Large-scale transformer area short-term load forecasting method

Abstract

The application provides a large-scale transformer area short-term load prediction method, based on historical data of each transformer area, the transformer area is divided into temperature sensitive transformer area and temperature inert transformer area, and for the temperature sensitive transformer area, a dynamic temperature characteristic cumulative effect correction model is constructed according to the difference of the fluctuation law of the load and the temperature characteristics, the temperature data is corrected, and then the short-term load of the transformer area is predicted based on the characteristics of 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.

Description

Technical Field

[0001] This invention relates to the field of power system load forecasting technology, specifically to a large-scale short-term load forecasting method for distribution areas based on cumulative effect feature correction and incremental learning. Background Technology

[0002] Load forecasting, as a sensing method for predicting and judging future load changes, is an important component of the smart grid data platform. Meanwhile, low-voltage distribution transformer areas are the final link in achieving intelligent and refined management of the power grid. Accurate load forecasting for regional distributed transformer areas can provide important data for transformer capacity allocation, low-voltage distribution network planning, line loss reduction, and risk warning.

[0003] Unlike system-level and bus-side regional load forecasting, short-term load forecasting for distribution transformer areas is significantly influenced by meteorological and human factors, resulting in prominent nonlinearity and non-stationarity of load sequences. Against the backdrop of new power system construction, the level of terminal electrification is continuously improving, leading to a more complex power load composition and making its patterns more difficult to predict. Furthermore, with the development of the power industry and changes in the power market, new technologies and roles such as demand response, the sharing economy, load aggregators, and virtual power plants are constantly emerging, causing power loads to exhibit more complex and volatile characteristics and forms: On the one hand, the randomness and uncertainty of load changes are increasing, making them highly susceptible to extreme weather conditions. In recent years, the frequent occurrence of extreme high and low temperatures has made load fluctuations difficult to predict and regulate, resulting in power supply gaps; on the other hand, due to the increasingly complex load structure, the power system places higher demands on the spatial granularity and temporal accuracy of load forecasting. Simultaneously, the large number and high heterogeneity of distribution transformer area loads make it difficult for traditional forecasting models to capture the complex fluctuation characteristics of distribution transformer area loads and lack the ability to process massive amounts of distribution transformer area data.

[0004] Currently, there are many methods for power load forecasting. For example, patent CN202411662478.0 proposes a small-sample power load forecasting method based on transfer learning, which can extract common knowledge from similar tasks and transfer it to the training process of the target task, thereby improving the model's performance in handling the target task. Patent CN202411761457.4 proposes an ultra-short-term load forecasting method and system based on a hybrid expert model, which optimizes prediction accuracy, shortens overall training time, reduces costs, and allows for parallel training of various expert sub-models, greatly shortening the prediction cycle. However, these methods are generally difficult to apply well to large-scale distribution area load forecasting, and this requirement still faces many technical challenges. First, with the intelligent development of power systems, more and more electrical equipment is connected to the power system, causing changes in distribution area loads. In particular, short-term load changes in low-voltage distribution areas are more easily affected by factors such as weather conditions, electricity prices, electricity consumption habits, social time, and holidays. Therefore, how to extract highly relevant feature datasets to help deep network models predict is a major challenge that current short-term load forecasting technology needs to consider. Secondly, there is a high degree of heterogeneity among different distribution transformer areas. The load data fluctuations in each area are influenced by different factors, and the power load exhibits significant seasonality, periodicity, and uncertainty, posing a challenge to load forecasting for large-scale clustered distribution transformer areas. Finally, within a large regional area, there are massive numbers of distribution transformers, with extremely large volumes of historical load data and characteristic data. Faced with this influx of massive data, conventional load forecasting models often lack the ability to process large datasets, significantly weakening the timeliness of load forecasting. Therefore, it is urgent to construct efficient load forecasting models to process massive amounts of data and preserve the timeliness of load forecasting to the greatest extent possible. Summary of the Invention

[0005] This invention is made to solve the above-mentioned problems, and its purpose is to provide a method that can accurately predict the short-term load of large-scale transformer areas. The technical solution adopted by this invention is as follows:

[0006] This invention provides a method for large-scale short-term load forecasting of transformer substations, which includes the following steps: Step S1, acquiring historical data for each transformer substation and preprocessing it to obtain a transformer substation dataset, wherein the historical data includes at least temperature data and load data; Step S2, performing feature analysis based on the transformer substation dataset, and dividing the transformer substations into temperature-sensitive and temperature-inertial substations based on the feature analysis results; Step S3, for the temperature-sensitive substations, constructing a cumulative effect correction model, and using the cumulative effect correction model to correct the temperature data of the temperature-sensitive substations to eliminate the impact of extreme temperatures on load forecasting; Step S4, for each transformer substation, constructing a personalized feature set for the substation based on the feature analysis results, wherein, for the temperature-sensitive substations, the personalized feature set includes temperature features corresponding to the corrected temperature data; Step S5, constructing a short-term load forecasting model for the transformer substations that combines an incremental learning method with an inverted transformer (iTransformer), and using the short-term load forecasting model for the transformer substations based on the personalized feature set to obtain the short-term load forecasting results for each transformer substation.

