features and the target variable, and use the Granger Causality Test to verify the causal relationship between input features and the prediction target. Retain features with strong correlations as model inputs.

[0239] 1.5 Data standardization:

[0240] Perform standardization processing on dynamic and static features to normalize the features to the range of [0, 1].[[]]

[0241] II. Data imputation:

[0242] Design a time series imputation method based on the self-attention mechanism, which captures temporal dependencies and correlations between features through the diagonal masked self-attention module to generate high-precision imputed data.

[0243] 2.1 Joint optimization training method:

[0244] 2.1.1 Missing mask matrix preparation:

[0245] Input the load / photovoltaic prediction model as a multivariate time series matrix, denoted as , , where represents the input feature dimension, represents the time step, represents the input feature at the -th step, , and each value may be missing. To represent the missing variables in the input time series matrix , generate a missing mask matrix according to the data missing situation in Step 1.2.1, and denote it as , , where the observed value is 1 and the missing value is 0, which can be expressed as follows:

[0246]

[0247] 2.1.2 Joint optimization training method:

[0248] Design a joint optimization training method based on imputation and reconstruction, which specifically includes two parts: the masked imputation task and the observed reconstruction task, as Figure 4 shown.

[0249] First, randomly set a part of the observed values in the input data to the "missing" state to obtain the masked input time series , the corresponding missing mask vector , and denote the artificial marked mask as to distinguish the artificial mask from the original missing values. and are expressed as follows:

[0250] ,

[0251] Input and into the model, and denote the imputation result output by the model as .

[0252] The masked imputation task is used to predict the randomly masked observed values. By training the model to learn how to fill in the missing values, the imputation ability of the model for real missing values is improved. The imputation results of the artificially marked missing values are compared with the real values, the imputation error is calculated, and the mean absolute error (MAE) is used as the loss function to calculate the masked imputation task loss :

[0253]

[0254]

[0255] where represents the input feature dimension, represents the time step, represents the element-wise multiplication, represents the imputation result, represents the original input data, represents the artificially marked mask.

[0256] The observation reconstruction task is used to reconstruct the original observed values to ensure that the model can learn the distribution characteristics of the observed data. The mean absolute error is also used as the loss function to calculate the masked imputation task loss :

[0257]

[0258] 2.2 Imputation Model Based on Self-Attention Mechanism:

[0259] The imputation model based on self-attention mechanism is as Figure 5 shown, mainly including the first diagonal masked self-attention module (DMSA_1), the second diagonal masked self-attention module (DMSA_2), and the dynamic weighted combination module. The self-attention mechanism can be used to explicitly capture the time dependence and feature correlation, and at the same time, combined with the dynamic weight adjustment module to optimize the imputation result, realizing the efficient imputation of complex multivariate time series.

[0260] 2.2.1 The First Diagonal Masked Self-Attention Module DMSA_1:

[0261] This module is used to extract the basic features of the time series and generate the preliminary imputation result. The time series feature matrix and the missing mask matrix can be input (as in step 2.1).

[0262] ① First, and are concatenated as the input, and the result is denoted as , , then is projected to Dimension, and add position encoding information, which can be expressed as follows:

[0263]

[0264] Among them Represents the feature concatenation operation, And Respectively represent the weight and bias parameters, , , Represents the position encoding matrix.

[0265] ② Use the multi-head self-attention mechanism to capture temporal dependencies and feature interaction relationships, and use the feed-forward network to perform non-linear transformation on the attention results to enhance the feature expression ability and generate the first-stage interpolation result.

[0266]

[0267] Among them, Represents the diagonal masked multi-head attention mechanism, Represents the feed-forward network, Represents stacking Layers. Its specific calculation formula is as follows:

[0268]

[0269] Represents the diagonal masking operation. For the input feature , set the diagonal elements to .

[0270]

[0271] Represents the diagonal masked self-attention mechanism, Represents the query vector, Represents the key vector, Represents the value vector, Is the normalization function, Represents the dimension. Specifically, for a time series with a data length of 5, the calculation process of the diagonal masked multi-head attention mechanism is as Figure 6 Shown.

[0272]

[0273]

[0274] Represents the diagonal masked multi-head attention mechanism, Represents the feature concatenation operation, Represents the th head, Indicates the number of heads, Indicates the weight matrix of the output layer, , , respectively represent , , 's projection matrix.

[0275] The feedforward network consists of two fully connected layers and a ReLU activation function, which can be expressed as follows:

[0276]

[0277] where , are weight parameters, , are bias parameters.

