Large-scale transformer area short-term load prediction method

Abstract

The invention provides a large-scale transformer area short-term load prediction method, and the method comprises the steps: dividing transformer areas into temperature sensitive transformer areas and temperature inert transformer areas based on the historical data of each transformer area, constructing a dynamic temperature feature cumulative effect correction model for the temperature sensitive transformer areas according to the fluctuation rule difference of load and temperature features, and carrying out the prediction of the short-term load of each transformer area. According to the method, the temperature data is corrected, and the short-term load of the transformer area is predicted by using the model based on the characteristics corresponding to the modified temperature data, so that the influence of the extreme temperature on the load prediction can be eliminated or reduced, and the load prediction precision is improved. In addition, a transformer area short-term load prediction model in which the incremental learning method and the iTransform are combined is constructed, so that the load prediction efficiency can be effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system load forecasting, in particular to a large-scale substation short-term load forecasting method based on cumulative effect feature correction and incremental learning. BACKGROUND

[0002] Load forecasting, as a means of perception for predicting and judging future load changes, is an important part of the intelligent power grid data platform. At the same time, low-voltage distribution substation is the terminal link of intelligent fine management of power grid. Accurate prediction of regional distributed substation load can provide important basis for distribution capacity configuration, low-voltage distribution network planning, line loss reduction, risk early warning and other work.

[0003] Different from regional load forecasting at the system level and bus side, substation short-term load forecasting is greatly influenced by meteorological, human and other factors, and the nonlinearity and non-stationarity of load sequence are prominent. Under the background of new power system construction, the terminal electrification level is continuously improved, and the power load composition is more complex, making it more difficult to grasp the rules. In addition, with the development of the power industry and the changes in the power market, new technologies and roles such as demand response, sharing economy, load aggregator, and virtual power plant are constantly emerging, making the power load present new characteristics and new forms of more complex and variable: on the one hand, the randomness and uncertainty of load changes are enhanced, and they are easily affected by extreme weather and other factors. In recent years, due to the frequent occurrence of extreme high and low temperature weather conditions, load fluctuations are difficult to predict and adjust, resulting in a lack of power supply. On the other hand, due to the increasingly complex load structure, the power system has higher requirements for the spatial granularity and time accuracy of load forecasting. At the same time, the large number and strong heterogeneity of substation loads make it difficult for traditional forecasting models to capture the complex fluctuation characteristics of substation loads and lack the ability to handle massive substation data.

[0004] At present, there are many power load prediction methods, such as patent CN202411662478.0 proposes a small sample power load prediction method based on transfer learning, which can extract common knowledge from similar tasks and migrate it to the training process of the target task, thereby improving the performance of the model in processing the target task; Patent CN202411761457.4 proposes a super short-term load prediction method and system based on a hybrid expert model, which optimizes the prediction accuracy, has shorter overall training time and lower cost, and each expert sub-model can be trained in parallel, greatly shortening the prediction cycle. But these methods are generally difficult to be well applied to large-scale transformer area load prediction. For this demand, there are still many technical challenges. First, with the development of intelligent power systems, more and more electrical equipment is connected to the power system, resulting in changes in transformer area load, especially the short-term load change of low-voltage distribution transformer area is more easily affected by weather conditions, electricity price, power usage habits, social time, holidays and other factors, so how to extract a high correlation feature dataset to help the deep network model prediction is a big difficulty that needs to be considered in the current short-term load prediction technology. Secondly, there is high heterogeneity between different distribution transformer areas, and the load data of each transformer area fluctuates, and the influencing factors are different, and the power load has obvious seasonality, periodicity and uncertainty, which brings difficulties to large-scale cluster transformer area load prediction. Finally, in a large area, there are a large number of transformer distribution transformers, and their historical load data and feature data are of large scale. In the face of the influx of massive data, conventional load prediction models often lack the ability to process big data, greatly weakening the timeliness of load prediction, so it is urgent to build an efficient load prediction model to process massive data and maximize the timeliness of load prediction. SUMMARY

[0005] The present application is to solve the above problems, and aims to provide a method capable of accurately predicting large-scale transformer short-term load. The present application adopts the following technical solutions:

[0006] The application provides a large-scale transformer area short-term load prediction method, which has the technical features that the method comprises the following steps: step S1, obtaining historical data of each transformer area and preprocessing the historical data to obtain a transformer area data set, wherein the historical data at least comprises temperature data and load data; step S2, performing feature analysis based on the transformer area data set, and dividing the transformer area into a temperature sensitive transformer area and a temperature inert transformer area based on the feature analysis result; step S3, for the temperature sensitive transformer area, constructing a cumulative effect correction model, and correcting the temperature data of the temperature sensitive transformer area by using the cumulative effect correction model to eliminate the influence of extreme temperature on load prediction; step S4, for each transformer area, constructing a personalized feature set of the transformer area based on the feature analysis result, wherein for the temperature sensitive transformer area, the personalized feature set contains temperature features corresponding to the corrected temperature data; step S5, constructing a transformer area short-term load prediction model combined with an incremental learning method and an inverter transformer (iTransformer), and obtaining short-term load prediction results of each transformer area based on the personalized feature set by using the transformer area short-term load prediction model.

