CN110837888A - Traffic missing data completion method based on bidirectional cyclic neural network - Google Patents

Abstract

The invention provides a traffic missing data completion method based on a bidirectional cyclic neural network, which belongs to the field of traffic. This method firstly utilizes the time-series characteristics of data, and also considers the impact of data before and after the completion time point on the current time point, which greatly improves the utilization and completion accuracy of data, and secondly considers the external The influence of features and adjacent sensor data on the current sensor data is added to the completion model, which greatly improves the completion accuracy. The method of the present invention not only greatly improves the completion accuracy in the case of a low data missing rate, but also improves the completion accuracy in the case of a high data missing rate.

Description

A Complete Method for Traffic Missing Data Based on Bidirectional Recurrent Neural Network

technical field

The invention belongs to the field of traffic, and in particular relates to a traffic missing data completion method based on a bidirectional cyclic neural network.

Background technique

The traffic flow data of road coils have periodicity, time series and trend. At this stage, the method for completing traffic flow data is mainly based on its timing.

Based on the completion of time series traffic flow data, the data of a period of time before the current missing point is taken, and the missing point data is completed through the neural network. For example, if you want to complete the traffic flow data at 16:00 today, then take the data from 8:00 to 15:00 that day as input, and get the data at the next time point - 16:00 through the recurrent neural network. This historical data-based completion method makes good use of the time-series characteristics of data for completion, and the completion results are relatively good, but this method has limitations. When a special event occurs, the current missing point is also preceded by a series of missing points, such as a power outage, which will result in the loss of a continuous segment of data. The full effect is very poor in this case.

The neural network was originally inspired by the biological nervous system and appeared in order to simulate the biological nervous system. It is composed of a large number of nodes (or neurons) connected with each other. The neural network adjusts the weights according to changes in the input, improves the behavior of the system, and automatically learns a model that can solve the problem. LSTM (Long Short-Term Memory Network) is a special form of RNN (Recurrent Neural Network), which can effectively solve the gradient disappearance and gradient explosion problems of multi-layer neural network training, and can handle long-term time-dependent sequences. LSTM can capture the time series characteristics of charging data, and the use of LSTM model can effectively improve the completion accuracy.

An LSTM network consists of LSTM cells, which consist of cells, input gates, output gates, and forget gates.

Forget gate: decides how much information to discard from the output state of the previous unit, the formula is as follows:

f _t =σ _g (W _f x _t +U _f h _t-1 +b _f )

where f _t is the output of the forget gate, x _t is the input sequence, h _t-1 is the output of the previous unit, σ _g represents the sigmoid function, W _f represents the input weight parameter matrix, and U _f represents the output of the previous unit Weight parameter matrix, b _f represents the bias parameter vector.

Input gate: Determine how much new information is added to the Cell state, and update the cell state C, the formula is as follows:

i _t =σ _g (W _i x _t +U _i h _t-1 +b _i )

表示矩阵乘积，W_i表示输入的权重参数矩阵，U_i表示上一个单元输出的权重参数矩阵，b_i表示偏差参数向量，f_t是遗忘门的输出，c_t-1是上一个单元的单元状态，表示矩阵乘积，W_c表示输入的权重参数矩阵，U_c表示上一个单元输出的权重参数矩阵， b_c表示偏差参数向量。where c _t represents the cell state of the current cell, σ _g and σ _c represent the sigmoid function,

represents the matrix product, Wi represents the input weight parameter matrix, U _i represents the weight parameter matrix output by the previous unit, _{bi represents} the bias parameter vector, f _t is the output of the forget gate, and c _t-1 is the unit of the previous unit state, represents the matrix product, W _c represents the input weight parameter matrix, U _c represents the weight parameter matrix output by the previous unit, and b _c represents the bias parameter vector.