[0278] The first-stage interpolation process is as shown in Figure 7 , and the result of the first-stage interpolation can be expressed as follows: , are weight parameters, is the bias parameter, and then the missing points in are replaced into the original data to obtain the complete feature vector , , represents element-wise multiplication, represents the result of the first-stage interpolation, represents the feature matrix, represents the missing mask matrix (as described in step 2.1).

[0279] 2.2.2 Second Diagonal Mask Self-Attention Module DMSA_2:

[0280] The main steps are the same as those of the first diagonal mask self-attention block, and its input includes the output of the first diagonal mask self-attention block , the mask matrix , and the specific calculation process is as follows:

[0281]

[0282]

[0283]

[0284] For specific parameter descriptions, refer to 2.2.1, and the specific steps are as shown in Figure 8 .

[0285] 2.2.3 Weighted Combination Module:

[0286] First, extract the weight matrix from the multi-head attention of DMSA_2 , and calculate the multi-head average attention weight , as shown in the following formula:

[0287]

[0288] As described in 2.2.1, represents the number of heads, represents the -th head's weight matrix.

[0289] Then, calculate the weight factor according to the multi-head average attention weight :

[0290]

[0291] represents the activation function, represents the missing mask matrix, represents the weight parameter, represents the bias parameter.

[0292] Dynamically fuse the imputation results of DMSA_1 and DMSA_2 to generate a weighted combined imputation result:

[0293]

[0294] Then replace the missing points in with the original data to obtain the final output result :

[0295]

[0296] 2.3 Loss function:

[0297] The data imputation loss is shown in the following formula:

[0298]

[0299] Where represents the masked imputation loss, represents the observed reconstruction loss, is a hyperparameter used to adjust the weights of the two tasks. Specifically:

[0300]

[0301]

[0302] III. Prediction model construction:

[0303] To fully exploit the spatial features and temporal dependencies of multi-source data, this application designs a parallel architecture based on a convolutional neural network (CNN) and a gated recurrent unit (GRU) to extract the local features and long-term dependencies of multi-source data, explore their internal relationships, and improve the prediction accuracy.

[0304] 3.1 Convolutional neural network:

[0305] Since the convolutional layer can effectively capture static features such as periodic changes in time series and extract the spatial local dependence relationship of data, this application uses a convolutional neural network to process static features. Specifically, the convolutional neural network includes two convolutional layers, two max pooling layers, and one fully connected layer, and its structure is as Figure 9 shown.

[0306] Its output can be expressed as follows:

[0307]

[0308] where is the fully connected layer, is the activation function, is the fully connected layer, , are the weight parameters, , are the bias terms.

[0309] 3.2 Multi-layer gated recurrent unit:

[0310] The multi-layer gated recurrent unit can use the gating mechanism (update gate and reset gate) to control the flow of information, update or forget past information according to the information at the current time step, so as to capture the dynamic features and temporal dependencies of the time series. The gated recurrent unit mainly consists of a forget gate, an update gate, and a hidden state, as Figure 10 shown.

[0311] Among them, represents the input information at the current moment, represents the hidden state at the previous moment, represents the hidden state passed to the next moment (new weight matrix), represents the hidden candidate state, represents the fractional forget gate, represents the update gate, represents the activation function, represents the hyperbolic tangent function. Specifically:

[0312]

[0313]

[0314]

[0315]

[0316] Among them, represents element-wise multiplication. By using a multi-layer gated recurrent unit to obtain temporal features, in this application, the hidden layer of the multi-layer gated recurrent unit is set to two layers, and the extracted features are mapped through a fully connected layer and converted into the target dimension.

[0317] 3.3 Model Output:

[0318] Concatenate the output results of 3.1 and 3.2, and use the fully connected layer to output the final model prediction result.

[0319] IV. Model Training and Validation:

[0320] After the data imputation in step 2, the data set is split into a training set and a validation set according to 8:2. The training set is iteratively trained in the prediction model constructed in step 3 to optimize the model parameters, and finally the optimal model is obtained. The specific steps are as follows:

[0321] 4.1 Forward Propagation:

[0322] Divide the input data features into static features and dynamic features, and input them into the parallel architecture in step 3 respectively to obtain the model prediction result.

[0323] 4.2 Loss Calculation:

[0324] Use the mean squared error (MSE) as the loss function. After obtaining the prediction result, calculate the error between the predicted value and the actual value:

[0325]

[0326] Among them, represents the number of samples, represents the model predicted value, represents the true value.

