A tropical cyclone rapid intensification prediction method based on random forest and transfer learning

Abstract

This invention discloses a method for predicting rapid intensification of tropical cyclones based on random forest and transfer learning, comprising: S1, determining TC data and environmental factor data for input to the classification model; S2, performing data preprocessing: preprocessing the TC data and environmental factor data; S3, constructing the observation dataset required for training the model; S4, constructing a rapid intensification prediction model for tropical cyclones based on the random forest classification algorithm on the observation dataset; S5, establishing a real-time prediction system for rapid intensification of tropical cyclones using transfer learning based on data from the Global Forecasting System (GFS) model. According to this invention, a prediction model with high accuracy and low false alarm rate for rapid intensification events of tropical cyclones in the Northwest Pacific is established, which can improve the level of tropical cyclone intensity prediction in the Northwest Pacific and has good application prospects.

Description

Technical Field

[0001] This invention relates to the technical field of rapid intensification forecasting of tropical cyclones, and particularly to a method for rapid intensification forecasting of tropical cyclones based on random forest and transfer learning. Background Technology

[0002] Tropical cyclones (TCs) are among the most destructive weather systems in the world, and TC disasters are closely related to their intensity. Over the past 20 years, progress in TC intensity forecasting has been slow. The International Meteorological Organization (IMO) has pointed out that rapid TC intensification is one of the main reasons for large errors in TC intensity forecasts. Improving TC rapid intensification forecasts is one of the top priorities for major global tropical cyclone forecasting centers. Current forecasting methods for rapid TC intensification mainly include: the "rapid intensification index" which relies on environmental factors; satellite detection methods that rely on core shape; probabilistic forecasting models improved based on the "rapid intensification index"; and 24-hour lead-through rapid intensification forecasting methods based on machine learning. However, the overall 24-hour forecast performance of these methods is not optimistic; for example, it is only 21% for the North Atlantic and 31% for the Northeast Pacific. Although some methods can improve the 24-hour forecast accuracy to around 60%, the accuracy is even lower for longer forecast leads, especially those exceeding 48 hours, and the false alarm rate is higher.

[0003] Chinese patents CN111695473A and CN113642475A disclose methods for estimating tropical cyclone intensity based on long short-term memory (LSTM) network models and convolutional neural network (CNN) models, respectively. Both methods utilize deep learning technology to analyze and learn from a large amount of infrared satellite cloud imagery, extracting complex features related to typhoon intensity, and then performing intensity estimation. These methods reduce the subjectivity of manually defined feature factors and improve the accuracy of intensity estimation.

[0004] Chinese patent CN104932035A discloses a method and system for forecasting tropical cyclone intensity. This method is based on historical data analyzed statistically, selects predictive factors and calculates relevant parameter data, establishes a multiple regression model, inputs relevant environmental factor data from historical cyclone (TC) data into the regression model for training, and obtains the TC intensity forecast equation based on the training results, thereby forecasting TC intensity. Chinese patent CN111461427A discloses a method and system for generating tropical cyclone intensity forecast information. This method first obtains data to be processed according to different data dimensions, then preprocesses the data to obtain initial data, and finally uses an XGBoost model and a classification regression tree to train the initial data to obtain tropical cyclone intensity forecast information. Furthermore, Chinese patent CN108983320A discloses a numerical weather prediction-artificial intelligence coupled prediction method for extreme wind speeds of coastal typhoons. This method trains a deep neural network using measured data and a numerical weather prediction-based model, then downscales the calculation using a mesoscale meteorological data model to predict wind speed and frictional wind speed over the next 24 hours. Combined with an optimized peak factor, it obtains the extreme wind speed within 24 hours. This method overcomes the limitations of statistical prediction. The aforementioned patents use different methods to reduce human subjectivity and overcome the shortcomings of traditional statistical methods from different perspectives, improving the accuracy of tropical cyclone intensity prediction and its general applicability. However, it does not specifically address the forecasting of rapid intensification processes of tropical cyclones. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a rapid intensification forecasting method for tropical cyclones based on random forests and transfer learning. A forecasting model with high accuracy and low false alarm rate for rapid intensification events of tropical cyclones in the Northwest Pacific has been established, which can improve the level of tropical cyclone intensity forecasting in the Northwest Pacific and has good application prospects. To achieve the above-mentioned objectives and other advantages of the present invention, a rapid intensification forecasting method for tropical cyclones based on random forests and transfer learning is provided, comprising:

[0006] S1. Determine the TC data and environmental factor data to be used as input to the classification model;

[0007] S2. Perform data preprocessing: Preprocess the TC data and environmental factor data;

[0008] S3, The observation dataset required to build the training model;

[0009] S4. Construct a rapid intensification forecast model for tropical cyclones based on the random forest classification algorithm on the observation dataset;

[0010] S5. Based on data from the Global Forecast System (GFS) model, establish a real-time forecasting system for rapid intensification of tropical cyclones using migrating forests.

