Load sequence scene generation method and device under rainstorm weather

Abstract

The invention provides a load sequence scene generation method and device under a rainstorm weather, and relates to the technical field of data processing. The method comprises the following steps: acquiring historical load data; wherein the historical load data comprises historical load data in a rainstorm day and historical load data in a non-rainstorm day; based on correlation information between the historical load data, performing preliminary expansion on the historical load data to obtain a preliminary expansion sample set; determining residual information between the historical load data in the rainstorm day and the historical load data in the non-rainstorm day in the preliminary expansion sample set; expanding the preliminary expanded sample set according to the residual information to obtain a target expanded sample set; and training the target expansion sample set to obtain a load sequence scene under the rainstorm weather. According to the method, first sample expansion is performed through relevance, and related interpretation of a scene causal relationship is generated. The second sample expansion is carried out through the residual error information, so that the accuracy of the generated load sequence scene can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular, to a method and device for generating a load sequence scenario under a rainstorm weather condition. Background Art

[0002] With the rapid development of new energy and the large-scale access of flexible loads represented by electric vehicles to the power grid, both the energy side and the load side of the power system show significant strong uncertainty characteristics. Under this background, accurately generating load scenarios has become one of the key bases for ensuring the balance between power supply and demand.

[0003] The load change trend is greatly affected by meteorological conditions. Especially in the North China region with complex underlying surfaces, the characteristics of suddenness, extremity, and large rainfall volume of such rainstorm processes not only seriously affect people's lives and property safety. Since the historical records of extreme weather such as rainstorms are often scarce, it is difficult to accurately generate load scenarios under corresponding meteorological conditions, seriously affecting the power and electricity balance calculation, safety verification, and preventive control of the power system under extreme meteorological conditions. Therefore, how to accurately generate load scenarios under rainstorm meteorological conditions has very positive significance for effective power and electricity balance calculation and related regulation decisions to ensure and improve the safe and stable operation of the power grid.

[0004] Currently, the methods for generating load sequence scenarios under rainstorm meteorological conditions can be roughly divided into two categories: probability model algorithms and artificial intelligence algorithms.

[0005] Probability model algorithms use parameter estimation or non-parametric estimation methods to fit historical real data, obtain its statistical probability distribution law and characteristics, and generate the required scenarios by combining technologies such as sampling, clustering, Markov chain, and reduction. The probability model method uses an explicit probability model to fit a random process, with the advantages of low complexity and strong interpretability; however, the parameter estimation method in this type of method requires prior knowledge such as the probability distribution type of new energy output that is not easily obtained in advance, and the non-parametric estimation method requires a sufficient number of historical samples, otherwise its estimation accuracy is low.

[0006] Artificial intelligence algorithms utilize the powerful learning ability of artificial intelligence to automatically extract the statistical laws contained in historical data and generate corresponding scenarios, with the advantage of simple and convenient modeling. There are mainly two methods for generating load sequence scenarios under rainstorm meteorological conditions based on artificial intelligence algorithms, namely the scenario generation method based on data expansion and the scenario generation method with causal interpretation.

[0007] In the scenario generation method based on data expansion, in order to overcome the problem of insufficient training data for artificial intelligence models, there are currently mainly solutions such as adjusting the model, data augmentation, and transfer learning. However, all these existing methods utilize the characteristic that the Generative Adversarial Network (GAN) has a strong sample augmentation ability to successfully achieve effective augmentation of fewer samples; but it still requires that the original samples must have a considerable quantity; when the initial samples are very few, such as only a few or dozens, due to the very small amount of statistical law information contained in the extremely few samples, and at the same time GAN cannot converge effectively, the augmented samples obtained based on this cannot fully reproduce the statistical laws of the real samples and cannot achieve effective sample augmentation. In addition, for the sample expansion method based on the GAN method, when facing very few samples of some heavy rain types, it cannot achieve effective sample expansion, and when obtaining an extended sample set by migrating or varying from the nearby sample set, uncontrollable errors are inevitably introduced.

[0008] In the scenario generation method with causal explanation, the explainable methods for scenario generation based on artificial intelligence technology basically measure the accuracy of the model generation results based on the accuracy of a single scenario, as well as the degree of regulation of different features on a single generated scenario, lacking the overall accuracy evaluation of scenario generation and the evaluation of the overall regulation degree of different features on the generated scenario. And because heavy rain events are rare events, so far, no relevant scenario feature modeling and generation methods have been seen for either probability model algorithms or artificial intelligence algorithms.

[0009] It can be seen that the current load sequence scenario generation method under heavy rain meteorology requires a large amount of effective data during training and learning, and the corresponding historical sample data including extreme events such as heavy rain are often very few. In addition, there is also a problem that there is a lack of relevant explanations for the causal relationship of the generated scenario, and it is impossible to give information such as the accuracy of the generated scenario and the influence and regulation intensity of each input feature on the generated scenario in an explainable manner, resulting in a low accuracy of the finally obtained load sequence scenario. Summary of the Invention

[0010] An embodiment of the present invention provides a load sequence scenario generation method and device under heavy rain meteorology to solve the problem that the load sequence scenarios generated by the current load sequence scenario generation method under heavy rain meteorology have low accuracy.

[0011] In a first aspect, an embodiment of the present invention provides a load sequence scenario generation method under heavy rain meteorology, including:

[0012] Obtain historical load data; wherein, the historical load data includes historical load data on heavy rain days and historical load data on non - heavy rain days;

[0013] Based on the correlation information between historical load data, the historical load data is preliminarily expanded to obtain a preliminarily expanded sample set;

[0014] Determine the residual information between the historical load data on rainy days and the historical load data on non-rainy days in the preliminarily expanded sample set;

[0015] Expand the preliminarily expanded sample set according to the residual information to obtain a target expanded sample set;

[0016] Train the target expanded sample set to obtain the load sequence scenario under rainy weather.

[0017] In an optional implementation manner, based on the correlation information between historical load data, the historical load data is preliminarily expanded to obtain a preliminarily expanded sample set, including:

[0018] Determine the correlation between each collection moment within each collection day in the corresponding historical stage of the historical load data to obtain the horizontal correlation;

[0019] Determine the correlation between the same collection moment on different collection days in the corresponding historical stage of the historical load data to obtain the vertical correlation;

[0020] Preliminarily expand the historical load data according to the horizontal correlation and the vertical correlation to obtain a preliminarily expanded sample set.

