A method for predicting time-space series of power outages during typhoons

Abstract

The present invention discloses a method for predicting a time-space sequence of power outages during a typhoon, comprising: S1, data collection; S2, pre-processing the data: dividing the target area into grid units, assigning values ​​to the collected data according to the grid units, and obtaining a data set of the target area; S3, creating a training set, processing the training set through a random forest classification model, and constructing a power outage time series prediction model based on a gated cyclic unit model and a random forest classification model; S4, model training: training the gated cyclic unit model in the power outage time series prediction model through the processed training set, and training the random forest classification model in the power outage time series prediction model through the training set before processing; S5, model testing: performing a time-space sequence prediction test on the power outage time series prediction model. The present invention can predict the time and space of large-scale power outages before a typhoon lands, and has high prediction accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of disaster prevention and reduction, and in particular to a method for predicting a time-space sequence of power outages during a typhoon. Background Art

[0002] The prediction of power outages in the distribution system throughout the entire process before and after the typhoon lands can guide power grid companies to formulate disaster response plans, rationally plan and allocate various personnel and resources, minimize the impact of disasters on the distribution system, and thus improve the resilience of the distribution system.

[0003] There are many factors that cause power outages in distribution systems during typhoon disasters, and they are highly uncertain. Power outage prediction is very challenging. Research on power outage prediction before typhoon disasters can be mainly divided into physical-driven research and data-driven research. The power outage prediction model in physical-driven research mainly analyzes the damage mechanism of the distribution system under typhoon disasters and constructs a mathematical model between the reaction environment and the loss of the distribution system. When there are many influencing factors, the construction of the physical model is extremely complex and difficult to solve, which is not suitable for power outage prediction in complex environments during typhoon disasters.

[0004] With the continuous improvement of power grid data management systems and the continuous development of data mining technology, data-driven methods are widely used in power outage prediction. Statistical learning methods such as generalized linear models and generalized additive models are used to establish power outage prediction models, but these statistical learning models strongly rely on historical data distribution, making it difficult to mine the causes of typhoon-induced power outages and cannot be used in new typhoon events.

[0005] Compared with statistical learning methods, machine learning methods have good performance in data mining and establishing feature relationships, and have begun to be widely used in power outage prediction research. For example, the random forest algorithm has been proven to have good power outage prediction performance in typhoon disasters. However, the existing machine learning-based power outage prediction models generally use traditional machine learning methods, which make it difficult to consider the relationship between feature time series and can only predict the spatial distribution of power outages. During typhoon disasters, it is also very important to pay attention to the time distribution of power outages. Accurate prediction of the spatiotemporal distribution of power outages can better serve the emergency response of power grid companies in typhoon disasters.

[0006] Deep learning is a branch of machine learning, in which recurrent neural networks (RNNs) can automatically extract temporal features and perform inference, and are widely used in time series prediction tasks. Some optimization models based on recurrent neural networks, such as the Long Short-Term Memory (LSTM) model and the Gated Recurrent Unit (GRU) model, have been applied to power outage prediction. However, these models usually only focus on the prediction of power outage time series in small areas and are not suitable for power outage prediction in large areas. It is difficult to meet the actual needs of power grid companies. The application of deep learning technology in the prediction of power outage time series in typhoon disasters needs further study. Summary of the invention

[0007] In order to overcome the deficiencies of the above technologies, the present invention provides a method for predicting the spatiotemporal sequence of power outages during typhoons.

[0008] Terminology explanation:

[0009] 1. DEM: Digital Elevation Model.

[0010] 2. MSE: mean-square error.

[0011] 3. MAE: Mean Absolute Error, mean absolute error.

[0012] 4. RMSE: Root Mean Square Error.

[0013] The technical solution adopted by the present invention to overcome the technical problems is:

[0014] A method for predicting time-space series of power outages during typhoons comprises the following steps:

[0015] S1. Data collection, wherein the data at least includes meteorological data, geographical data and power distribution system data of the target area;

[0016] S2, pre-processing the data collected in step S1: dividing the target area into grid units, assigning values ​​to the collected data according to the grid units, and obtaining a data set of the target area;

[0017] S3, creating a training set through the data set of step S2, processing the training set through a random forest classification model, and building a power outage time series prediction model based on the gated recurrent unit model and the random forest classification model;

[0018] S4, model training: training the gated recurrent unit model in the power outage time series prediction model using the training set processed in step S3, and training the random forest classification model in the power outage time series prediction model using the training set before processing in step S3;

[0019] S5. Model testing: Performing a spatiotemporal series prediction test on the power outage time series prediction model trained by the model in step S4.

[0020] Furthermore, in step S1, the meteorological data includes at least the maximum wind speed, hourly rainfall and typhoon path observed at the meteorological observation site; the geographic data includes at least digital elevation model data; and the distribution system data includes at least the spatial distribution of distribution transformers and power outage monitoring data of distribution transformers.

[0021] Furthermore, step S2 specifically includes:

[0022] 1) Pre-processing the meteorological data collected in step S1, specifically including:

[0023] According to the maximum wind speed and hourly rainfall of the meteorological observation station, the maximum wind speed and hourly rainfall of the center point of the grid unit are estimated by using the inverse distance interpolation method, and the estimated maximum wind speed and hourly rainfall of the center point of the grid unit are assigned to the corresponding grid unit;

[0024] The typhoon path is calculated to obtain the distance to the typhoon center. For the distance to the typhoon center, the distance from the center point of the grid unit to the current measured typhoon center is calculated every hour, and the distance between the center point of the grid unit and the current typhoon center is assigned to the corresponding grid unit;

[0025] 2) Pre-processing the geographic data collected in step S1, specifically including:

[0026] Based on the collected digital elevation model data of the target area, ArcGIS software was used to extract the slope and aspect, and the mean elevation, slope and aspect values ​​within the grid cells were calculated and assigned to the grid cells.

[0027] 3) Pre-processing the power distribution system data collected in step S1, specifically including:

[0028] Based on the collected spatial distribution of distribution transformers in the target area, the total number of distribution transformers in the grid unit is counted and assigned to the grid unit;

[0029] Based on the collected power outage monitoring data of distribution transformers in the target area, the total number of distribution transformers that experience power outages in the grid unit every hour is counted and assigned to the grid unit hour by hour.