[0007] The large-scale transformer area short-term load prediction method provided by the application can also have the technical features that the transformer area data set at least contains load data and temperature data, and step S2 comprises the following sub-steps: step S2-1, performing statistical analysis based on the load data and the temperature data, and constructing a daily load sequence and a temperature sequence of the transformer area; step S2-2, performing feature analysis on the daily load sequence and the temperature sequence to obtain the feature analysis result, which contains load entropy spectrum, time period feature and lag correlation feature of temperature relative to load lag, wherein the load entropy spectrum contains daily load approximate entropy feature; step S2-3, dividing the transformer area into the temperature sensitive transformer area and the temperature inert transformer area by using an extreme gradient boosting tree based on the daily load approximate entropy feature, the time period feature and the lag correlation feature.

[0008] The large-scale transformer area short-term load prediction method provided by the application can also have the technical features that in step S2-1, the daily load sequence is represented as:

[0009]

[0010] In the formula, x i is the load value at the i th moment, and m is the total amount of load sequence data. In step S2-2, the daily load approximate entropy feature is represented as:

[0011]

[0012] wherein r is a tolerance coefficient related to a standard deviation of the daily load sequence, N is a number of sampling points, is an average logarithmic probability, obtained based on a proportion of similar pattern pairs in a sub-sequence of the daily load sequence, obtained based on the tolerance coefficient. The period feature comprises a night load ratio and a temperature sensitivity, the night load ratio being expressed as:

[0013]

[0014] wherein n is a start point of daytime load, k is an end point of daytime load, P i is a load value at the i th time of night in the daily load sequence. The temperature sensitivity is expressed as:

[0015]

[0016] wherein ΔP is a load change amount, ΔT is a temperature change amount. The lag correlation feature is expressed as:

[0017]

[0018] wherein P t is a load value at t th time in the daily load sequence, T t-k is a temperature value at t-k th time in the temperature sequence, D is a variance of load and temperature.

[0019] The large-scale area short-term load prediction method provided by the present application can further have the following technical features: in step S2-2, the maximum distance between each pair of sub-sequences in the daily load sequence is calculated:

[0020]

[0021] The tolerance coefficient is set as wherein σ p is a standard deviation of the daily load sequence. The pair of sub-sequences with the maximum distance less than the tolerance coefficient is taken as the similar pattern pair, and the proportion of the similar pattern pair is:

[0022] .

[0023] The large-scale transformer area short-term load prediction method provided by the application can further have the following technical features: step S3 comprises the following sub-steps: step S3-1, extracting the transformer area data set of the temperature-sensitive transformer area from the plurality of transformer area data sets; step S3-2, fitting a temperature-load curve based on the transformer area data set, and taking the temperature at which the load elasticity coefficient is maximum as the limit temperature; step S3-3, calculating the correlation coefficient of the cumulative temperature and the load of the predetermined number of days by using a sliding window based on the temperature-load curve, and taking the day in which the correlation coefficient is maximum as the maximum cumulative day; step S3-4, constructing a temperature correction term and an objective function based on the limit temperature and the maximum cumulative day, wherein the temperature correction term is used to correct the temperature data of the temperature-sensitive transformer area, and the objective function is used to maximize the correlation between the temperature correction term and the corresponding load; step S3-5, solving the objective function to obtain a cumulative effect coefficient; step S3-6, judging whether the objective function has converged to the minimum according to a predetermined convergence rule, and returning to step S3-5 when the judgment is no; and step S3-7, constructing a cumulative effect correction model based on the cumulative effect coefficient and a dynamic weight coefficient when the judgment in step S3-6 is yes, and correcting the temperature data of the temperature-sensitive transformer area by using the cumulative effect correction model.

[0024] The large-scale transformer area short-term load prediction method provided by the application can further have the following technical features: in step S3-2, the temperature-load curve is expressed as:

[0025]

[0026] In the formula, T t is the temperature at time t, L t is the load at time t, n is the order of the polynomial, β0, β1, β2, …, β n are undetermined coefficients. The load elasticity coefficient is expressed as:

[0027]

[0028] The limit temperature is expressed as:

[0029]

[0030] In the formula, T represents the temperature value at a certain time, is the index for finding the maximum value of an array. In step S3-3, the maximum cumulative day is expressed as:

[0031]

[0032] In the formula, T is a correlation coefficient between average temperature and response load, and is expressed as:

[0033]

[0034] wherein L t is a response load, is an average temperature of the past several days, and is expressed as:

[0035]

[0036] wherein T t-j is a temperature at t-j. In step S3-4, the temperature correction term is expressed as:

[0037]

[0038] wherein T t is a temperature at t, and T t-j,24 is a temperature at the previous t-j. The objective function is expressed as:

[0039]

[0040] wherein, represents a correlation between two variables, and k j is a cumulative effect coefficient. In step S3-5, the objective function is solved by using a sequential least squares programming numerical optimization algorithm.

[0041] The large-scale area short-term load forecasting method provided by the present applicationapplicationalso have the following technical features: in step S3-7, the dynamic weight coefficient is expressed as:

[0042]

[0043] wherein, represents a sliding average result of the temperature within a window range of a predetermined length centered at t. The cumulative effect correction model is expressed as:

[0044]

[0045] wherein L min , L max are the minimum load and the maximum load in the temperature-load curve, respectively.

[0046] The large-scale transformer area short-term load prediction method provided by the present application can further have the following technical features: in step S5, the incremental learning method is used to train a student model based on a teacher model, wherein the teacher model is trained based on the accumulated historical data, and the trained student model is used as the transformer area short-term load prediction model. The incremental learning method includes: first step learning, training the student model based on the output of the teacher model, the incremental data, and a distillation loss function; and second step learning, matching the incremental data with the historical data to obtain matching data, and training the student model based on the matching data and a mean absolute error loss function to achieve balanced learning.