[0327] 4.3 Parameter Update:

[0328] First, calculate the gradient of the loss function with respect to the model parameters, and then use the Adam optimization algorithm to update the model parameters according to the gradient information.

[0329] 4.4 Iterative Training:

[0330] During the training process, the performance of the model is evaluated using the validation set to prevent overfitting. Through multiple iterative trainings, the model parameters are continuously adjusted until the model converges or its accuracy on the validation set no longer improves, and then the model weights are saved.

[0331] V. Regulation Scheme Generation:

[0332] Using the daily power curves of each user node predicted by the source-load prediction model in Step 4, it is judged whether there is reverse overload in the distribution area according to the prediction results. When there is reverse overload, a day-ahead scheduling scheme is formulated.

[0333] Specifically, the load rate calculation formula is:

[0334]

[0335] Among them, represents the load consumption power, represents the photovoltaic output power, represents the rated power, represents the load rate. Combining the distribution area archives, with the goal of minimizing the three-phase unbalance degree and the line loss, an optimal control model is constructed based on the genetic algorithm.

[0336] Obtain the three-phase current of the user node. When the current unbalance degree exceeds the threshold and the current load rate exceeds the threshold, solve the optimal control model to generate the optimal phase change strategy. At the same time, it is necessary to perform rolling source-load prediction (that is, after new data is collected, the model input and output are rolled back one time point), so as to ensure accurate source-load prediction. Finally, adjust the phase of the single-phase user node according to the optimal phase change strategy.

[0337] Compared with the prior art, the present application has the following advantages:

[0338] 1. Multi-source data feature extraction and fusion: The present application effectively fuses dynamic features (such as power load power, photovoltaic power generation power) with static features (such as meteorological data, time information), uses the STL decomposition method to extract features such as trends and seasonality, and explores the internal relationships of multi-source data to improve the model prediction accuracy.

[0339] 2. An efficient data imputation method based on the self-attention mechanism: To solve the problem of the impact of data missing, the present application proposes a data imputation method based on the self-attention mechanism, uses the diagonal masked self-attention module to capture time dependence and the correlation between features, and improves the imputation accuracy through a joint optimization training strategy to ensure the integrity of the input data.

[0340] 3. Parallel architecture enhances model generalization ability: A parallel architecture based on convolutional neural network and gated recurrent unit is proposed. By combining the advantages of CNN in spatial feature extraction and GRU in temporal dependence modeling, local features and long-term temporal dependence relationships in the data are fully exploited to enhance the model generalization ability.

[0341] 4. Three-phase unbalance regulation and line loss optimization: Based on the short-term load and photovoltaic power generation prediction results, three-phase unbalance regulation is implemented to adjust the grid load, reduce line losses, and improve the grid operation efficiency.

[0342] It should be understood that although the steps in the flowcharts involved in the above-described embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise clearly stated in this article, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-described embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be executed alternately or in turn with at least a part of other steps or steps or stages in other steps.

[0343] Based on the same inventive concept, an embodiment of the present application further provides a device for reducing line loss by adjusting the phase of single-phase user nodes for implementing the above-mentioned method for reducing line loss by adjusting the phase of single-phase user nodes. The solution provided by this device for solving problems is similar to the solution described in the above method. Therefore, the specific limitations in one or more embodiments of the device for reducing line loss by adjusting the phase of single-phase user nodes provided below can refer to the limitations on the method for reducing line loss by adjusting the phase of single-phase user nodes in the above text, and will not be repeated here.

[0344] In an exemplary embodiment, as Figure 11 shown, a device for reducing line loss by adjusting the phase of single-phase user nodes is provided, including: an acquisition module 1101, an interpolation module 1102, a prediction module 1103, and an adjustment module 1104, where:

[0345] The acquisition module 1101 is configured to acquire the source-load power data, meteorological information, and time information of the substation area, and acquire the initial source-load power data features corresponding to the source-load power data, as well as the meteorological data features and time data features corresponding to the meteorological information and time information respectively.

[0346] An interpolation module 1102, configured to obtain a data interpolation result according to the initial source-load power data characteristics and a pre-constructed data interpolation model based on a self-attention mechanism, and update the initial source-load power data characteristics to source-load power data characteristics by using the data interpolation result;

[0347] A prediction module 1103, configured to input the source-load power data characteristics, meteorological data characteristics, and time data characteristics into a pre-constructed source-load power data prediction model, obtain static data characteristics through a convolutional neural network unit in the source-load power data prediction model, and obtain dynamic data characteristics through a gated recurrent unit in the source-load power data prediction model, and obtain the predicted source-load power data corresponding to the substation area according to the static data characteristics and the dynamic data characteristics; the static data characteristics are used to characterize data characteristics with stability and repeatability, and the dynamic data characteristics are used to describe data characteristics reflecting time series dependence and change trends;

[0348] An adjustment module 1104, configured to formulate a control scheme based on the predicted source-load power data and adjust the phases of single-phase user nodes in the substation area to achieve line loss reduction.