Output gates: output results based on the current cell state.

o _t =σ _g (W _o x _t +U _o h _t-1 +b _o )

表示矩阵乘积， W_o表示输入的权重参数矩阵，U_o表示上一个单元输出的权重参数矩阵，b_o表示偏差参数向量。where h _t represents the output of the current unit, σ _g and σ _h represent the sigmoid function,

represents the matrix product, W _o represents the input weight parameter matrix, U _o represents the weight parameter matrix output by the previous unit, and b _o represents the bias parameter vector.

SUMMARY OF THE INVENTION

The invention proposes a traffic missing data completion method based on a bidirectional cyclic neural network, which is a deep learning completion method based on time series, periodicity and space, and aims to improve the completion accuracy of road traffic flow data.

Technical scheme of the present invention:

A traffic missing data completion method based on bidirectional recurrent neural network, the steps are as follows:

The first step is to preprocess the traffic flow data

The preprocessing includes time granularity division and data standardization;

The second step is to perform random data point loss processing on the preprocessed data, construct a data set with missing points, and then record the location information of the missing points as a verification value to verify the completion effect of the method.

At the same time, the construction time dimension affects the decay matrix. Due to the lack of data, there will be continuous deletion. For example, the damage of the power supply element of the sensor will lead to the loss of data for a period of time. With the accumulation of time, the impact of historical data on the data of the missing points will become smaller and smaller. It affects the completion accuracy, so it is necessary to record the attenuation of the influence of time dimension data. The time dimension influence decay matrix is defined as follows:

Among them, n _t represents the current moment, is defined as follows:

The third step is to divide the lost traffic data into training set, validation set and test set. In each dataset, the data used by the different models are of the following types:

Data used by the forward time series deep learning module:

Data for the reverse time series deep learning module:

External feature data used in the external feature module: F _n ;

Periodic sequence data used in periodic feature module:

表示包括第n时刻的前t个时刻的车流量数据的集合，

表示包括第n时刻当天的前p天日内相同时刻的车流量数据集合，F_n表示在第n时刻的外部特征，包括节假日、位置区域、天气和气温。Among them, n represents the current moment, t represents the step size of the time series, and p represents the step size of the periodic sequence. S represents the traffic flow data, and T represents the reverse sequence of S in the time dimension. s _i represents the traffic flow data at time n, represents the traffic flow data at the same time in the day i days before the nth time,

represents the set of traffic flow data including the first t moments of the nth moment,

Represents the traffic flow data set at the same time in the p days before the nth time, and _Fn represents the external features at the nth time, including holidays, location area, weather and temperature.

The fourth step is to build a completion model. The completion model includes a forward time series deep learning module, a reverse time series deep learning module, a periodic feature module and an external feature module. The structure and training mechanism of each module are as follows:

(1) Forward time series deep learning module: It is a combination of a linear regression network and a multi-layer long-short-term memory network LSTM model. Through a layer of linear regression network, the continuity information of the current missing points in time is added to deal with long-term In the case of missing sequences, the completion accuracy is improved.

输入LSTM网络中，对当前时刻输入值x_t，如果数据点没有缺失，则直接输入，当数据点缺失时，将上一个时刻的隐含状态作为当前时刻的输入，在处理完输入后，对深度学习网络进行训练，在不断的迭代更新中得到最终的前向序列深度学习模块的输出。The implementation details of the forward sequence deep learning module: first input the time dimension decay matrix into the linear regression network, and then combine the output of the linear regression network with the forward time series data

In the LSTM network, input the value x _t at the current moment. If the data point is not missing, input it directly. When the data point is missing, use the hidden state of the previous moment as the input at the current moment. The deep learning network is trained, and the output of the final forward sequence deep learning module is obtained in the continuous iterative update.

(2) Reverse time series deep learning module: The network structure is consistent with the forward series deep learning module, the difference is that the input of the forward time series deep learning module is reversely processed in the time dimension, as the module's input enter.

(3) Periodic feature learning module: It is a module composed of a three-layer fully connected network. Through the extraction of periodic data features, the change rule of traffic flow in the historical data, the same sensor, and the same time period is obtained, and then the Extracted feature output. Implementation details: The periodic sequence data is input into the fully connected layer, and after three fully connected layers, the time series features of the periodic data are extracted, and then output.