[0011] Preferably, the TC data in step S1 is the TC best path dataset for 6 hours from the Shanghai Typhoon Research Institute of the China Meteorological Administration, including the TC center location, the maximum 2-minute average continuous surface wind speed and the minimum central sea level pressure.

[0012] The environmental factor data used were ERA-Interim reanalysis data, which employed atmospheric and oceanic variables, including sea surface temperature, wind field, relative humidity, eddy current, and divergence.

[0013] Preferably, in step S2, by screening TC data in the Northwest Pacific and using ERA-Interim reanalysis data, an influencing factor related to the rapid intensification of TC is calculated and synthesized as a forecasting factor within a fixed radius range of the TC center.

[0014] Preferably, step S3 includes creating classification labels based on the definition of rapid augmentation, and constructing input training datasets corresponding to different forecast lead times based on different forecast lead times.

[0015] Preferably, step S4 further includes the following steps:

[0016] S41. Determine the training set samples and test set samples;

[0017] S42. Balance the number of fast-enhancement events and non-fast-enhancement events in the training set samples;

[0018] S43. Adjust the model parameters to obtain the optimal parameters;

[0019] S44. Conduct an importance analysis of forecast factors;

[0020] S45. Establish a forecast model.

[0021] Preferably, step S5 further includes the following steps:

[0022] S51. Determine the source domain dataset X t and target domain dataset X a The dataset constructed based on TC data and reanalysis data will be used as the source domain dataset, and the real-time GFS data will be used as the target domain dataset.

[0023] S52, Equal amounts from two domains X t and X a S samples are randomly selected from the data to form a subset X', which is used to train the decision tree;

[0024] S53. In the subset, the decision tree is trained using covariate loss, with each training sample weighted by covariate loss.

[0025] S54. Add the decision tree obtained from the previous training step to the candidate set of the migration forest;

[0026] S55. Repeat S52 to S54 until a large number of candidate decision trees for the migration forest are obtained;

[0027] S56. Select the latter half of the candidate decision tree as the migration forest.

[0028] S57 starts by inputting an unknown sample x into each decision tree generated by the algorithm. At each node of each decision tree, a classification function is used to classify x into the left or right branch. The class probability is output and saved at the leaf nodes. The average value of the output values ​​of the second half of the decision trees of the migration forest is calculated.

[0029] Compared with existing technologies, the beneficial effects of this invention are: to improve the current slow-progressing TC intensity forecasting level, enhance the forecasting of rapid TC intensification events in the Northwest Pacific, and fill the gap in current domestic operational forecasting of rapid TC intensification. Furthermore, compared with conventional meteorological numerical weather prediction models, this scheme features faster computation speed, lower computational requirements, and better prediction results. Attached Figure Description

[0030] Figure 1 The flowchart shows the real-time forecasting system for rapid intensification of tropical cyclones according to the present invention, which is based on the method of rapid intensification forecasting of tropical cyclones using random forest and transfer learning. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Reference Figure 1 A method for rapid intensification forecasting of tropical cyclones based on random forest and transfer learning includes: S1, determining TC data and environmental factor data to be used as input to the classification model;

[0033] S2. Perform data preprocessing: Preprocess the TC data and environmental factor data;

[0034] S3, The observation dataset required to build the training model;

[0035] S4. Construct a rapid intensification forecast model for tropical cyclones based on the random forest classification algorithm on the observation dataset;

[0036] S5. Based on data from the Global Forecast System (GFS) model, establish a real-time forecasting system for rapid intensification of tropical cyclones using migrating forests.

[0037] Furthermore, in step S1, the TC data is the 6-hourly TC best path dataset from the Shanghai Typhoon Research Institute of the China Meteorological Administration, including the TC center location, maximum 2-minute average continuous surface wind speed, and minimum central sea level pressure.

[0038] The environmental factor data used were ERA-Interim reanalysis data, which employed atmospheric and oceanic variables, including sea surface temperature, wind field, relative humidity, eddy current, and divergence.