[0007] The large-scale transformer area short-term load forecasting method provided by this invention may also have the following technical features, wherein the transformer area dataset includes at least load data and temperature data, and step S2 includes the following sub-steps: Step S2-1, performing statistical analysis based on the load data and the temperature data, and constructing the daily load sequence and 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 results, which include load entropy spectrum, time period characteristics, and lag correlation characteristics of temperature relative to load, wherein the load entropy spectrum includes the daily load approximate entropy characteristics; Step S2-3, based on the daily load approximate entropy characteristics, the time period characteristics, and the lag correlation characteristics, using an extreme gradient boosting tree to divide the transformer area into the temperature-sensitive transformer area and the temperature-inert transformer area.

[0008] The large-scale transformer area short-term load forecasting method provided by this invention may also have the following technical feature, wherein, in step S2-1, the daily load sequence is represented as:

[0009]

[0010] In the formula, x i Let 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:

[0011]

[0012] 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, whereby the nighttime load ratio is expressed as:

[0013]

[0014] 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 be the load value at time i during the night in the daily load sequence. The temperature sensitivity is expressed as:

[0015]

[0016] In the formula, ΔP represents the load change, and ΔT represents the temperature change. The hysteresis characteristic is expressed as:

[0017]

[0018] 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 variance of the load and temperature.

[0019] The large-scale transformer area short-term load forecasting method provided by this invention may also have the following technical feature, wherein, in step S2-2, the maximum distance between each pair of subsequences in the daily load sequence is calculated:

[0020]

[0021] Set the tolerance coefficient to , where σ p The standard deviation of the daily load sequence is defined as follows: Subsequence pairs whose maximum distance is less than the tolerance coefficient are considered as similar pattern pairs, and the proportion of such similar pattern pairs is:

[0022] .

[0023] The large-scale transformer area short-term load forecasting method provided by this invention may also have the following technical features, wherein step S3 includes the following sub-steps: Step S3-1, extracting the transformer area dataset of the temperature-sensitive transformer area from multiple transformer area datasets; Step S3-2, fitting a temperature-load curve based on the transformer area dataset, and taking the temperature at the point where the load elasticity coefficient is the largest as the threshold temperature; Step S3-3, calculating the correlation coefficient between the cumulative temperature and load for a predetermined number of days based on the temperature-load curve using a sliding window, and taking the number of days with the largest correlation coefficient as the maximum cumulative number of days; Step S3-4, based on the threshold temperature and the maximum cumulative number of days... The process involves several steps: First, constructing a temperature correction term and an objective function. The temperature correction term is used to correct the temperature data of the temperature-sensitive substation, and the objective function is used to maximize the correlation between the temperature correction term and the corresponding load. Second, solving the objective function to obtain the cumulative effect coefficient. Third, determining whether the objective function has converged to its minimum according to a predetermined convergence rule; if not, returning to step S3-5. Fourth, if the determination in step S3-6 is yes, constructing a cumulative effect correction model based on the cumulative effect coefficient and dynamic weighting coefficient, and using this cumulative effect correction model to correct the temperature data of the temperature-sensitive substation.

[0024] The large-scale transformer area short-term load forecasting method provided by this invention may also have the following technical feature, wherein, in step S3-2, the temperature-load curve is represented as:

[0025]

[0026] 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 These are undetermined coefficients. The load elasticity coefficient is expressed as:

[0027]

[0028] The threshold temperature is expressed as:

[0029]

[0030] In the formula, This represents the temperature value at a certain moment. This is the index used to find the maximum value in the array. In step S3-3, the maximum cumulative number of days is represented as:

[0031]

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

[0033]

[0034] In the formula, L t In response to load, The average temperature over the past several days is expressed as:

[0035]

[0036] In the formula, T t-j Let be the temperature at time tj. In steps S3-4, the temperature correction term is expressed as:

[0037]

[0038] In the formula, T t Let T be the temperature at time t. t-j,24 Let be the temperature at time tj. The objective function is expressed as:

[0039]

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

[0041] The large-scale transformer area short-term load forecasting method provided by this invention may also have the following technical feature, wherein, in steps S3-7, the dynamic weighting coefficient is expressed as:

[0042]

[0043] In the formula, This represents the moving average of the temperature over a predetermined length of window centered at t. The cumulative effect correction model is expressed as:

[0044]

[0045] In the formula, L min L max These are the minimum and maximum loads in the temperature-load curve, respectively.