[0047] The large-scale transformer area short-term load prediction method provided by the present application can further have the following technical features: in step S5, the inverse transformer includes: an input module including an input layer and an embedding layer, the input layer being used to receive original input variables, and the embedding layer being used to map the original input variables into high-dimensional feature tokens; a feature extraction module including a multi-head self-attention unit, a first layer normalization unit, a feedforward neural network, and a second layer normalization unit, the self-attention unit being used to extract multiple features in a time series from different subspaces in parallel, and the feedforward neural network being used to extract nonlinear relationships in the time series; and an output module including a projection layer and an output layer, the projection layer being used to filter and retain key features extracted by the feature extraction module, and map the key features into transformer area short-term load prediction results, and the output layer being used to output the transformer area short-term load prediction results.

[0048] Effects of the present application

[0049] According to the large-scale transformer area short-term load prediction method provided by the present application, the transformer areas are divided into temperature-sensitive transformer areas and temperature-inert transformer areas based on the historical data of each transformer area, and for the temperature-sensitive transformer areas, a dynamic temperature feature cumulative effect correction model is constructed according to the differences in the fluctuation rules of load and temperature characteristics, the temperature data is corrected, and the transformer area short-term load is predicted based on the features corresponding to the modified temperature data using the model, so that the influence of extreme temperature on load prediction can be eliminated or reduced, and the load prediction accuracy can be improved. In addition, since the incremental learning method and the iTransformer combined transformer area short-term load prediction model are constructed, the efficiency of load prediction can also be effectively improved. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 is a flowchart of the large-scale transformer area short-term load prediction method in the embodiments of the present application;

[0051] Figure 2is a schematic diagram of a mixed substation classification model in an embodiment of the present application;

[0052] Figure 3 is a flowchart of step S2 in an embodiment of the present application;

[0053] Figure 4 is a flowchart of step S3 in an embodiment of the present application;

[0054] Figure 5 is a schematic diagram of the structure of an inverter in an embodiment of the present application;

[0055] Figure 6 is a schematic diagram of the framework of an incremental learning method in an embodiment of the present application;

[0056] Figure 7 is a graph of the actual load curve, the temperature prediction curve before and after temperature correction, and the load prediction curve of a type 1 substation in an embodiment of the present application;

[0057] Figure 8 is a graph of the actual load curve, the temperature prediction curve before and after temperature correction, and the load prediction curve of a type 2 substation in an embodiment of the present application;

[0058] Figure 9 is a graph of the actual load curve, the temperature prediction curve before and after temperature correction, and the load prediction curve of a type 3 substation in an embodiment of the present application;

[0059] Figure 10 is a graph of the actual load curve and the predicted load curve of a type 1 substation in an embodiment of the present application;

[0060] Figure 11 is a graph of the actual load curve and the predicted load curve of a type 2 substation in an embodiment of the present application;

[0061] Figure 12 is a graph of the actual load curve and the predicted load curve of a type 3 substation in an embodiment of the present application. DETAILED DESCRIPTION

[0062] In order to make the technical means, creative features, purposes and effects realized by the present application easy to understand, the present application is described in detail below in combination with embodiments and drawings.

[0063] The present embodiment provides a large-scale substation short-term load prediction method for short-term load prediction of large-scale substations, wherein the substation is a basic power supply unit of a distribution network in a power system, and a large-scale substation generally refers to a large number of substations, a wide coverage of multiple substations, and / or multiple complex load characteristics of substations.

[0064] Figure 1is a flowchart of the large-scale transformer short-term load forecasting method in this embodiment.

[0065] As shown in Figure 1 the large-scale transformer short-term load forecasting method includes the following steps:

[0066] Step S1, obtain the historical data of each transformer and pre-process the historical data to form a transformer dataset.

[0067] Step S2, perform feature analysis based on the transformer dataset and divide the transformer into temperature-sensitive and temperature-inert types based on the feature analysis results.

[0068] Step S3, for the temperature-sensitive transformer, build a cumulative effect correction model of dynamic temperature characteristics based on the cumulative effect of extreme high or low temperature on the transformer load, and correct the temperature data of the temperature-sensitive transformer using the model.

[0069] Step S4, for each transformer, build a personalized feature set of the transformer based on the feature analysis results, wherein for the temperature-sensitive transformer, the features include temperature features corresponding to the corrected temperature data.

[0070] Step S5, build a transformer short-term load forecasting model combining incremental learning method and iTransformer, and obtain the short-term load forecasting results of each transformer based on the personalized feature set using the model.

[0071] The above steps will be described in detail below.

[0072] Step S1, obtain the historical data of each transformer and pre-process the historical data to form a transformer dataset.

[0073] The historical data of the transformer includes historical load, weather, social, and attribute data, and the weather data includes temperature data. In this step, the pre-processing of the historical data may include data cleaning, missing data completion, etc.

[0074] Step S2, perform feature analysis based on the transformer dataset and divide the transformer into temperature-sensitive and temperature-inert types based on the feature analysis results.

[0075] In this step, the daily load curve and temperature load curve of the transformer are obtained based on the historical weather data and historical load data, and feature analysis is performed based on the inherent rules and time sequence differences of the daily load curve and temperature load curve, to obtain the feature analysis results, and then the transformer is preliminarily classified as temperature-sensitive or temperature-inert (conventional) using XGBoost based on the feature analysis results.