[0039] Furthermore, in step S2, the TC data in the Northwest Pacific is filtered, and the intensity change rate (ΔV) of TC within 24 hours is calculated based on the filtered data. 24 ), and use all TCΔV 24 The 95th percentile of the event further defines ΔV. 24 ≥15ms -1 To quickly enhance events and ΔV 24 <15ms -1 As a non-rapidly enhanced event;

[0040] Based on ERA-Interim reanalysis data, influencing factors related to the rapid intensification of the TC were calculated and synthesized within a fixed radius of the TC center as forecast factors. These factors include: the maximum wind speed (Vmax), longitude (Lon), and latitude (Lat) of the TC center at different lead times from the initial moment of rapid intensification (the initial moment of rapid intensification is denoted as 0, and n hours in advance [n=6,12,18,24…96] is denoted as -n hours). The vertical wind shear (VWS), tropospheric outflow temperature (T0), sea surface temperature (SST), 200 hPa divergence (DIV200), 600 hPa relative humidity (RH600), 850 hPa vorticity (VOR850), and the maximum potential intensity (MPI) of the TC are calculated.

[0041] Furthermore, step S3 includes creating classification labels based on the definition of rapid reinforcement, i.e., ΔV 24 ≥15ms -1 The rapid enhancement event is marked as "1", ΔV 24 <15ms -1 Non-fast-enhancement events are marked as "0";

[0042] Based on different forecast lead times, input training datasets corresponding to different forecast lead times are constructed. For example, if the forecast at 00:00 on August 1, 2018 predicts whether a rapid intensification will occur in the next 24 hours at 12:00 on August 2, the forecast lead time is 36 hours. Then, the classification labels of the forecast factor data at 00:00 corresponding to the 36-hour forecast lead time are used as the training dataset.

[0043] Furthermore, step S4 also includes the following steps:

[0044] S41. Determine the training set samples and test set samples. In this invention, the dataset from 1980 to 2015 is used as the training set, and the dataset from 2016 to 2018 is used as the test set.

[0045] S42. Balance the number of fast-enhancement events and non-fast-enhancement events in the training set samples, and randomly select an equal number of the two types of event samples in the training set as the final input training set.

[0046] S43. Adjust the model parameters to obtain the optimal parameters, and determine the optimal number of decision trees based on the out-of-bag mean squared error (OOB-MSE). In this algorithm, when there are k decision trees in an ensemble, k samples are drawn with replacement from the training dataset to train each of the k decision trees. If the training set size is N (i.e., there are N sets of data), then the sample size drawn in each of the k rounds is also N. Therefore, the probability that a single data point is not drawn in each round is... In each round, approximately one-third of the training data is not selected; this data is called out-of-bag (OOB) data. This data has its own value; it does not participate in the model's training process and can replace cross-validation to calculate the generalization error, thus evaluating the model's performance. This significantly reduces computation, a major advantage of this algorithm. After each tree is trained, there are k weak learning machines, which are then combined into a strong learning machine. For classification problems, the vote of the majority of decision trees is used as the final result. In this model, the optimal number of decision trees is determined to be 1000.

[0047] S44. After constructing the random forest classification model, the value of the m-th predictor in the training set is changed. The mean squared error of the out-of-bag data after the perturbation is calculated. The increase in the mean squared error of the out-of-bag data after the perturbation is obtained for each decision tree. This increase is then averaged across all decision trees. The final value is used as an important indicator to measure the importance of the m-th predictor. For each predictor, the larger this value is, the more important the predictor is.

[0048] S45. Establish a forecasting model using the RandomForestClassifier function from the sklearn.ensemble open-source library in Python to complete the classification forecast. In this invention, forecasts from different timeframes were performed 100 times and then aggregated.

[0049] Furthermore, step S5 also includes the following steps:

[0050] S51. Determine the source domain dataset X t and target domain dataset X a The dataset constructed based on TC data and reanalysis data will be used as the source domain dataset, and the real-time GFS data will be used as the target domain dataset.

[0051] S52, Equal amounts from two domains X t and X a S samples are randomly selected from the data to form a subset X', which is used to train the decision tree;

[0052] S53. In the subset, the decision tree is trained using covariate loss. Each training sample is weighted by covariate loss, which is defined as: Where P a (C k |X j ) is the class label probability P of the random forest previously trained using source domain samples. t (C k |X j ) represents the probability of the class label in the transfer forest built using two domains. After constructing the decision tree, it is re-estimated using the formula, and then the transfer forest is updated to construct the next decision tree. Here, some source domain samples that are far from the target domain are given a particularly small weight. The role of covariate loss is to use it to divide the training samples during the training of the transfer forest to adapt to the target domain and prevent performance degradation;

[0053] S54. Add the decision tree obtained from the previous training step to the candidate set of the migration forest. Update the covariate loss using the migration forest and the random forest previously trained using the source domain.

[0054] S55. Repeat S52 to S54 until a large number of candidate decision trees for the migration forest are obtained;

[0055] S56. The latter half of the candidate decision tree is selected as the transfer forest. The covariate loss parameter of transfer learning changes gradually during training. When constructing the first tree, the data distribution cannot be fully estimated, so these values ​​are unreliable. The algorithm selects the latter half of the decision tree, which is reliable.