[0021] In an optional implementation manner, according to the horizontal correlation and the vertical correlation, the historical load data is preliminarily expanded to obtain a preliminarily expanded sample set, including:

[0022] Preliminarily expand the historical load data according to the horizontal correlation and the vertical correlation respectively to obtain a horizontally expanded sample set and a vertically expanded sample set;

[0023] Fuse the horizontally expanded sample set and the vertically expanded sample set to obtain a preliminarily expanded sample set.

[0024] In an optional implementation manner, the historical load data includes historical load sequence data; determining the residual information between the historical load data on rainy days and the historical load data on non-rainy days in the preliminarily expanded sample set includes:

[0025] In the preliminarily expanded sample set, extract the historical load sequence data of the target non-rainy day adjacent to the target rainy day;

[0026] Calculate the residual between the historical load sequence data on the target rainy day and the historical load sequence data on the target non-rainy day to obtain the residual information.

[0027] In an alternative embodiment, the residual between the historical load sequence data on the target heavy rain day and the historical load sequence data on the target non - heavy rain day is calculated to obtain residual information, including:

[0028] Calculate the residual between the historical load sequence data on the target heavy rain day and the historical load sequence data on the target non - heavy rain day to obtain the target residual sequence;

[0029] Calculate the kernel density corresponding to the target residual sequence and the historical load sequence data on the target non - heavy rain day respectively to obtain the empirical probability density function of the target residual sequence and the empirical probability density function on the target non - heavy rain day;

[0030] Obtain the empirical density function according to the empirical probability density function of the target residual sequence and the empirical probability density function on the target non - heavy rain day;

[0031] Sample the empirical density function to obtain a sampled residual sequence, and use the sampled residual sequence as the residual information.

[0032] In an alternative embodiment, the preliminary augmented sample set is expanded according to the residual information to obtain the target augmented sample set, including:

[0033] Randomly sample and generate a sampled sample sequence on the target non - heavy rain day according to the empirical probability density function on the target non - heavy rain day and the empirical density function;

[0034] Overlay the sampled sample sequence and the sampled residual sequence to obtain the load augmented sample on the target heavy rain day;

[0035] Based on the load augmented sample on the target heavy rain day, obtain the target augmented sample set.

[0036] In an alternative embodiment, before the historical load data is preliminarily augmented based on the correlation information between the historical load data to obtain the preliminary augmented sample set, it further includes:

[0037] Obtain the heavy rain data corresponding to the historical load data; wherein, the heavy rain data includes the heavy rain influence area, rainfall data, and the position information of the daily load in the heavy rain cycle;

[0038] Normalize the historical load data, and classify the normalized historical load data according to the heavy rain data.

[0039] In an alternative embodiment, the historical load data includes the historical load day type; normalizing the historical load data includes:

[0040] Calculate the load level coefficient corresponding to the historical load data;

[0041] Based on the load level factor, the historical load data that does not belong to the working day is corrected to the historical load data corresponding to the working day;

[0042] The corrected historical load data is normalized.

[0043] In an optional implementation, training is performed on the target expanded sample set to obtain a load sequence scenario under heavy rain weather, including:

[0044] Divide the target expanded sample set into training set and test set, and initialize the network parameters;

[0045] Training is performed based on the initialized network parameters, training set, and test set to obtain the load sequence scenario under heavy rain weather.

[0046] In a second aspect, an embodiment of the present invention provides a device for generating a load sequence scenario under heavy rain weather, comprising:

[0047] An acquisition module is used to acquire historical load data; wherein the historical load data includes historical load data on rainstorm days and historical load data on non-rainstorm days;

[0048] An expansion module is used to preliminarily expand the historical load data based on the correlation information between the historical load data to obtain a preliminary expanded sample set;

[0049] A determination module is used to determine the residual information between the historical load data on rainstorm days and the historical load data on non-rainstorm days in the preliminary expanded sample set;

[0050] The expansion module is further used to expand the preliminary expanded sample set according to the residual information to obtain the target expanded sample set;

[0051] The training module is used to train the target expanded sample set to obtain the load sequence scenario under heavy rain weather.

[0052] In a third aspect, an embodiment of the present invention provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the method in the first aspect or any possible implementation of the first aspect is implemented.

[0053] Embodiments of the present invention provide a method and apparatus for generating load sequence scenarios under heavy rainfall conditions. In these embodiments, a two-stage sample expansion process addresses the issue of insufficient valid sample data. Furthermore, during sample expansion, the present embodiment considers the correlations between data and the corresponding causal relationships. The first sample expansion based on correlation enables interpretation of the causal relationships in the generated load sequence scenarios. A second sample expansion using residual information improves the accuracy of the generated load sequence scenarios. Description of the Drawings

[0054] Figure 1 is the implementation flowchart of the method for generating a load sequence scenario under a rainstorm weather provided by an embodiment of the present invention;

[0055] Figure 2 is the quality comparison chart of the sample set provided by an embodiment of the present invention;

[0056] Figure 3 is the change curve of the discriminator loss function of the method for generating a load sequence scenario under a rainstorm weather provided by an embodiment of the present invention;

[0057] Figure 4a and Figure 4b is the box plot of the autocorrelation coefficient and partial autocorrelation coefficient provided by an embodiment of the present invention;

[0058] Figure 5a and Figure 5b is the comparison chart of the root mean square error of the probability density function and cumulative probability density function provided by an embodiment of the present invention;

[0059] Figure 6a and Figure 6b is the comparison chart of the influence degree of the sample sets under different labels on the generated load sequence scenario under a rainstorm weather provided by an embodiment of the present invention;

[0060] Figure 7 is the implementation flowchart of the method for generating a load sequence scenario under a rainstorm weather provided by another embodiment of the present invention;

[0061] Figure 8 is the structural schematic diagram of the device for generating a load sequence scenario under a rainstorm weather provided by an embodiment of the present invention;

[0062] Figure 9 is the schematic diagram of the electronic device provided by an embodiment of the present invention. Detailed Embodiment

[0063] Next, embodiments of the present invention will be described in detail with reference to the drawings.

[0064] Figure 1 is the implementation flowchart of the method for generating a load sequence scenario under a rainstorm weather provided by an embodiment of the present invention. As Figure 1 shown, the method may include:

[0065] Step 110: Obtain historical load data; wherein, the historical load data includes historical load data on rainstorm days and historical load data on non-rainstorm days.

[0066] In this embodiment, the historical load data may be a historical load sequence in years. Each historical load data corresponds to a historical load day type. Among them, the historical load day types may include weekends, holidays, and working days, etc.