[0046] The large-scale transformer area short-term load forecasting method provided by this invention may also have the following technical features: 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 based on accumulated historical data, and the trained student model is used as the transformer area short-term load forecasting model. The incremental learning method includes: a first step of learning, training the student model based on the output of the teacher model, the incremental data, and a distillation loss function; and a second step of 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 short-term load forecasting method for transformer substations provided by this invention may also have the following technical features, wherein, in step S5, the inverted transformer includes: an input module comprising an input layer and an embedding layer, wherein the input layer is used to receive the original input variables, and the embedding layer is used to map the original input variables to high-dimensional feature words; a feature extraction module comprising a multi-head self-attention unit, a first-layer normalization unit, a feedforward neural network, and a second-layer normalization unit, wherein 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; and an output module comprising a projection layer and an output layer, wherein 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 forecasting result for the transformer substation, and the output layer is used to output the short-term load forecasting result for the transformer substation.

[0048] The role and effect of invention

[0049] The large-scale short-term load forecasting method for transformer substations provided by this invention divides substations into temperature-sensitive and temperature-inertial substations based on historical data. For temperature-sensitive substations, a dynamic temperature characteristic cumulative effect correction model is constructed to correct the temperature data, taking into account the differences in the fluctuation patterns of load and temperature characteristics. Then, the model is used to forecast the short-term load of the substation based on the characteristics corresponding to the modified temperature data. Therefore, the impact of extreme temperatures on load forecasting can be eliminated or reduced, improving the accuracy of load forecasting. In addition, since a short-term load forecasting model for transformer substations is constructed that combines incremental learning and iTransformer, the efficiency of load forecasting can also be effectively improved. Attached Figure Description

[0050] Figure 1 This is a flowchart of the large-scale transformer area short-term load forecasting method in an embodiment of the present invention;

[0051] Figure 2This is a schematic diagram of the hybrid stage differentiation model in an embodiment of the present invention;

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

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

[0054] Figure 5 This is a schematic diagram of the inverter converter in an embodiment of the present invention;

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

[0056] Figure 7 These are the actual load curves, temperature prediction curves before and after temperature correction, and load prediction curves for Type 1 transformer substations in this embodiment of the invention.

[0057] Figure 8 These are the actual load curves, temperature prediction curves before and after temperature correction, and load prediction curves for type 2 transformer substations in this embodiment of the invention.

[0058] Figure 9 These are the actual load curves, temperature prediction curves before and after temperature correction, and load prediction curves for type 3 transformer substations in this embodiment of the invention.

[0059] Figure 10 These are the actual load curves and predicted load curves for Type 1 transformer substations in this embodiment of the invention.

[0060] Figure 11 These are the actual load curves and predicted load curves for Type 2 transformer substations in this embodiment of the invention.

[0061] Figure 12 This is a diagram showing the actual load curve and predicted load curve of type 3 transformer area in this embodiment of the invention. Detailed Implementation

[0062] To make the technical means, creative features, objectives and effects of this invention easy to understand, the following describes in detail the large-scale short-term load forecasting method for transformer areas based on cumulative effect feature correction and incremental learning, in conjunction with embodiments and accompanying drawings.

[0063] This embodiment provides a method for short-term load forecasting of large-scale distribution areas, which is used to perform short-term load forecasting for large-scale distribution areas. A distribution area is a basic power supply unit of the power distribution network in a power system. A large-scale distribution area usually refers to a large number of distribution areas, a wide coverage area of ​​multiple distribution areas, and / or multiple distribution areas with diverse and complex load characteristics.

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

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

[0066] Step S1: Obtain historical data for each transformer area and preprocess the historical data to form a transformer area dataset.

[0067] Step S2: Perform feature analysis based on the transformer area dataset, and divide the transformer area into temperature-sensitive transformer areas and temperature-inert transformer areas based on the feature analysis results.

[0068] Step S3: For temperature-sensitive transformer substations, a cumulative effect correction model of dynamic temperature characteristics is constructed based on the cumulative effect of extreme high or low temperatures on the substation load, and the temperature data of the temperature-sensitive transformer substations is corrected using this model.

[0069] Step S4: For each transformer substation, construct a personalized feature set for the substation based on the feature analysis results. For temperature-sensitive substations, the features include the temperature features corresponding to the corrected temperature data.

[0070] Step S5: Construct a short-term load forecasting model for transformer areas that combines incremental learning methods with an inverted transformer (iTransformer), and use this model to obtain short-term load forecasting results for each transformer area based on a personalized feature set.

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

[0072] Step S1: Obtain historical data for each transformer area and preprocess the historical data to form a transformer area dataset.