[0030] Furthermore, in step S3, a training set is created using the data set of step S2, and the training set is processed using a random forest classification model, including:

[0031] Collect data on historical typhoon events and create a training set a;

[0032] Adjust the form of training set a to obtain training set b;

[0033] The random forest classification model is trained through the training set b to obtain the power outage samples caused by non-wind and rain factors contained in the training set a and filter them to obtain the training set c.

[0034] Further, in step S3:

[0035] 1) Collect data on historical typhoon events and create a training set a, including:

[0036] Collect data of N historical typhoon events to create a training set a, where N ≥ 3 and N is an integer, and the grid cells containing distribution transformers are set as target grid cells;

[0037] From the spatial dimension, the number, elevation, slope and aspect of distribution transformers of the target grid unit are extracted as the input of the power outage time series prediction model;

[0038] From the time dimension, a sample is created by sliding the time window every hour for each target grid unit, and the hourly time series of the maximum wind speed, hourly rainfall, distance to the typhoon center, and the cumulative number of distribution transformers with power outages for the past g hours and the next h hours (a total of g+h hours) are extracted at each moment, where the value range of g is [3,12], the value range of h is [3,12], and both g and h are integers. The time series of the cumulative number of distribution transformers with power outages in the past g hours, the time series of the maximum wind speed, hourly rainfall, and distance to the typhoon center in the past g hours, and the time series of the maximum wind speed, hourly rainfall, and distance to the typhoon center in the next h hours are the inputs of the power outage time series prediction model, and the time series of the cumulative number of distribution transformers with power outages in the next h hours is the output of the power outage time series prediction model.

[0039] 2) Adjust the form of training set a to obtain training set b, including:

[0040] For all samples in training set a, four features, namely, the number of distribution transformers, elevation, slope, and slope direction, are used as static feature inputs of the random forest classification model; the maximum value, minimum value, and mean value of the time series of the maximum wind speed, hourly rainfall, and distance to the typhoon center in the past g hours are extracted respectively, and the maximum value, minimum value, and mean value of the time series of the maximum wind speed, hourly rainfall, and distance to the typhoon center in the future h hours are extracted respectively, and the maximum value and mean value of the time series of the number of distribution transformers with cumulative power outages in the past g hours are extracted. A total of 20 features are used as dynamic feature inputs of the random forest classification model. A total of 24 features are input into the random forest classification model, and the 24 features include 4 static features used as inputs to the random forest classification model and 20 dynamic features used as inputs to the random forest classification model;

[0041] The time series of the number of distribution transformers with cumulative power outages in the next h hours is converted into whether the target grid unit will have a power outage in the next h hours: if the cumulative power outage time series in the next h hours increases, the target grid unit will have a power outage; otherwise, the target grid unit will not have a power outage; whether the grid unit will have a power outage in the next h hours is used as the output of the random forest classification model;

[0042] After processing the input and output of training set a, training set b is obtained;

[0043] 3) The random forest classification model is trained with training set b to obtain the power outage samples caused by factors other than wind and rain contained in training set a and filter them to obtain training set c, which specifically includes:

[0044] The samples of training set b are randomly divided into two training sets: training set b1 and training set b2; the random forest classification models are trained using training set b1 and training set b2 respectively, and they are named RF_F1 model and RF_F2 model respectively. The RF_F1 model is set as model one and the RF_F2 model is set as model two respectively. Then, prediction is performed in another training set. The power outage samples with correct prediction are regarded as power outage samples caused by wind and rain factors and retained in the training set, and the power outage samples with incorrect prediction are regarded as power outage samples caused by non-wind and rain factors and removed from the training set; the numbers of the samples in the filtered training set b1 and training set b2 are recorded, and the training set c is obtained by selecting the corresponding samples from the training set a according to the numbers.

[0045] Furthermore, in step S3, a power outage time series prediction model is constructed based on the gated recurrent unit model and the random forest classification model, specifically including:

[0046] The random forest classification model is divided into RF_F model and RF_C model according to different functions, among which RF_F model is a general term for model one and model two. The RF_F model is set as model three and the RF_C model is set as model four respectively. Model three is configured to filter the training set to eliminate power outage samples caused by non-wind and rain factors. Model four is configured to perform output correction on the final output module of the power outage time series prediction model. The structure, input and output of model three and model four are the same. Each sample inputs the 24 features of the grid unit respectively, and the output is whether a power outage will occur in the grid unit in the next h hours.

[0047] Further, in step S3, the gated recurrent unit model is recorded as a GRU_T model, and the GRU_T model includes a dynamic feature extraction module, a static feature extraction module, and a feature extraction result output module; wherein the dynamic feature extraction module adopts a coding prediction structure, and inputs a time series of four channels, namely, the maximum wind speed, hourly rainfall, distance to the typhoon center, and the cumulative number of distribution transformers with power outages in the past g hours and the next h hours, totaling g+h hours of the grid unit, wherein the time series of the cumulative number of distribution transformers with power outages in the next h hours adopts 0 placeholder processing; the static feature extraction module inputs the number of distribution transformers, mean elevation, mean slope, and mean slope aspect of the grid unit; the results of the static feature extraction module and the results of the dynamic feature extraction module are spliced ​​at the channel, and then input into the feature extraction result output module for final feature extraction processing, and the GRU_T model finally outputs the time series of the cumulative number of distribution transformers with power outages in the next h hours of the grid unit.

[0048] Furthermore, in step S3, the power outage time series prediction model is composed of the RF_C model and the GRU_T model, which is recorded as the RF_GRU_RF model and set as model five, wherein the GRU_T model is used as a time series prediction module to predict the time series of the cumulative number of distribution transformers with power outages, and the RF_C model is used as the final output module to perform output correction on the time series prediction module.

[0049] Furthermore, step S4 specifically includes:

[0050] Training of RF_GRU_RF model: Among them, RF_C model is trained in training set b; GRU_T model is trained on training set c after filtering by RF_F model. Specifically, training set c is randomly divided into training set ct and verification set cv according to a preset ratio, GRU_T model is trained in training set ct, and verified in verification set cv. The training adopts early stopping mechanism. When the loss function no longer decreases in verification set cv, the GRU_T model is considered to have been trained.

[0051] Furthermore, step S5 specifically includes:

[0052] The power outage time series prediction model is tested in the training set of typhoon events outside the training set created in step S3. Let the training set used to test the power outage time series prediction model outside the training set created in step S3 be d. The mean absolute error MAE and root mean square error RMSE are selected to evaluate the prediction performance of the power outage time series prediction model. The calculation formula of the indicators is as follows:

[0053] ,

[0054] ,

[0055] In the above formula, represents the number of samples in the training set d, represents the prediction time step, represents the observed value, Represents the predicted value.

