CN115099322A - Dynamic classification method of time-varying sectors for traffic state - Google Patents

Abstract

The invention discloses a dynamic classification method of time-varying sectors for traffic states, which comprises the following steps: collecting historical operating data of sectors, calculating dynamic operating characteristic variable data and performing data preprocessing to obtain clustering input characteristic vectors; clustering analysis is carried out on sector traffic running states; constructing a dynamic classification model of a time-varying sector of a neural network based on a multilayer perceptron; and carrying out dynamic classification of the time-varying sectors facing the sector traffic running state. The method of the invention optimizes the flow distribution of the airway route according to the traffic states of different periods of the sector, fully utilizes air traffic control resources, and is favorable for creating a civil aviation airspace system with dynamic and interoperable resources, adjustable capacity elasticity, autonomous and efficient operation and strong performance.

Description

Dynamic classification method of time-varying sectors for traffic state

Technical Field

The invention belongs to the field of air traffic management and airspace planning, and particularly relates to a dynamic classification method for time-varying sectors oriented to traffic states.

Background

With the rapid development of air transportation in China, the flight demand is increasing day by day, the airspace environment is more complex, and the airspace use mode is diversified, complicated and flexible. The existing sector classification is divided into area sectors, approaching sectors and the like based on the inherent structure and function classification of airspace, the classification method is physical classification according to the flight process of an aircraft, is only suitable for the basic sector classification in the control process, lacks comprehensive consideration on the structural characteristics and traffic flow characteristics of the sectors, and cannot meet the requirements of flow management and airspace management.

From patent and research literature on sector classification at home and abroad, the sector classification based on sector operation characteristics is less researched. Meanwhile, the integral classification of the sectors based on the operation characteristics is beneficial to macroscopically knowing the sectors and executing corresponding management measures. However, the sector classification research should also consider the dynamic operation state, i.e. dynamic category, of the sector having time-varying traffic flow assigned thereto. Analysis mining through historical data in past sector classification studies, given a fixed classification for each period of the sector, appears to be inflexible and objective. When the traffic flow changes due to the influence of weather and flow control strategies in a certain time period, the category corresponding to the time period may not be consistent with the given fixed category. Thus, such sector classification studies are less helpful in making decisions to managers of air traffic at the pre-tactical level. In addition, home and abroad scholars mostly adopt a hard clustering mode (such as a system clustering method, K-means and the like) on a clustering algorithm based on sector classification research in the aspects of sector complexity and capacity, and classification results and interpretability of clustering are unreasonable to a certain extent.

Disclosure of Invention

Aiming at the defects of the prior art, the invention aims to provide a dynamic classification method for time-varying sectors facing traffic states, so as to solve the problems that the dynamic characteristics of sector traffic operation cannot be reflected and a controller is assisted to formulate a control strategy at a pre-tactical level because the sectors are given fixed categories in the prior art, and a hard clustering algorithm in the traditional classification is unreasonably classified in the aspect of categories; the method of the invention optimizes the flow distribution of the airway route according to the traffic states of different periods of the sector, fully utilizes air traffic control resources, and is favorable for creating a civil aviation airspace system with dynamic and interoperable resources, adjustable capacity elasticity, autonomous and efficient operation and strong performance.

In order to achieve the purpose, the technical scheme adopted by the invention is as follows:

the invention relates to a dynamic classification method of time-varying sectors facing traffic states, which comprises the following steps:

1) collecting historical operating data of sectors, calculating dynamic operating characteristic variable data and performing data preprocessing to obtain clustering input characteristic vectors;

2) clustering analysis is carried out on sector traffic running states; adopting a GA-KFCM clustering algorithm to perform clustering division on the clustering input characteristic vectors obtained in the step 1) on the sector traffic running states, obtaining a data set of various traffic running states of the sector according to the division result, and taking the data set as a corresponding traffic running state measurement standard;

3) constructing a dynamic classification model of a time-varying sector of a neural network based on a multilayer perceptron;

4) and carrying out dynamic classification of the time-varying sectors facing the sector traffic running state.

