CN115238963A - Traffic prediction method, device, electronic device and storage medium - Google Patents

Abstract

The present invention provides a traffic prediction method, device, electronic device and storage medium, wherein the method includes: determining historical traffic data and a missing position matrix of the historical traffic data; inputting the historical traffic data and the missing position matrix into a prediction completion model, Obtain the prediction result output by the prediction completion model; based on the weight parameters of the data completion module in the initial model, the prediction completion model applies the sample data, the sample missing data, the expected value of the sample data and the sample missing position matrix to carry out the analysis on the initial model. obtained by training; the weight parameters of the data completion module in the initial model are obtained by pre-training based on sample data, sample missing data and sample missing position matrix, and the method provided by the present invention realizes the prediction module and the data completion module in the prediction model The information exchange between them completes the end-to-end prediction, which improves the accuracy of the prediction results while improving the immediacy.

Description

Traffic prediction method, device, electronic device and storage medium

technical field

The present invention relates to the technical field of artificial intelligence, and in particular, to a traffic prediction method, device, electronic device and storage medium.

Background technique

With the development of big data and artificial intelligence technology, the increasingly accumulated traffic big data and various excellent deep learning algorithms provide a realistic basis for alleviating traffic congestion and improving traffic pollution. However, in real traffic scenarios, due to the difficulty in maintaining physical detection equipment, as well as possible accidental factors (such as data loss, errors, etc.) in the process of data transmission and storage, the obtained traffic data is often of poor data quality, or There are different proportions of missing data.

Although the existing methods can achieve better prediction performance in the scenario of missing data, they still have the following shortcomings:

First, the existing methods regard data completion as an independent task, and traffic prediction in a data-missing scenario is divided into two stages: data completion and prediction, which is less immediate;

Second, there is no effective interaction between data completion and prediction tasks, which will cause errors in data completion to further accumulate in the prediction task, which will have a negative impact on the prediction accuracy.

SUMMARY OF THE INVENTION

The present invention provides a traffic prediction method, device, electronic device and storage medium, which are used to solve the defects of poor immediacy and low prediction accuracy of traffic prediction in the prior art in the scenario of missing traffic data.

The present invention provides a traffic prediction method, comprising:

determining historical traffic data and a matrix of missing locations for said historical traffic data;

Inputting the historical traffic data and the missing position matrix into the prediction completion model to obtain the prediction result output by the prediction completion model;

The prediction completion model is obtained by training the initial model by applying sample data, sample missing data, sample data expected value and sample missing position matrix on the basis of the weight parameters of the data completion module in the initial model; The weight parameters of the data completion module in the initial model are pre-trained based on the sample data, the sample missing data and the sample missing position matrix.

According to a traffic prediction method provided by the present invention, the training steps of the prediction completion model include:

Determine the initial model; the initial model includes a data completion module and a prediction module;

Pre-training the data completion module based on the sample data, the sample missing data, the expected value of the sample data and the sample missing position matrix to obtain the weight parameter of the data completion module;

Based on the weight parameter, the sample data, the sample missing data, the expected value of the sample data, and the sample missing position matrix, the data completion module and the prediction module are jointly trained to obtain the prediction Complete the model.

According to a traffic prediction method provided by the present invention, the joint training of the data completion module and the prediction module includes:

Based on the joint training loss function, the data completion module and the prediction module are jointly trained; the joint training function is based on the expected value of the sample data and the initial prediction result of the sample data output by the prediction module. and the difference between the missing data of the sample and the complementary data of the sample data output by the data completion module.

According to a traffic prediction method provided by the present invention, the step of obtaining the initial prediction result includes:

determining sample reconstruction data based on the sample data and the complement data of the sample data;

Based on the sample reconstruction data and each time granularity, determine the sample reconstruction data corresponding to each time granularity;

Determine the sample fusion feature based on the sample reconstruction data corresponding to each time granularity;

Based on the sample fusion feature, it is determined that the initial prediction result is obtained.

According to a traffic prediction method provided by the present invention, the determination of sample fusion features based on the sample reconstruction data corresponding to each time granularity includes:

Perform spatiotemporal feature extraction on the sample reconstruction data corresponding to each time granularity to obtain the sample spatiotemporal feature corresponding to each time granularity;

Determine the sample spatiotemporal fusion feature based on the sample spatiotemporal features and adaptive weights corresponding to each time granularity;

The sample fusion characteristic is determined based on the sample spatiotemporal fusion characteristic and the external factor characteristic; the external factor characteristic includes the weather factor characteristic and/or the time period factor characteristic.

According to a traffic prediction method provided by the present invention, the data completion module is pre-trained based on the sample data, the sample missing data, the expected value of the sample data and the sample missing position matrix, and the obtained data is obtained by pre-training the data completion module. The weight parameters of the data completion module, including:

determining the sample data, the sample missing data and the sample missing location matrix;

Perform spatiotemporal feature extraction on the sample data, and complete the sample data based on the obtained sample spatiotemporal features corresponding to the sample data, to obtain complementary data corresponding to the sample data;

Based on the completion data corresponding to the sample data, the sample missing data and the sample missing position matrix, the loss of the completion module is determined, and based on the loss of the completion module, the parameters of the data completion module are iterated update to obtain the weight parameters of the data completion module.

According to a traffic prediction method provided by the present invention, the steps of determining the sample data, the sample missing data and the sample missing location matrix are as follows:

determine raw traffic data;

Randomly delete each preset proportion of traffic data from the original traffic data, and obtain sample data corresponding to each preset proportion and sample missing data corresponding to each preset proportion;

Based on the spatiotemporal position of the original traffic data and the spatiotemporal position of the sample missing data corresponding to each preset proportional quantity in the original traffic data, determining a sample missing position matrix corresponding to each preset proportional quantity;

Based on the sample data corresponding to each preset proportional quantity, the sample missing data corresponding to each preset proportional quantity, and the sample missing position matrix corresponding to each preset proportional quantity, determine the sample data, the sample missing data and the sample missing location matrix.