[0056] The beneficial effects of the present invention are:

[0057] The method for predicting the time-space sequence of power outages during typhoons proposed in this invention can predict the time and space of large-scale power outages before the typhoon lands, and serve the emergency decision-making of power grid companies. Specifically:

[0058] 1. The power outage time series prediction model of the present invention can complete the spatiotemporal prediction of power outages caused by typhoons in the target area based on the time series prediction results of the maximum wind speed, hourly rainfall and distance to the typhoon center of the target grid unit.

[0059] 2. The power outage time series prediction model of the present invention can be generalized to typhoon events. When the training set does not contain events similar to typhoon trajectories, it can quickly predict large-scale power outages, and the prediction results are highly consistent with the actual observation results.

[0060] 3. The training data created in the present invention is screened using a random forest classification model, which can filter out power outage samples caused by factors other than wind and rain, so that the power outage time series prediction model can better grasp the characteristics of power outages caused by strong winds and heavy rains, and improve the accuracy of the power outage time series prediction model in predicting these power outages. In addition, the output of the power outage time series prediction model is corrected using a random forest classification model, which allows the power outage time series prediction model to output a power outage time series with a higher correlation with the input variables, and the prediction results are more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 This is a schematic diagram of a method for predicting the spatiotemporal sequence of power outages during a typhoon according to an embodiment of the present invention.

[0062] Figure 2A schematic diagram of the input and output of the random forest classification model described in an embodiment of the present invention.

[0063] Figure 3 Schematic diagram of the structure of the gated recurrent unit model described in an embodiment of the present invention.

[0064] Figure 4 It is a schematic diagram of different power outage time series prediction models described in the embodiments of the present invention.

[0065] Figure 5 This is a spatiotemporal distribution diagram of the accumulated power outages of distribution transformers actually observed in an embodiment of the present invention.

[0066] Figure 6 This is a spatiotemporal distribution diagram of the accumulated power outage distribution transformers predicted by model five according to an embodiment of the present invention.

[0067] Figure 7 This is a curve of the number of distribution transformers with cumulative power outages in the target area predicted and observed by Model 5, Model 6, Model 7, and Model 8 respectively according to the embodiment of the present invention.

[0068] Figure 8 These are curves of the MAE predicted by model five, model six, model seven, and model eight respectively according to the embodiments of the present invention.

[0069] Fig. 9 These are curves of the RMSE predicted by model five, model six, model seven, and model eight respectively according to the embodiments of the present invention.

[0070] Fig.10 It is a box plot of the maximum wind speed of samples without power outages, samples with power outages, samples among the power outage samples correctly predicted by model five, and samples among the power outage samples incorrectly predicted by model five during the entire prediction process of all target grid units described in the embodiment of the present invention.

[0071] Fig.11 A box plot of the average rainfall of samples without power outages, samples with power outages, samples among the power outage samples correctly predicted by model five, and samples among the power outage samples incorrectly predicted by model five during the entire prediction process of all target grid units described in an embodiment of the present invention.

[0072] Fig.12 It is a box plot of the minimum distances to the typhoon center for samples without power outages, samples with power outages, samples among the power outage samples correctly predicted by model five, and samples among the power outage samples incorrectly predicted by model five during the entire prediction process of all target grid units described in the embodiment of the present invention.

[0073] Fig.13 These are the prediction results for the last hour predicted by model six, model eight, model seven, and model five respectively as described in the embodiments of the present invention.

[0074] Fig.14 It is a box plot of the absolute error and relative error of the last hour prediction results predicted by model six, model eight, model seven, and model five respectively according to the embodiment of the present invention. DETAILED DESCRIPTION

[0075] In order to facilitate those skilled in the art to better understand the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The following is only exemplary and does not limit the protection scope of the present invention.

[0076] The present embodiment provides a method for predicting the time-space sequence of power outages during a typhoon. Figure 1 As shown, the steps include:

[0077] S1. Data collection, wherein the data at least includes meteorological data, geographical data and power distribution system data of the target area;

[0078] S2, pre-processing the data collected in step S1: dividing the target area into grid units, assigning values ​​to the collected data according to the grid units, and obtaining a data set of the target area;

[0079] S3, creating a training set through the data set of step S2, processing the training set through a random forest classification model, and building a power outage time series prediction model based on the gated recurrent unit model and the random forest classification model;

[0080] S4, model training: training the gated recurrent unit model in the power outage time series prediction model using the training set processed in step S3, and training the random forest classification model in the power outage time series prediction model using the training set before processing in step S3;

[0081] S5. Model testing: Performing a spatiotemporal series prediction test on the power outage time series prediction model trained by the model in step S4.

[0082] In order to better understand the above technical solution, the exemplary embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings. These are only exemplary embodiments of the present invention. However, it should be understood that the present invention can also be implemented in various forms and is not limited to the embodiments described herein. These embodiments are intended to enable those skilled in the art to understand the present invention more clearly and thoroughly.

[0083] This embodiment takes a coastal province of a certain country as an example, set as Province B of Country A, and collects power outage monitoring data during typhoons when a total of 4 typhoons affected Province B from 2019 to 2022. The proposed power outage time series prediction model can quickly complete the spatiotemporal prediction of power outages in Province B for the next 48 hours during the remaining 1 typhoon after training on the data set of 3 typhoon events filtered by the random forest classification model. The specific embodiments are as follows:

[0084] 1. Data Collection

[0085] During typhoons, the correlation of power outages caused by meteorological factors such as wind and rain is higher than that in daily situations. Disasters such as strong winds, heavy rains and floods accompanying typhoons may cause problems such as the collapse of distribution towers and damage to transformers in the distribution system, resulting in power outages for users in the distribution area. Therefore, in order to predict power outages caused by typhoons, this embodiment collects three types of data: meteorological data, geographic data and distribution system data in the target area. Specifically, the meteorological data at least includes the maximum wind speed, hourly rainfall and typhoon path observed by the meteorological observation site; the geographic data at least includes digital elevation model data; the distribution system data at least includes the spatial distribution of distribution transformers and the power outage monitoring data of distribution transformers. The original data and sources of meteorological data, geographic data and distribution system data are shown in Table 1; the dynamic data of the four typhoons used are shown in Table 2, where the four typhoons are named Typhoon T1, Typhoon T2, Typhoon T3 and Typhoon T4 respectively.

