CN111709553B - Subway flow prediction method based on tensor GRU neural network - Google Patents

Abstract

The invention discloses a subway flow prediction method based on tensor GRU neural network, which comprises the steps of counting original data, filling the data according to an output time period, dividing the data into a training set and a testing set, and inputting the training set and the testing set into the network day by day according to a date sequence; current inputX ^t Output from the last timeH ^t‑1 Generating an output at that time via operations of the GRU update gate and the reset gateH ^t The method comprises the steps of carrying out a first treatment on the surface of the Will last momentH ^t Performing loss calculation and back propagation with the normalized label, and outputting the back normalized label in the final step of epochH ^t As the final output of the network. The invention enables the network to process higher order data and obtain higher order results, so that the network can grasp the structural characteristics in the predicted data results.

Description

A Metro Flow Forecasting Method Based on Tensor GRU Neural Network

technical field

The invention relates to a subway flow prediction method, in particular to a subway flow prediction method based on a tensor GRU neural network.

Background technique

The traffic monitoring of every city hopes to obtain real-time and accurate traffic in various regions, and then conduct inter-regional traffic analysis based on these existing traffic data, which will help to grasp the traffic conditions of the entire city and enrich the decision-making ability of the city brain . The data generated at each moment in the city is diverse, and the data can be the density of the crowd, the flow of traffic, or even the gender of an individual. This paper proposes a method for predicting subway traffic using tensor GRU (an evolution of RNN) network for the subway monitoring scenario.

Subway traffic is based on a comprehensive evaluation of the number of people entering and leaving each subway station along each subway line in the city in a time unit. In order to predict the direction of the subway flow in a period of time, other data of the subway flow, such as the location of the subway station, the time period of the subway station, etc., need to be considered. This information is naturally structural data. Traditional methods often split the data of each dimension for feature extraction, and finally perform feature fusion, which directly damages the structural relationship in the data. Therefore, the tensor operation method is adopted, and the neural network matrix operation is changed to full tensor operation. According to the data characteristics of the subway, the time series network is used to accept the input of different time periods, and finally the position of each element in the data is considered according to the tensor distance. In order to achieve the purpose of directly using its data structure, it can also expand the high-level thinking to the output end, so that the output end is no longer limited to the judgment of the vector level, and relies on the calculation of the subscript of the element in the tensor distance to achieve the grasp The purpose of tensor structure features, so that the output can also achieve high-level expansion, which is of great significance in predicting different results under different development trends in the future.

Contents of the invention

Aiming at the deficiencies of the prior art, the present invention proposes a subway flow prediction method based on a tensor GRU neural network.

The technical problem to be solved by the present invention is how to use the tensor GRU to solve the problem of predicting the flow of people in the subway under the scenario of a high-order data set.

In order to solve this problem, the present invention is realized through the following technical solutions:

The subway flow prediction method based on tensor GRU includes the following steps:

1) Take the specific data of card swiping at each subway station at each time period in a certain period of time as a data set, clean and filter the data set, find the subway station, subway line, and holiday type as the dimensions of each level of data, and form a high-level Zhang quantity.

2) Before entering the GRU neural network, normalize each input according to the order, and at the same time ensure that the label data and the input data share the same set of mean variance; according to the dimension of the network output, z-score normalizes the input data to obtain the mean value Tensor M and standard deviation tensor St ; K ₁ ×K ₂ ×…×K _N -shaped tensors of order N are expanded according to order.

3) Divide the data processed in step 2) into each time step X and input it to the GRU neural network. In each tensor GRU operation unit, the tensor of the hidden layer is obtained by multiplying the weight W and X by Einstein Quantity state, the specific expression of the Einstein product is:

Among them, A and B represent an N-order tensor and an N+M-order tensor respectively, k _N represents the dimension size of the N-th order of the tensor, and the corresponding subscript elements will be sequentially processed from k ₁ to k _N in order of order Multiply and then perform a summation operation to reduce the N-order to 1-order so that the result is an M-order tensor; the final hidden layer state H is obtained through the operation of the update gate and reset gate inside the tensor GRU, which is used as an output or as the next Input for a time step. After the calculation of the specified time step, the resulting tensor H is obtained, which is the obtained hidden layer state H.