The present invention also provides a traffic prediction device, comprising:

a determining module for determining historical traffic data and a missing location matrix of the historical traffic data;

A prediction module, for inputting the historical traffic data and the missing position matrix into a prediction completion model to obtain a prediction result output by the prediction completion model;

The prediction completion model is obtained by training the initial model by applying sample data, sample missing data, sample data expected value and sample missing position matrix on the basis of the weight parameters of the data completion module in the initial model; The weight parameters of the data completion module in the initial model are pre-trained based on the sample data, the sample missing data and the sample missing position matrix.

The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, when the processor executes the program, the traffic prediction method as described above is implemented by the processor .

The present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, implements any one of the traffic prediction methods described above.

The present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements any one of the traffic prediction methods described above.

The traffic forecasting method, device, electronic equipment and storage medium provided by the present invention, through the forecasting complementation model obtained by training on the basis of the data complementation model, predicts according to historical traffic data, obtains the forecasting result, and realizes the forecasting complementation model The end-to-end prediction process, and during the training of the prediction completion model, the parameters of the initial model of the prediction completion network model will be updated according to the output initial prediction results, and the prediction module and data completion module in the prediction completion model are implemented. The information exchange between them completes the end-to-end prediction, which improves the accuracy of the prediction results while improving the immediacy.

Description of drawings

In order to explain the present invention or the technical solutions in the prior art more clearly, the following will briefly introduce the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are the For some embodiments of the invention, for those of ordinary skill in the art, other drawings can also be obtained according to these drawings without any creative effort.

1 is a schematic flowchart of a traffic prediction method provided by the present invention;

Fig. 2 is one of the schematic flow charts of the training method of prediction completion model provided by the present invention;

3 is a schematic flowchart of a method for obtaining an initial prediction result provided by the present invention;

4 is a schematic flowchart of a method for obtaining a sample fusion feature provided by the present invention;

5 is a network structure diagram of a spatiotemporal feature extraction network provided by the present invention;

Fig. 6 is the second schematic flow chart of the prediction completion model training method provided by the present invention;

Fig. 7 is the network structure diagram of the prediction completion model provided by the present invention;

8 is a schematic structural diagram of a traffic prediction device provided by the present invention;

FIG. 9 is a schematic structural diagram of an electronic device provided by the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention. , not all examples. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

At present, the existing data completion is mainly through classical data interpolation methods such as historical average and linear regression, or through traditional machine learning methods such as Markov Chain Monte Carlo (MCMC) and deep learning-based methods. Data completion method. Although the existing methods can achieve better prediction performance in the scenario of missing data, due to the fact that data completion is an independent task independent of prediction, the immediacy is poor, and there is no interaction between data completion and prediction. , which will cause errors in data completion to accumulate in the prediction, resulting in a prediction accuracy angle.

Therefore, how to improve the immediacy and prediction accuracy of traffic data prediction in a data-missing scenario is a technical problem to be solved urgently by those skilled in the art.

In view of the above technical problems, embodiments of the present invention provide a traffic prediction method. FIG. 1 is a schematic flowchart of a traffic prediction method provided by the present invention. As shown in Figure 1, the method includes:

Step 110, determining the historical traffic data and the missing location matrix of the historical traffic data;

It should be noted that historical traffic data is a sequence of traffic data in an area within a time period, and the traffic data can be data such as traffic flow, speed, or demand, where the area can be one street or multiple streets, or The city is divided into multiple regions based on longitude and latitude, which is not limited in this embodiment of the present invention.

In addition, the historical traffic data may be complete traffic data in a time period in an area, or may be traffic data with missing data in a time period in an area, which is not limited in this embodiment of the present invention. The complete traffic data means that the traffic data of each time step is not lost from the start time to the end time of the time period, where the time step represents the time interval, for example: if the time step is 1 second, then the complete traffic in one minute The data includes 60 pieces of data, that is, one piece of data per second. The missing location matrix of historical traffic data is constructed with the number of complete traffic data as the size, and the missing flag value is applied to record the missing data in the historical traffic data, for example: the time step is 1 second, constructing one minute of historical traffic data. The missing position matrix can be a 60x1 matrix. If the traffic data of the 15th second in the historical traffic data is missing, the 15th index bit in the matrix is marked as 0, and the rest of the matrix bits are marked as 1, which is not performed in this embodiment of the present invention. limit.

Step 120, input the historical traffic data and the missing position matrix into the prediction completion model, and obtain the prediction result output by the prediction completion model;

The prediction completion model is obtained by training the initial model by applying sample data, sample missing data, sample data expected value and sample missing position matrix on the basis of the weight parameters of the data completion module in the initial model; data completion in the initial model The weight parameters of the module are pre-trained based on sample data, sample missing data and sample missing location matrix.

The prediction process of the current traffic prediction model in the scenario of missing data is to first complete the missing data through the data completion model alone to obtain the completed traffic data, and then the prediction model predicts based on the completed traffic data. To get the prediction result, the prediction process is not an end-to-end process, and the information interaction between the data completion module and the prediction module cannot be realized. It can be seen that the prediction completion model needs to have the ability to exchange information between the data completion module and the prediction module. Only by realizing the end-to-end prediction process can improve the immediacy of prediction and alleviate the further accumulation of errors in data completion results in the prediction module, thereby improving the accuracy of the prediction results. Therefore, in the embodiment of the present invention, the data completion module is used to provide the prediction completion model with the capability of data completion, and by designing a joint training mechanism, the information interaction between the data completion module and the prediction module is realized, and the prediction is realized. Completing the model end-to-end prediction process.