[0076] Figure 2 is a schematic diagram of the mixed substation classification model in this embodiment, Figure 3 is a flowchart of step S2 in this embodiment.

[0077] As shown in Figure 2 and Figure 3 , step S2 specifically includes the following sub-steps:

[0078] Step S2-1, based on the load data and temperature data of the substation, statistical analysis is performed, and the daily load sequence and temperature sequence of the substation are constructed.

[0079] For example, based on historical daily load data and a predetermined time granularity, the following daily load sequence can be constructed, that is, an m-dimensional vector is generated:

[0080]

[0081] In the formula, x i is the load value at the i-th moment, , m is the total amount of load sequence data, and N is the number of sampling points.

[0082] Similarly, the temperature sequence can be constructed based on historical meteorological data and a predetermined time granularity.

[0083] Step S2-2, feature analysis is performed on the daily load sequence and temperature load sequence, and feature analysis results are obtained, including load entropy spectrum, time period feature, and lag correlation feature.

[0084] Among them, the load entropy spectrum includes daily load approximate entropy feature, daily load curve and temperature load curve.

[0085] The daily load approximate entropy feature is used to measure the complexity and predictability of the load sequence, which can be obtained in the following manner: the maximum distance between each pair of sub-sequences in the daily load sequence is calculated , and the calculation formula is as follows:

[0086]

[0087] Set the similar mode ratio, set the tolerance coefficient , where σ p is the standard deviation of the daily load sequence. Among the multiple pairs of sub-sequences of the daily load sequence, the pairs of sub-sequences with a distance less than the tolerance coefficient r are recorded as similar mode pairs, and the ratio of similar mode pairs is :

[0088]

[0089] The definition of the daily load approximate entropy feature is:

[0090]

[0091] wherein, is the average logarithmic probability, which can be calculated based on the proportion of similar pattern pairs

[0092] The period feature includes two features, i.e., a night load ratio and a temperature sensitivity.

[0093] The night load ratio The calculation formula is as follows:

[0094]

[0095] wherein, m is the total amount of load sequence data, n is the start point of daytime load, k is the end point of daytime load, P i is the load value at the i th moment of night in the daily load sequence.

[0096] The temperature sensitivity is the elasticity coefficient of load change with temperature, and the calculation formula is as follows:

[0097]

[0098] wherein, ΔP is the load change amount, and ΔT is the temperature change amount.

[0099] The lag correlation feature is used to represent the difference in time sequence between the daily load sequence and the temperature sequence. For the daily load sequence and the temperature sequence with a 24h time granularity, the lag correlation of the daily load sequence and the temperature sequence can be regarded as the correlation coefficient of the daily load sequence and the temperature, which is calculated as follows:

[0100]

[0101] wherein, P t is the load value at the t th moment in the daily load sequence, T t-k is the temperature value at the t-k th moment in the temperature sequence, i.e., the temperature lags for k time units, and D is the variance of load and temperature.

[0102] In step S2-3, based on the daily load approximate entropy feature, the period feature, and the lag correlation feature, an XGBoost (Extreme Gradient Boosting) is used to divide the transformer area into a temperature sensitive transformer area or a temperature inert transformer area.

[0103] In step S2-4, based on the daily load approximate entropy feature, the period feature, and the lag correlation feature, a hierarchical clustering algorithm is used to cluster multiple transformer area data sets to obtain multiple clustering clusters with different load time sequence features, and the transformer area is divided into multiple types with different load features based on the clustering clusters.

[0104] ​Specifically, the clustering algorithm is a non-supervised machine learning method, which can classify the given data set according to its data distribution. Among them, the agglomerative hierarchical clustering is a clustering method based on the similarity of cluster data, which can construct a hierarchical tree clustering result by merging similar clusters layer by layer.

[0105] In this step, the hierarchical clustering algorithm includes a data processing process, a hierarchical clustering process and a cluster formation process.

[0106] In the data processing process, a large number of transformer area data sets are constructed, and missing values and outliers in each transformer area data set are detected and corrected, data cleaning is performed to avoid the influence of data anomalies on the subsequent clustering effect.

[0107] In the hierarchical clustering process, each transformer area data set is regarded as an independent cluster, the distance between each cluster is calculated, and Euclidean distance can be selected as the distance measurement formula. According to the calculation result of Euclidean distance, the two clusters with the shortest distance are merged into a new cluster.

[0108] For example, for two points in n-dimensional space , The Euclidean distance between the two points is:

[0109]

[0110] In the formula, x i , y i are the i-th elements of sequences x and y, respectively.

[0111] In the cluster formation process, the above hierarchical clustering process is repeatedly cycled to continuously merge similar clusters and gradually form cluster clusters with similar characteristics until the preset loop stop condition is met. The hierarchical clustering algorithm is completed, and finally a tree-shaped multiple clustering cluster is formed. Multiple transformer areas in the same clustering cluster have similar characteristics in load time sequence.

[0112] In this embodiment, the hierarchical clustering algorithm is used to divide the multiple transformer area data sets into three types, each type having corresponding load time sequence characteristics. The three types are type 1 (daytime peak type), type 2 (nighttime peak type) and type 3 (double peak type). For example, Figure 2As shown, the load curve of Type 1, for example, a transformer area containing manufacturing plants, etc., has a peak in the daytime, and the power load is relatively stable; the load curve of Type 2, for example, a transformer area containing staggered operation of manufacturing plants, or a transformer area containing energy storage systems for energy storage at night, etc., has a peak at night; the load curve of Type 3, for example, a transformer area mixed with business and residence, or a transformer area as an industrial park, etc., has a peak at about 9:00-12:00 in the daytime and at about 16:00-20:00 at night.