[0067] In addition, historical rainstorm data corresponding to the historical load data can also be obtained; among them, the historical rainstorm data may include the rainstorm impact area, rainfall data, and the position information of the daily load in the rainstorm cycle, etc.

[0068] In this embodiment, in machine learning or data analysis, the normalized data can accelerate the convergence speed of model training and improve the model performance. Therefore, the historical load data can be normalized according to the obtained data. Rainstorms may have a significant impact on the load (for example, rainstorms may cause the electricity load in some areas to increase or decrease). Through classification, the influence law of rainstorms on the load can be analyzed more accurately. Therefore, in this embodiment, the normalized historical load data can be classified according to the rainstorm data.

[0069] The following will separately describe the normalization and the classification of the normalized historical load data according to the rainstorm data.

[0070] In an alternative embodiment, the normalization process of the historical load data includes:

[0071] Calculate the load level coefficient corresponding to the historical load data.

[0072] Based on the load level coefficient, correct the historical load data that does not belong to the working day to the historical load data corresponding to the working day.

[0073] Normalize the corrected historical load data.

[0074] In this embodiment, since the historical load data can be divided into historical load data of working days, historical load data of weekends, and historical load data of holidays. In order to eliminate the influence of working days and unify the data benchmark, the embodiments of the present invention calculate two load level coefficients of working day - weekend and working day - holiday based on the load day type corresponding to the historical load data.

[0075] Among them, the two load level coefficients of working day - weekend and working day - holiday can be calculated through the following formulas respectively:

[0076]

[0077] Among them, δ m,work-weekend and δ m,work-holThey are the load level coefficients for working days - weekends and working days - holidays in the m-th year; V is the total number of load samples corresponding to working days; Q is the total number of load samples corresponding to weekends; is the v-th load sample corresponding to working days in the m-th year; is the q-th load sample corresponding to weekends in the m-th year; is the r-th load sample corresponding to holidays in the m-th year.

[0078] Based on the load level coefficient corresponding to working days - weekends, the historical load data corresponding to weekends is corrected to working days. Based on the load level coefficient corresponding to working days - holidays, the historical load data corresponding to holidays is corrected to working days.

[0079] Then, in order to eliminate annual differences, the historical load data can be normalized annually. The normalization formula can be:

[0080]

[0081] where l m ' is the normalized data; l m is the historical load data in the m-th year, m = 1, 2... M, and M is the total number of years in the historical load data record; l m,min , l m,max are the minimum load data and the maximum load data in the historical load data of the m-th year, respectively.

[0082] The normalized historical load data can be classified according to the rainstorm data. It can be classified according to the rainstorm impact area, rainfall data, and the position information of the daily load in the rainstorm cycle. Specifically:

[0083] The rainstorm impact area is used to represent the location where the rainstorm makes landfall. Based on the rainstorm impact area, the historical load data is divided into Z categories, as the first label, denoted as z ∈ [1, 2,..., Z].

[0084] Based on the rainfall data, the rainfall level can be divided into three levels, namely strong rainfall, heavy rainfall, and extreme rainfall. Correspondingly, the historical load data can be divided into three categories according to the rainfall data. This label can be the first label, denoted as T y ∈ [1, 2, 3]. In this embodiment, strong rainfall can be between [50 - 80 mm / h] in rainfall, heavy rainfall can be between [80 - 100 mm / h] in rainfall, and extreme rainfall can be above 100 mm / h in rainfall.

[0085] Based on the position of the rainstorm daily load in the rainstorm cycle, it can be divided into the first day of the rainstorm, the middle day of the rainstorm, and the last day of the rainstorm. Based on this information, it can be used as the third label, denoted as C ∈ [1, 2, 3].

[0086] Step 120: Based on the correlation information between historical load data, preliminarily expand the historical load data to obtain a preliminarily expanded sample set.

[0087] In this embodiment, the correlation information between historical load data may include the correlation between load data at adjacent moments on the same day and the correlation between load data at the same moment on different days. Based on these correlations, the historical load data can be preliminarily expanded to obtain a preliminarily expanded sample set. In this way, the subsequent training accuracy can be improved, the representativeness of the data can be enhanced, and the uncertainty of the data can be reduced.

[0088] Step 130: Determine the residual information between the historical load data on rainstorm days and the historical load data on non-rainstorm days in the preliminarily expanded sample set.

[0089] In this embodiment, since there are significant differences in the load data between rainstorm days and non-rainstorm days, the load on rainstorm days may be affected by extreme weather and exhibit characteristics different from normal conditions. By calculating the residual information, the deviation between the load data on rainstorm days and the load data on non-rainstorm days can be identified, so as to better understand the special impact of extreme weather on the load.

[0090] Step 140: Expand the preliminarily expanded sample set according to the residual information to obtain a target expanded sample set.

[0091] In this embodiment, by expanding the sample set through residual information, more samples related to the original data but with differences can be introduced. This helps to increase the diversity of the data, enables the model to learn a wider range of features and patterns, and thus improves its generalization ability. In addition, by expanding the sample set through residual information, more samples related to the original data but with differences can be introduced. This helps to increase the diversity of the data, enables the model to learn a wider range of features and patterns, and thus improves its generalization ability.

[0092] Figure 2 is a comparison graph of the quality of the sample set provided by the embodiment of the present invention. As Figure 2 shown, its abscissa is time and its ordinate is the same amount of load data. Figure 2 includes the original sample set, that is, the sample set obtained in step 110 including historical load data, the sample set expanded based on the comparison algorithm, and the expanded sample (target expanded sample set) after being expanded by the algorithm provided by the embodiment of the present invention. It can be seen from Figure 2 that the load data curve in the target expanded sample set obtained by the method provided in this embodiment is generally smoother, has fewer sawtooth waves, and has better load data quality than the original data and the load data obtained by the traditional expansion method (comparison algorithm).

[0093] Step 150: Train the target augmented sample set to obtain a load sequence scenario under rainstorm weather.

[0094] In this embodiment, a Conditional Generative Adversarial Network (CGAN) can be used to train the target augmented sample set to generate a corresponding load sequence scenario under rainstorm weather.

[0095] In this embodiment, during the preprocessing stage, the historical load data has been classified and labeled. Therefore, the load data in the target augmented sample set also has corresponding labels.

[0096] Divide the target augmented sample set into a training set and a test set. Among them, 20% of the samples can be selected as the test set, and 80% as the training set. Of course, in the actual training process, it can also be set according to needs.