Specifically, the data completion module in the initial model is pre-trained according to the sample data, the sample missing data and the sample missing position matrix, and the weight parameters of the data completion module are obtained, so that the initial model itself has the ability of data completion. . On this basis, the parameters of the initial model can be iteratively updated with sample data, sample missing data, sample data expected value and sample missing position matrix. It will be updated at the same time, so that the data completion module in the initial model will be affected by the initial prediction result output by the prediction module, so that the prediction completion model for the missing data scenario has the data completion capability and the data in the prediction completion model. The ability to form information interaction between the completion module and the prediction module.

The resulting prediction completion model has the end-to-end prediction process and the ability to exchange information between the data completion module and the prediction module in the prediction completion model. Therefore, historical traffic data can be directly input into the prediction model. Obtain accurate and reliable prediction results output by the prediction model.

It should be noted that the sample data can be incomplete traffic data in a region with different missing proportions within a time period, and the sample labeled data is the traffic data missing from the sample data, and the missing proportion represents the number of missing traffic data. The proportion of complete traffic data, for example: complete traffic data for one minute with a time step of 1 second, and the missing rate is 10%, then 6 pieces of data are randomly missing in the sample data determined by the missing rate. This is not limited. In addition, the expected value of the sample data refers to the traffic data value within the same time period in the area in the future, such as a flow value or a speed value, which is not limited in the embodiment of the present invention.

In the traffic prediction method provided by the embodiment of the present invention, a prediction completion model obtained by training based on the weight parameters of the data completion module obtained by pre-training is used to predict historical traffic data, and a prediction result is obtained. During the training process, the parameters of the initial model of the prediction completion network model will be updated according to the output initial prediction results, realizing the information interaction between the prediction module and the data completion module in the prediction model, and completing the end-to-end prediction. While improving immediacy, the accuracy of prediction results is improved.

Based on the above embodiment, FIG. 2 is one of the schematic flowcharts of the training method of the prediction completion model provided by the present invention. As shown in Figure 2, the training steps of the prediction model in step 120 include:

Step 210, determine an initial model; the initial model includes a data completion module and a prediction module;

Step 220, pre-training the data completion module based on the sample data, the sample missing data, the expected value of the sample data and the sample missing position matrix, to obtain the weight parameters of the data completion module;

Step 230: Jointly train the data completion module and the prediction module based on the weight parameters, sample data, missing sample data, expected value of sample data, and sample missing position matrix to obtain a prediction completion model.

In order to enable the prediction completion model to have the ability to complete data and predict results in the scenario of missing data, the data completion module and the prediction module are used to build the initial model, and the data completion module is pre-trained to ensure Get the weight parameters of the data completion module.

Compared with the initial model before the data completion module pre-training, the initial model after the data completion module pre-training already has certain data completion capabilities.

Then apply the weight parameters, sample data, sample missing data, sample data expected value and sample missing position matrix to jointly train the data completion module and the prediction module in the initial model to obtain the prediction completion model, that is, the initial prediction based on the output of the initial model As a result, the parameters of the data completion module and the prediction module in the initial model are iteratively updated to obtain the prediction completion model. Here, the parameters of the data completion module are also updated along with the initial prediction result, so that the ability of information interaction is formed between the data completion module and the prediction module in the prediction completion model.

In the traffic prediction method provided by the embodiment of the present invention, after the weight parameters of the data completion module in the initial model are obtained by pre-training, the data completion module and the prediction module are jointly trained, so that the prediction and completion model obtained by training has the data In addition to the completion ability and prediction ability, it has the ability to exchange information between the data completion module and the prediction module, thereby reducing the accumulation of errors during training, thereby improving the accuracy of the prediction results of the prediction completion model.

Based on the above embodiment, the joint training of the data completion module and the prediction module in step 220 includes:

Based on the joint training loss function, the data completion module and the prediction module in the initial model are jointly trained; the joint training function is the difference between the initial prediction result of the sample data output by the prediction module in the initial model and the expected value of the sample data, and It is constructed from the difference between the missing data in the sample and the completion data of the sample data output by the data completion module in the initial model.

Considering that information interaction needs to be formed between the data completion module and the prediction module in the prediction completion model, and considering that the gradient between the data completion module and the prediction module in the initial model is interrupted, the gradient of the data completion module cannot be It is passed to the prediction module. Therefore, in the embodiment of the present invention, the data completion module and the prediction module in the initial model are jointly trained with the goal of minimizing the joint training loss.

Specifically, the joint training loss function is jointly determined based on the loss function of the data completion module and the loss function of the prediction module in the initial model. For example, it may be a weighted sum of the two, or a nonlinear combination. The present invention implements The example does not limit this.

Among them, the loss function of the data completion module in the initial model is constructed with the goal of minimizing the difference between the completed data of the sample data and the missing data of the sample. Further, the loss of the data completion module in the initial model can be calculated by the following formula:

为样本数据的补全数据，m为样本缺失位置矩阵，在计算数据补全的误差时只保留样本缺失位置矩阵中缺失标识的位置的数据参与计算，β为超参数。where x is the sample missing data and

is the completion data of the sample data, m is the sample missing position matrix, when calculating the error of the data completion, only the data of the position of the missing mark in the sample missing position matrix is retained to participate in the calculation, and β is the hyperparameter.

The loss function of the prediction module in the initial model is constructed with the goal of minimizing the difference between the initial prediction result and the expected value of the sample data.

Further, the loss of the prediction module in the initial model can be calculated by the following formula:

为初始预测结果，β为超参数。where y is the expected value of the sample data and

is the initial prediction result, and β is a hyperparameter.

Further, the joint loss can be calculated by the following formula:

分别为数据补全和预测结果的损失值，λ∈[0.0,1.0]是一个调节因子，用于调节联合损失函数中数据补全模块损失的最大占比，γ是一个缩放因子，用于对联合损失函数中数据补全模块损失的初始比例进行调节。In the formula,

are the loss values of data completion and prediction results, respectively, λ∈[0.0,1.0] is an adjustment factor used to adjust the maximum proportion of the loss of the data completion module in the joint loss function, γ is a scaling factor, used to The initial scale of the data completion module loss in the joint loss function is adjusted.