(4) External feature module: It consists of two parts: the first part deals with holiday and weather features, and is a feature encoding layer. Implementation details: Input the external feature data into the feature encoding layer, convert the data into vector form, and then combine the obtained vector with the outputs of the above three modules.

The second part deals with spatial features. In order to take into account the information on the road space, all the sensors on the road segment are input into the second part at the same time, and then the hidden states of other sensors at the same time as the missing point of the current sensor are used as input, and after calculating the weight through the Softmax network, we get output, which is fed into the forward and reverse time series deep learning modules.

Finally, the outputs of the above four modules are combined into a one-dimensional vector, and the final completion result is obtained through a layer of fully connected network.

Step 5: Use the training set data to pre-train the pre-training part of the forward time series deep learning module and the reverse time series deep learning module, optimize the parameters of the time series deep learning model in advance, and avoid optimizing the parameters during the overall training. to the local optimum.

Step 6: Use the training set data and the validation set data to perform overall training on the four modules established in Step 4:

The preprocessed data are input into the corresponding modules respectively, and the overall training is performed on all modules at the same time. Calculate the loss function value of the complement value after each training and the true value of the traffic flow data, and train the parameters of the model to the target value. According to the effect of the model on the training set and validation set, the hyperparameters of the model are continuously debugged, and the completion accuracy is improved under the condition of reducing overfitting.

(前t₁小时的车流量数据)、反向时间序列数据

(后t₂小时的车流量数据)、周期序列数据

(前t₃天同一时刻的车流量数据)、时间维度影响衰减性矩阵

外部特征数据F_n(第n时刻的节假日、区域、天气和气温外部特征数据)和车流量数据的真值

(当前时刻的车流量数据)。The input data includes: forward time series data

(traffic data for the first t ₁ hour), reverse time series data

(traffic flow data in the last t ₂ hours), periodic series data

(traffic flow data at the same time in the first t ₃ days), time dimension influence attenuation matrix

The true value of external feature data _Fn (holiday, area, weather and temperature external feature data at the nth time) and traffic flow data

(traffic data at the current moment).

After one iteration, the traffic flow data after one completion operation is obtained. The data after this iteration is used as the input of the next iteration. Although the missing points have completed values before, but because the labels still indicate missing, in the subsequent iteration process, the goal is to complete the data for these missing points, but due to the The existence of data that is relatively close to the true value provides prior knowledge, which can improve the speed of model convergence and completion accuracy.

The seventh step is to use the test set to use the model trained in the sixth step to complete the traffic flow data.

反向时间序列数据周期序列数据时间维度影响衰减性矩阵

外部特征数据

和车流量数据的真值

The input data is: forward time series data

reverse time series data Periodic sequence data Time Dimension Influence Decay Matrix

External feature data

and the ground truth of the traffic flow data

The complement value of the missing traffic flow data is obtained through the model in the sixth step, and is compared with the verification value obtained after the loss processing in the second step to verify the complement effect of the model.

In the first step, the specific process of preprocessing is:

(1) Time granularity division: All traffic flow data are processed into traffic flow data per k minutes according to the time granularity of k minutes;

(2) Standardize the data: Use the minimum and maximum values to standardize the traffic flow data, the formula is as follows:

Among them, x represents the original value, x _min represents the minimum value of the original value, x _max represents the maximum value of the original value, max is the upper limit of normalization, min is the lower limit of normalization, [min,max] Indicates the normalized interval, and x ^* is the normalized result.

In the fourth step, consider the road space information part (Softmax processing): set the hidden states of all sensors at the current moment h=<h ₁ , h ₂ , h ₃ ,..., _hi ,..., h _t >, h _i is the hidden state of the i-th sensor at the current moment, and then the weight is calculated for each h _i to obtain the new hidden state h′ _i of the current sensor.