[0097] Initialize the CGAN network parameters, where the label serial number h is 0. Determine whether h is 0; if h is not 0, delete the h-th input in the training set.

[0098] Generate random noise that conforms to the standard normal distribution, and combine the one-hot encoding of the rainstorm label to input into the generator to obtain generated samples.

[0099] Input the generated samples and their corresponding labels, and the samples and their corresponding labels in the training set into the discriminator.

[0100] Calculate the loss functions of the generator and the discriminator, and update the network parameters of the generator and the discriminator according to the loss functions. If the training is not over, return to the step of generating random noise that conforms to the standard normal distribution and combining the one-hot encoding of the rainstorm label to input into the generator to obtain generated samples for the next round of training. If the training is over, call the generator module, input random noise and the rainstorm label to generate a set of Q rainstorm load sequence scenarios.

[0101] In this embodiment, CGAN includes a generator G and a discriminator D. G generates samples according to the conditional information, while D distinguishes real samples and generated samples under the action of the conditional information. Among them, Figure 3 is the change curve of the discriminator loss function of the rainstorm weather load sequence scenario generation method provided by the embodiment of the present invention; from Figure 3 it can be seen that G generates samples according to the conditional information, while D distinguishes real samples and generated samples under the action of the conditional information.

[0102] Detect the label serial number h at this time. If h = 0, use the generated set of rainstorm load sequence scenarios and the test set to calculate the accuracy coefficient δ of the generated rainstorm load sequence scenarioacc and define the generated scenario set at this time as the complete input generated scenario set; otherwise, calculate the influence coefficient δ of the sub-sample set under the h-th label on the network output result by using the generated scenario set and the complete input generated scenario set at this time h ;

[0103] Let h = h + 1; if h = H (where H represents the total number of labels in the training set), the calculation ends; otherwise, return to the step of judging whether h is 0 until h = H

[0104] In this embodiment, the purpose of generating the rainstorm load sequence scenario based on CGAN is to generate samples that are similar but not identical to the real samples, that is, having statistical laws similar to historical scenarios. Therefore, the embodiments of the present invention adopt distribution characteristic indicators such as probability density function (PDF), cumulative distribution function (CDF), autocorrelation function (ACF), partial autocorrelation function (PACF), and range indicators such as critical ratio (CR) to measure the distribution characteristics and value ranges of the scenarios respectively, and then construct a comprehensive index of rainstorm scenario generation quality to measure the degree to which the comprehensive random characteristics of the generated scenarios are close to the historical scenarios. The closer to the random characteristics of the historical scenarios themselves, the higher the quality of the generated scenarios

[0105] By analyzing whether deleting a certain feature in the model input has a significant impact on the decision result and comparing the impact degrees of different features on the result, the degree of regulation of the model output by each input can be obtained

[0106] In the process of generating the rainstorm load sequence scenario, it is necessary to calculate the corresponding accuracy coefficient and influence coefficient according to the distribution characteristic indicators. The accuracy coefficient measures the degree of closeness between the generated scenario set of the rainstorm day load sequence and the test set (historical sample set). The closer to 0, the closer the statistical characteristics of the generated scenario are to the statistical characteristics contained in the historical samples, and the higher the quality of the corresponding generated scenario. The influence coefficient accurately depicts the influence of the presence or absence of the sub-sample set under the h-th label in the neural network input on the output result. Through these indicators, it is possible to further evaluate which features in the training set mainly regulate the generated scenarios and the differences in the regulation strengths of different features on the final result, so as to give causal explanation information of the generated scenarios

[0107] Exemplarily, the autocorrelation function and the partial autocorrelation function are illustrated below through an optional embodiment Figure 4a andFigure 4b is the box plot of the autocorrelation coefficient and the partial autocorrelation coefficient provided by the embodiment of the present invention. Among them, Figure 4a is the box plot of the autocorrelation coefficient, Figure 4b is the box plot of the partial autocorrelation coefficient. The abscissa in the figure represents time, and the ordinate represents the magnitude of the coefficient value. It can be seen from Figure 4a that as the time interval becomes longer, the autocorrelation coefficient of the load sequence gradually decreases. It can be seen from Figure 4b that the load sequence has a certain partial autocorrelation within the time interval of 1h, and the partial autocorrelation coefficient fluctuates around 0 after 1h. At the same time, it can also be seen that the autocorrelation coefficient and the partial autocorrelation coefficient of the real load (the symbol in the figure is ×) are both included in the generated load scenario set.

[0108] Exemplarily, the probability density function and the cumulative distribution function are described below through an optional embodiment; Figure 5a and Figure 5b is the comparison chart of the root mean square error of the probability density function and the cumulative probability density function provided by the embodiment of the present invention; it shows the root mean square error of the probability density function and the cumulative distribution function corresponding to the original data, the comparison algorithm, and the algorithm given in this embodiment.

[0109] Figure 5a and Figure 5b By using kernel density estimation, the PDFs of the daily load sequence scenarios generated by the algorithm given in this embodiment and the comparison algorithm are respectively fitted with the label {1, 2, 1} (the meaning of this label can be determined according to the explanations given in the above related embodiments and will not be elaborated here), Figure 5a and the corresponding PDF curves are given; Figure 5b gives the CDF curve obtained by integrating the PDF.

[0110] It can be seen that compared with the comparison algorithm, the PDF and CDF curves of the daily load sequence scenarios generated by the algorithm in the text are closer to the PDF and CDF curves of the real scenario. Table 1 gives the root mean square errors between the PDFs, CDFs of the daily load scenarios generated by the algorithm in the text, the comparison algorithm with the label {1, 2, 1} and the PDFs, CDFs of the real scenario with the label {1, 2, 1}. <{

[0111] Table 1

[0112] Root Mean Square Error PDF CDF The algorithm in the text 0.110 0.007 Comparison algorithm 0.307 0.020

[0113] From Table 1 and Figure 5a and Figure 5bIt can be seen that the root mean square error of the algorithm in the text is significantly smaller than that of the comparative algorithm, that is, the root mean square error between the PDF and CDF of the daily load sequence scenarios generated by the algorithm in the text and the PDF and CDF of the real scenarios is significantly smaller. This shows that the statistical characteristics of the daily load sequence scenarios with the label {1, 2, 1} generated by the algorithm in this paper are significantly closer to the statistical characteristics of the real scenarios, and the same is true for the daily load sequence scenario sets under other labels. Accordingly, the quality of the daily load sequence scenarios under the complete rainstorm cycle obtained based on this combination is higher than that of the comparative algorithm. The reason is that the comparative algorithm directly uses GAN for sample expansion, and when the number of original samples is small (only 634 days in the calculation example of this paper), the sample expansion effect is poor, resulting in this situation.