Further, the step 1) specifically includes:

11) analyzing real ASDB radar track data and sector historical operation data obtained by AirTop computer simulation by taking 15 minutes as a time sample, and acquiring time-varying state characterization index data related to a sector traffic operation state as a basic data set, wherein the time-varying state characterization index data comprises a static structure factor and a dynamic traffic factor of a sector;

12) normalizing the sample data by adopting Z-Score, wherein the mean value of the normalized data is 0, and the standard deviation is 1; and (4) performing relevance analysis on the indexes, further performing protocol processing on the data through PCA analysis, and acquiring clustering input feature vectors.

Further, the step 2) specifically includes:

21) determining the number of the sector traffic running state categories: carrying out a plurality of groups of comparison experiments on the clustering number c of the GA-KFCM clustering algorithm by taking different values to determine the optimal natural clustering number; analyzing the overall contour coefficient value and V of the traffic flow sample data set clustering under the condition of different clustering cluster number values _KXB The index value is used for selecting the optimal clustering number c value to obtain the optimal clustering effect;

22) clustering analysis of sector traffic running states: initializing GA-KFCM clustering algorithm parameters, inputting clustering input feature vectors into a clustering algorithm, and dividing to obtain various sector traffic running states; analyzing traffic operation states of various sectors by combining traffic flow representation index distribution conditions reflected by nuclear density estimation graphs of partial indexes in different traffic states; and acquiring historical state data sets of various traffic running states in the sector through the division result, and taking the historical state data sets as corresponding traffic running state measurement standards.

Further, the step 3) specifically includes:

31) separating the historical state data set into a training data set and an evaluation data set; verifying that the segmentation parameters are set to be 0.2 of the size of the original data set to realize the segmentation of the data set, namely 80% of the original data are used as a training set and 20% of the original data are used as a test set;

32) selecting an activation function and an error function, wherein the activation function is divided into a hidden layer activation function and an output layer activation function;

33) setting the neuron number of an input layer as the number of characteristic variables of an input data set, and setting the neuron number of an output layer as an expected target dimension of dynamic sector classification; designing a plurality of groups of comparison optimization experiments under parameter combinations of different hidden layer numbers and neuron numbers, and monitoring the training condition of the model through a TensorBoard machine learning visualization tool provided by TensorFlow to obtain the influence condition of the different hidden layer numbers on model training; reading comparison experiment logs, outputting an iteration change comparison graph of the accuracy and the loss function in each group of experiments, and determining the number of hidden layers in the dynamic classification model of the sector and the number of neurons contained in each hidden layer; the empirical formula for hidden layer neuron number determination is as follows:

in the formula, N _i Is the number of input signals; n is a radical of ₀ Outputting dimensions for the expected target; n is a radical of _s The number of samples contained in the training set; alpha is any variable that can be taken from, and typically can range from 2 to 10.

Further, the step 32) specifically includes:

321) selecting a Relu function as a hidden layer activation function, wherein the expression is as follows:

322) selecting a softmax function as an output layer activation function, wherein the expression is as follows:

wherein z is a vector, z _i And z _j Is an element thereof;

323) using the cross entropy error as a loss measure of the sector dynamic classification model; the formula for the cross entropy is as follows: the parameters of the loss function are set as classified cross entropy logarithmic loss functions, and the expression is as follows:

wherein, log represents the natural logarithm taking e as a base number; y is _k For model output, t _k For each label, t _k Only the label of the correct solution is 1, and the rest are 0; therefore, the larger the output corresponding to correct label solving is, the closer the value of the cross entropy is to 0; when the output is 1, crossThe entropy error is 0; conversely, if the output corresponding to the correct de-labeling is smaller, the value of the cross entropy is larger.