[0073] The historical data for the distribution area includes historical load, meteorological, social, and attribute data, with meteorological data including temperature data. Preprocessing of the historical data in this step may include data cleaning, missing data completion, etc.

[0074] Step S2: Perform feature analysis based on the transformer area dataset, and divide the transformer area into temperature-sensitive transformer areas and temperature-inert transformer areas based on the feature analysis results.

[0075] In this step, the daily load curve and temperature load curve of the station area are obtained based on historical meteorological data and historical load data. Based on the objective laws and temporal differences of the daily load curve and temperature load curve, feature analysis is performed to obtain feature analysis results. Then, based on the feature analysis results, the station area is preliminarily classified using Extreme Gradient Boosting Tree (XGBoost) to classify it as either temperature-sensitive or temperature-inert (normal).

[0076] Figure 2 This is a schematic diagram of the hybrid stage classification model in this embodiment. Figure 3 This is a flowchart of step S2 in this embodiment.

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

[0078] Step S2-1: Perform statistical analysis based on the load and temperature data of the transformer area, and construct the daily load sequence and temperature sequence of the transformer area.

[0079] For example, a daily load sequence can be constructed based on historical daily load data and a predetermined time granularity, which generates an m-dimensional vector:

[0080]

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

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

[0083] Step S2-2 involves performing feature analysis on the daily load sequence and temperature load sequence to obtain the feature analysis results, including load entropy spectrum, time period characteristics, and lag correlation characteristics.

[0084] The load entropy spectrum includes the approximate entropy characteristics of daily load, the daily load curve, and the temperature load curve.

[0085] The approximate entropy feature of daily loads is used to measure the complexity and predictability of a load sequence, and can be obtained as follows: calculate the maximum distance between each pair of subsequences in the daily load sequence. The calculation formula is as follows:

[0086]

[0087] Set the similarity mode ratio, set the tolerance coefficient. , where σ p Let be the standard deviation of the daily load series. Among multiple sub-sequence pairs of the daily load series, sub-sequence pairs whose distance is less than the tolerance coefficient r are denoted as similar pattern pairs. The proportion of similar pattern pairs is... for:

[0088]

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

[0090]

[0091] In the formula, The average log probability can be based on the proportion of similar pattern pairs. Calculated.

[0092] The time-period characteristics include two features: nighttime load ratio and temperature sensitivity.

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

[0094]

[0095] In the formula, m is the total amount of load sequence data, n is the starting point of the daytime load, k is the ending point of the daytime load, and P i Let be the load value at time i during the night in the daily load sequence.

[0096] Temperature sensitivity The elastic coefficient of the load as a function of temperature is calculated using the following formula:

[0097]

[0098] In the formula, ΔP is the load change and ΔT is the temperature change.

[0099] The lag correlation characteristic is used to characterize the temporal differences between daily load and temperature sequences. For daily load and temperature sequences with a 24-hour time granularity, the lag correlation can be considered as calculating the correlation coefficient between the daily load sequence and temperature.

[0100]

[0101] 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 sequence, that is, the temperature lags by k time units, and D is the variance of load and temperature.

[0102] Step S2-3: Based on the approximate entropy characteristics, time period characteristics, and hysteresis correlation characteristics of the daily load, the transformer area is divided using Extreme Gradient Boosting Tree (XGBoost) to classify the transformer area into temperature-sensitive transformer area or temperature-inert transformer area.

[0103] Step S2-4: Based on the approximate entropy characteristics, time period characteristics, and lag correlation characteristics of daily load, hierarchical clustering algorithm is used to cluster multiple transformer area datasets to obtain multiple clusters with different load time series characteristics. Based on the clusters, the transformer area is divided into multiple types with different load characteristics.

[0104] Specifically, clustering algorithms are unsupervised machine learning methods that classify a given dataset based on its data distribution. Among them, agglomerative hierarchical clustering is a clustering method based on the similarity of cluster data, which can construct a hierarchical tree-like clustering result by merging similar clusters layer by layer.

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

[0106] During the data processing, a large number of transformer area datasets are constructed, and missing and outlier values ​​in each transformer area dataset are detected and corrected. Data cleaning is performed to avoid data anomalies affecting the subsequent clustering results.

[0107] In the hierarchical clustering process, each substation dataset is treated as an independent cluster, and the distance between each cluster is calculated separately. Euclidean distance can be used as the distance metric. Based on the Euclidean distance calculation results, the two closest clusters are merged into a new cluster.

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

[0109]

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

[0111] During the formation of clusters, the hierarchical clustering process described above is repeatedly cycled to continuously merge similar clusters and gradually form clusters with similar characteristics until the preset cycle stopping condition is met. The hierarchical clustering algorithm is then complete, and multiple tree-like clusters are finally formed. Multiple transformer areas within the same cluster have similar characteristics in terms of load timing.