4) Calculate the loss of the obtained output H and the labels that are also normalized before input, and use the tensor distance as the loss function to capture the structural characteristics of high-order tensors:

表示对应下标的元素值的欧式距离，通过反向传播机制更新每一个时刻的W_i 权重张量,i代表输出的某一个时间步，Where l and m represent the subscripts of the two tensors respectively, x _l , y _l , x _m , y _m represent the element values of the corresponding subscripts of the two tensors respectively,

Represents the Euclidean distance of the element value corresponding to the subscript, and updates the W _i weight tensor at each moment through the back propagation mechanism, i represents a certain time step of the output,

For different epochs, repeat steps 3) to 4) until the end of the epoch before performing the next step.

5) When the epoch is over, obtain the result of step 3), denormalize the output of the network according to M and St , and obtain the effective dimension in the denormalized tensor as the final predicted result value of the network, that is, get The final traffic forecast.

Effect that the present invention has relative to prior art:

1. The neural network based on vector and matrix operations can be operated based on the tensor structure, and the dimension of the neural network can be improved. In the application of subway traffic, high-level information can be input and high-level data fusion can be performed. The computing power is vector Incomparable.

2. In the case of the tensor distance as the loss function, the characteristics of the structure can be learned during the network learning process, and the output results of the traffic can predict the multi-modal scene prediction value, broadening the network prediction or regression dimension boundaries.

Description of drawings

Figure 1 is a structural diagram of the GRU network;

Fig. 2 is the overall flow chart of the present invention.

Detailed ways

The present invention will be further described below in conjunction with accompanying drawing and specific embodiment:

1) Obtain the specific data of card swiping at each subway station at each time period of the month and make statistics on it. Take ten minutes as a time unit, divide a day into 144 time units, and count the number of people leaving and entering the station for each time unit. Get the flow data of each subway station every day from the 1st to the 26th of the month, and then construct a fifth-order tensor X according to the subway line where the subway station is located and the holidays on the date, and X ∈ ^{R 26×144 ×80×3×2} , where 26 means there are 26 days of data, 144 means 144 time periods in a day, 80 means the number of subway stations, 3 means the number of subway lines, 2 means working days and not Two cases on weekdays.

2) Take the 1-25th as the training data and the 26th as the test data. This paper chooses 144 as the training batch, which will help the network to randomly learn the rules of each time period. When predicting the data of a certain time period, this paper will take the traffic data of the first 50 minutes and the last 40 minutes of the time period plus its own ten minutes for a total of 100 minutes as input, that is, at each time step of the GRU, the number of features The flow data of 100 minutes is 10 numbers.

3) In a single training session, choose a period from 144 time periods, and then expand the data of 25 days from 1 to 25 days, and transfer them to the GRU on a daily basis, so the tensor order of the input of each GRU time step is R ^80×3×2×10 .

4) Before inputting into the network operation, normalize the input and labels, and perform z-score normalization on the input data according to the dimension of the network output. If the dimension of the output is R ⁸⁰ , then the R ^3×2× of the input data will be merged ¹⁰ These three dimensions get the mean tensor M and standard deviation tensor St of R ⁸⁰ , which is convenient for denormalizing the output data. At the same time, regardless of the distribution characteristics of the label itself, the label is calculated using the mean standard deviation obtained from the input. Normalized.