After processing with Softmax, all weights sum to 1. Among them, l represents the number of sensors, and h _ij represents the implicit state of the jth sensor i at the moment.

In the sixth step, the mean square error MAE between the complemented data obtained in each iteration and the true value of the traffic flow data is calculated, and the Adam method is used to minimize the MAE.

Among them, x′ _i represents the real value of the sensor at the ith moment, and _xi represents the sensor complement value at the ith moment.

Beneficial effects of the present invention: The difference between the present invention and the existing method is that, first of all, it is an improvement in the use of the time-series characteristics of data. When using the time-series characteristics of data, the previous method often considers the difference between historical data and current time point data. However, in the application of the completion of traffic flow data, the information of the subsequent time point has an impact on the data of the current time point. The present invention simultaneously considers the forward time series and the reverse time series, which greatly improves the completion accuracy. Secondly, considering the influence of external characteristics, holidays and sensor adjacent areas on traffic flow data, it is added to the completion model, which greatly improves the completion accuracy and the completion of special values. Finally, the attenuation of the impact of missing data in the time dimension is also considered, which improves the completion accuracy. The method of the invention not only greatly improves the completion accuracy of the low missing rate traffic flow data, but also can achieve a good completion effect under the condition of a high data missing rate.

Description of drawings

FIG. 1 is a structural diagram of a complementation model according to the present invention.

Figure 2 is a comparison diagram of the low-missing rate completion result with a data missing rate of 20% and the true value.

Figure 3 is a comparison diagram of the high missing rate completion result with a data missing rate of 50% and the true value.

Specific implementation method

The technical solutions of the present invention will be further described below with reference to specific embodiments and accompanying drawings.

A traffic missing data completion method based on bidirectional recurrent neural network, the steps are as follows:

The first step is to preprocess the traffic flow data

(1) Time granularity division: All traffic flow data are processed into traffic flow data every 5 minutes according to the time granularity of 5 minutes;

(2) Standardize the data: Use the minimum and maximum values to standardize the traffic flow data, the formula is as follows:

Among them, x represents the original value, x _min represents the minimum value of the original value, x _max represents the maximum value of the original value, max is the upper limit of normalization, min is the lower limit of normalization, [min,max] Indicates the normalized interval, and x ^* is the normalized result.

The second step is to lose random data points in the preprocessed data. Using the method of random numbers, a certain proportion of the data (set by yourself according to the experimental requirements) is labeled as missing points, which are used as missing points, and then record the data of these points. The value, as the true value, is used to verify the final completion effect of the model.

At the same time, the time dimension influence decay matrix is established. Due to the lack of data, there will be continuous loss. For example, a power outage may cause the sensor to fail to collect data within a few hours. With the accumulation of time, the impact of historical data on the data of missing points will become smaller and smaller. , which will affect the completion accuracy, so it is necessary to record the attenuation of the influence of time dimension data. The time dimension influence decay matrix is defined as follows:

的定义如下：Among them, n _t represents the current moment,

is defined as follows:

The third step is to divide the preprocessed traffic flow data into training set, validation set and test set, and divide them according to the ratio of 8:1:1. In each dataset, the data used by the different models are of the following types:

Data used by the forward time series deep learning module:

Data for the reverse time series deep learning module:

External feature data used in the external feature model: F _n ;

Periodic sequence data used in periodic feature module:

表示第n时刻的前i天的日内相同时刻的车流量数据，表示包括第n时刻的前t个时刻的车流量数据的集合，

表示包括第n时刻当天的前p天日内相同时刻的车流量数据集合，F_n表示在第n时刻的外部特征，包括节假日、位置区域、天气和气温。Among them, n represents the current moment, t represents the step size of the time series, and p represents the step size of the periodic sequence. S represents the traffic flow data, and T represents the reverse sequence of S in the time dimension. s _i represents the traffic flow data at time n,

represents the traffic flow data at the same time in the day i days before the nth time, represents the set of traffic flow data including the first t moments of the nth moment,

Represents the traffic flow data set at the same time in the p days before the nth time, and _Fn represents the external features at the nth time, including holidays, location area, weather and temperature.