[0349] In one embodiment, the data interpolation model includes a first diagonal masked self-attention module, a second diagonal masked self-attention module, and a dynamic weighted combination module; the interpolation module 1102 further includes a masking processing sub-module, a splicing sub-module, a first interpolation sub-module, an updating sub-module, a second interpolation sub-module, and an interpolation sub-module, where:

[0350] The masking processing sub-module is configured to obtain a missing mask matrix pre-constructed according to the source-load power data, and perform masking processing on the initial source-load power data characteristics to obtain primary source-load power data characteristics.

[0351] The splicing sub-module is configured to splice the missing mask matrix and the primary source-load power data characteristics to obtain secondary source-load power data characteristics.

[0352] The first interpolation sub-module is configured to input the secondary source-load power data characteristics into the first diagonal masked self-attention module to obtain a first-stage interpolation result.

[0353] The updating sub-module is configured to update the primary source-load power data characteristics by using the first-stage interpolation result, and splice the updated primary source-load power data characteristics and the missing mask matrix to obtain high-level source-load power data characteristics.

[0354] The second interpolation sub-module is configured to input the high-level source-load power data characteristics into the second diagonal masked self-attention module to obtain a second-stage interpolation result.

[0355] An interpolation sub-module for inputting the missing mask matrix, the first-stage interpolation result, and the second-stage interpolation result into a dynamic weighted combination module to obtain a data interpolation result.

[0356] In one embodiment, the first diagonal mask self-attention module includes a first feature projection unit, a first feature enhancement unit, and a first feature restoration unit; the first feature enhancement unit is constructed based on the multi-head self-attention mechanism; the first interpolation sub-module is further configured to input the secondary source-load power data feature into the first feature projection unit to obtain the secondary source-load power data feature in the target dimension; input the secondary source-load power data feature in the target dimension into the first feature enhancement unit to obtain the feature-enhanced secondary source-load power data feature; and input the feature-enhanced secondary source-load power data feature into the first feature restoration unit to obtain the first-stage interpolation result.

[0357] In an exemplary embodiment, the second diagonal mask self-attention module includes a second feature projection unit, a second feature enhancement unit, and a second feature restoration unit; the second feature enhancement unit is constructed based on the multi-head self-attention mechanism; the second interpolation sub-module is further configured to input the high-level source-load power data feature into the second feature projection unit to obtain the high-level source-load power data feature in the target dimension; input the high-level source-load power data feature in the target dimension into the second feature enhancement unit to obtain the feature-enhanced high-level source-load power data feature; and input the feature-enhanced high-level source-load power feature into the second feature restoration unit to obtain the second-stage interpolation result.

[0358] In one embodiment, the second-stage interpolation result carries the weight information of the second feature enhancement unit; the data interpolation sub-module is further configured to obtain a weight factor according to the weight information and the missing mask matrix; and use the weight factor to perform weighted summation on the first-stage interpolation result and the second-stage interpolation result to obtain the data interpolation result.

[0359] In one of the embodiments, the initial source-load power data feature includes a source-load trend feature, a source-load seasonal feature, and an average daily source-load feature; the acquisition module 1101 further includes a time series construction sub-module, a detrending sub-module, a period decomposition sub-module, a deseasonalization sub-module, and a feature acquisition sub-module, where:

[0360] The time series construction sub-module is configured to generate a corresponding time series based on the initial source-load power data and obtain the average daily source-load feature according to the time series; the time series includes a source-load trend component, a source-load seasonal component, and a source-load residual component.

[0361] The detrending sub-module is configured to subtract a preset source-load trend component from the time series to obtain a detrended time series.

[0362] A periodic decomposition sub-module is used to decompose the detrended time series according to a preset period to obtain multiple time sub-series, and obtain a new source-load seasonal component based on the time sub-series.

[0363] A deseasonalization sub-module is used to subtract the new source-load seasonal component from the detrended time series to obtain a deseasonalized time series.