[0086] Table 1 Original data and sources of meteorological data, geographic data and distribution system data

[0087]

[0088] Table 2 Dynamic data of the four typhoons used

[0089]

[0090] 2. Data Pre-processing

[0091] Dividing the target area into grid units can simplify the data processing and analysis process, and can also increase data samples from the spatial dimension. The goal of this embodiment is to predict the hot spots of power outages caused by typhoons, and provide a reference for the formulation of disaster response plans. The accuracy of the grid units should meet the needs of emergency management. Too small grid units contain fewer substations, and meteorological factors such as wind and rain have no significant impact on the power distribution system in the grid unit. In addition, there will be too many grid unit samples without power outages, resulting in extremely unbalanced samples; and too large grid units cannot meet the prediction space accuracy requirements. According to the collected data conditions and research objectives, in this embodiment, the target area is divided into grid units with a resolution of 4km×4km, and a total of 6142 grid units are obtained. The collected data is assigned according to the grid units to obtain a data set for the target area.

[0092] In this embodiment, the three aspects of the collected meteorological data, geographical data and power distribution system data are pre-processed, specifically including:

[0093] 1) Pre-processing of meteorological data, including:

[0094] According to the maximum wind speed and hourly rainfall of the meteorological observation station, the maximum wind speed and hourly rainfall of the center point of the grid unit are estimated by the inverse distance interpolation method, and the estimated maximum wind speed and hourly rainfall of the center point of the grid unit are assigned to the grid unit. The mathematical expression of the inverse distance interpolation method is as follows:

[0095] (1)

[0096] In formula (1), Represents the estimated value of the target interpolation point; Represents the observed value of each observation point; Indicates the distance from the observation point to the interpolation point; Represents the number of observation points closest to the interpolation point, where In this embodiment, it is preferred ; Represents a tunable parameter, In this embodiment, it is preferred .

[0097] The typhoon path is calculated to obtain the distance to the typhoon center. For the distance to the typhoon center, the distance from the center point of the grid unit to the actual typhoon center at the current time is calculated every hour, and the distance between the center point of the grid unit and the current typhoon center is assigned to the corresponding grid unit.

[0098] 2) Pre-processing of geographic data, including:

[0099] Based on the collected digital elevation model data of the target area, ArcGIS software was used to extract the slope and aspect, and the mean elevation, slope and aspect within the grid cell were calculated and assigned to the grid cell.

[0100] 3) Pre-process the distribution system data, including:

[0101] Based on the collected spatial distribution of distribution transformers in the target area, the total number of distribution transformers in the grid cells is counted and assigned to the grid cells.

[0102] Based on the collected power outage monitoring data of distribution transformers in the target area, the total number of distribution transformers that experience power outages in the grid unit every hour is counted and assigned to the grid unit hour by hour.

[0103] 3. Create a training set and build a power outage time series prediction model

[0104] A training set is created using the data set in step S2, the training set is processed using a random forest classification model, and a power outage time series prediction model is constructed based on the gated recurrent unit model and the random forest classification model.

[0105] (i) Creating a training set using the data set in step S2, and processing the training set using a random forest classification model, specifically including:

[0106] 1) Collect data on historical typhoon events and create a training set a, including:

[0107] In the data collected in this embodiment, there are 5096 grid cells with distribution transformers distributed in the target area (i.e., target grid cells), and the data of three typhoon events from 2019 to 2021 are used to create a training set a. From the spatial dimension, the number of distribution transformers, elevation, slope, and slope aspect of the target grid cell are extracted as the input of the power outage time series prediction model; from the time dimension, each target grid cell creates samples by sliding the time window every hour. In this embodiment, g=6 and h=6 are preferably selected, and the hourly time series of the maximum wind speed, hourly rainfall, distance to the typhoon center, and cumulative number of distribution transformers with power outages for the past 6 hours and the next 6 hours, totaling 12 hours, are extracted at each moment, wherein the time series of the cumulative number of distribution transformers with power outages in the past 6 hours, the time series of the maximum wind speed, hourly rainfall, and distance to the typhoon center in the past 6 hours, and the time series of the maximum wind speed, hourly rainfall, and distance to the typhoon center in the next 6 hours are the input of the power outage time series prediction model, and the time series of the cumulative number of distribution transformers with power outages in the next 6 hours is the output of the power outage time series prediction model. For each typhoon event, data from 60 hours before and after the typhoon landfall was used, and 305,760 samples were created in the sliding time window every hour. The total number of samples in the training set a for the three events was 917,280.

[0108] 2) Adjust the form of training set a to obtain training set b, including:

[0109] Training set a contains power outage samples caused by factors other than wind and rain. In order to use the random forest model to filter these samples, the form of training set a needs to be adjusted. The adjusted form is called training set b.

[0110] For all samples in training set a, four features including the number of distribution transformers, elevation, slope and aspect are used as static feature inputs of the random forest classification model; the maximum, minimum and mean values ​​of the time series of the maximum wind speed, hourly rainfall and distance to the typhoon center in the past 6 hours are extracted respectively, and the maximum, minimum and mean values ​​of the time series of the maximum wind speed, hourly rainfall and distance to the typhoon center in the next 6 hours are extracted respectively, and the maximum and mean values ​​of the time series of the cumulative number of distribution transformers with power outages in the past 6 hours are extracted. A total of 20 features are used as dynamic feature inputs of the random forest classification model. A total of 24 features are input into the random forest classification model, and the 24 features include 4 static features as inputs to the random forest classification model and 20 dynamic features as inputs to the random forest classification model.

[0111] The time series of the number of distribution transformers that will experience cumulative power outages in the next 6 hours is converted into whether the target grid unit will experience a power outage in the next 6 hours: if the cumulative power outage time series in the next 6 hours increases, the target grid unit will experience a power outage; otherwise, the target grid unit will not experience a power outage; whether the grid unit will experience a power outage in the next 6 hours is used as the output of the random forest classification model.

[0112] After processing the input and output of training set a, training set b is obtained.