[0112] In this embodiment, a hierarchical clustering algorithm is used to divide multiple transformer area datasets into three types, each with corresponding load time-series characteristics. The three types are Type 1 (daytime peak type), Type 2 (nighttime peak type), and Type 3 (dual-peak type). Figure 2As shown, the load curve of type 1 mainly forms a peak during the day. Type 1 includes, for example, transformer substations containing manufacturing plants, where the electricity load is relatively high and stable during the day. The load curve of type 2 mainly forms a peak at night. Type 2 includes, for example, transformer substations containing manufacturing plants operating during off-peak hours, or transformer substations containing energy storage systems that store energy at night. The load curve of type 3 has two peaks, one during the day from about 9:00 to 12:00 and the other at night from about 16:00 to 20:00. Type 3 includes, for example, transformer substations that are mixed commercial and residential areas, or transformer substations that are industrial parks.

[0113] Step S3: For temperature-sensitive transformer substations, a cumulative effect correction model of dynamic temperature characteristics is constructed based on the cumulative effect of extreme high or low temperatures on the substation load, and the temperature data of the temperature-sensitive transformer substations is corrected using this model.

[0114] Since people's electricity consumption behavior is affected by historical extreme high or low temperatures and exhibits a certain inertia, this step considers the cumulative effect of extreme high or low temperatures on the load of the transformer area. A dynamic cumulative effect correction model is constructed that considers the difference in the fluctuation patterns of load and temperature characteristics on a daily time scale. This model is used to correct the temperature in the transformer area data of temperature-sensitive areas in order to eliminate the influence of historical extreme temperatures on the forecast.

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

[0116] like Figure 4 As shown, step S3 includes the following sub-steps:

[0117] Step S3-1: Extract the temperature-sensitive transformer dataset from multiple transformer datasets.

[0118] Step S3-2: For each extracted transformer area dataset, fit the temperature-load curve of the corresponding transformer area based on the transformer area dataset, and take the temperature at the point where the load elasticity coefficient is the maximum as the boundary temperature.

[0119] Among them, the temperature point in the data set of the transformer area where the load is most affected by temperature (i.e. the load is most sensitive to temperature changes) 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 distribution area, the temperature-load curve is first fitted using an nth-order polynomial:

[0121]

[0122] In the formula, T t Let L be the temperature at time t. tLet be the load at time t, and n be the order of the polynomial; β0, β1, β2, …, β n These are coefficients to be determined.

[0123] Load elasticity coefficient e t for:

[0124]

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

[0126]

[0127] In the formula, This represents the temperature value at a certain moment. This is the index used to find the maximum value in the array.

[0128] 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 day with the largest correlation coefficient as the maximum cumulative number of days.

[0129] Wherein, the past d is defined max Average temperature over 1 day for:

[0130]

[0131] In the formula, d max T represents the maximum cumulative number of days to be calculated. t-j Let tj be the temperature at time tj, which means the temperature lags behind the load at time t by j time units.

[0132] Calculate the response load L in response to temperature. t The correlation with average temperature is used to determine the maximum cumulative number of days (d) based on the number of days with the highest correlation. max :

[0133]

[0134]

[0135] In this embodiment, the predetermined number of days is from day 1 to day 5.

[0136] 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 of the temperature-sensitive area, and the objective function is used to maximize the correlation between the temperature correction term and the corresponding load, which includes the corresponding correlation coefficient (cumulative effect coefficient).

[0137] Among them, the temperature correction term T t corr Defined as:

[0138]

[0139] In the formula, T t Let T be the temperature at time t, and take the temperature of the latest day; t-j,24 The temperature at time tj; d max T represents the maximum cumulative number of days calculated using the method described above. min The limit temperature is calculated using the method described above.

[0140] The corresponding objective function is defined as:

[0141]

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

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

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

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

[0146] 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 and solve again; if the determination is positive, output the cumulative effect coefficient k. j .

[0147] One example of a convergence rule is to compare the results of multiple solutions (i.e., the cumulative effect coefficients), and determine that the objective function has converged to the minimum when the change in the most recent solutions is less than a predetermined threshold.

[0148] Step S3-7: If the determination in step S3-6 is yes, construct a cumulative effect correction model based on the cumulative effect coefficient and the dynamic weight coefficient, and use the model to correct the temperature of the temperature-sensitive substation.

[0149] To fully consider the characteristics of the load curve and the impact of temperature changes at different time points on the load, a dynamic weighting coefficient ω is further introduced. t It is used to dynamically adjust the weight of temperature in the temperature correction term.

[0150] For a load curve segment, its normalized load is:

[0151]

[0152] In the formula, L min L max These represent the minimum and maximum loads on the temperature-load curve, respectively; L t This indicates the current load.