[0114] Figure 6a and Figure 6b are the comparison diagrams of the influence degrees of the sample sets under different labels provided by the embodiments of the present invention on the generated load sequence scenarios under rainstorm weather; as Figure 6a and Figure 6b shown.

[0115] In this embodiment, Figure 6a is the influence degree of the sample set obtained when the label is [1, 2, 1] on the generated load sequence scenarios under rainstorm weather, Figure 6b is the influence degree of the sample set obtained when the label is [3, 3, 2] on the generated load sequence scenarios under rainstorm weather.

[0116] To visually observe the influence degrees of the sample sets under different labels on the generated daily load sequence scenario sets under the target label, let the target label be Then the distance d dis between the target label and other labels is defined as:

[0117]

[0118] The influence coefficients of each label in Table 3 are plotted as a color block diagram, and sorted according to the size order of the distance d dis between each label and the target label (random sorting if the distances are the same), and the obtained influence coefficient color block diagram is as shown in Figure 9 . The abscissa represents the distance between each label and the target label, the ordinate represents different confidence levels, and the color of the color block represents the size of the influence coefficient. The closer the color is to the warm color system, the higher the influence coefficient, and the closer the color is to the cold color system, the lower the influence coefficient.

[0119] Among them, Table 2 is:

[0120] Table 2

[0121]

[0122] In Table 2, after separately deleting the subsample sets under each label in the training set and training the CGAN, after the training is completed, the influence coefficients on the daily load sequence scenario sets with generated labels {1, 2, 1} and {3, 3, 2} are calculated respectively. It can be found that for the label {1, 2, 1}, the influence coefficients of the label {1, 2, 1} in the training set on it at each confidence level are 0.051, 0.038, and 0.017 respectively, which are significantly greater than the influence coefficients of other labels (features); for the label {3, 3, 2}, the influence coefficients of the label {3, 3, 2} in the training set on it at each confidence level are 0.053, 0.036, and 0.025 respectively, which are significantly greater than the influence coefficients of other labels (features). Because the subsample set under the same label as the target label in the training set naturally has the greatest regulatory strength for the sample generation under the target label.

[0123] At the same time, observing Table 2, it can be found that when generating the daily load sequence scenario set with the label {1, 2, 1}, the influence coefficients of labels close to {1, 2, 1} such as {1, 1, 1}, {1, 3, 1}, and {1, 2, 2} in the training set at each confidence level are greater than those of labels slightly farther from {1, 2, 1} such as {3, 1, 3}, {3, 2, 3}, and {3, 3, 3} at each confidence level. That is, the influence coefficients of the labels in the training set that are closer to the target label are larger than those of the labels that are farther away. This is completely correct. The subsample set under the label that is relatively close naturally has a greater regulatory strength for the generation of the daily load sequence scenario under the target label.

[0124] From Figure 6a and Figure 6b it can be seen that no matter at which confidence level, as the abscissa changes, the overall color block changes from the warm color system to the cold color system, that is, as the distance between the label and the target label increases, its influence coefficient generally shows a decreasing trend. Therefore, Figure and ​ intuitively shows that the daily load sequence scenario set under the target label is mainly regulated by the subsample set under its own label and the subsample sets under other labels that are close, while the regulatory strength of the subsample sets under the labels that are farther away is relatively smaller. The above causal explanation of the regulatory strength of the generation of the daily load sequence scenario under the target label is reasonable.

[0125] In summary, in the embodiments of the present invention, historical load data is first preprocessed to eliminate the differences in limited historical load data due to the number of years and day types, and improve the availability of scarce historical load data. Then, a two-stage sample expansion strategy based on empirical statistics is proposed to effectively expand scarce samples; based on the effectively expanded rainstorm load sample set, a rainstorm load sequence generation method based on CGAN and a corresponding causal explanation method are proposed. Specifically, it includes constructing a generation scenario accuracy index to evaluate the quality of the generation scenario, constructing a feature influence index to evaluate the regulation strength of different features on the generation scenario, and combining with CGAN training to realize the generation of rainstorm load sequence scenarios based on CGAN and reasonable and effective causal explanations.

[0126] In an alternative embodiment, in step 120, based on the correlation information between historical load data, the historical load data is preliminarily expanded to obtain a preliminarily expanded sample set, which may include:

[0127] Step 121: Determine the correlation between each acquisition moment within each acquisition day in the corresponding historical stage of the historical load data to obtain the horizontal correlation.

[0128] Step 122: Determine the correlation between the same acquisition moment on different acquisition days in the corresponding historical stage of the historical load data to obtain the vertical correlation.

[0129] Step 123: According to the horizontal correlation and the vertical correlation, the historical load data is preliminarily expanded to obtain a preliminarily expanded sample set.

[0130] In this embodiment, for any category set, the correlation between the load data at adjacent moments on the same day in the category set can be used as the horizontal correlation; the correlation between the load data at the same moment on different days can be used as the vertical correlation. Taking a load data as an example below, it shows how to preliminarily expand the historical load data according to the horizontal correlation and the vertical correlation:

[0131] The preliminary expansion of the historical load data according to the horizontal correlation can be expressed as:

[0132] D h,a,t = D a,t-1 + r and,h × (D a,t - D a,t-1 )

[0133] Where D h,a,t is the load value obtained by horizontally expanding the sample load data at the t-th moment and the (t - 1)-th moment on the a-th day in the category set; D a,t-1 and D a,tThey are the load data at the (t-1)th moment and the tth moment on the ath day in the category set respectively; r and,h is a random number between (0, 1).

[0134] According to the longitudinal correlation, the preliminary expansion of historical load data can be expressed as:

[0135] D z,a,t = D a,t + r and,z × (D a+1,t - D a,t )

[0136] where D z,a,t is the load value obtained by longitudinally expanding the load data at the tth moment on the ath day and the (a + 1)th day in the category set; D a+1,t is the data at the tth moment on the (a + 1)th day in the category set; r and,z is a random number between (0, 1).