Further, the step 4) specifically includes: training the time-varying dynamic sector classification model according to the neural network topological structure determined in the step 3) and the selected excitation function and loss function, and predicting to obtain the dynamic sector category.

Further, the step 4) specifically includes:

41) setting multi-layer perceptron algorithm parameters and format codes; carrying out one-hot coding processing on the dynamic classification data, and when y classes exist, coding the output target value of the sample into a y-dimensional vector, wherein the identifier of the position of the sample belonging to a certain class in the vector is 1, and the identifiers of the rest positions are 0; the multi-layer perceptron algorithm parameters are set as follows:

an Optimizer: using an adam algorithm in a high-efficiency gradient descent method as an optimization algorithm for searching optimal weight;

metric: the measurement index is designated as accuracy in the multivariate classification problem;

epochs: setting the cycle times of training the whole training data set as i times;

batchsize: specifying a batch size at which the weight is updated;

42) the performance of the sector classification model is verified and evaluated by the segmentation samples, the time-varying state characterization index data related to the traffic running state of the sector is used as a training sample generation model, and the test is carried out on the reserved sample; according to the segmented test set and verification set, determining an optimal network topology structure and corresponding multi-layer perceptron algorithm parameter setting starting training model and carrying out performance evaluation; the data contained in the output log is used for observing the change trend of the data in each training period, and whether the model is normally trained and whether the training needs to be stopped is judged;

43) evaluating the performance of the classification algorithm by using k-fold cross validation; randomly dividing a data set to be learned into m disjoint subsets, wherein the subsets have the same number of instances; randomly selecting 1 subset as an evaluation data set, using the other m-1 subsets to train the model, and sequentially using each subsequent subset as a test set of the model obtained by training the other m-1 subsets to evaluate the effect of the model;

44) inputting time-varying state characterization index data related to a sector traffic running state, performing dynamic classification of the time-varying sectors facing the sector traffic running state by using a sector dynamic classification model and taking 15 minutes as a time segment according to the traffic flow index data in the sectors at 96 time intervals in one day, and acquiring the category of the sector in every 15 minutes; and comparing the obtained dynamic category of the sector in the 15-minute time period with the actual category of the sector in the sample data in the 15-minute time period, so as to show the accuracy and the availability of the dynamic classification model of the sector.

The invention has the beneficial effects that:

1. the method introduces the GA-KFCM soft clustering algorithm based on the ambiguity of the sector time-varying traffic state, overcomes the problem that the traditional hard clustering algorithm is unreasonable in classification, and ensures the partition result of the sector traffic state. The mass sector traffic flow operation index data is utilized to carry out cluster analysis, a sector historical traffic state sample data set is obtained, and reliable historical sample data input can be provided for a sector dynamic classification model.

2. Compared with the prior art, the sector classification method is only given to a fixed category for each period of the sector through analysis of historical data. According to the invention, the sector dynamic classification facing the time-varying performance of the traffic state is researched by constructing a sector traffic state classification discrimination algorithm. And obtaining traffic flow characteristics and complex situations corresponding to the traffic states of different sectors by analyzing the clustering. And further combining a deep learning neural network algorithm to establish an efficient and accurate sector traffic running state distinguishing and predicting model. Finally, the traffic state of the pre-tactical level of the output sector is predicted through the input sector operation data, the dynamic classification of the time-varying sector facing the traffic state is realized, and the defect that the classification effect in the prior art is less helpful to the pre-tactical level is overcome.

3. The sector dynamic classification realized by the invention can provide refined sector dynamic operation information in 15min (minute) as a time period, and can assist a controller to make control strategies under different traffic operation states so as to strengthen the study and judgment on future traffic scenes.