Based on the foregoing embodiment, FIG. 3 is a schematic flowchart of a method for obtaining an initial prediction result provided by the present invention. As shown in Figure 3, the steps for obtaining the initial prediction result include:

Step 310, based on the sample data and the complement data of the sample data, determine the sample reconstruction data;

Specifically, according to the spatiotemporal position of the missing data in the sample data, the complementary data of the sample data is inserted into the corresponding spatiotemporal position of the missing data to obtain the sample reconstruction data.

It should be noted that since the sample data comes from a complete traffic data sequence arranged in time order collected in time steps within one area and one time period, the spatiotemporal position represents the position of the traffic data in the traffic data sequence, for example: time step is 1 second and the time period is 1 minute, then there are 60 traffic data in the complete traffic data sequence arranged in chronological order, then the traffic data whose spatiotemporal position is 15 represents the traffic data whose index is 14 in the traffic data sequence. The index number starts from 0.

Step 320, based on the sample reconstruction data and each time granularity, determine the sample reconstruction data corresponding to each time granularity;

Step 330: Determine the sample fusion feature based on the sample reconstruction data corresponding to each time granularity;

Step 340, based on the sample fusion feature, determine to obtain an initial prediction result.

Considering that if the sampling time can be changed from short to long from different time granularities, the prediction completion model can learn a variety of temporal trend features at different time scales, which can improve the accuracy of the prediction results of the prediction completion model. sex.

Specifically, reconstruct the sample reconstruction data at each time granularity to obtain the sample reconstruction data corresponding to each time granularity, perform spatiotemporal feature extraction on the sample reconstruction data corresponding to each time granularity, and extract the sample reconstruction data corresponding to each time granularity. The spatiotemporal features corresponding to the sample reconstruction data are fused to obtain the sample fusion features, and finally prediction is performed according to the sample fusion features to obtain the initial prediction result.

It should be noted that each time granularity in each time granularity is different, and each time granularity covers a time interval ranging from small to large, for example, time granularity minutes, hours, and days, which are not limited in this embodiment of the present invention. The reconstruction of sample reconstruction data at each time granularity can be expressed as the aggregation of sample data in adjacent time granularities. For example, if the time granularity is 1 hour, then the current sample data is located two hours after the time point. The sample data within is merged into the sample reconstruction data corresponding to the current sample data.

In addition, the spatiotemporal features corresponding to the sample reconstruction data corresponding to each time granularity are fused, which can be weighted based on a preset weight, and can also be weighted based on an adaptive weight, wherein the weight value of the adaptive weight can be based on the fusion loss function. The obtained loss is adjusted accordingly, which is not limited in the embodiment of the present invention.

Based on the foregoing embodiment, FIG. 4 is a schematic flowchart of a method for acquiring a sample fusion feature provided by the present invention. As shown in Figure 4, step 330 includes:

Step 331 , perform spatiotemporal feature extraction on the sample reconstruction data corresponding to each time granularity, to obtain the sample spatiotemporal feature corresponding to each time granularity;

Step 332, based on the sample spatiotemporal features and adaptive weights corresponding to each time granularity, determine the sample spatiotemporal fusion features;

Step 333: Determine the sample fusion characteristics based on the sample spatiotemporal fusion characteristics and external factor characteristics; the external factor characteristics include weather factor characteristics, and/or time period factor characteristics.

Considering that the adaptive weight can allow the prediction module in the initial model to adjust adaptively according to the loss calculated by the fusion loss function, that is, the weight ratio can be adjusted by data-driven, so that the trained prediction completion model can know which time granularity corresponds to. Spatiotemporal features can more affect the prediction results, thereby improving the accuracy of the prediction results.

At the same time, considering that the traffic data will be affected by some external factors, such as: rainy days are more congested than sunny days or more congested in the morning and evening rush hours, etc. Therefore, the embodiment of the present invention further integrates the external factor characteristics and the sample spatiotemporal characteristics, which can Further improve the accuracy of prediction results.

Specifically, first perform spatiotemporal feature extraction on the sample reconstruction features corresponding to each time granularity to obtain the sample spatiotemporal features corresponding to each time granularity; Fusion features, and finally the sample spatiotemporal fusion features and external factor features are fused again to obtain sample fusion features. The external factors include: weather factors, and/or time period factors. The time period factors may include morning and evening peak periods, holiday periods, etc., and the external factors may also include temperature factors, wind speed factors, and the like, which are not limited in this embodiment of the present invention.

It should be noted that the spatiotemporal feature extraction for the sample reconstruction features corresponding to each time granularity is extracted according to the spatiotemporal feature extraction network corresponding to each time granularity in the prediction module, and the spatiotemporal feature extraction corresponding to each time granularity in the prediction module is performed. The parameters of the network are iteratively updated according to the loss value calculated by the fusion loss function.

基于预测模块中时粒度的时空特征提取网络，对时粒度对应的样本重构数据进行时空特征提取，得到时粒度对应的样本时空特征

基于预测模块中天粒度的时空特征提取网络，对天粒度对应的样本重构数据进行时空特征提取，得到天粒度对应的样本时空特征

基于

和

以及

的自适应权重W_m、

的自适应权重W_h和

的自适应权重W_d，通过融合公式进行融合得到样本时空融合特征

具体公式如下：Further, the time granularity includes: sub-granularity, time-granularity and day-granularity, respectively, based on the spatial-temporal feature extraction network of sub-granularity in the prediction module, perform spatio-temporal feature extraction on the sample reconstruction data corresponding to sub-granularity, and obtain samples corresponding to sub-granularity. spatiotemporal features

Based on the spatiotemporal feature extraction network of the time granularity in the prediction module, perform spatiotemporal feature extraction on the sample reconstruction data corresponding to the time granularity, and obtain the sample spatiotemporal features corresponding to the time granularity.