[0153] To highlight the impact of temperature during periods of high load, a dynamic weighting coefficient ω is defined. t for:

[0154]

[0155] In the formula, This represents the moving average of the temperature over a window of predetermined length centered at t, for example, a window length of 3.

[0156] Therefore, the final cumulative temperature correction formula with weighted coefficients is (i.e., the cumulative effect correction model can be expressed as):

[0157]

[0158] Step S4: For each transformer substation, construct a personalized feature set for the substation based on the feature analysis results. For temperature-sensitive substations, the features include the temperature features corresponding to the corrected temperature data.

[0159] For each transformer substation, a personalized feature set can be constructed by combining its time-series characteristics, load characteristics, meteorological characteristics, and attribute characteristics. For temperature-sensitive substations, the meteorological characteristics include the temperature features corresponding to the corrected temperature data.

[0160] Step S5: Construct a short-term load forecasting model for transformer areas that combines incremental learning with iTransformer, and use this model to obtain short-term load forecasting results for each transformer area based on a personalized feature set.

[0161] In this embodiment, considering the significant temporal dependencies in transformer area load data, the short-term load forecasting model for transformer areas is built based on iTransformer. iTransformer is a deep learning model based on the Transformer architecture, which can effectively capture the global dependencies between time tokens in time series, and exhibits excellent performance, especially in multivariate time series forecasting.

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

[0163] like Figure 5 As shown, iTransformer adopts a Transformer structure that only includes an encoder, and the whole consists of three parts: an input module, a feature extraction module, and an output module.

[0164] The input module comprises an input layer and an embedding layer. The input layer receives the raw input variables. The embedding layer maps the raw input variables into high-dimensional feature tokens, fully extracting the temporal dependencies between variables.

[0165] The feature extraction module comprises a multi-head self-attention unit, a first-layer normalization unit, a feedforward neural network, and a second-layer normalization unit, connected sequentially. The multi-head self-attention unit and the feedforward neural network have residual connections. The multi-head self-attention unit contains multiple self-attention heads, enabling it to capture various features from different subspaces in parallel. The feedforward neural network (typically a fully connected layer) enhances the network's representational power, extracting 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 their attention calculation formulas are as follows:

[0167]

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

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

[0170] Meanwhile, each lexical unit was standardized individually using layer normalization units to eliminate the influence of differences in feature dimensions. The normalization formula is as follows:

[0171]

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

[0173] The output module consists of a projection layer and an output layer. The projection layer filters and retains key features while discarding redundant or interfering information, thereby improving the accuracy of the final prediction result. It also maps key features (hidden states) to values ​​at future time steps, which are the short-term load prediction results for the transformer area. The output layer outputs this result.

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

[0175] like Figure 6 As 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, using the influx of incremental data, for the prediction task after the next data update. The teacher model is trained on a large amount of accumulated historical data (i.e., raw transformer load data) and can achieve high-precision prediction for existing transformer load data; the trained student model is the short-term transformer load prediction model.

[0176] Furthermore, the incremental learning method in this embodiment employs a two-step learning mechanism, or two-step training method, which first allows the student model to learn from incremental data and then performs balanced learning.

[0177] Specifically, in the first learning step, the student model is trained only on new incremental data, based on the output of the teacher model, the incremental data, and the distillation loss function. Gradient updates are used to adapt the student model to the new incremental data.

[0178] In the second learning step, incremental data is matched with 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, enabling the student model to adapt to both historical and incremental data. Since a large amount of historical data contains varying quality, making it difficult to mix them into the same dataset for training, this embodiment calculates the Euclidean distance between each sequence in the incremental data and each sequence in the historical data, and selects sequences with similar distributions for matching based on the Euclidean distance, forming the matched data.

[0179] Figure 7 This embodiment shows the actual load curve, temperature prediction curve before and after temperature correction, and load prediction curve for Type 1 transformer substation. Figure 8 This embodiment shows the actual load curve, temperature prediction curve before and after temperature correction, and load prediction curve for Type 2 transformer substation. Figure 9 This embodiment shows the actual load curve, temperature prediction curve before and after temperature correction, and load prediction curve for type 3 transformer area.

[0180] like Figures 7 to 9 As shown, the temperature curves before and after correction differ significantly, especially at peak load periods. The temperature curve corrected for the cumulative effect better matches the fluctuation and trend of the load. For all three types of transformer substations, compared to the original temperature curve ("Temperature Before Correction" curve in the figure, marked with a dotted line), the corrected temperature curve ("Temperature After Correction" curve in the figure, marked with a dotted line) better reflects the fluctuation of the load curve.