[0137] According to the above formula, the historical load data is preliminarily expanded respectively according to the horizontal correlation and the longitudinal correlation, and then a horizontally expanded sample set and a longitudinally expanded sample set are obtained. Each data in the horizontally expanded sample set and the longitudinally expanded sample set corresponds one by one. For any pair of data among them, the following steps are executed:

[0138] For D z,a,t and D h,a,t adopt the fusion based on the Kalman gain coefficient to improve the effectiveness of the expanded load value, that is:

[0139] D r,a,t = D h,t + b r (D z,a,t - D h,a,t )

[0140]

[0141] where D r,a,t is the preliminary expanded load data at the tth moment on the ath day obtained by fusing the load data at the tth moment and the (t - 1)th moment on the ath day and the (a + 1)th day in the category set; b r is the Kalman gain coefficient.

[0142] In this embodiment, 96 data can be collected in one day. According to the above method, the preliminary expanded load data in each category set is obtained. For each preliminarily expanded load data obtained by expansion, its corresponding category label is supplemented, and then a preliminarily expanded sample set is obtained.

[0143] In an optional embodiment, in step 130, the residual information between the historical load data on rainy days and the historical load data on non-rainy days in the preliminary augmented sample set may include:

[0144] Step 131: In the preliminary augmented sample set, extract the historical load sequence data of the target non-rainy day adjacent to the target rainy day.

[0145] Step 132: Calculate the residual between the historical load sequence data on the target rainy day and the historical load sequence data on the target non-rainy day to obtain the residual information.

[0146] In this embodiment, in the preliminary augmented sample set, the rainstorm load samples with the class label [z,T y ,C] can be used as the historical load sequence data corresponding to the target rainy day; where D is the number of target rainy days in this class label. In this embodiment, the d-th target rainy day can be taken as an example for illustration. Therefore, d = 1 can be set.

[0147] In the preliminary augmented sample set, extract the historical load sequence data adjacent to the target rainy day where Let b = 1 to obtain the historical load sequence data of the b-th target non-rainy day.

[0148] Taking any moment as an example, the residual between the historical load sequence data on the target rainy day and the historical load sequence data on the target non-rainy day is determined by the following formula:

[0149]

[0150] where represents the residual at the e-th moment; represents the load data of the target rainy day at the e-th moment; represents the load data of the target non-rainy day at the e-th moment.

[0151] Based on the above formula, determine all the corresponding residuals in the preliminary augmented sample set, and then obtain the residual information.

[0152] In an optional embodiment, in step 132, calculating the residual between the historical load sequence data on the target rainy day and the historical load sequence data on the target non-rainy day to obtain the residual information may include:

[0153] Calculate the residual between the historical load sequence data on the target rainy day and the historical load sequence data on the target non-rainy day to obtain the target residual sequence.

[0154] Calculate the kernel densities corresponding to the target residual sequence and the historical load sequence data on the target non-rainstorm days respectively, to obtain the empirical probability density function of the target residual sequence and the empirical probability density function on the target non-rainstorm days.

[0155] According to the empirical probability density function of the target residual sequence and the empirical probability density function on the target non-rainstorm days, obtain the empirical density function.

[0156] Sample the empirical density function to obtain a sampled residual sequence, and use the sampled residual sequence as the residual information.

[0157] In this embodiment, taking a target rainstorm day and its corresponding target non-rainstorm day as an example, according to the residuals corresponding to each moment between the target rainstorm day and its corresponding target non-rainstorm day, obtain the target residual sequence between the target rainstorm day and its corresponding target non-rainstorm day.

[0158] The empirical probability density function f r (r) of the target residual sequence and the empirical probability density function f l (l C ) under the target non-rainstorm days can be calculated respectively using the kernel density estimation formula; among them, the kernel density estimation formula can be expressed as:

[0159]

[0160] Among them, B is the total number of bandwidths; h is the optimal bandwidth determined by h = 1.06σn -1 / 5 ; x is a random variable; X is a sample point.

[0161] According to the empirical probability density function of the target residual sequence and the empirical probability density function on the target non-rainstorm days, obtain their cumulative distributions, that is, F r (r) and F l (l C ).

[0162] Use the Copula function C(·) to describe the empirical joint distribution function of r and l C , that is:

[0163] F rl (r, l C ) = C(F r (r), F l (l C ))

[0164] Among them, F rl (r, l C ) is the joint distribution function of r and l C , and C(·) is the Copula function.

[0165] Taking the derivative of the empirical joint distribution function gives:

[0166]

[0167] where c(·) is the density function of the Copula function C(·). In the embodiments of the present invention, the Frank-Copula function can be selected. Correspondingly, the empirical density function can be expressed as:

[0168]

[0169] where U and V are random variables respectively; θ is a parameter.

[0170] Based on Bayes' formula, the empirical joint distribution function is processed to obtain the empirical conditional probability density function:

[0171]

[0172] Sampling the empirical density function to obtain a sampling residual sequence, and using the sampling residual sequence as the residual information R g =[r g,1 , r g,2 ... r g,e ... r g,96 .

[0173] In an alternative embodiment, in step 140, expanding the preliminary augmented sample set according to the residual information to obtain the target augmented sample set may include:

[0174] Randomly sampling to generate a sampling sample sequence on the target non-rainstorm day according to the empirical probability density function and the empirical density function on the target non-rainstorm day.

[0175] Superimposing the sampling sample sequence and the sampling residual sequence to obtain the load augmented sample on the target rainstorm day.

[0176] Based on the load augmented sample on the target rainstorm day, obtain the target augmented sample set.

[0177] In this embodiment, according to the empirical probability density function and the empirical density function on the target non-rainstorm day, that is, f l (l C ) and c(U, V), randomly sample to generate the sampling sample sequence on the target non-rainstorm day

[0178] Superimpose the sampling sample sequence and the sampling residual sequence R g =[r g,1 , r g,2 ... r g,e ... r g,96Perform superposition to obtain the load expansion samples under the target heavy rain days. Its expression is:

[0179]

[0180] Based on the load expansion samples under the target heavy rain days, obtain the target expansion sample set.

[0181] ​ is the implementation flowchart of the load sequence scenario generation method under heavy rain weather provided by another embodiment of the present invention. As ​ shown, the method may include:

[0182] First, obtain historical load data, correct the historical load level; classify the corrected historical load data and determine the corresponding labels according to the classification results. Perform two-stage expansion on the historical load data to obtain the target expansion sample set.