[0113] In step S3, a cumulative effect correction model of dynamic temperature characteristics is constructed based on the cumulative effect of extreme high temperature or extreme low temperature on the load of the transformer area, and the temperature data of the temperature-sensitive transformer area is corrected using the model.

[0114] Since the power consumption behavior of people will be affected by the extreme high temperature or extreme low temperature of the historical day and show certain inertia, in this step, the cumulative effect of extreme high temperature or extreme low temperature on the load of the transformer area is considered, and a dynamic cumulative effect correction model considering the difference in fluctuation rule of load and temperature characteristics at the daily time scale is constructed, and the temperature in the transformer area data set of the temperature-sensitive transformer area is corrected using the model to eliminate the influence of historical extreme temperature on the prediction.

[0115] Figure 4 is the flowchart of step S3 in this embodiment.

[0116] As shown in Figure 4 Step S3 includes the following sub-steps:

[0117] Step S3-1, extract the transformer area data set of the temperature-sensitive transformer area from the plurality of transformer area data sets.

[0118] Step S3-2, for each extracted transformer area data set, fit the temperature-load curve of the corresponding transformer area based on the transformer area data set, and take the temperature at the maximum load elasticity coefficient as the limit temperature.

[0119] Among them, the temperature point in the transformer area data set that is most affected by temperature (i.e. the maximum sensitivity of load to temperature change) is defined as the limit temperature, and the limit temperature is determined according to the load-temperature fitting curve and the load elasticity coefficient.

[0120] Specifically, based on the historical load data and historical meteorological data of the transformer area, first, the temperature-load curve is fitted by an n-order polynomial:

[0121]

[0122] In the formula, T t is the temperature at time t, L tis the load at time t, n is the order of the polynomial; β0, β1, β2, …, β n is the undetermined coefficient.

[0123] The load elasticity coefficient e t is:

[0124]

[0125] The limit temperature T min is the temperature corresponding to the maximum load elasticity coefficient in the temperature-load curve:

[0126]

[0127] In the formula, represents the temperature value at a certain time, is the index for finding the maximum value of the array.

[0128] Step S3-3, based on the temperature-load curve, the correlation coefficient of the cumulative temperature and the load of the predetermined number of days is calculated by using the sliding window, and the day with the maximum correlation coefficient is taken as the maximum cumulative day.

[0129] wherein the average temperature of the past d max +1 days is defined as:

[0130]

[0131] In the formula, d max is the maximum cumulative day to be calculated, T t-j is the temperature at time t-j, that is, the temperature lags the load at time t by j time units.

[0132] The response load L t corresponding to the temperature is calculated, and the correlation with the average temperature is calculated, and the day with the maximum correlation is taken as the maximum cumulative day d max .

[0133]

[0134]

[0135] In this embodiment, the predetermined number of days is the first day to the fifth day.

[0136] Step S3-4, based on the limit temperature and the maximum cumulative day, a temperature correction term and an objective function are constructed, wherein the temperature correction term is used to correct the temperature of the temperature-sensitive transformer area, and the objective function is used to maximize the correlation between the temperature correction term and the corresponding load, which contains the corresponding correlation coefficient (cumulative effect coefficient).

[0137] ​wherein, the temperature correction term T t corr is defined as:

[0138]

[0139] wherein, T t is the temperature at time t, taking the temperature of the latest day; T t-j,24 is the temperature at time t-j; d max is the maximum cumulative number of days calculated by the above method; T min is the limit temperature calculated by the above method.

[0140] The corresponding objective function is defined as:

[0141]

[0142] wherein, represents the correlation between two variables, k j is the cumulative effect coefficient.

[0143] Solving the above objective function to obtain the cumulative effect coefficient k j , which is the negative correlation coefficient of temperature and load.

[0144] Step S3-5, solving the objective function to obtain the cumulative effect coefficient.

[0145] wherein, the Sequential Least Squares Programming (SLSQP) numerical optimization algorithm is used to solve the objective function to obtain the cumulative effect coefficient k j that can minimize the objective function.

[0146] Step S3-6, judging whether the objective function converges to the minimum according to the predetermined convergence rule, returning to step S3-5 for solving again when the judgment is no, and outputting the cumulative effect coefficient k j when the judgment is yes.

[0147] wherein, the convergence rule may be, for example, comparing the results of multiple solutions (i.e., cumulative effect coefficients), and determining that the objective function has converged to the minimum when the variation of the results of the latest several solutions is less than a predetermined variation threshold.

[0148] Step S3-7, when the judgment in step S3-6 is yes, constructing a cumulative effect correction model based on the cumulative effect coefficient and the dynamic weight coefficient, and correcting the temperature of the temperature-sensitive area using the model.

[0149] wherein, in order to fully consider the load curve characteristics and the influence of temperature change at different time points on the load, a dynamic weight coefficient ω is further introduced t for dynamically adjusting the weight of temperature in the temperature correction term.

[0150] For a period of load curve, the normalized load is:

[0151]

[0152] In the formula, L min , L max are the minimum load and the maximum load in the temperature-load curve respectively; L t represents the current load.

[0153] In order to highlight the temperature influence in the high load period, a dynamic weight coefficient ω t is defined as:

[0154]

[0155] In the formula, represents the sliding average result of temperature in the window range with t as the center and a predetermined length, for example, 3.