[0181] Meanwhile, it can be seen that the predicted load curve based on the corrected temperature (the predicted load after correction curve, marked with a solid line of triangles in the figure) is closer to the actual load (marked with a solid line of squares in the figure) than the predicted load curve based on the original temperature (the "predicted load - before correction" curve in the figure). This is true for all three types of transformer substations, proving the effectiveness of the method in this embodiment.

[0182] Figure 10 This is a diagram showing the actual load curve and predicted load curve for Type 1 transformer area in this embodiment. Figure 11 This is a graph showing the actual load curve and the predicted load curve for Type 2 transformer substation in this embodiment. Figure 12 This diagram shows the actual load curve and predicted load curve for Type 3 transformer substations in this embodiment. The dashed line represents the actual load curve, and the solid line represents the predicted load curve. The vertical dashed line in the middle of the diagram is a dividing line; the line to the left of the dividing line represents the prediction result of the model trained using only historical data, and the line to the right represents the prediction result of the model trained using the method described above in this embodiment, incorporating incremental data.

[0183] like Figures 10 to 12 As shown, for all three types of transformer substations, the predicted load curves show a high degree of agreement with the actual load curves, demonstrating the effectiveness of the method in this embodiment. Furthermore, it can be observed from the figure that the predicted load curve to the right of the dividing line closely matches the actual load curve. This is because the incremental learning loss function and the two-step learning mechanism effectively help the model adapt to new incremental data, further proving the effectiveness of the incremental learning method in this embodiment.

[0184] According to the large-scale transformer substation short-term load forecasting method based on cumulative effect feature correction and incremental learning provided in this embodiment, transformer substations are divided into temperature-sensitive and temperature-inertial substations based on historical data. For temperature-sensitive substations, a dynamic temperature feature cumulative effect correction model is constructed to correct the temperature data, taking into account the differences in the fluctuation patterns of load and temperature characteristics. Then, the model is used to predict the short-term load of the substation based on the features corresponding to the modified temperature data. Therefore, the impact of extreme temperatures on load forecasting can be eliminated or reduced, improving the accuracy of load forecasting. In addition, since a short-term load forecasting model for transformer substations is constructed by combining incremental learning and iTransformer, the efficiency of load forecasting can also be effectively improved.

[0185] In this embodiment, daily load sequences and temperature load sequences were constructed through statistical analysis of the original historical data. Further feature analysis yielded approximate entropy characteristics, time-period characteristics, and lag correlation characteristics of the daily load. Based on these characteristics, XGBoost was used to initially classify temperature-sensitive or temperature-inertia-type transformer areas. Furthermore, a hierarchical clustering algorithm was used to divide the transformer area dataset into three categories according to their load characteristics. This achieved a two-layer refined classification of large-scale mixed transformer areas. Combined with a correction model to correct the temperature characteristics of temperature-sensitive transformer areas, a multi-type transformer area highly correlated feature dataset was formed. This dataset can be used for model training in this embodiment, as well as for training other models or further statistical analysis, providing a better data foundation for the study of large-scale transformer area loads.

[0186] Furthermore, in this embodiment, the incremental learning method employs a two-step learning mechanism. First, the student model learns based solely on incremental data, and then it learns based on the matching data between the incremental data and historical data. This allows the prediction model to quickly adapt to new incremental data and enables the prediction model to perform balanced learning, allowing it to achieve more accurate predictions on both incremental and historical data.

[0187] The above embodiments are merely illustrative of specific implementations of the present invention, and the present invention is not limited to the scope of the description of the above embodiments. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are only for illustrating the principles of the present invention. Various changes and modifications can be made to the present invention without departing from the spirit and scope thereof, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A large-scale transformer area short-term load forecasting method, characterized in that, The method comprises the following steps: Step S1, obtaining historical data of each substation area and pre-processing the historical data to obtain a substation area data set, wherein the historical data at least includes temperature data and load data; Step S2, performing feature analysis based on the substation area data set, and dividing the substation area into a temperature sensitive substation area and a temperature inert substation area based on the feature analysis result; Step S3, for the temperature sensitive substation area, constructing a cumulative effect correction model, and correcting the temperature data of the temperature sensitive substation area by using the cumulative effect correction model to eliminate the influence of extreme temperature on load prediction; Step S4, for each substation area, constructing a personalized feature set of the substation area based on the feature analysis result, wherein for the temperature sensitive substation area, the personalized feature set contains temperature features corresponding to the corrected temperature data; Step S5, constructing a substation short-term load prediction model combined with an incremental learning method and an inverter, and obtaining short-term load prediction results of each substation area based on the personalized feature set by using the substation short-term load prediction model, wherein the substation area data set at least contains load data and temperature data, 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 substation 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, based on the daily load approximate entropy feature, the time period feature and the lag correlation feature, the substation area is divided into the temperature sensitive substation area and the temperature inert substation area by using an extreme gradient boosting tree, In step S5, the inverter comprises: an input module comprising an input layer and an embedding layer, the input layer being used for receiving original input variables, and the embedding layer being used for mapping the original input variables into high-dimensional feature tokens; a feature extraction module comprising 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 for extracting multiple features in a time sequence from different subspaces in parallel, and the feedforward neural network being used for extracting nonlinear relationships in the time sequence; and an output module comprising a projection layer and an output layer, the projection layer being used for screening and retaining key features of the features extracted by the feature extraction module, and mapping the key features into substation short-term load prediction results, and the output layer being used for outputting the substation short-term load prediction results.