Drawings

FIG. 1 is a flow chart of the method of the present invention;

FIG. 2 is a diagram illustrating a visualization result of a linear correlation analysis between sector time-varying state characterization indicators obtained by the correlation analysis according to the present invention;

FIG. 3a is a graph of a nuclear density estimate for traffic flow indicators in different dynamic categories;

FIG. 3b is a graph of the kernel density estimates for potential conflict indicators in different dynamic categories;

FIG. 3c is a graph of the kernel density estimates for the mean time-of-flight indicator under different dynamic categories;

FIG. 3d is a graph of a nuclear density estimate for descending traffic flow indicators in different dynamic categories;

FIG. 4a is a diagram illustrating a trend of accuracy variation in an implicit layer number optimization experiment;

FIG. 4b is a graph of loss function variation trend in an optimization experiment of the number of hidden layers;

FIG. 5 is a comparison graph of sector historical operating period class and sector dynamic classification model prediction class results.

Detailed Description

In order to facilitate understanding of those skilled in the art, the present invention will be further described with reference to the following examples and drawings, which are not intended to limit the present invention.

Referring to fig. 1, a traffic state-oriented time-varying sector dynamic classification method according to the present invention includes the following steps:

step 1: collecting historical operation data of the sector, calculating dynamic operation characteristic variable data and performing data preprocessing;

selecting historical sector operation data, analyzing the sector operation data by taking 15 minutes as a time sample, and acquiring basic data describing the dynamic traffic state of the sector; the dynamic factors representing the sector traffic state comprise the number of times of entering the aircraft into the rack, the number of times of leaving the rack, the total number of occupied aircrafts in the sector, the maximum number of occupied aircrafts in the sector within 15 minutes, the average flight time of the aircrafts, the average flight distance, the proportion of aircrafts with different flight attitudes, the entropy of the airway flow and the number of potential conflicts in the sector and the like.

The correlation analysis refers to the analysis of two or more variable elements with correlation; the Pearson correlation coefficient rho (a, b) introduced based on covariance is defined as:

wherein X in the matrix X _a Y in column sum matrix Y _b The columns of the image data are,

n is the length of each column; when the absolute rho absolute is less than or equal to 0.3, the linear correlation does not exist; 0.3<The rho | is less than or equal to 0.5, which represents that the variables are low-degree linear correlation; 0.5<The rho | is less than or equal to 0.8, and the expression variable is obviously linear correlation; | rho | fly>0.8, which indicates that there is a highly linear correlation between the variables; according to the positive and negative of the numerical value, the positive correlation and the negative correlation are simultaneously divided, and 0 represents mutual independence; the result of the linear correlation analysis visualization between sector time-varying state characterization indicators is shown in fig. 2.

By carrying out correlation analysis on the selected index variables, aiming at the linear correlation with different degrees among different indexes, the information reflected by the sample data can be overlapped.

And further extracting principal components with the accumulative contribution rate of more than 85% through principal component analysis, and taking the principal components as feature vectors input by the clustering analysis of the sector running traffic states.

Step 2: clustering analysis is carried out on sector traffic running states;

carrying out a plurality of groups of comparison experiments on the clustering number c of the clustering algorithm by taking different values to determine the optimal natural clustering number; too many or too few sector traffic running state categories are not suitable, too few sector traffic running state categories cannot reflect the variability of sector traffic states, and the sector traffic running state categories cannot play a role in accurately distinguishing different categories; too many categories also have the problem of unobvious differentiation, and simultaneously, the complexity of a controller in the process of using the sector dynamic category to make a control strategy and open and close sector decisions is increased.

Optimizing the clustering number in a proper interval, and acquiring and analyzing the overall contour coefficient value and V of the traffic flow sample data set clustering under the condition of different clustering number values _KXB An index value; when the number of the clustered clusters is c, the overall contour coefficient is taken to be the maximum value MAX, V _KXB And taking the index value to the minimum MIN, and determining the clustering number as c according to the clustering number selection standard.