Based on the spatiotemporal feature extraction network of the sky granularity in the prediction module, the spatiotemporal feature extraction of the sample reconstruction data corresponding to the sky granularity is carried out, and the spatiotemporal characteristics of the sample corresponding to the sky granularity are obtained.

based on

and

as well as

The adaptive weight W _m ,

The adaptive weights W _h and

The adaptive weight W _d of , and the fusion formula is used to obtain the sample spatio-temporal fusion features

The specific formula is as follows:

where ⊙ represents the Hadamard product, and W _m , W _h and W _d are adaptively adjusted in a data-driven manner.

Further, FIG. 5 is a network structure diagram of a spatiotemporal feature extraction network provided by the present invention. As shown in Figure 5, first, according to the spatiotemporal information in the sample data, the feature extraction is performed on the sample data by means of residual error, and then the feature dimension of the extracted features is adjusted, and finally the adjusted features are input into the dilated causal convolution , the sample spatiotemporal features corresponding to the output sample data are obtained. In the figure, BN is the labeling, Conv2d, Conv2d_1 and Conv2d_2 are the two-dimensional convolutional layers, ReLu is the activation function, and ResBlock 1...L is the residual module unit.

The traffic prediction method provided by the embodiment of the present invention reconstructs the sample reconstruction data through multi-time granularity, so that the prediction completion model can learn multi-trend spatio-temporal features, and fuse the external factor features with the sample spatio-temporal features of each time granularity. The fusion enables the prediction completion model to learn the influence of external factors on the prediction results, and further improves the accuracy of the prediction results output by the prediction completion model.

Based on the above embodiment, step 220 includes:

Step 410, determine sample data, sample missing data and sample missing position matrix;

It should be noted that the sample data can be expressed as sample traffic data with different data missing ratios, sample missing data and sample traffic data missing in the corresponding sample data, and sample missing location matrix can be used to mark the corresponding sample data. The missing sample traffic data corresponds to the spatiotemporal location, wherein the sample missing data can be used as the label of its corresponding sample data for the final calculation of the loss.

Step 420, extracting the spatiotemporal features of the sample data, and completing the sample data based on the sample spatiotemporal features corresponding to the obtained sample data, to obtain the complementary data corresponding to the sample data;

Step 430: Determine the loss of the completion module based on the completion data corresponding to the sample data, the sample missing data and the sample missing position matrix, and based on the loss of the completion module, iteratively update the parameters of the data completion module to obtain the data completion module. weight parameter.

Specifically, the spatiotemporal feature extraction is performed on the sample data to obtain the sample spatiotemporal feature corresponding to the sample data, and the sample spatiotemporal feature corresponding to the sample data obtains the initial completion result through the convolution layer with the activation layer. Completion network loss function calculates the loss of the completion module based on the initial completion result, sample missing data and sample missing position matrix. Through the completion module loss, the parameters of the data completion module are iteratively updated until the training is completed, and the data complement is obtained. The weight parameter of the full module. Further, the complete network loss function is:

为样本数据的补全数据，m为样本缺失位置矩阵，在计算数据补全的误差时只保留样本缺失位置矩阵中缺失标识的位置的数据参与计算，β为超参数。where x is the sample missing data and

is the completion data of the sample data, m is the sample missing position matrix, when calculating the error of the data completion, only the data of the position of the missing mark in the sample missing position matrix is retained to participate in the calculation, and β is the hyperparameter.

Based on the above embodiment, the steps for determining sample data, sample labeling data and missing position matrix are as follows:

Step 510, determine the original traffic data;

It should be noted that the original traffic data may be represented as a complete sequence of traffic data in a time sequence in a region within a time period. Among them, the traffic data in the traffic data sequence is the data subjected to standardized preprocessing, and further, the standardized preprocessing is processed based on the following formula:

In the formula, x' is the traffic data before normalization, x _max and x _min are the maximum and minimum values of the traffic data in the corresponding traffic data sequence, respectively, x' is the traffic data after normalization, and is based on the missing location coding matrix M. , and fill the position with missing data in x' with -1.

Step 520: Randomly delete traffic data of each preset proportion from the original traffic data, and obtain sample data corresponding to each preset proportion and sample missing data corresponding to each preset proportion;

Step 530: Determine the sample missing position matrix corresponding to each preset proportional quantity based on the spatiotemporal position of the original traffic data and the temporal and spatial position of the sample missing data corresponding to each preset proportional quantity in the original traffic data;

Step 540: Determine sample data, sample missing data, and sample missing position matrix based on sample data corresponding to each preset proportional quantity, sample missing data corresponding to each preset proportional quantity, and sample missing position matrix corresponding to each preset proportional quantity.

Specifically, according to each preset proportion, the traffic data corresponding to the preset proportion is randomly deleted from the original traffic data. In this case, the traffic data deleted from the original traffic data is the sample missing data corresponding to the preset proportion, The traffic data that is not deleted in the original traffic data is the sample data corresponding to the preset proportion. Then, according to the temporal and spatial positions of the original traffic data and the temporal and spatial positions of the sample missing data corresponding to each preset proportion in the original traffic data, determine the sample missing position matrix corresponding to each preset proportion, and finally determine the sample missing position matrix corresponding to each preset proportion. The sample data, the sample missing data corresponding to each preset proportion, and the sample missing position matrix corresponding to each preset proportion, determine the sample data, sample missing data and Sample missing location matrix.