2. The large-scale substation short-term load prediction method according to claim 1, wherein: wherein, in step S2-1, the daily load sequence is represented as: In the formula, x i is the load value at the i th moment, m is the total amount of load sequence data, in step S2-2, the daily load approximate entropy feature is represented as: where r is a tolerance factor related to the standard deviation of the daily load sequence, N is the number of sampling points, is the average log probability, based on the proportion of similar pattern pairs in the subsequence of the daily load sequence, which is based on the tolerance factor, the time period feature contains night load ratio and temperature sensitivity, the night load ratio is represented as: where n is the start of the daytime load, k is the end of the daytime load, P i is the load value at the i-th time of the night in the sequence of daily loads, the temperature sensitivity is represented as: wherein ΔP is a load variation amount, and ΔT is a temperature variation amount, The hysteresis correlation feature is expressed as: wherein P t is the load value at time t in the daily load sequence, T t-k is the temperature value at time t-k in the temperature sequence, and D is the deviation product sum of load and temperature.

3. The large-scale transformer area short-term load forecasting method according to claim 2, characterized in that: wherein In step S2-2, the maximum distance between each pair of sub-sequences in the daily load sequence is calculated: The tolerance factor is set to where σ p is the standard deviation of the daily load sequence, The pair of sub-sequences with a maximum distance less than the tolerance coefficient is taken as the similar pattern pair, and the proportion of the similar pattern pair is: 。 4. The method of claim 1, wherein the short-term load forecasting method of a large-scale transformer area, characterized in that: wherein, Step S3 includes the following sub-steps: Step S3-1, extracting the transformer data set of the temperature-sensitive transformer area from a plurality of transformer data sets; Step S3-2, fitting a temperature-load curve based on the transformer data set, and taking the temperature at which the load elasticity coefficient is maximum as the limit temperature; Step S3-3, based on the temperature-load curve, calculating the correlation coefficient of cumulative temperature and load for a predetermined number of days using a sliding window, and taking the day with the maximum correlation coefficient as the maximum cumulative day; Step S3-4, based on the limit temperature and the maximum cumulative day, constructing a temperature correction term and an objective function, 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, determining whether the objective function has converged to a minimum according to a predetermined convergence rule, and returning to step S3-5 if the determination is negative; Step S3-7, when the determination in step S3-6 is positive, constructing a cumulative effect correction model based on the cumulative effect coefficient and a dynamic weight coefficient, and correcting the temperature data of the temperature-sensitive transformer area using the cumulative effect correction model.

5. The large-scale transformer area short-term load forecasting method according to claim 4, characterized in that: wherein, In step S3-2, the temperature-load curve is expressed as: 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: The limit temperature is expressed as: In the formula, denotes 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: wherein is the correlation coefficient of the average temperature with the response load, expressed as: wherein L t in response to the load, is the average temperature of the past few days, expressed as: In the formula, T t-j is the temperature at time t-j, In step S3-4, the temperature correction term is expressed as: In the formula, T t is the temperature at time t, T t-j,24 is the temperature at time t-j The objective function is expressed as: wherein represents the correlation between two variables, k j is the cumulative effect coefficient, In step S3-5, the sequential least squares programming numerical optimization algorithm is used to solve the objective function.

6. The large-scale transformer area short-term load forecasting method according to claim 5, characterized in that: wherein, In step S3-7, the dynamic weight coefficient is expressed as: In the formula, represents a sliding average result of the temperature in a window range of a predetermined length centered on t, The cumulative effect correction model is expressed as: wherein L min , L max are the minimum load and the maximum load in the temperature-load curve, respectively.

7. The method of claim 1, wherein the short-term load forecasting method is applied to a large-scale transformer area. characterized in that: wherein, In step S5, the incremental learning method is used to train a student model based on a teacher model using incremental data, wherein the teacher model is trained based on cumulative historical data, and the trained student model is taken as the transformer area short-term load forecasting 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 In the second step, the incremental data is matched with the historical data to obtain matched data, and the student model is trained based on the matched data and a mean absolute error loss function to achieve balanced learning.

Images