[0156] Therefore, the final cumulative temperature correction formula with the weight coefficient is (that is, the cumulative effect correction model can be expressed as):

[0157]

[0158] Step S4, for each area, based on the feature analysis result, a personalized feature set of the area is constructed, wherein for the temperature sensitive area, the features include temperature features corresponding to the corrected temperature data.

[0159] For each area, the personalized feature set of the area can be constructed in combination with its time sequence features, load features, weather features and attribute features. For the temperature sensitive area, the weather features include temperature features corresponding to the corrected temperature data.

[0160] Step S5, a short-term load prediction model of the area is constructed by combining the incremental learning method and the iTransformer, and the model is used to obtain the short-term load prediction result of each area based on the personalized feature set.

[0161] In this embodiment, considering that the substation load data has significant time sequence dependence, a substation short-term load prediction model is constructed based on iTransformer. iTransformer is a deep learning model based on the Transformer architecture, which can effectively capture the global dependence between time tokens in time series, and especially shows excellent performance in multivariate time series prediction.

[0162] Figure 5 is the schematic diagram of the inverter in this embodiment.

[0163] As shown in Figure 5 , iTransformer adopts a Transformer structure containing only an encoder, which is composed of an input module, a feature extraction module, and an output module.

[0164] The input module includes an input layer and an embedding layer. The input layer is used to receive the original input variables. The embedding layer maps the original input variables to high-dimensional feature tokens, fully extracting the time dependence between variables.

[0165] The feature extraction module includes a multi-head self-attention unit, a first normalization unit, a feedforward neural network, and a second normalization unit connected in turn, and the multi-head self-attention unit and the feedforward neural network have a residual connection structure. The multi-head self-attention unit includes multiple self-attention heads, which can capture multiple features in the time series from different subspaces in parallel. The feedforward neural network (usually a fully connected layer) is used to enhance the representation ability of the network and extract more complex nonlinear relationships.

[0166] Specifically, in the multi-head self-attention unit, the query vector Q, the key vector K, and the value vector V are defined respectively, and the attention calculation formula is as follows:

[0167]

[0168] In the formula, d K is the dimension of the key vector K.

[0169] When the query vector Q, the key vector K, and the value vector V come from the same input, it constitutes a self-attention mechanism, which can capture the feature relationship within a single time series. By constructing multiple self-attention heads in parallel, a multi-head self-attention mechanism is formed, which not only improves the capacity of model parameters, but also enhances the ability of the model to mine potential deep associations in data.

[0170] Meanwhile, each word is normalized by a layer normalization unit to eliminate the influence of feature dimension difference, and the normalization formula is:

[0171]

[0172] In the formula, H is the feature matrix to be normalized; h n is the nth sequence in the feature matrix; is a mean function used to calculate the average of the sequence; is a variance function, where .

[0173] The output module includes a projection layer and an output layer. The projection layer is used to filter and retain key features, discard redundant or interfering information, thereby improving the accuracy of the final prediction result, and map the key features (hidden states) to the values of future time steps, i.e. the substation short-term load prediction result. The output layer is used to output the result.

[0174] Figure 6 is a schematic diagram of the framework of the incremental learning method in the embodiment.

[0175] As Figure 6 shown, in the incremental learning method, the goal is to quickly train a student model that can adapt to the incremental data stream based on the teacher model and the incoming incremental data stream, for the prediction task after the next data update. The teacher model is trained based on a large amount of accumulated historical data (i.e. original substation load data), and can achieve high-precision prediction for existing substation load data; the trained student model is the substation short-term load prediction model.

[0176] Moreover, in the incremental learning method of the embodiment, a two-step learning mechanism, or two-step training method, is adopted. First, the student model learns from the incremental data, and then performs balanced learning.

[0177] Specifically, in the first step of learning, only the new incremental data is used to train the student model based on the output of the teacher model, the incremental data, and the distillation loss function, and the student model is updated through gradient to adapt to the new incremental data.

[0178] In the second step of learning, the incremental data is matched with the historical data, and the student model is trained based on the matched data and the mean absolute error (MAE) loss function to achieve balanced learning, so that the student model can adapt to the historical data and the incremental data. Since there are also good and bad data in a large amount of historical data, it is difficult to mix them into the same data set for training. Therefore, in the embodiment, the Euclidean distance between each sequence in the incremental data and each sequence in the historical data is calculated, and sequences with similar distributions are selected for matching to form the matched data.

[0179] Figure 7 is a graph of the actual load curve, the temperature prediction curve before and after temperature correction, and the load prediction curve of a type 1 transformer area in the embodiment, Figure 8 is a graph of the actual load curve, the temperature prediction curve before and after temperature correction, and the load prediction curve of a type 2 transformer area in the embodiment, Figure 9 is a graph of the actual load curve, the temperature prediction curve before and after temperature correction, and the load prediction curve of a type 3 transformer area in the embodiment.

[0180] As shown in Figures 7 to 9 , the difference between the temperature curves before and after correction is large, especially at the load peak. The fluctuation degree and change trend of the temperature after the cumulative effect correction are more consistent with the load. For the three types of transformer areas, the corrected temperature curve (the curve of “temperature_corrected” in the figure, indicated by a dotted line with round dots) is more consistent with the fluctuation state of the load curve than the original temperature curve (the curve of “temperature_corrected” in the figure, indicated by a dotted line with square dots).