It should be noted that, according to the temporal and spatial positions of the original traffic data and the temporal and spatial positions of the sample missing data corresponding to each preset proportional quantity in the original traffic data, the sample missing position matrix corresponding to each preset proportional quantity may be determined specifically, based on the original traffic data. The spatiotemporal position of the traffic data determines the initial missing position matrix. Based on the spatiotemporal position of the sample missing data corresponding to each preset proportion in the original traffic data, the missing labeling is performed at the index position corresponding to the initial missing position matrix, and each prediction is obtained. Set the sample missing position matrix corresponding to the proportional quantity. Further, each element in the initial missing position matrix is 1, and the element of the index position corresponding to the missing data in the sample missing position matrix is missing and marked as 0.

Further, random deletion is performed in the original traffic data with different preset proportions of 10%, 20%, 30%, 40% and 50%, and 10%, 20%, 30%, 40% and 50% different The sample data and sample missing data corresponding to the preset proportions.

FIG. 6 is the second schematic flowchart of the training method of the prediction completion model provided by the present invention. FIG. 7 is a network structure diagram of the prediction completion model provided by the present invention. As shown in Figure 6 and Figure 7, the training method includes:

Step 610 , according to the originally collected traffic data, perform aggregation in time sequence to obtain the original traffic data, and perform standardization processing on the original traffic data.

Step 620, performing missing data construction on the original traffic data, specifically:

Random deletion in raw traffic data with different preset ratio numbers of 10%, 20%, 30%, 40% and 50%, resulting in different preset ratios of 10%, 20%, 30%, 40% and 50% The number of corresponding sample data and sample missing data, and by 10%, 20%, 30%, 40% and 50% of different preset proportions corresponding to the sample data and sample missing data, construct the corresponding sample missing position matrix.

Step 630: Pre-train the data completion model according to different preset proportions of 10%, 20%, 30%, 40%, and 50% corresponding to sample data, sample missing data, and sample missing position matrix, and train the completed model. The model parameters of the data completion model, namely the weight parameter W ^pre and the bias parameter b ^pre , are transplanted into the data completion module of the initial model.

Step 640, reconstruct the sample data and the complement data of the sample data output by the data complement module of the initial model to obtain the sample reconstruction data, and the specific reconstruction formula is:

X ^* = X · M + X ^im · M _re

In the formula, X and X ^im represent the original missing data and the completed data respectively, M is the sample missing position matrix, which is a 0-1 binary adjacency matrix, M _re is the corresponding inverse code matrix, X ^* is the reconstruction result of the data after completion.

Step 650, applying multi-trends of sub-granularity, time-granularity and day-granularity to reconstruct the result obtained by reconstruction, and obtain reconstructed data of samples corresponding to each time granularity;

基于预测模块中时粒度的时空特征提取网络，对时粒度对应的样本重构数据进行时空特征提取，得到时粒度对应的样本时空特征

Step 660: Apply the reconstruction data of the samples corresponding to each time granularity to the spatiotemporal feature extraction network in the prediction module in the corresponding initial model to perform feature extraction, and obtain the sample spatiotemporal features corresponding to each time granularity. Among them, the sample spatiotemporal features corresponding to the granularity

Based on the spatiotemporal feature extraction network of the time granularity in the prediction module, perform spatiotemporal feature extraction on the sample reconstruction data corresponding to the time granularity, and obtain the sample spatiotemporal features corresponding to the time granularity.

具体公式为：Step 670: Perform adaptive weight-weighted fusion of the sample space-time corresponding to each time granularity to obtain the sample space-time fusion feature

The specific formula is:

where ⊙ represents the Hadamard product, and W _m , W _h and W _d are adaptively adjusted in a data-driven manner.

其中，外部因素特征是先将外部因素进行One-hot编码，并输入到特征嵌入网络中得到的，具体为，将样本数据中不同时间步对应的外部因素输入特征嵌入网络进行编码，具体步骤为将天气、节假日、温度、风速数据以及时间编码输入到环境特征嵌入网络，公式为：Step 680, splicing the sample spatiotemporal fusion features and external factor features to obtain the sample fusion features, and the specific formula for splicing is:

Among them, the external factor features are obtained by first performing One-hot encoding on the external factors and inputting them into the feature embedding network. Specifically, the external factors corresponding to different time steps in the sample data are input into the feature embedding network for encoding. The specific steps are as follows: Input weather, holiday, temperature, wind speed data and time code into the environmental feature embedding network, the formula is:

X _ef =Relu(WX _e +b)

In the formula, X _e represents the external factor, X _ef represents the extracted external factor feature, W represents the preset weight, b represents the preset offset, Relu is the activation function, and its formula is:

where α is a hyperparameter.

Step 690, the sample fusion feature obtains a preliminary prediction result through multi-step regression data, and calculates the loss according to the preliminary prediction result and the expected value of the sample data, and iteratively updates the parameters of the initial model based on the calculated loss to obtain a prediction completion model.

The traffic prediction device provided by the present invention is described below, and the traffic prediction device described below and the traffic prediction method described above can be referred to each other correspondingly.

FIG. 8 is a schematic structural diagram of a traffic prediction device provided by the present invention. As shown in FIG. 8 , the apparatus includes: a determination module 810 and a prediction module 820 .

in,

a determining module 810, configured to determine historical traffic data and a missing location matrix of the historical traffic data;

The prediction module 820 is used to input the historical traffic data and the missing position matrix into the prediction completion model to obtain the prediction result output by the prediction completion model;

The prediction completion model is obtained by training the initial model by applying sample data, sample missing data, sample data expected value and sample missing position matrix on the basis of the weight parameters of the data completion module in the initial model; data completion in the initial model The weight parameters of the module are pre-trained based on sample data, sample missing data and sample missing location matrix.

In the embodiment of the present invention, the determination module is used to determine the historical traffic data and the missing position matrix of the historical traffic data; the prediction module is used to input the historical traffic data and the missing position matrix into the prediction completion model to obtain the prediction supplement The prediction result output by the full model; the prediction completion model is based on the weight parameters of the data completion module in the initial model, and applies the sample data, sample missing data, sample data expected value and sample missing position matrix to train the initial model. ;The weight parameters of the data completion module in the initial model are pre-trained based on sample data, sample missing data and sample missing position matrix, and the prediction completion will be updated according to the output initial prediction results during the training of the prediction completion model. The parameters of the initial model of the network model realize the information interaction between the prediction module and the data completion module in the prediction model, and complete the end-to-end prediction, which improves the accuracy of the prediction results while improving the immediacy.