Initializing and setting GA-KFCM algorithm parameters, and starting sector traffic state cluster analysis:

a. setting genetic algorithm parameters, population number n and initial cross probability p _c0 Initial mutation probability p _m0 Genetic evolution algebra T and genetic algorithm convergence threshold value delta;

b. setting basic parameters of a KFCM algorithm, a Gaussian kernel function scale parameter sigma, a target function termination tolerance epsilon, a fuzzy clustering algebra k, a clustering number c and a fuzzy index m;

and inputting the characteristic vectors into a clustering algorithm, and clustering and dividing to obtain the traffic state of the c-type sector. And analyzing the clustering result by combining the historical traffic flow representation parameter distribution and the kernel density estimation graph of the part indexes in the c-type traffic state shown in the figures 3 a-3 d. The different traffic states of the type c at which the sectors operate in different time periods are obviously different, the distribution of the traffic flow, the potential conflict times, the average flight time and the descending traffic flow proportion in the corresponding sectors shows the sector operation characteristics of different operation states, and simultaneously, the distribution also corresponds to the traffic concentration and the workload of controllers in different route pairs in different sectors. In conclusion, in the class c traffic state, a low-load traffic state at night is identified, in which the traffic flow in the sector is low and there is no potential conflict among aircrafts; concentrating into a descending traffic situation of the aircraft where the percentage of the highly descending traffic flow is higher than the mean and median of the whole population; the flight flow characteristic is more obvious in a traffic state, the traffic flow is high in the state, more traffic flows with changed heights and more scattered route pair flow distribution exist in the sector, and the capacity of the sector is close to a saturated state; a sector overload running state, wherein the traffic entropy of the air route, the number of times of entering the rack in 15 minutes and the number of potential conflicts among aircrafts are higher than those of other types of traffic states, and meanwhile, a large number of aircrafts generate altitude changes in the vertical direction; the controller workload substantially exceeds the 70% threshold.

The rationality of the traffic state division is further verified, the sector flight flow is counted at 15 minute intervals throughout the day (24 hours) according to historical operating data, and the sector flow space-time distribution characteristics are analyzed. And comparing the proportion of the c-type traffic state in the historical typical day operation of the sector with the clustering sample proportion of the c-type traffic state obtained by the clustering result, and providing reliable baseline data for subsequent experiments.

And step 3: constructing a dynamic classification model of a time-varying sector of a neural network based on a multilayer perceptron;

acquiring a data sample at 15-minute intervals, wherein the total number of the data samples is N; 80% of the raw data were segmented as training set and 20% of the raw data as test set.

Selecting a relu function which is convenient for processing back propagation to be called at the hidden layer as an activation function; and if the sector runs in c different states, the output layer comprises c neurons, for the output layer of the multi-classification problem, softmax is adopted as an activation function, and the loss function parameter is set as a classification cross entropy logarithmic loss function.

As shown in fig. 4a and 4b, designing multiple sets of comparison experiments under different parameter combinations of the number of hidden layers and the number of neurons, and outputting iterative change conditions of accuracy and loss functions in each set of experiments; and determining the number of hidden layers of the constructed sector traffic state discrimination prediction model as j according to the comparison experiment result, wherein the number of neurons contained in each hidden layer is k.

And 4, step 4: selecting a neural network topological structure determined in the step 3 and a related excitation function and a related loss function; and setting other parameters of the multi-layer perceptron algorithm, further training a dynamic sector classification model, and dynamically classifying the traffic-state-oriented denatured sectors.