[0364] A feature acquisition sub-module is used to obtain source-load trend features and source-load seasonal features based on the deseasonalized time series and the new source-load seasonal component.

[0365] In an exemplary embodiment, the feature acquisition sub-module is further used to perform locally weighted regression filtering on the deseasonalized time series to extract a new source-load trend component; if the similarity between the new source-load trend component and the source-load trend component is greater than or equal to a preset first similarity, and the new source-load seasonal component and the source-load seasonal component are greater than or equal to a preset second similarity, then the new source-load trend component is determined as the source-load trend feature, and the new source-load seasonal component is determined as the source-load seasonal feature.

[0366] Each module in the above loss reduction device for adjusting the phase of a single-phase user node can be implemented in whole or in part by software, hardware, and their combination. The above modules can be embedded in the processor of the computer device in hardware form or independent of it, or stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to the above modules.

[0367] In an exemplary embodiment, a computer device is provided. The computer device can be a server, and its internal structure diagram can be as Figure 12 shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O), and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store source-load power data, meteorological information, time information, initial source-load power data features, meteorological data features, time data features, and source-load power data features. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals through a network connection. When the computer program is executed by the processor, it implements a loss reduction method for adjusting the phase of a single-phase user node.

[0368] Those skilled in the art can understand that Figure 12 the structure shown in Figure 12 is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0369] In an exemplary embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the loss reduction method for adjusting the phase of a single-phase user node in the above embodiment is implemented.

[0370] In an embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the loss reduction method for adjusting the phase of a single-phase user node in the above embodiment is implemented.

[0371] In an embodiment, a computer program product is provided, including a computer program. When the computer program is executed by a processor, the loss reduction method for adjusting the phase of a single-phase user node in the above embodiment is implemented.

[0372] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data that have been authorized by the user or fully authorized by all parties, and the collection, use, and processing of relevant data need to comply with relevant regulations.

[0373] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium used in the embodiments provided in the present application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The databases involved in the embodiments provided in the present application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., without limitation. The processors involved in the embodiments provided in the present application can be general-purpose processors, central processors, graphics processors, digital signal processors, programmable logic devices, data processing logics based on quantum computing, artificial intelligence (AI) processors, etc., without limitation.

[0374] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in the present application.

[0375] The above-described embodiments merely represent several implementation manners of the present application. The description thereof is relatively specific and detailed, but it should not be construed as a limitation to the patent scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all fall within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the appended claims.

Claims

1. A loss reduction method for adjusting the phase of single-phase user nodes, characterized in that, The method includes: Obtaining the source-load power data, meteorological information, and time information of the power grid area, and obtaining the initial source-load power data features corresponding to the source-load power data, as well as the meteorological data features and time data features corresponding to the meteorological information and the time information respectively; According to the initial source-load power data features and a pre-constructed data interpolation model based on the self-attention mechanism, obtaining a data interpolation result, and using the data interpolation result to update the initial source-load power data features to source-load power data features; Inputting the source-load power data features, the meteorological data features, and the time data features into a pre-constructed source-load power data prediction model, obtaining static data features through the convolutional neural network unit in the source-load power data prediction model, and obtaining dynamic data features through the gated recurrent unit in the source-load power data prediction model, and obtaining the predicted source-load power data corresponding to the power grid area according to the static data features and the dynamic data features; the static data features are used to characterize data features with stability and repeatability, and the dynamic data features are used to describe data features reflecting temporal dependence and change trends; Formulating a regulation plan based on the predicted source-load power data, and adjusting the phases of the single-phase user nodes in the power grid area to achieve line loss reduction.

2. The method according to claim 1, wherein The data interpolation model includes a first diagonal masked self-attention module, a second diagonal masked self-attention module, and a dynamic weighted combination module; The obtaining of the data interpolation result according to the initial source-load power data features and a pre-constructed data interpolation model based on the self-attention mechanism includes: Obtaining a missing mask matrix pre-constructed according to the source-load power data, and performing mask processing on the initial source-load power data features to obtain primary source-load power data features; Concatenating the missing mask matrix and the primary source-load power data features to obtain secondary source-load power data features; Inputting the secondary source-load power data features into the first diagonal masked self-attention module to obtain a first-stage interpolation result; Updating the primary source-load power data features using the first-stage interpolation result, and concatenating the updated primary source-load power data features and the missing mask matrix to obtain high-level source-load power data features; Inputting the high-level source-load power data features into the second diagonal masked self-attention module to obtain a second-stage interpolation result; Inputting the missing mask matrix, the first-stage interpolation result, and the second-stage interpolation result into the dynamic weighted combination module to obtain the data interpolation result.