Based on the above-mentioned embodiment, the traffic prediction device further includes: a prediction completion model training module, the prediction completion model training module includes:

The initial model determination sub-module is used to determine the initial model; the initial model includes a data completion module and a prediction module;

The pre-training sub-module of the completion module is used to pre-train the data completion module based on sample data, sample missing data, sample data expected value and sample missing position matrix, and obtain the weight parameters of the data completion module;

The initial model training sub-module is used to jointly train the data completion module and the prediction module based on the weight parameters, sample data, sample missing data, sample data expected value and sample missing position matrix to obtain a prediction completion model.

Based on any of the above embodiments, the initial model training sub-module includes:

The joint loss calculation sub-module is used to jointly train the data completion module and the prediction module based on the joint training loss function; the joint training function is the difference between the expected value of the sample data and the initial prediction result of the sample data output by the prediction module, and the difference between the sample missing data and the completed data of the sample data output by the data completion module.

Based on any of the above embodiments, the initial prediction sub-module includes:

The reconstruction data determination sub-module determines the sample reconstruction data based on the sample data and the complementary data of the sample data;

The multi-time granularity reconstruction data determination submodule is used to determine the sample reconstruction data corresponding to each time granularity based on the sample reconstruction data and each time granularity;

The feature fusion sub-module is used to reconstruct the data based on the samples corresponding to each time granularity to determine the sample fusion features;

The prediction sub-module is used to fuse the features based on the sample to determine the initial prediction result.

Based on any of the above embodiments, the feature fusion sub-module includes:

The spatiotemporal feature extraction sub-module is used to extract the spatiotemporal features of the sample reconstruction data corresponding to each time granularity, and obtain the sample spatiotemporal features corresponding to each time granularity;

The spatiotemporal feature fusion sub-module is used to determine the sample spatiotemporal fusion features based on the sample spatiotemporal features and adaptive weights corresponding to each time granularity;

The external factor splicing sub-module is used to determine the sample fusion characteristics based on the sample spatiotemporal fusion characteristics and the external factor characteristics; the external factor characteristics include weather factor characteristics and/or time period factor characteristics.

Based on any of the above embodiments, the pre-training sub-module of the completion module includes:

The sample information determination sub-module is used to determine the sample data, the sample missing data and the sample missing position matrix;

The completion data acquisition sub-module is used to extract the spatiotemporal features of the sample data, and based on the sample spatiotemporal features corresponding to the obtained sample data, complete the sample data to obtain the completed data corresponding to the sample data;

The model training sub-module is used to determine the loss of the completion module based on the completion data, sample missing data and sample missing position matrix corresponding to the sample data, and based on the loss of the completion module, iteratively update the parameters of the data completion module to obtain The weight parameter of the data completion module.

Based on any of the above embodiments, the sample information determination submodule includes:

The original data determination sub-module is used to determine the original traffic data;

Each proportional sample information determination submodule is used to randomly delete traffic data of each preset proportion from the original traffic data, and obtain sample data corresponding to each preset proportion and sample missing data corresponding to each preset proportion;

Each proportional sample matrix determination submodule is used to determine the sample missing position matrix corresponding to each preset proportional quantity based on the temporal and spatial position of the original traffic data and the temporal and spatial position of the sample missing data corresponding to each preset proportional quantity in the original traffic data;

The sample information and matrix determination sub-module is used to determine the sample data and sample missing based on the sample data corresponding to each preset proportion, the sample missing data corresponding to each preset proportion, and the sample missing position matrix corresponding to each preset proportion. Data and sample missing location matrix.

FIG. 9 illustrates a schematic diagram of the physical structure of an electronic device. As shown in FIG. 9 , the electronic device may include: a processor (processor) 910, a communication interface (Communications Interface) 920, a memory (memory) 930, and a communication bus 940, The processor 910 , the communication interface 920 , and the memory 930 communicate with each other through the communication bus 940 . The processor 910 may invoke logic instructions in the memory 930 to execute a traffic prediction method, the method comprising: determining historical traffic data and a missing location matrix of the historical traffic data; for inputting the historical traffic data and the missing location matrix to the prediction completion In the model, the prediction result output by the prediction completion model is obtained; the prediction completion model applies sample data, sample missing data, sample data expected value and sample missing position matrix on the basis of the weight parameters of the data completion module in the initial model, to It is obtained by training the initial model; the weight parameters of the data completion module in the initial model are pre-trained based on sample data, sample missing data and sample missing position matrix.

In addition, the above-mentioned logic instructions in the memory 930 can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present invention can be embodied in the form of a software product in essence, or the part that contributes to the prior art or the part of the technical solution. The computer software product is stored in a storage medium, including Several instructions are used to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, Read-Only Memory (ROM, Read-Only Memory), Random Access Memory (RAM, Random Access Memory), magnetic disk or optical disk and other media that can store program codes .

In another aspect, the present invention also provides a computer program product, the computer program product includes a computer program, the computer program can be stored on a non-transitory computer-readable storage medium, and when the computer program is executed by a processor, the computer can Execute the traffic prediction method provided by the above methods, the method includes: determining the historical traffic data and the missing position matrix of the historical traffic data; for inputting the historical traffic data and the missing position matrix into the prediction completion model to obtain the prediction completion The prediction result output by the model; the prediction completion model is obtained by training the initial model by applying sample data, sample missing data, sample data expected value and sample missing position matrix based on the weight parameters of the data completion module in the initial model; The weight parameters of the data completion module in the initial model are pre-trained based on sample data, sample missing data and sample missing location matrix.