[0181] At the same time, it can be seen that the predicted load curve based on the corrected temperature (the curve of “predicted load_corrected” in the figure, indicated by a solid line with triangular dots) is closer to the actual load (indicated by a solid line with square dots in the figure) than the predicted load curve based on the original temperature (the curve of “predicted load_corrected” in the figure, indicated by a solid line with round dots), and this is true for the three types of transformer areas, proving the effectiveness of the method of the embodiment.

[0182] Figure 10 is a graph of the actual load curve and the predicted load curve of a type 1 transformer area in the embodiment, Figure 11 is a graph of the actual load curve and the predicted load curve of a type 2 transformer area in the embodiment, Figure 12 is a graph of the actual load curve and the predicted load curve of a type 3 transformer area in the embodiment. The dashed line in the figure is the actual load curve, and the solid line is the predicted load curve. The vertical dashed line in the middle of the figure is the dividing line, and the left side of the dividing line is the prediction result of the model trained only with historical data, and the right side of the dividing line is the prediction result of the model trained by adding incremental data and through the above-mentioned method of the embodiment.

[0183] AsFigures 10 to 12 As shown in the figure, for the three types of transformer area, the predicted load curve has a high degree of coincidence with the actual load curve, proving the effectiveness of the method of the embodiment. And from the figure, it can be observed that the predicted load curve on the right side of the boundary line is more fitted with the actual load curve, which is because the loss function of incremental learning and the two-step learning mechanism can effectively help the model adapt to new incremental data, which also proves the effectiveness of the incremental learning method of the embodiment.

[0184] According to the large-scale transformer area short-term load prediction method based on cumulative effect feature correction and incremental learning provided in the embodiment, the transformer area is divided into temperature sensitive transformer area and temperature inert transformer area based on the historical data of each transformer area, and for the temperature sensitive transformer area, a dynamic temperature feature cumulative effect correction model is constructed according to the difference in the fluctuation rule of the load and temperature characteristics, the temperature data is corrected, and then the short-term load of the transformer area is predicted based on the features corresponding to the modified temperature data using the model, so that the influence of extreme temperature on load prediction can be eliminated or reduced, and the load prediction accuracy can be improved. In addition, since the incremental learning method and the iTransformer combined transformer area short-term load prediction model are constructed, the efficiency of load prediction can also be effectively improved.

[0185] In the embodiment, the daily load sequence and temperature load sequence are constructed through statistical analysis of the original historical data, and the daily load approximate entropy feature, time period feature and lag correlation feature are obtained through feature analysis. Based on these features, the temperature sensitive transformer area or temperature inert transformer area is preliminarily divided using XGBoost, and in addition, the transformer area data set is divided into three categories according to its load characteristics through hierarchical clustering algorithm, so as to realize the double-layer fine classification of large-scale mixed transformer area. The temperature feature of the temperature sensitive transformer area is corrected by combining the correction model, and a high correlation feature data set of multiple types of transformer area is formed, which can be used for model training of the embodiment, and also can be used for other model training or further statistical analysis, providing a better data basis for the research of large-scale transformer area load.

[0186] Further, in the embodiment, the incremental learning method adopts a two-step learning mechanism, which first allows the student model to learn based on the incremental data, and then learns based on the matching data of the incremental data and the historical data, so as to enable the prediction model to quickly adapt to new incremental data, and enable the prediction model to balance learning, so that it can achieve more accurate prediction on both incremental data and historical data.

[0187] The above examples are only used for illustrating the specific embodiments of the present application, and the present application is not limited to the description range of the above examples. It should be understood by those skilled in the art that the present application is not limited by the above examples, and the above examples and the description in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for short-term load forecasting of large-scale transformer substations, characterized in that, Includes the following steps: Step S1: Obtain historical data for each transformer area and preprocess it to obtain a transformer area dataset, wherein the historical data includes at least temperature data and load data; Step S2: Perform feature analysis based on the data set of the transformer area, and divide the transformer area into temperature-sensitive transformer areas and temperature-inert transformer areas based on the feature analysis results; Step S3: For the temperature-sensitive transformer area, construct a cumulative effect correction model and use the cumulative effect correction model to correct the temperature data of the temperature-sensitive transformer area in order to eliminate the impact of extreme temperatures on load forecasting. Step S4: For each of the transformer substations, a personalized feature set for the transformer substation is constructed based on the feature analysis results. For the temperature-sensitive transformer substation, the personalized feature set includes the temperature features corresponding to the corrected temperature data. Step S5: Construct a short-term load forecasting model for the transformer area that combines the incremental learning method with the inverted converter, and use this model to obtain the short-term load forecasting results for each transformer area based on the personalized feature set.

2. The large-scale transformer area short-term load forecasting method according to claim 1, Its features are: The data set for the transformer substation includes at least load data and temperature data. Step S2 includes the following sub-steps: Step S2-1: Perform statistical analysis based on the load data and temperature data, and construct the daily load sequence and temperature sequence of the transformer area; Step S2-2: Perform feature analysis on the daily load sequence and the temperature sequence to obtain the feature analysis results, which include load entropy spectrum, time period characteristics, and lag correlation characteristics of temperature relative to load. The load entropy spectrum includes approximate entropy characteristics of daily load. Step S2-3: Based on the approximate entropy characteristics of the daily load, the time period characteristics, and the hysteresis correlation characteristics, the transformer area is divided into the temperature-sensitive transformer area and the temperature-inert transformer area using an extreme gradient boosting tree.