Inputting time-varying state characterization index data related to the traffic running state of the sector, predicting and judging the dynamic prediction sector category of the sector by using a dynamic classification model of the sector, comparing the accuracy and the availability of the dynamic classification model of the sector with the category of the actual period of the sector, and selecting a traffic flow data sample 24 hours in one day to obtain the dynamic category of the sector for displaying. As shown in fig. 5, sector categories in 96 time periods of a day are distinguished and predicted by taking 15 minutes as a time segment, and sector categories in 3 time periods in the classification result do not match the real category, but the traffic state and complexity degree under the category which does not match the actual situation in the prediction are not much different from those under the real category. The dynamic classification model of the sector constructed by the invention has stronger prediction and discrimination accuracy, and the extracted dynamic classification index of the sector has good universality. The method can provide refined sector dynamic information in 15 minutes as a time period, and can assist a controller to make control strategies under different traffic running states so as to strengthen research and judgment on future traffic scenes. Meanwhile, as management assistance, when the category fluctuation of the sector is small in a period of time in the future, and the corresponding traffic flow is low and the control difficulty is small in such a state, the sector in the same state can be considered to be subjected to sector combination operation. In addition, an efficient and safe sector operation configuration scheme and a human resource configuration scheme can be provided according to different dynamic types of sectors, and a novel post mode is innovatively developed.

While the invention has been described in terms of its preferred embodiments, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention.

Claims (7)

1. A dynamic classification method of time-varying sectors facing traffic states is characterized by comprising the following steps:

1) collecting historical operation data of sectors, calculating dynamic operation characteristic variable data and performing data preprocessing to obtain clustering input characteristic vectors;

2) clustering analysis is carried out on sector traffic running states; clustering division of sector traffic running states is carried out on the clustering input characteristic vectors obtained in the step 1), and a data set of various traffic running states of the sector is obtained through division results and is used as a corresponding traffic running state measurement standard;

3) constructing a dynamic classification model of a time-varying sector of a neural network based on a multilayer perceptron;

4) and carrying out dynamic classification of the time-varying sectors facing the sector traffic running state.

2. The traffic state-oriented time-varying sector dynamic classification method according to claim 1, wherein the step 1) specifically comprises:

11) analyzing real radar track data and sector historical operation data obtained by simulation by taking a time sample of 15 minutes, and acquiring time-varying state characterization index data related to a sector traffic operation state as a basic data set, wherein the time-varying state characterization index data comprises a static structure factor and a dynamic traffic factor of a sector;

12) normalizing the sample data by adopting Z-Score, wherein the mean value of the normalized data is 0, and the standard deviation is 1; and (4) performing relevance analysis on the indexes, further performing protocol processing on the data through PCA analysis, and acquiring clustering input feature vectors.

3. The traffic state-oriented time-varying sector dynamic classification method according to claim 1, wherein the step 2) specifically comprises:

21) determining the number of the sector traffic running state categories: carrying out a plurality of groups of comparison experiments on the clustering number c of the GA-KFCM clustering algorithm by taking different values to determine the optimal natural clustering number; analyzing the overall contour coefficient value and V of the traffic flow sample data set clustering under the condition of different clustering cluster number values _KXB The index value is used for selecting the optimal clustering number c value to obtain the optimal clustering effect;

22) clustering analysis of sector traffic running states: initializing GA-KFCM clustering algorithm parameters, inputting clustering input feature vectors into a clustering algorithm, and dividing to obtain various sector traffic running states; analyzing traffic operation states of various sectors by combining traffic flow representation index distribution conditions reflected by nuclear density estimation graphs of partial indexes in different traffic states; and acquiring historical state data sets of various traffic running states in the sector through the division result, and taking the historical state data sets as corresponding traffic running state measurement standards.

4. The traffic state-oriented time-varying sector dynamic classification method according to claim 1, wherein the step 3) specifically comprises:

31) separating the historical state data set into a training data set and an evaluation data set; verifying that the segmentation parameters are set to be 0.2 of the size of the original data set to realize the segmentation of the data set, namely 80% of the original data are used as a training set and 20% of the original data are used as a test set;

32) selecting an activation function and an error function, wherein the activation function is divided into a hidden layer activation function and an output layer activation function;

33) setting the neuron number of an input layer as the number of characteristic variables of an input data set, and setting the neuron number of an output layer as an expected target dimension of dynamic sector classification; designing a plurality of groups of comparison optimization experiments under parameter combinations of different hidden layer numbers and neuron numbers, and monitoring the training condition of the model through a TensorBoard machine learning visualization tool provided by TensorFlow to obtain the influence condition of the different hidden layer numbers on model training; reading comparison experiment logs, outputting iteration change comparison graphs of accuracy and loss functions in each group of experiments, and determining the number of hidden layers in the sector dynamic classification model and the number of neurons contained in each hidden layer; the empirical formula for hidden layer neuron number determination is as follows:

in the formula, N _i The number of input signals; n is a radical of ₀ Outputting dimensions for the expected target; n is a radical of _s The number of samples contained in the training set; alpha is any value variable that can be taken from.