[0113] 3) The random forest classification model is trained with training set b to obtain the power outage samples caused by factors other than wind and rain contained in training set a and filter them to obtain training set c:

[0114] The samples of training set b are randomly divided into two sub-training sets: training set b1 and training set b2; the random forest classification models are trained using training set b1 and training set b2 respectively, and they are named RF_F1 model and RF_F2 model respectively. The RF_F1 model is set as model one and the RF_F2 model is set as model two respectively, and then prediction is performed in another training set. For example, if RF_F1 is trained in training set b1, it is predicted in training set b2. The power outage samples predicted correctly are regarded as power outage samples caused by wind and rain factors and retained in the training set, and the samples predicted incorrectly are regarded as power outage samples caused by non-wind and rain factors and removed from the training set. The numbers of the samples in training set b1 and training set b2 after filtering are recorded, and the training set c is obtained by selecting the corresponding samples from training set a according to the numbers. The total number of samples in training set c is 864910, that is, 52370 samples are removed.

[0115] (II) Construct a power outage time series prediction model based on the gated recurrent unit model and the random forest classification model, including:

[0116] The prediction target of this embodiment is the power outage caused by the typhoon, which mainly refers to the power outage caused by strong winds, heavy rains and other wind and rain-related factors. In the actual collected distribution transformer power outage monitoring data, some power outages are not caused by wind and rain factors, and contain a large number of false alarms of power outages caused by monitoring system signals during typhoons. It is difficult to predict these power outages based on the features collected in this embodiment. These power outages are not within the scope of discussion in this embodiment. During model training, these power outage samples should be deleted, otherwise it will confuse the training of the power outage time series prediction model and reduce the accuracy of model prediction.

[0117] The random forest classification model is divided into RF_F model and RF_C model according to different functions, among which RF_F model is a general term for model one and model two. The RF_F model is set as model three and the RF_C model is set as model four respectively. Model three is configured to filter the training set to eliminate power outage samples caused by factors other than wind and rain. Model four is configured to perform output correction on the final output module of the power outage time series prediction model to ensure the correlation between the predicted power outage and wind and rain.

[0118] Model 3 and Model 4 have the same structure, both using the default hyperparameters of RandomForestClassifier in the Scikit-Learn library (n_estimators = 100, max_depth = None, min_samples_split = 2, max_features = "sqrt(n_features)"); the input and output of Model 3 and Model 4 are also the same, such as Figure 2As shown, each sample inputs the 24 features of the grid unit respectively. The 24 features mentioned here are the 24 features mentioned in "(i) creating a training set through the data set of step S2, and processing the training set through the random forest classification model", and the output is whether a power outage will occur in the grid unit in the next 6 hours.

[0119] In this embodiment, the gated recurrent unit model is recorded as a GRU_T model, and the GRU_T model includes a dynamic feature extraction module, a static feature extraction module, and a feature extraction result output module. Figure 3 As shown, N represents the number of samples, L represents the length of the time series, and C represents the number of channels (which represents the number of features when the model is input). Specifically, the dynamic feature extraction module includes at least 4 layers of gated recurrent unit layers, which are respectively set as gated recurrent unit layer 1, gated recurrent unit layer 2, gated recurrent unit layer 3 and gated recurrent unit layer 4; the static feature extraction module includes at least two layers of fully connected layers, which are respectively set as fully connected layer 1 and fully connected layer 2; the feature extraction result output module includes at least 3 layers of fully connected layers and cumulative sum functions, and the 3 layers of fully connected layers are respectively set as fully connected layer 3, fully connected layer 4 and fully connected layer 5. In addition, in this embodiment, the rectified linear unit is preferably used as the activation function in the fully connected layer. The structural parameters of the GRU_T model are shown in Table 3. Among them, the dynamic feature extraction module adopts a coding prediction structure, and inputs the time series of 4 channels, including the maximum wind speed, hourly rainfall, distance to the typhoon center and the cumulative number of distribution transformers with power outages in the past 6 hours and the next 6 hours for a total of 12 hours. Among them, the time series of the cumulative number of distribution transformers with power outages in the next 6 hours adopts 0 placeholder processing; the static feature extraction module inputs the number of distribution transformers, mean elevation, mean slope and mean slope aspect of the grid unit; the results of the static feature extraction module and the results of the dynamic feature extraction module are spliced ​​at the channel, and then input into the feature extraction result output module for the final feature extraction processing. The output layer of the feature extraction result output module is the cumulative sum function to ensure that the output is the cumulative sum. The GRU_T model finally outputs the time series of the cumulative number of distribution transformers with power outages in the next 6 hours of the grid unit.

[0120] Table 3 Structural parameters of the GRU_T model

[0121]

[0122] In this embodiment, the power outage time series prediction model is composed of the RF_C model and the GRU_T model, which is recorded as the RF_GRU_RF model and set as model five, among which the GRU_T model is used as a time series prediction module to predict the time series of the cumulative number of distribution transformers with power outages, and the RF_C model is used as the final output module to perform output correction on the time series prediction module.

[0123] 4. Training of power outage time series prediction model

[0124] The gated recurrent unit model in the power outage time series prediction model is trained using the training set processed in step S3, and the random forest classification model in the power outage time series prediction model is trained using the training set before processing in step S3. Specifically:

[0125] Training GRU_T model: GRU_T model is trained on the training set c filtered by RF_F model to learn the relationship between dynamic variables and static variables such as wind and rain in grid units and the cumulative power outage time series in the next 6 hours. Specifically, the training set c is randomly divided into training set ct and verification set cv according to a preset ratio, the GRU_T model is trained in the training set ct, and the GRU_T model is verified in the verification set cv. In this embodiment, the training set c is preferably randomly divided into training set ct and verification set cv according to 8:2, the number of samples (batch size) for each iteration is set to 64, and the Adam optimizer (learning rate is 10 -4 ) to minimize the loss function MSE. The training adopts the early stopping mechanism. When the MSE value on the validation set does not decrease after 10 iterations, the training is considered to have basically converged. In this embodiment, the GRU_T model is trained for 160 times and the training is stopped. It takes 1.4 hours. The iteration round with the minimum loss function value of the validation set is taken as the training result of the GRU_T model. This model is named GRU_T. Among them, the expression of MSE is as follows:

[0126] (2)

[0127] In formula (2), represents the number of samples in the training set d, represents the prediction time step, represents the observed value, Represents the predicted value.

[0128] Training RF_C model: The RF_C model is trained on training set b to ensure that as many dynamic variables as possible such as wind and rain in the grid unit and the relationship between static variables and whether there will be a power outage in the next 6 hours are learned.