The fourth step is to build a completion model. The completion model includes a forward sequence deep learning module, a reverse time series deep learning module, a periodic feature module and an external feature module. The structure and training mechanism of each module are as follows:

(1) Forward sequence deep learning module: It is a combination of a linear regression network and a multi-layer long-short-term memory network LSTM model. Through a layer of linear regression network, the temporal continuity information of the current missing point is added to deal with long-term sequences. In the case of missing, improve the completion accuracy.

输入LSTM网络中，对当前时刻输入值x_t，如果数据点没有缺失，则直接输入，当数据点缺失时，将上一个时刻的隐含状态作为当前时刻的输入，在处理完输入后，对深度学习网络进行训练，在不断的迭代更新中得到最终的前向序列深度学习模块的输出。The implementation details of the forward sequence deep learning module: first input the time dimension decay matrix into the linear regression network, and then combine the output of the linear regression network with the forward time series data

In the LSTM network, input the value x _t at the current moment. If the data point is not missing, input it directly. When the data point is missing, use the hidden state of the previous moment as the input at the current moment. The deep learning network is trained, and the output of the final forward sequence deep learning module is obtained in the continuous iterative update.

(2) Reverse sequence deep learning module: The network structure is consistent with the forward sequence deep learning module, the difference is that the input of the forward sequence deep learning module is reversely processed in the time dimension as the input of the module.

(3) Periodic feature module: It is a module composed of a three-layer fully connected network. Through the extraction of periodic data features, it can obtain the change law of the traffic flow of the same sensor and the same time period in the historical data, and then extract the to the feature output. Implementation details: The periodic sequence data is input into the fully connected layer, and after three fully connected layers, the time series features of the periodic data are extracted, and then output.

(4) External feature module: it is a layer of feature coding layer; implementation details: input external feature data into the feature coding layer, and describe the weather, holidays and other external features in text, through the method of grading: for example, according to whether it is a holiday or not , which will be represented by 1 for holidays and 0 for non-holidays, convert the periodic sequence data into vector form, and then output the resulting vector to the next step.

In order to take into account the information on the road space, a spatial feature learning module is also added to input all sensors on the road segment into the model at the same time, and then use the hidden state of other sensors at the same time as the missing point of the current sensor as input, through After the Softmax network calculates the weights, the output is obtained, which will be input to the forward sequence module and the reverse sequence module.

Finally, the outputs of each module are combined into a one-dimensional vector, and then the final completion result is obtained through a layer of fully connected network.

The fifth step is to use the training set data to pre-train the pre-training part of the time series deep learning model, optimize the parameters of the time series deep learning model in advance, and avoid optimizing the parameters to the local optimum during the overall training.

Step 6: Use the training set data and the validation set data to perform overall training on the four modules established in Step 4 (replace the points with missing data with complementary values, and keep the original data unchanged if there is no missing data):

The preprocessed data are input into the corresponding modules respectively, and the overall training is performed on all modules at the same time. Calculate the loss function value of the complement value after each training and the true value of the traffic flow data, and train the parameters of the model to the target value. According to the effect of the model on the training set and validation set, the hyperparameters of the model are continuously debugged, and the completion accuracy is improved under the condition of reducing overfitting. During the training process, the MAE (mean square error) of the complemented data obtained in each iteration and the true value of the traffic flow data is calculated, and the Adam method is used to minimize the MAE.

Among them, x′ _i represents the real value of the sensor at the ith moment, and _xi represents the sensor complement value at the ith moment.