5. The traffic state-oriented time-varying sector dynamic classification method according to claim 4, characterized in that said step 32) specifically comprises:

321) selecting a Relu function as a hidden layer activation function, wherein the expression is as follows:

322) selecting a softmax function as an output layer activation function, wherein the expression is as follows:

wherein z is a vector, z _i And z _j Is an element thereof;

323) using the cross entropy error as a loss measure of the sector dynamic classification model; the formula for the cross entropy is as follows: the parameters of the loss function are set as classified cross entropy logarithmic loss functions, and the expression is as follows:

in the formula, log represents a natural logarithm based on e; y is _k For model output, t _k For each label, t _k In the method, only the label with correct solution is 1, and the rest are 0; the larger the output corresponding to correct label solving is, the closer the value of the cross entropy is to 0; when the output is 1, the cross entropy error is 0; conversely, if the output corresponding to the correct de-labeling is smaller, the value of the cross entropy is larger.

6. The traffic state-oriented time-varying sector dynamic classification method according to claim 1, wherein the step 4) specifically comprises: training the time-varying dynamic sector classification model according to the neural network topological structure determined in the step 3) and the selected excitation function and loss function, and predicting to obtain the dynamic sector category.

7. The traffic state-oriented time-varying sector dynamic classification method according to claim 1, wherein the step 4) specifically comprises:

41) setting multi-layer perceptron algorithm parameters and format codes; carrying out one-hot coding processing on the dynamic classification data, and when y classes exist, coding the output target value of the sample into a y-dimensional vector, wherein the identifier of the position of the sample belonging to a certain class in the vector is 1, and the identifiers of the rest positions are 0; the multi-layer perceptron algorithm parameters are set as follows:

an Optimizer: using an adam algorithm in a high-efficiency gradient descent method as an optimization algorithm for searching optimal weight;

metric: the measurement index is designated as accuracy in the multivariate classification problem;

epochs: setting the cycle times of training the whole training data set as i times;

batchsize: specifying a batch size at which the weight is updated;

42) the performance of a sector classification model is verified and evaluated by a segmentation sample, time-varying state representation index data related to the traffic running state of a sector is used as a training sample generation model, and a test is carried out on a reserved sample; according to the segmented test set and verification set, determining an optimal network topology structure and corresponding multi-layer perceptron algorithm parameter setting starting training model and carrying out performance evaluation; the data contained in the output log is used for observing the change trend of the data in each training period, and whether the model is normally trained and whether the training needs to be stopped is judged;

43) evaluating the performance of the classification algorithm by using k-fold cross validation; randomly dividing a data set to be learned into m disjoint subsets, wherein the subsets have the same number of instances; randomly selecting 1 subset as an evaluation data set, using the other m-1 subsets to train the model, and sequentially using each subsequent subset as a test set of the model obtained by training the other m-1 subsets to evaluate the effect of the model;

44) inputting time-varying state characterization index data related to a sector traffic running state, performing dynamic classification of the time-varying sectors facing the sector traffic running state by using a sector dynamic classification model and taking 15 minutes as a time segment according to the traffic flow index data in the sectors at 96 time intervals in one day, and acquiring the category of the sector in every 15 minutes; and comparing the obtained dynamic category of the sector in the 15-minute time period with the actual category of the sector in the sample data in the 15-minute time period, so as to show the accuracy and the availability of the dynamic classification model of the sector.

Images