[0129] In order to compare the model performance and prove the performance superiority of the RF_GRU_RF power outage time series prediction model composed of the RF_C model and the GRU_T model proposed in the present invention, three other power outage time series prediction models (GRU model, RF_GRU model, GRU_RF model) are also set in this embodiment for comparative analysis. Figure 4As shown. The GRU model is set to model six, and the model six only contains the gated recurrent unit model (GRU_T model), and the gated recurrent unit model is trained on the training set a that has not been filtered by the RF_F model (model three); the RF_GRU model is set to model seven, and the model seven also only contains the gated recurrent unit model (GRU_T model), and the gated recurrent unit model is trained on the training set c that has been filtered by the RF_F model (model three); the GRU_RF model is set to model eight, and the model eight is composed of the gated recurrent unit model (GRU_T model) and the RF_C model (model four), and the gated recurrent unit model is trained on the training set a that has not been filtered by the RF_F model (model three).

[0130] 5. Testing of the power outage time series prediction model

[0131] The RF_GRU_RF model is essentially a regression model. After training, it is necessary to conduct performance tests in typhoon events outside the training set. The mean absolute error (MAE) and root mean square error (RMSE) are used to evaluate the prediction performance of the power outage time series prediction model. The calculation formula of the indicators is as follows:

[0132] (3)

[0133] (4)

[0134] In the above formula, represents the number of samples in the training set d, represents the prediction time step, represents the observed value, Represents the predicted value.

[0135] This embodiment tests the prediction performance of the power outage time series prediction model in the typhoon T4 event. Typhoon T4 landed in City C, Province B at around 20:30 on September 14, 2022. This embodiment sets the power outage time series prediction model to predict the spatial and temporal distribution of the distribution transformers with cumulative power outages in the province in the next 48 hours at 00:00 on September 14. The power outage time series prediction model inputs the meteorological observation time series of the static characteristics of the grid unit, the maximum wind speed in the next 48 hours, the hourly rainfall, and the distance to the typhoon center. Each prediction is 6 hours, and a total of 8 iterations are performed, of which the RF_GRU_RF model takes 2.6 seconds.

[0136] 6. Test Results Evaluation

[0137] 1. Evaluation of the prediction process

[0138] Figure 5 is the spatial and temporal distribution diagram of the cumulative power outage distribution transformers actually observed, where: Figure 5(a) in the figure represents the spatial and temporal distribution diagram of the cumulative power outage distribution transformers at 09:00 on September 14, 2022, as actually observed. Figure 5 (b) in the figure represents the spatial and temporal distribution diagram of the cumulative power outage distribution transformers at 21:00 on September 14, 2022. Figure 5 (c) in the figure represents the spatial and temporal distribution diagram of the cumulative power outage distribution transformers at 09:00 on September 15, 2022, as actually observed. Figure 5 (d) in the figure represents the spatial and temporal distribution map of the cumulative power outage distribution transformers at 21:00 on September 15, 2022 as actually observed. Figure 6 is the spatiotemporal distribution diagram of the cumulative power outage distribution transformers predicted by model 5, where Figure 6 (a) in the figure represents the spatiotemporal distribution diagram of the cumulative power outage distribution transformers at 09:00 on September 14, 2022 predicted by model 5. Figure 6 (b) in the figure represents the spatiotemporal distribution diagram of the cumulative power outage distribution transformers at 21:00 on September 14, 2022 predicted by model 5. Figure 6 (c) in the figure represents the spatiotemporal distribution diagram of the cumulative power outage distribution transformers at 09:00 on September 15, 2022 predicted by model 5. Figure 6 (d) in the figure represents the spatial and temporal distribution of the cumulative power outage distribution transformers at 21:00 on September 15, 2022 predicted by Model 5. It can be seen that the power outage hotspots predicted by Model 5 are basically consistent with the observations. Only the development of the more sporadic power outage hotspots far away from the typhoon landing site was not captured by Model 5.

[0139] from Figure 7It can be seen that the curve of the number of distribution transformers with cumulative power outages in the target area mainly increases significantly 6 hours before and after the typhoon lands, and the number of distribution transformers with power outages almost does not increase after the typhoon leaves. Among the four power outage time series prediction models, model 6 predicts that the number of distribution transformers with cumulative power outages is generally higher, while model 5 predicts that the number of distribution transformers with cumulative power outages is generally lower. The curve of the number of distribution transformers with cumulative power outages in the province predicted by model 5 is closer to the power outage observation. Comparing Model 6 and Model 7, it can be seen that after the training set of the random forest classification model is filtered, the power outage time series prediction model predicts significantly fewer power outages; comparing Model 6 and Model 8, it can be seen that after the output correction of the random forest classification model, the power outage time series prediction model predicts fewer power outages, and the predictions of Model 8 and Model 7 are similar in the early stage, but in the later period after the typhoon leaves, the cumulative power outages predicted by Model 7 are still increasing, and the cumulative power outages predicted by Model 8 have not increased, which is more in line with the actual situation. This phenomenon can also be observed from the prediction curve of Model 5. Therefore, it can be obtained that the output correction of the RF_C model (Model 4) can make the model prediction more in line with the actual situation as a whole. After the typhoon is far away, the number of distribution transformers with power outages is almost no longer predicted to increase. From the predictions of Model 7 and Model 5, it can be seen that the power outage time series prediction model trained on the basis of the training set filtered by the RF_F model (Model 3) predicts more power outages, which not only includes power outage samples caused by wind and rain factors, but also needs to be corrected by the output of the RF_C model (Model 4) to output power outages that are more directly related to the typhoon.

[0140] from Figure 8 and Fig. 9 It can be seen that as the prediction time increases, the error of the power outage time series prediction model is gradually increasing. From the perspective of MAE, the prediction value of model six is ​​the largest and does not converge, while the prediction value of model five is the smallest and converges; from the perspective of RMSE, the prediction value of model six is ​​the largest and does not converge, and the prediction values ​​of model five and model seven are not much different, both converge, but because the RMSE of model five is smaller in the last 6 hours, the prediction of model five is considered to be the smallest. Therefore, from the MAE and RMSE process curves, it can be seen that the overall prediction error of model five is smaller than that of the other three models (i.e., model six, model eight, and model seven). In addition, the evaluation indicators MAE and RMSE of model five remained basically unchanged 6 hours after the typhoon landed, which shows that model five has recognized that the time of large-scale power outages has passed and almost no longer predicts that new power outages will occur.