[0183] Step 1: For the target expansion sample set, divide 20% of it into the test set and 80% into the training set, initialize the CGAN network parameters, and let h = 0, where h represents the label sequence number;

[0184] Step 2: If h = 0, go to Step 3; otherwise, delete the h-th input in the training set and go to Step 3;

[0185] Step 3: Generate random noise that conforms to the standard normal distribution, and combine it with the one-hot encoding of the heavy rain label to input the generator to obtain the generated samples;

[0186] Step 4: Input the generated samples and their corresponding labels, and the training set samples and their corresponding labels into the discriminator;

[0187] Step 5: Calculate the loss functions of the generator and the discriminator, and update the network parameters of the generator and the discriminator according to the loss functions. If the training is not over, return to Step 3 for the next round of training; otherwise, perform Step 6;

[0188] Step 6: Call the generator module, input random noise and heavy rain labels to generate Q sets of heavy rain load sequence scenarios;

[0189] Step 7: If h = 0, calculate the accuracy coefficient δ acc of the generated scenarios of the heavy rain load sequence using the generated scenario set and the test set, and call the generated scenario set at this time the complete input generated scenario set; otherwise, calculate the influence coefficient δ h of the sub-sample set under the h-th label on the network output result using the generated scenario set at this time and the complete input generated scenario set; determine whether h is equal to H;

[0190] Step 8: If not, set h = h + 1 and return to step 2; if h = H (H represents the total number of labels in the training set), the calculation ends.

[0191] In summary, the present embodiment solves the problem of insufficient valid sample data through a two-stage sample expansion. Furthermore, during sample expansion, the present embodiment considers the correlations between data and the corresponding causal relationships. The first sample expansion based on correlations enables interpretation of the causal relationships in the generated load sequence scenarios. The second sample expansion, using residual information, improves the accuracy of the generated load sequence scenarios.

[0192] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0193] The following are device embodiments of the present invention. For details not fully described therein, reference may be made to the corresponding method embodiments described above.

[0194] ​ The following is a schematic diagram showing the structure of a load sequence scene generation device under heavy rain weather according to an embodiment of the present invention. For ease of explanation, only the parts related to the embodiment of the present invention are shown, which are described in detail as follows:

[0195] like ​ As shown, the load sequence scene generation device 8 under heavy rain weather includes:

[0196] An acquisition module 81 is configured to acquire historical load data, wherein the historical load data includes historical load data on rainstorm days and historical load data on non-rainstorm days;

[0197] An expansion module 82 is used to preliminarily expand the historical load data based on the correlation information between the historical load data to obtain a preliminary expanded sample set;

[0198] A determination module 83 is used to determine residual information between historical load data on rainstorm days and historical load data on non-rainstorm days in the preliminary expanded sample set;

[0199] The expansion module 82 is further configured to expand the preliminary expanded sample set according to the residual information to obtain a target expanded sample set;

[0200] The training module 84 is used to train the target expanded sample set to obtain a load sequence scenario under heavy rain weather.

[0201] In an optional embodiment, the expansion module 82 is specifically configured to:

[0202] Determine the correlation between each collection moment within each collection day in the corresponding historical stage of the historical load data to obtain the horizontal correlation;

[0203] Determine the correlation between the same collection moment in different collection days in the corresponding historical stage of the historical load data to obtain the vertical correlation;

[0204] According to the horizontal correlation and the vertical correlation, preliminarily expand the historical load data to obtain a preliminarily expanded sample set.

[0205] In an alternative embodiment, the expansion module 82 is specifically configured to:

[0206] Preliminarily expand the historical load data according to the horizontal correlation and the vertical correlation respectively to obtain a horizontally expanded sample set and a vertically expanded sample set;

[0207] Fuse the horizontally expanded sample set and the vertically expanded sample set to obtain a preliminarily expanded sample set.

[0208] In an alternative embodiment, the historical load data includes historical load sequence data; the determination module 83 is specifically configured to:

[0209] Extract the historical load sequence data of the target non-rainstorm day adjacent to the target rainstorm day in the preliminarily expanded sample set;

[0210] Calculate the residual between the historical load sequence data on the target rainstorm day and the historical load sequence data on the target non-rainstorm day to obtain residual information.

[0211] In an alternative embodiment, the determination module 83 is specifically configured to:

[0212] Calculate the residual between the historical load sequence data on the target rainstorm day and the historical load sequence data on the target non-rainstorm day to obtain a target residual sequence;

[0213] Calculate the kernel densities corresponding to the target residual sequence and the historical load sequence data on the target non-rainstorm day respectively to obtain the empirical probability density function of the target residual sequence and the empirical probability density function on the target non-rainstorm day;

[0214] According to the empirical probability density function of the target residual sequence and the empirical probability density function on the target non-rainstorm day, obtain an empirical density function;

[0215] Sample the empirical density function to obtain a sampled residual sequence, and use the sampled residual sequence as the residual information.

[0216] In an alternative embodiment, the expansion module 82 is specifically configured to:

[0217] According to the empirical probability density function and empirical density function on the target non-rainstorm days, a sampling sample sequence on the target non-rainstorm days is randomly generated;

[0218] The sampling sample sequence and the sampling residual sequence are superimposed to obtain a load expansion sample on the target rainstorm days;

[0219] Based on the load expansion sample on the target rainstorm days, a target expansion sample set is obtained.

[0220] In an alternative embodiment, the device further includes a processing module 85; the processing module 85 is specifically configured to:

[0221] Obtain rainstorm data corresponding to historical load data; wherein, the rainstorm data includes the rainstorm impact area, rainfall data, and the position information of the daily load in the rainstorm cycle;

[0222] Perform normalization processing on the historical load data, and classify the normalized historical load data according to the rainstorm data.

[0223] In an alternative embodiment, the processing module 85 is specifically configured to:

[0224] Calculate the load level coefficient corresponding to the historical load data;

[0225] Based on the load level coefficient, correct the historical load data that does not belong to working days to the historical load data corresponding to working days;

[0226] Perform normalization processing on the corrected historical load data.

[0227] In an alternative embodiment, the training module 84 is specifically configured to:

[0228] Divide the target expansion sample set into a training set and a test set, and initialize the network parameters;

[0229] Based on the initialized network parameters, the training set and the test set, perform training to obtain a load sequence scenario under rainstorm weather.

[0230] ​ It is a schematic diagram of an electronic device provided by an embodiment of the present invention. As ​ shown, the electronic device 9 in this embodiment includes: a processor 90 and a memory 91. The memory 91 stores a computer program 92. When the processor 90 executes the computer program 92, the steps in the above-mentioned various method embodiments are implemented. Alternatively, when the processor 90 executes the computer program 92, the functions of each module / unit in the above-mentioned various device embodiments are implemented.

[0231] Exemplarily, the computer program 92 can be divided into one or more modules / units, which are stored in the memory 91 and executed by the processor 90 to implement the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing specific functions, and these instruction segments are used to describe the execution process of the computer program 92 in the electronic device 9.