In another aspect, the present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored, the computer program is implemented by a processor to execute the traffic prediction method provided by the above methods, and the method includes: Determine the missing location matrix of historical traffic data and historical traffic data; it is used to input historical traffic data and missing location matrix into the prediction completion model, and obtain the prediction result output by the prediction completion model; the prediction completion model contains the data in the initial model Based on the weight parameters of the completion module, the initial model is trained by applying sample data, sample missing data, expected value of sample data and sample missing position matrix; the weight parameters of the data completion module in the initial model are based on sample data, Sample missing data and sample missing location matrix are pretrained.

The device embodiments described above are only illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in One place, or it can be distributed over multiple network elements. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution in this embodiment. Those of ordinary skill in the art can understand and implement it without creative effort.

From the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and certainly can also be implemented by hardware. Based on this understanding, the above-mentioned technical solutions can be embodied in the form of software products in essence or the parts that make contributions to the prior art, and the computer software products can be stored in computer-readable storage media, such as ROM/RAM, magnetic A disc, an optical disc, etc., includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to perform the methods described in various embodiments or some parts of the embodiments.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, but not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that it can still be The technical solutions described in the foregoing embodiments are modified, or some technical features thereof are equivalently replaced; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (10)

1. A traffic prediction method, comprising:

determining historical traffic data and a missing position matrix of the historical traffic data;

inputting the historical traffic data and the missing position matrix into a prediction completion model to obtain a prediction result output by the prediction completion model;

the prediction completion model is obtained by training the initial model by applying sample data, sample missing data, a sample data expected value and a sample missing position matrix on the basis of a weight parameter of a data completion module in the initial model; and the weight parameters of the data completion module in the initial model are obtained by pre-training based on the sample data, the sample missing data and the sample missing position matrix.

2. The traffic prediction method according to claim 1, wherein the training step of the predictive completion model comprises:

determining the initial model; the initial model comprises a data completion module and a prediction module;

pre-training the data completion module based on the sample data, the sample missing data, the sample data expected value and the sample missing position matrix to obtain a weight parameter of the data completion module;

and performing joint training on the data completion module and the prediction module based on the weight parameter, the sample data, the sample missing data, the sample data expected value and the sample missing position matrix to obtain the prediction completion model.

3. The traffic prediction method of claim 2, wherein the jointly training the data completion module and the prediction module comprises:

performing joint training on the data completion module and the prediction module based on a joint training loss function; the joint training function is constructed by the difference between the expected value of the sample data and the initial prediction result of the sample data output by the prediction module and the difference between the missing sample data and the complement data of the sample data output by the data complement module.

4. The traffic prediction method of claim 3, wherein the step of obtaining the initial prediction result comprises:

determining sample reconstruction data based on the sample data and the complementary data of the sample data;

determining sample reconstruction data corresponding to each time granularity based on the sample reconstruction data and each time granularity;

determining sample fusion characteristics based on the sample reconstruction data corresponding to each time granularity;

and determining to obtain the initial prediction result based on the sample fusion characteristics.

5. The traffic prediction method according to claim 4, wherein the determining a sample fusion feature based on the sample reconstruction data corresponding to each time granularity comprises:

performing space-time feature extraction on the sample reconstruction data corresponding to each time granularity to obtain sample space-time features corresponding to each time granularity;

determining sample space-time fusion characteristics based on the sample space-time characteristics and the self-adaptive weight corresponding to each time granularity;

determining the sample fusion characteristics based on the sample spatio-temporal fusion characteristics and the external factor characteristics; the external factor characteristic includes a weather factor characteristic, and/or a time period factor characteristic.

6. The traffic prediction method according to any one of claims 2 to 5, wherein the pre-training the data completion module based on the sample data, the missing sample data, the expected sample data value, and the missing sample position matrix to obtain the weight parameter of the data completion module comprises:

determining the sample data, the sample missing data and the sample missing position matrix;

performing space-time feature extraction on the sample data, and completing the sample data based on the obtained sample space-time feature corresponding to the sample data to obtain complete data corresponding to the sample data;

and determining the loss of a completion module based on completion data corresponding to the sample data, the missing sample data and the missing sample position matrix, and iteratively updating the parameters of the data completion module based on the loss of the completion module to obtain the weight parameters of the data completion module.

7. The traffic prediction method of claim 6, wherein the sample data, the missing sample data and the missing sample location matrix are determined by:

determining original traffic data;

randomly deleting traffic data with each preset proportion quantity from the original traffic data to obtain sample data corresponding to each preset proportion quantity and sample missing data corresponding to each preset proportion quantity;

determining a sample missing position matrix corresponding to each preset proportion quantity based on the space-time position of the original traffic data and the space-time position of the sample missing data corresponding to each preset proportion quantity in the original traffic data;

and determining the sample data, the sample missing data and the sample missing position matrix based on the sample data corresponding to each preset proportion quantity, the sample missing data corresponding to each preset proportion quantity and the sample missing position matrix corresponding to each preset proportion quantity.

8. A traffic prediction apparatus, comprising:

a determination module for determining historical traffic data and a missing location matrix of the historical traffic data;

the prediction module is used for inputting the historical traffic data and the missing position matrix into a prediction completion model to obtain a prediction result output by the prediction completion model;

the prediction completion model is obtained by training the initial model by applying sample data, sample missing data, a sample data expected value and a sample missing position matrix on the basis of a weight parameter of a data completion module in the initial model; and the weight parameters of the data completion module in the initial model are obtained by pre-training based on the sample data, the sample missing data and the sample missing position matrix.

9. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor when executing the program implements the traffic prediction method according to any of claims 1 to 7.

10. A non-transitory computer-readable storage medium having stored thereon a computer program, which when executed by a processor implements the traffic prediction method according to any one of claims 1 to 7.

Images