(前t₁小时的车流量数据)、反向时间序列数据

(后t₂小时的车流量数据)、时间维度影响衰减性矩阵

周期序列数据

(前t₃天同一时刻的车流量数据)、外部特征数据F_n(第n时刻的节假日、区域、天气和气温外部特征数据)和车流量数据的真值

(当前时刻的车流量数据)。The input data includes: forward time series data

(traffic data for the first t ₁ hour), reverse time series data

(traffic flow data in the last t ₂ hours), time dimension influence attenuation matrix

Periodic sequence data

(traffic flow data at the same time in the previous t ₃ days), external feature data _Fn (holidays, area, weather and temperature external feature data at the nth time) and the true value of traffic flow data

(traffic flow data at the current moment).

The seventh step is to use the test set to use the model trained in the sixth step to complete the traffic flow data.

反向时间序列数据

周期序列数据

外部特征数据

和车流量数据的真值

时间维度影响衰减性矩阵

The input data is: forward time series data

reverse time series data

Periodic sequence data

External feature data

and the ground truth of the traffic flow data

Time Dimension Influence Decay Matrix

Figure 2 is a comparison chart of the completion result with the data missing rate of 20% and the real value. The mean square error MAE between the model completion result and the real value of traffic flow is 29.18. (Select the first 100 missing points in the figure)

Figure 3 is a comparison diagram of the completion result with the data missing rate of 50% and the real value. The mean square error MAE between the model completion result and the real value of traffic flow is 31.94. (The first 100 missing points are selected in the figure).

Claims (5)

1. a traffic missing data completion method based on bidirectional recurrent neural network, is characterized in that, step is as follows: The first step is to preprocess the traffic flow data The preprocessing includes time granularity division and data standardization; The second step is to perform random data point loss processing on the preprocessed data, construct a data set with missing points, and then record the location information of the missing points as the verification value; at the same time, construct the time dimension influence attenuation matrix:

Among them, n _t represents the current moment,

is defined as follows:

The third step is to divide the traffic flow data after loss processing into training set, validation set and test set; in each data set, the data used by different models have the following types: Data used by the forward time series deep learning module:

Data for the reverse time series deep learning module:

External feature data used in the external feature module: F _n ; Periodic sequence data used in periodic feature module:

Among them, _n represents the current moment, t represents the step size of the time series, p represents the step size of the periodic sequence; S represents the traffic flow data, T represents the reverse sequence of S in the time dimension; time traffic data,

represents the traffic flow data at the same time in the day i days before the nth time,

represents the set of traffic flow data including the first t moments of the nth moment,

Represents the traffic flow data set at the same time in the p days before the nth time, and _Fn represents the external characteristics at the nth time, including holidays, location area, weather and temperature; The fourth step is to build a completion model. The completion model includes a forward time series deep learning module, a reverse time series deep learning module, a periodic feature module and an external feature module. The structure and training mechanism of each module are as follows: (1) Forward time series deep learning module: It is a combination of a linear regression network and a multi-layer long-short-term memory network LSTM model. Through a layer of linear regression network, the continuity information of the current missing points in time is added to deal with long-term In the case of missing sequences, the completion accuracy is improved; The implementation details of the forward sequence deep learning module: first input the time dimension decay matrix into the linear regression network, and then combine the output of the linear regression network with the forward time series data