[0141] Fig.10 , Fig.11 , Fig.12The box plot distributions of the three dynamic variables, namely the maximum wind speed, average rainfall and minimum distance to the typhoon center, of the samples without power outages, the samples with power outages, the samples among the power outage samples correctly predicted by model five and the samples among the power outage samples incorrectly predicted by model five in the whole prediction process of all target grid units respectively show that the maximum wind speed and average rainfall of the power outage samples as a whole are greater than those of the samples without power outages, and the minimum distance to the typhoon center is overall less than that of the samples without power outages; among the power outage samples, the power outage samples that can be predicted by model five have greater maximum wind speed and average rainfall as a whole, and smaller minimum distance to the typhoon center than those that cannot be predicted, which indicates that model five tends to predict power outages under conditions of strong winds and heavy rains during a typhoon and close to the typhoon, while power outages with less correlation with these factors are difficult to be predicted, which is consistent with the objectives of the present invention.

[0142] 2. Final result evaluation

[0143] The cumulative power outage prediction results of the power outage time series prediction model in the 48th hour represent the total results of the power outage caused by the typhoon in the next 48 hours, which can be directly provided to decision makers for the power outage hotspots in the next 48 hours. Fig.13 As shown in FIG. 1 , the prediction results of the last hour predicted by model six, model eight, model seven, and model five are respectively shown in FIG. 1 , wherein: Fig.13 (a) is the prediction result of the last hour through model 6. Fig.13 (b) is the prediction result of the last hour predicted by model 8. Fig.13 (c) in the figure is the prediction result of the last hour through model 7. Fig.13 (d) in the figure is the last hour’s forecast result predicted by model 5. Fig.13 It can be seen that the actual observed power outage areas are mainly concentrated in the typhoon landing point, and there are more sporadic power outage areas in areas far from the typhoon landing. The cumulative number of distribution transformers with power outages in these areas is almost less than 15. The four power outage time series prediction models predict that the key power outage areas are mainly concentrated in the coastal areas close to the path of Typhoon T4. The key power outage areas include the areas where the typhoon landed. Among them, Model 6 predicts the largest area and number of distribution transformers with cumulative power outages, and there are some sporadic power outage area predictions in areas far from the landing of Typhoon T4, but these predictions are not very consistent with actual observations; Model 5 predicts the least cumulative area and number of distribution transformers with power outages in the 48th hour. The predicted key power outage areas are basically consistent with the observed results, except for a small number of units in the lower right corner.

[0144] As can be seen from Table 4, the three evaluation index values ​​of model 6 are the largest. After filtering the training set and output correction of the random forest classification model, the evaluation indexes are reduced. The MSE and RMSE of model 7 after training set filtering are lower, and the MAE of model 8 after output correction is lower, which shows that the power outage time series prediction model with only output correction has more outlier predictions than the power outage time series prediction model with only training set filtering. The three evaluation index values ​​of model 5, which performs training set filtering and output correction at the same time, are significantly lower than those of model 7 and model 8, which shows that the use of training set filtering and output correction at the same time can better reduce the model prediction error. The three evaluation index values ​​of the last hour prediction of model 5 are the lowest, which shows that the final prediction of model 5 is more consistent with the final observation results.

[0145] Table 4 MAE, MSE, and RMSE values ​​of the 48th hour prediction results of Model 5, Model 6, Model 7, and Model 8

[0146]

[0147] from Fig.14 It can be observed that the training set filtering and output correction of the random forest classification model can significantly reduce the overall error of the grid cells in the target area, and the output correction is more effective in reducing the relative error. By comparing Model 8 and Model 5, it can be found that the overall distribution of absolute error and relative error between the two is not much different. The relative error distribution of Model 8 is slightly larger than that of Model 5. Therefore, it can be considered that the final prediction error of Model 5 is smaller, that is, the final prediction error of the power outage time series prediction model proposed by the present invention is smaller.

[0148] The above only describes the basic principles and preferred embodiments of the present invention. Those skilled in the art may make many changes and improvements based on the above description, and these changes and improvements should fall within the protection scope of the present invention.