3. The large-scale transformer area short-term load forecasting method according to claim 2, characterized in that: in, In step S2-1, the daily load sequence is represented as follows: ， In the formula, x i Let m be the load value at time i, and m be the total amount of load sequence data. In step S2-2, the approximate entropy characteristic of the daily load is expressed as: ， In the formula, r is the tolerance coefficient related to the standard deviation of the daily load sequence, and N is the number of sampling points. The average log-probability is obtained based on the proportion of similar pattern pairs in the subsequences of the daily load sequence, and this proportion is based on the tolerance coefficient. The time-period characteristics include nighttime load ratio and temperature sensitivity. The nighttime load ratio is expressed as: ， In the formula, n is the starting point of the daytime load, k is the ending point of the daytime load, and P i Let i be the load value at time i during the night in the daily load sequence. The temperature sensitivity is expressed as: ， In the formula, ΔP represents the load change, and ΔT represents the temperature change. The hysteresis correlation feature is represented as follows: ， In the formula, P t Let T be the load value at time t in the daily load sequence. t-k Let tk be the temperature value at time tk in the temperature series, and D be the sum of the product of the load and the temperature deviations from the mean.

4. The large-scale transformer area short-term load forecasting method according to claim 3, characterized in that: in, In step S2-2, the maximum distance between each pair of subsequences in the daily load sequence is calculated: ， Set the tolerance coefficient to , where σ p The standard deviation of the daily load series is given. The subsequence pairs whose maximum distance is less than the tolerance coefficient are considered as similar pattern pairs, and the proportion of similar pattern pairs is: 。 5. The large-scale transformer area short-term load forecasting method according to claim 1, Its features are: in, Step S3 includes the following sub-steps: Step S3-1: Extract the temperature-sensitive transformer data set from the multiple transformer data sets; Step S3-2: Fit the temperature-load curve based on the data set of the transformer area, and take the temperature at the point where the load elasticity coefficient is the maximum as the limit temperature. Step S3-3: Based on the temperature-load curve, calculate the correlation coefficient between the cumulative temperature and load for a predetermined number of days using a sliding window, and take the number of days with the largest correlation coefficient as the maximum cumulative number of days. Step S3-4: Construct a temperature correction term and an objective function based on the limit temperature and the maximum cumulative number of days. The temperature correction term is used to correct the temperature data of the temperature-sensitive area, and the objective function is used to maximize the correlation between the temperature correction term and the corresponding load. Step S3-5: Solve the objective function to obtain the cumulative effect coefficient; Step S3-6: Determine whether the objective function has converged to the minimum according to the predetermined convergence rule. If the determination is negative, return to step S3-5. Step S3-7: If the determination in step S3-6 is yes, a cumulative effect correction model is constructed based on the cumulative effect coefficient and the dynamic weight coefficient, and the temperature data of the temperature-sensitive substation is corrected using the cumulative effect correction model.

6. The large-scale transformer area short-term load forecasting method according to claim 5, characterized in that: in, In step S3-2, the temperature-load curve is represented as follows: ， In the formula, T t Let L be the temperature at time t. t Let be the load at time t, and n be the order of the polynomial, β0, β1, β2, …, β n For undetermined coefficients, The load elasticity coefficient is expressed as: ， The threshold temperature is expressed as: ， In the formula, This represents the temperature value at a certain moment. The index used to find the maximum value in the array. In step S3-3, the maximum cumulative number of days is expressed as: ， In the formula, The correlation coefficient between average temperature and response load is expressed as: ， In the formula, L t In response to load, The average temperature over the past few days is expressed as: ， In the formula, T t-j The temperature at time tj. In step S3-4, the temperature correction term is expressed as: ， In the formula, T t Let T be the temperature at time t. t-j,24 The temperature at time tj. The objective function is expressed as: ， In the formula, k represents the correlation between two variables. j This is the cumulative effect coefficient. In steps S3-5, the objective function is solved using a sequential least squares programming numerical optimization algorithm.

7. The large-scale transformer area short-term load forecasting method according to claim 6, characterized in that: in, In steps S3-7, the dynamic weighting coefficient is expressed as: ， In the formula, This represents the moving average of temperatures over a window of predetermined length centered at t. The cumulative effect correction model is expressed as follows: ， In the formula, L min L max These are the minimum and maximum loads in the temperature-load curve, respectively.

8. The large-scale transformer area short-term load forecasting method according to claim 1, Its features are: in, In step S5, the incremental learning method is used to train a student model based on the teacher model using incremental data. The teacher model is trained using accumulated historical data, and the trained student model is used as the short-term load forecasting model for the transformer area. The incremental learning method includes: The first step of learning involves training the student model based on the output of the teacher model, the incremental data, and the distillation loss function. as well as The second step involves matching the incremental data with the historical data to obtain matching data, and then training the student model based on the matching data and the mean absolute error loss function to achieve balanced learning.

9. The large-scale transformer area short-term load forecasting method according to claim 1, Its features are: in, In step S5, the inverting converter includes: An input module includes an input layer and an embedding layer. The input layer is used to receive raw input variables, and the embedding layer is used to map the raw input variables into high-dimensional feature words. The feature extraction module includes a multi-head self-attention unit, a first-layer normalization unit, a feedforward neural network, and a second-layer normalization unit. The self-attention unit is used to extract multiple features from the time series in parallel from different subspaces, and the feedforward neural network is used to extract nonlinear relationships in the time series. The output module includes a projection layer and an output layer. The projection layer is used to filter the features extracted by the feature extraction module and retain key features, and to map the key features to the short-term load prediction results of the transformer area. The output layer is used to output the short-term load prediction results of the transformer area.

Images