5) The normalized data is sequentially used as the input of the GRU cell according to the time dimension of the 25th, according to the formula:

Z ^t = σ( U ^(Z) * _N X ^t + W ^(Z) * _M S ^t-1 ),

R ^t = σ( U ^(R) * _N X ^t + W ^(R) * _M S ^t-1 )

H ^t ＝tanh( U ^(H) * _N X ^t + W ^(H) * _M ( S ^t-1 ⊙ R ^t ))

S ^t ＝(1- Z ^t )⊙ H ^t + Z ^t ⊙ S ^t-1

Among them, U and W represent the weight tensor of the operation with the input tensor and the hidden layer tensor at the previous moment, Z ^t represents the tensor of the update gate obtained after calculation, R ^t represents the tensor of the reset gate, and H ^t represents The hidden layer tensor obtained by resetting the tensor S ^t-1 of the previous time step according to the reset gate, then the tensor S ^t of this time step is the Z ^t ⊙ S updated according to the update gate at the previous moment It is obtained by adding ^t-1 and (1- Z ^t )⊙ H ^t that does not participate in the update, and performs forward propagation. First, the 4th-order input tensor at time t and the 4+n-order weight tensor are processed by Einstein Stan product, n is the order of the hidden layer tensor represented by a hyperparameter, the data passes through the update gate and the reset gate, and according to the reserved weight value calculated by each gate, it is determined that some elements of the previous state and the current state are valid worth the stay.

的反向传播，主要是为了更新网络中各个门中的权重张量U ^(Z)，W ^(Z)，U ^(R)，W ^(R)，U ^(H)，W ^(H)。6) After the network obtains the network output result Y through forward propagation, use the normalized label value and output to perform tensor-based loss

The back propagation of is mainly to update the weight tensors U ^(Z) , W ^(Z) , U ^(R) , W ^(R) , U ^(H) , W ^(H) in each gate in the network.

7) Repeat step 4) to step 6) until the network converges, and denormalize the final output result according to M and St to obtain the final flow prediction.

Claims (1)

1. A subway flow prediction method based on tensor GRU neural network is characterized by comprising the following steps:

1) Taking specific data of each card swiping of each subway station in a certain period as a data set, cleaning and screening the data set, and respectively finding out the dimensions of each data of the subway station, the subway line and the holiday type to form a high-order tensor;

2) Before the GRU neural network is input, normalizing each input according to the order, and simultaneously ensuring the number of labelsSharing the same set of mean variances with the input data; performing z-score normalization on input data according to the dimension of network output to obtain a mean tensorMAnd standard deviation tensorSt；K ₁ ×K ₂ ×…×K _N Expanding the N-order tensors of the shape according to the order;

3) Dividing the data processed in the step 2) into each time stepXInput into GRU neural network, in each tensor GRU operation unit, by weightingWAnd (3) withXCarrying out Einstein multiplication to obtain the tensor state of the hidden layer, wherein the specific expression of Einstein product is as follows:

wherein the method comprises the steps ofAAndBrepresenting an N-order tensor and an N+M-order tensor, k, respectively _N Represents the dimension size of the nth order of the tensor, from k in order of order ₁ To k _N Sequentially multiplying corresponding index elements and then carrying out summation operation to reduce the N order to 1 order so that the result is M-order tensor; obtaining the final hidden layer state through the operation of the update gate and the reset gate in the tensor GRUHAs an output or as an input to a next time step; after calculation of a defined time step, the resulting tensor is obtainedHThe tensor is the resulting hidden layer stateH；

4) The obtained outputHThe loss calculation is performed with the label which is also normalized before input, and tensor distance is used as a loss function for grasping the structural characteristics of the higher-order tensor:

wherein l and m represent subscripts of two tensors, x, respectively _l ，y _l ，x _m ，y _m The element values of the corresponding subscripts representing the two tensors respectively,

euclidean distance representing element value of corresponding subscript, updating each moment by back propagation mechanism _i WThe weight tensor, i represents a certain time step of the output,

repeating steps 3) to 4) for different epochs until the epochs are over and then executing the next step;

5) After the end of epoch, the result of step 3) is obtained according toMAndStand carrying out inverse normalization on the output of the network, and obtaining the effective dimension in the tensor after inverse normalization as the final predicted result value of the network, thus obtaining the final flow predicted value.

Images