3. The method according to claim 2, wherein The first diagonal masked self-attention module includes a first feature projection unit, a first feature enhancement unit, and a first feature restoration unit; the first feature enhancement unit is constructed based on the multi-head self-attention mechanism; The inputting of the secondary source-load power data features into the first diagonal masked self-attention module to obtain a first-stage interpolation result includes: Inputting the secondary source-load power data features into the first feature projection unit to obtain secondary source-load power data features in the target dimension; Input the secondary source-load power data features of the target dimension into the first feature enhancement unit to obtain the enhanced secondary source-load power data features; Input the enhanced secondary source-load power data features into the first feature restoration unit to obtain the first-stage interpolation result.

4. The method according to claim 2, characterized in that, The second diagonal masked self-attention module includes a second feature projection unit, a second feature enhancement unit, and a second feature restoration unit; the second feature enhancement unit is constructed based on the multi-head self-attention mechanism; The step of inputting the high-level source-load power data features into the second diagonal masked self-attention module to obtain the second-stage interpolation result includes: Input the high-level source-load power data features into the second feature projection unit to obtain the high-level source-load power data features of the target dimension; Input the high-level source-load power data features of the target dimension into the second feature enhancement unit to obtain the enhanced high-level source-load power data features; Input the enhanced high-level source-load power features into the second feature restoration unit to obtain the second-stage interpolation result.

5. The method according to claim 4, wherein The second-stage interpolation result carries the weight information of the second feature enhancement unit; The step of inputting the missing mask matrix, the first-stage interpolation result, and the second-stage interpolation result into the dynamic weighted combination module to obtain the data interpolation result includes: Obtain the weight factor according to the weight information and the missing mask matrix; Use the weight factor to perform weighted summation on the first-stage interpolation result and the second-stage interpolation result to obtain the data interpolation result.

6. The method according to claim 1, characterized in that, The initial source-load power data features include source-load trend features, source-load seasonal features, and average daily source-load features; The step of obtaining the initial source-load power data features corresponding to the source-load power data includes: Generate a corresponding time series based on the initial source-load power data, and obtain the average daily source-load features according to the time series; the time series includes a source-load trend component, a source-load seasonal component, and a source-load residual component; Subtract the preset source-load trend component from the time series to obtain the detrended time series; Decompose the detrended time series according to a preset period to obtain multiple time subsequences, and obtain a new source-load seasonal component according to the time subsequences; Subtract the new source-load seasonal component from the detrended time series to obtain the deseasonalized time series; Obtain the source-load trend features and the source-load seasonal features according to the deseasonalized time series and the new source-load seasonal component.

7. The method according to claim 6, characterized in that, The step of obtaining the source-load trend features and the source-load seasonal features according to the deseasonalized time series and the new source-load seasonal component includes: Perform locally weighted regression filtering on the deseasonalized time series to extract a new source-load trend component; If the similarity between the new source-load trend component and the source-load trend component is greater than or equal to a preset first similarity, and the similarity between the new source-load seasonal component and the source-load seasonal component is greater than or equal to a preset second similarity, then the new source-load trend component is determined as the source-load trend feature, and the new source-load seasonal component is determined as the source-load seasonal feature.

8. A loss reduction device for adjusting the phase of a single-phase user node, characterized in that, The device includes: An acquisition module, configured to acquire the source-load power data, meteorological information, and time information of the substation area, and acquire the initial source-load power data feature corresponding to the source-load power data, as well as the meteorological data feature and time data feature corresponding to the meteorological information and the time information respectively; An interpolation module, configured to obtain an interpolation result according to the initial source-load power data feature and a pre-constructed data interpolation model based on the self-attention mechanism, and update the initial source-load power data feature to the source-load power data feature by using the interpolation result; A prediction module, configured to input the source-load power data feature, the meteorological data feature, and the time data feature into a pre-constructed source-load power data prediction model, obtain static data features through the convolutional neural network unit in the source-load power data prediction model, and obtain dynamic data features through the gated recurrent unit in the source-load power data prediction model, and obtain the predicted source-load power data corresponding to the substation area according to the static data features and the dynamic data features; the static data features are used to characterize data features with stability and repeatability, and the dynamic data features are used to describe data features reflecting temporal dependence and change trends; An adjustment module, configured to formulate a regulation plan based on the predicted source-load power data and adjust the phase of the single-phase user node of the substation area to achieve line loss reduction.

9. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

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

Images