In the LSTM network, input the value x _t at the current moment. If the data point is not missing, input it directly. When the data point is missing, use the hidden state of the previous moment as the input at the current moment. The deep learning network is trained, and the output of the final forward sequence deep learning module is obtained in the continuous iterative update; (2) Reverse time series deep learning module: The network structure is consistent with the forward series deep learning module, the difference is that the input of the forward time series deep learning module is reversely processed in the time dimension, as the module's input enter; (3) Periodic feature learning module: It is a module composed of a three-layer fully connected network. Through the extraction of periodic data features, the change rule of traffic flow in the historical data, the same sensor, and the same time period is obtained, and then the Extracted feature output; implementation details: input the periodic sequence data into the fully connected layer, through three fully connected layers, extract the time series features of the periodic data, and then output; (4) External feature module: This module consists of two parts: the first part deals with holiday and weather features, and is a feature coding layer; implementation details: input external feature data into the feature coding layer, convert the data into vector form, and then Combine the obtained vector with the output of the above three modules; The second part deals with spatial features. All sensors on the road section are input into the second part at the same time, and then the hidden state of other sensors at the same time as the missing point of the current sensor is used as input. After calculating the weight through the Softmax network, the output is obtained, Input this output into the forward and reverse time series deep learning modules; Finally, the outputs of the above four modules are combined into a one-dimensional vector, and the final completion result is obtained through a layer of fully connected network; Step 5: Use the training set data to pre-train the pre-training part of the forward time series deep learning module and the reverse time series deep learning module, optimize the parameters of the time series deep learning model in advance, and avoid optimizing the parameters during the overall training. to the local optimum; Step 6: Use the training set data and the validation set data to perform overall training on the four modules established in Step 4: Input the preprocessed data into the corresponding modules respectively, and conduct overall training for all modules at the same time; Target value; according to the effect of the model on the training set and validation set, continuously debug the hyperparameters of the model, and improve the completion accuracy under the condition of reducing overfitting; The input data includes: Forward time series data: traffic flow data for the first t ₁ hour

Reverse time series data: traffic flow data for the last t ₂ hours

Periodic series data: traffic flow data at the same time in the first t ₃ days The time dimension affects the decay matrix:

External feature data: holiday, area, weather and temperature external feature data F _n at the nth time; The true value of the traffic flow data: the traffic flow data at the current moment

After one iteration, the traffic flow data obtained after one completion operation is obtained; the data after this iteration is used as the input of the next iteration. Although the missing points have completed values before, but because the labels still indicate missing, In the subsequent iteration process, the goal is to complete data for these missing points; The seventh step is to use the test set to complete the traffic flow data with the model trained in the sixth step; The input data is: forward time series data reverse time series data

Periodic sequence data

Time Dimension Influence Decay Matrix

External feature data

and the ground truth of the traffic flow data

The complement value of the missing traffic flow data is obtained through the model in the sixth step, and is compared with the verification value obtained after the loss processing in the second step to verify the complement effect of the model. 2. a kind of traffic missing data completion method based on bidirectional cyclic neural network according to claim 1, is characterized in that, in the described first step, the concrete process of preprocessing is: (1) Time granularity division: All traffic flow data are processed into traffic flow data per k minutes according to the time granularity of k minutes; (2) Standardize the data: Use the minimum and maximum values to standardize the traffic flow data, the formula is as follows:

Among them, x represents the original value, x _min represents the minimum value of the original value, x _max represents the maximum value of the original value, max is the upper limit of normalization, min is the lower limit of normalization, [min,max] Indicates the normalized interval, and x ^* is the normalized result. 3. a kind of traffic missing data completion method based on bidirectional recurrent neural network according to claim 1 or 2, it is characterized in that, in described 4th step, the concrete process of processing spatial characteristic: set all sensors in the current The implicit state at the moment h=<h ₁ , h ₂ , h ₃ , ..., hi , ..., h _t >, _hi is the hidden state of the _ith sensor at the current moment, and then calculates for each _hi weight, get the new hidden state h′ _i of the current sensor;

Among them, l represents the number of sensors, and h _ij represents the implicit state of the jth sensor i at the moment. 4. a kind of traffic missing data completion method based on bidirectional cyclic neural network according to claim 1 and 2, is characterized in that, in the described 6th step, calculate the data after the completion obtained by each iteration and the mean square error MAE of the true value of the traffic flow data, using the Adam method to minimize the MAE; Among them, x′ _i represents the real value of the sensor at the ith moment, and _xi represents the sensor complement value at the ith moment. 5. a kind of traffic missing data completion method based on bidirectional cyclic neural network according to claim 3, is characterized in that, in described 6th step, calculate the data and vehicle after the completion obtained by each iteration The mean square error MAE of the true value of the flow data, using the Adam method to minimize the MAE;

Among them, x′ _i represents the real value of the sensor at the ith moment, and _xi represents the sensor complement value at the ith moment.

Images