[0056] Furthermore, in step S6, the migration forest classification algorithm for classifying and predicting the test samples inputs an unknown sample x into each decision tree generated by the algorithm. At each node of each decision tree, a classification function is used to classify x into the left or right branch. The class probability is output and saved at the leaf nodes. The average value of the output values ​​of the latter half of the decision trees in the migration forest is calculated, as shown in the formula. Finally, the category with the highest probability is determined: y = argmaxP(c|x).

[0057] The number of devices and processing scale described herein are for the purpose of simplifying the description of the invention, and applications, modifications and variations thereof will be apparent to those skilled in the art.

[0058] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A method for rapid intensification forecasting of tropical cyclones based on random forest and transfer learning, characterized in that, Includes the following steps: S1. Determine the TC data and environmental factor data to be used as input to the classification model; S2. Perform data preprocessing: Preprocess the TC data and environmental factor data, including standardization, missing value imputation, and outlier removal. S3, The observation dataset required to build the training model; S4. Construct a rapid intensification forecast model for tropical cyclones based on the random forest classification algorithm on the observation dataset; step S4 also includes the following steps: S41. Determine the training set samples and test set samples; S42. Balance the number of fast-enhancement events and non-fast-enhancement events in the training set samples, and randomly select an equal number of the two types of event samples in the training set as the final input training set. S43. Adjust the model parameters to obtain the optimal parameters; determine the optimal number of decision trees based on the mean squared error of the out-of-bag data. When there are k decision trees in an ensemble, k samples are drawn with replacement from the training dataset to train each of the k decision trees. If the training set size is N, then the sample size drawn in each of the k rounds is also N. Therefore, the probability that a single data point is not drawn in each round is... , Once each tree has been trained, there will be k weak learning machines, which will then be combined into a strong learning machine. S44. Conduct prediction factor importance analysis; After constructing the random forest classification model, change the value of the m-th prediction factor in the training set data, calculate the mean square error of the out-of-bag data of the dataset after perturbation, obtain the increase value of the mean square error of the out-of-bag data after perturbation for each decision tree, and then average it over all decision trees. The final value is used as an important indicator to measure the importance of the m-th prediction factor. S45. Establish a prediction model; use the RandomForestClassifier function from Python's sklearn.ensemble open-source library to complete the classification prediction; S5. Based on data from the Global Forecasting System (GFS) model, establish a real-time forecasting system for rapid intensification of tropical cyclones using migrating forests; step S5 also includes the following steps: S51. Determine the source domain dataset. and target domain dataset The dataset constructed based on TC data and reanalysis data will be used as the source domain dataset, and the real-time GFS data will be used as the target domain dataset. S52, Equal amounts from two domains S samples are randomly selected from the data to form a subset X', which is used to train the decision tree; S53. In the subset, the decision tree is trained using covariate loss l, with each training sample weighted by covariate loss l. S54. Add the decision tree obtained from the previous training step to the candidate set of the migration forest; S55. Repeat S52 to S54 until a large number of candidate decision trees for the migration forest are obtained; S56. Select the latter half of the candidate decision tree as the migration forest. S57. Initially, input an unknown sample x into each decision tree generated by the algorithm. At each node of each decision tree, use a classification function to classify x into the left or right branch. Output and save the class probability at the leaf node. Calculate the average value of the output values ​​of the second half of the decision tree of the migration forest.

2. The rapid intensification forecasting method for tropical cyclones based on random forest and transfer learning as described in claim 1, characterized in that, In step S1, the TC data is the 6-hourly TC best track dataset from the Shanghai Typhoon Research Institute of the China Meteorological Administration, including the TC center location, maximum 2-minute average continuous surface wind speed, and minimum central sea level pressure. The environmental factor data used were ERA-Interim reanalysis data, which employed atmospheric and oceanic variables, including sea surface temperature, wind field, relative humidity, eddy current, and divergence.

3. The rapid intensification forecasting method for tropical cyclones based on random forest and transfer learning as described in claim 1, characterized in that, In step S2, by screening TC data in the Northwest Pacific and using ERA-Interim reanalysis data, an influencing factor related to the rapid intensification of TC was calculated and synthesized as a forecasting factor within a fixed radius range of the TC center.

4. The rapid intensification forecasting method for tropical cyclones based on random forest and transfer learning as described in claim 1, characterized in that, Step S3 includes creating classification labels based on the definition of rapid augmentation, and constructing input training datasets corresponding to different forecast lead times based on different forecast lead times.

Images