[0232] The electronic device 9 may include, but is not limited to, a processor 90 and a memory 91. Those skilled in the art can understand that ​ merely examples of the electronic device 9, which do not constitute a limitation on the electronic device 9, may include more or fewer components than shown in the figure, or combine certain components, or different components. For example, the electronic device 9 may further include input / output devices, network access devices, a bus, etc.

[0233] For the convenience and brevity of description, only the above division of each functional module / unit is used as an example. In practical applications, the above functions can be allocated to different functional modules / units according to needs. The above modules / units can be implemented in the form of hardware, or in the form of software, or in the form of a combination of hardware and software.

[0234] In the above embodiments, the descriptions of the respective embodiments have their own focuses. For parts not detailed or recorded in a certain embodiment, reference can be made to the relevant descriptions of other embodiments. Without special instructions and logical conflicts, the terms and / or descriptions between different embodiments are consistent and can be mutually referred to. The technical features in different embodiments can be combined to form new embodiments according to their inherent logical relationships.

[0235] The above-described embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for generating load sequence scenarios under rainstorm weather, characterized in that, Including: Obtain historical load data; wherein, the historical load data includes historical load data on rainstorm days and historical load data on non-rainstorm days; Based on the correlation information between the historical load data, preliminarily expand the historical load data to obtain a preliminarily expanded sample set; Determine the residual information between the historical load data on rainstorm days and the historical load data on non-rainstorm days in the preliminarily expanded sample set; Expand the preliminarily expanded sample set according to the residual information to obtain a target expanded sample set; Train the target expanded sample set to obtain a load sequence scenario under rainstorm weather.

2. The method for generating a load sequence scenario under a rainstorm weather condition according to claim 1, wherein The step of based on the correlation information between the historical load data, preliminarily expanding the historical load data to obtain a preliminarily expanded sample set includes: Determine the correlation between each acquisition moment within each acquisition day in the corresponding historical stage of the historical load data to obtain the horizontal correlation; Determine the correlation between the same acquisition moment in different acquisition days in the corresponding historical stage of the historical load data to obtain the vertical correlation; Preliminarily expand the historical load data according to the horizontal correlation and the vertical correlation to obtain a preliminarily expanded sample set.

3. The method for generating a load sequence scenario under heavy rain weather according to claim 2, wherein The step of preliminarily expanding the historical load data according to the horizontal correlation and the vertical correlation to obtain a preliminarily expanded sample set includes: Preliminarily expand the historical load data according to the horizontal correlation and the vertical correlation respectively to obtain a horizontally expanded sample set and a vertically expanded sample set; Fuse the horizontally expanded sample set and the vertically expanded sample set to obtain a preliminarily expanded sample set.

4. The method for generating a load sequence scenario under a heavy rain weather condition according to claim 1, wherein The historical load data includes historical load sequence data; the step of determining the residual information between the historical load data on rainstorm days and the historical load data on non-rainstorm days in the preliminarily expanded sample set includes: In the preliminarily expanded sample set, extract the historical load sequence data of the target non-rainstorm day adjacent to the target rainstorm day; Calculate the residual between the historical load sequence data on the target rainstorm day and the historical load sequence data on the target non-rainstorm day to obtain the residual information.

5. The method for generating a load sequence scenario under a rainstorm weather condition according to claim 4, wherein The step of calculating the residual between the historical load sequence data on the target rainstorm day and the historical load sequence data on the target non-rainstorm day to obtain the residual information includes: Calculate the residual between the historical load sequence data on the target rainstorm day and the historical load sequence data on the target non-rainstorm day to obtain a target residual sequence; Calculate the kernel density corresponding to the target residual sequence and the historical load sequence data on the target non-rainstorm day respectively to obtain the empirical probability density function of the target residual sequence and the empirical probability density function of the target non-rainstorm day; Obtain an empirical density function according to the empirical probability density function of the target residual sequence and the empirical probability density function of the target non-rainstorm day; Sample the empirical density function to obtain a sampled residual sequence, and use the sampled residual sequence as the residual information.

6. The method for generating a load sequence scenario under heavy rain weather according to claim 5, characterized in that Expanding the preliminary augmented sample set according to the residual information to obtain a target augmented sample set, including: Randomly sampling to generate a sampling sample sequence under the target non-rainstorm days according to the empirical probability density function and the empirical density function under the target non-rainstorm days; Superimposing the sampling sample sequence and the sampling residual sequence to obtain a load augmented sample under the target rainstorm days; Based on the load augmented sample under the target rainstorm days, obtaining a target augmented sample set.

7. The method for generating a load sequence scenario under heavy rain weather according to claim 1, wherein Before the historical load data is preliminarily augmented based on the correlation information between the historical load data to obtain a preliminary augmented sample set, it further includes: Obtaining rainstorm data corresponding to the historical load data; wherein, the rainstorm data includes a rainstorm impact area, rainfall data, and position information of the daily load in the rainstorm cycle; Normalizing the historical load data, and classifying the normalized historical load data according to the rainstorm data.

8. The method for generating a load sequence scenario under heavy rain weather according to claim 7, characterized in that, The historical load data includes historical load day types; the normalizing of the historical load data includes: Calculating a load level coefficient corresponding to the historical load data; Based on the load level coefficient, correcting historical load data that does not belong to working days to historical load data corresponding to working days; Normalizing the corrected historical load data.

9. The method for generating a load sequence scenario under heavy rain weather according to claim 8, wherein Training the target augmented sample set to obtain a load sequence scenario under rainstorm weather, including: Dividing the target augmented sample set into a training set and a test set, and initializing network parameters; Training based on the initialized network parameters, the training set, and the test set to obtain a load sequence scenario under rainstorm weather.

10. A load sequence scenario generation device under rainstorm weather conditions, characterized in that, Including: An acquisition module, configured to acquire historical load data; wherein, the historical load data includes historical load data on rainstorm days and historical load data on non-rainstorm days; An augmentation module, configured to preliminarily augment the historical load data based on the correlation information between the historical load data to obtain a preliminary augmented sample set; A determination module, configured to determine the residual information between the historical load data on rainstorm days and the historical load data on non-rainstorm days in the preliminary augmented sample set; The augmentation module is further configured to expand the preliminary augmented sample set according to the residual information to obtain a target augmented sample set; A training module, configured to train the target augmented sample set to obtain a load sequence scenario under rainstorm weather.

Images