Claims

1. A method for predicting the spatiotemporal sequence of power outages during typhoons, characterized in that: The steps include: S1. Data collection, wherein the data at least includes meteorological data, geographical data and power distribution system data of the target area; S2, pre-processing the data collected in step S1: dividing the target area into grid units, assigning values ​​to the collected data according to the grid units, and obtaining a data set of the target area; S3. Create a training set using the data set of step S2, process the training set using a random forest classification model, and build a power outage time series prediction model based on a gated recurrent unit model and the random forest classification model; specifically: 1) Collect data on historical typhoon events and create a training set a, including: Collect data of N historical typhoon events to create a training set a, where N ≥ 3 and N is an integer, and the grid cells containing distribution transformers are set as target grid cells; From the spatial dimension, the number, elevation, slope and aspect of distribution transformers of the target grid unit are extracted as the input of the power outage time series prediction model; From the time dimension, a sample is created by sliding the time window every hour for each target grid unit, and the hourly time series of the maximum wind speed, hourly rainfall, distance to the typhoon center, and the cumulative number of distribution transformers with power outages for the past g hours and the next h hours (a total of g+h hours) are extracted at each moment, where the value range of g is [3,12], the value range of h is [3,12], and both g and h are integers. The time series of the cumulative number of distribution transformers with power outages in the past g hours, the time series of the maximum wind speed, hourly rainfall, and distance to the typhoon center in the past g hours, and the time series of the maximum wind speed, hourly rainfall, and distance to the typhoon center in the next h hours are the inputs of the power outage time series prediction model, and the time series of the cumulative number of distribution transformers with power outages in the next h hours is the output of the power outage time series prediction model. 2) Adjust the form of training set a to obtain training set b, including: For all samples in training set a, four features, namely, the number of distribution transformers, elevation, slope, and slope direction, are used as static feature inputs of the random forest classification model; the maximum value, minimum value, and mean value of the time series of the maximum wind speed, hourly rainfall, and distance to the typhoon center in the past g hours are extracted respectively, and the maximum value, minimum value, and mean value of the time series of the maximum wind speed, hourly rainfall, and distance to the typhoon center in the future h hours are extracted respectively, and the maximum value and mean value of the time series of the number of distribution transformers with cumulative power outages in the past g hours are extracted. A total of 20 features are used as dynamic feature inputs of the random forest classification model. A total of 24 features are input into the random forest classification model, and the 24 features include 4 static features used as inputs to the random forest classification model and 20 dynamic features used as inputs to the random forest classification model; The time series of the number of distribution transformers with cumulative power outages in the next h hours is converted into whether the target grid unit will have a power outage in the next h hours: if the cumulative power outage time series in the next h hours increases, the target grid unit will have a power outage; otherwise, the target grid unit will not have a power outage; whether the grid unit will have a power outage in the next h hours is used as the output of the random forest classification model; After processing the input and output of training set a, training set b is obtained; 3) The random forest classification model is trained with training set b to obtain the power outage samples caused by factors other than wind and rain contained in training set a and filter them to obtain training set c, which specifically includes: The samples of training set b are randomly divided into two training sets: training set b1 and training set b2; the random forest classification models are trained using training set b1 and training set b2 respectively, and are named RF_F1 model and RF_F2 model respectively, and the RF_F1 model is set as model one and the RF_F2 model is set as model two respectively, and then prediction is performed in another training set, and the power outage samples predicted correctly are regarded as power outage samples caused by wind and rain factors and retained in the training set, and the power outage samples predicted incorrectly are regarded as power outage samples caused by non-wind and rain factors and are removed from the training set; the numbers of the samples in the filtered training set b1 and training set b2 are recorded, and the training set c is obtained by selecting the corresponding samples from the training set a according to the numbers; The random forest classification model is divided into RF_F model and RF_C model according to different functions, where RF_F model is a general term for model one and model two. The RF_F model is set as model three and the RF_C model is set as model four. Model three is configured to filter the training set to remove power outage samples caused by non-wind and rain factors. Model four is configured to perform output correction on the final output module of the power outage time series prediction model. The structure, input and output of model three and model four are the same. Each sample is input into the 24 features of the grid unit, and the output is whether the grid unit will have a power outage in the next h hours. The gated recurrent unit model is recorded as a GRU_T model, and the GRU_T model includes a dynamic feature extraction module, a static feature extraction module, and a feature extraction result output module; wherein the dynamic feature extraction module adopts a coding prediction structure, and inputs a time series of four channels, namely, the maximum wind speed, hourly rainfall, distance to the typhoon center, and the cumulative number of distribution transformers with power outages in the past g hours and the next h hours, totaling g+h hours, of the grid unit, wherein the time series of the cumulative number of distribution transformers with power outages in the next h hours adopts 0 placeholder processing; the static feature extraction module inputs the number of distribution transformers, the mean elevation, the mean slope, and the mean slope aspect of the grid unit; the results of the static feature extraction module and the results of the dynamic feature extraction module are spliced ​​at the channel, and then input into the feature extraction result output module for final feature extraction processing, and the GRU_T model finally outputs the time series of the cumulative number of distribution transformers with power outages in the next h hours of the grid unit; The power outage time series prediction model is composed of the RF_C model and the GRU_T model, which is denoted as the RF_GRU_RF model and set as model five. The GRU_T model is used as the time series prediction module to predict the time series of the cumulative number of distribution transformers with power outages. The RF_C model is used as the final output module to correct the output of the time series prediction module. S4, model training: training the gated recurrent unit model in the power outage time series prediction model using the training set processed in step S3, and training the random forest classification model in the power outage time series prediction model using the training set before processing in step S3; S5. Model testing: Performing a spatiotemporal series prediction test on the power outage time series prediction model trained by the model in step S4.

2. The method for predicting power outage time-space series during typhoon according to claim 1, characterized in that: In step S1, the meteorological data includes at least the maximum wind speed, hourly rainfall and typhoon path observed at the meteorological observation site; the geographic data includes at least digital elevation model data; and the distribution system data includes at least the spatial distribution of distribution transformers and power outage monitoring data of distribution transformers.

3. The method for predicting the time-space sequence of power outages during typhoons according to claim 2, characterized in that: Step S2 specifically includes: 1) Pre-processing the meteorological data collected in step S1, specifically including: According to the maximum wind speed and hourly rainfall of the meteorological observation station, the maximum wind speed and hourly rainfall of the center point of the grid unit are estimated by using the inverse distance interpolation method, and the estimated maximum wind speed and hourly rainfall of the center point of the grid unit are assigned to the corresponding grid unit; The typhoon path is calculated to obtain the distance to the typhoon center. For the distance to the typhoon center, the distance from the center point of the grid unit to the current measured typhoon center is calculated every hour, and the distance between the center point of the grid unit and the current typhoon center is assigned to the corresponding grid unit; 2) Pre-processing the geographic data collected in step S1, specifically including: Based on the collected digital elevation model data of the target area, ArcGIS software was used to extract the slope and aspect, and the mean elevation, slope and aspect values ​​within the grid cells were calculated and assigned to the grid cells. 3) Pre-processing the power distribution system data collected in step S1, specifically including: Based on the collected spatial distribution of distribution transformers in the target area, the total number of distribution transformers in the grid unit is counted and assigned to the grid unit; Based on the collected power outage monitoring data of the distribution transformers in the target area, the total number of distribution transformers with power outages in the grid unit every hour is counted and assigned to the grid unit hour by hour.

4. The method for predicting power outage time-space sequence during typhoon according to claim 3, characterized in that: In step S3, a training set is created using the data set in step S2, and the training set is processed using a random forest classification model, including: Collect data on historical typhoon events and create a training set a; Adjust the form of training set a to obtain training set b; The random forest classification model is trained through the training set b to obtain the power outage samples caused by non-wind and rain factors contained in the training set a and filter them to obtain the training set c.

5. The method for predicting the time-space sequence of power outages during typhoons according to claim 4, characterized in that: Step S4 specifically includes: Training of RF_GRU_RF model: Among them, RF_C model is trained in training set b; GRU_T model is trained on training set c after filtering by RF_F model. Specifically, training set c is randomly divided into training set ct and verification set cv according to a preset ratio, GRU_T model is trained in training set ct, and verified in verification set cv. The training adopts early stopping mechanism. When the loss function no longer decreases in verification set cv, the GRU_T model is considered to have been trained.

6. The method for predicting the time-space sequence of power outages during typhoons according to any one of claims 1 to 5, characterized in that: Step S5 specifically includes: The power outage time series prediction model is tested in the training set of typhoon events outside the training set created in step S3. Let the training set used to test the power outage time series prediction model outside the training set created in step S3 be d. The mean absolute error MAE and root mean square error RMSE are selected to evaluate the prediction performance of the power outage time series prediction model. The calculation formula of the indicators is as follows: ， ， In the above formula, represents the number of samples in the training set d, represents the prediction time step, represents the observed value, Represents the predicted value.

Images