Passenger flow prediction method based on optimized ICEEMDAN and bidirectional time convolution network

Through the combination of improved adaptive noise complete ensemble empirical modal decomposition and bidirectional time convolution network, complex nonlinear data processing and frequency aliasing problems in subway passenger flow prediction are solved, and higher prediction accuracy and adaptability are achieved. Future trend forecasts for subway systems.

CN119066490APending Publication Date: 2024-12-03XUCHANG MUNICIPAL RAIL TRANSIT CO LTD

Patent Information

Application Number
CN202411105802.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-13
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The existing passenger flow prediction methods are difficult to effectively process complex nonlinear data, resulting in insufficient prediction accuracy and generalization capabilities. Especially in subway systems, they face the problems of frequency aliasing and modal mixing, which affects the prediction effect.

Method used

Using an improved adaptive noise complete ensemble empirical modal decomposition (Improved ICEEMDAN) and bidirectional time convolution network (BiTCN) method, the subway passenger flow data is decomposed and predicted, and the adaptive white noise mechanism and iterative decomposition strategy are introduced. , solve the frequency aliasing problem, and extract features through bidirectional time convolution network to improve prediction accuracy.

Benefits of technology

Improve the accuracy and adaptability of passenger flow prediction, can effectively process future trend prediction under different data situations, enhance the model's modeling ability of long sequence data, avoid future information leakage and gradient disappearance problems, improve prediction accuracy and stability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119066490A_ABST
    Figure CN119066490A_ABST
Patent Text Reader

Abstract

A subway passenger flow prediction method based on an optimized ICEEMDAN-bidirectional time convolution network model comprises the following steps: collecting passenger number data of subway passenger flow, and carrying out missing value processing; the data subjected to missing value processing are input into the optimized ICEEMDAN decomposition, and the data are decomposed; carrying out batch normalization on the data after data decomposition, inputting the obtained data into a bidirectional time convolution network model to carry out model parameter training, and carrying out feature extraction on the data through a causal expansion convolution layer, a linear activation layer, a discarding layer and residual connection; the full connection layers respectively output predicted values; and accumulating the predicted values of the decomposed data after reverse normalization to obtain a final prediction result. And inputting the test set into the trained prediction model to generate a final subway passenger flow prediction result. According to the subway passenger flow prediction method based on the optimized ICEEMDAN-bidirectional time convolution network model, the future change trend of the subway passenger flow can be well predicted, the passenger flow prediction precision is improved, and good adaptability is achieved. And decision support is provided for making a passenger flow management strategy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the field of subway transportation, and in particular to a passenger flow prediction method (I-ICEEMDAN-BiTCN) based on an optimized improved adaptive noise complete ensemble empirical mode decomposition (Improved-ICEEMDAN) and a bidirectional temporal convolutional network (BiTCN). Background Art

[0002] Urban rail transit is an effective means of transportation to ease traffic congestion and accelerate the development of smart cities worldwide. As an important part of urban public transportation, accurate prediction of passenger flow is of great significance to improving traffic efficiency and reducing congestion. In recent years, investment and construction of subways have been increasing in both developing and developed countries. The improvement of subway systems has facilitated passenger travel, thereby stimulating the rapid growth of passenger volume.

[0003] The explosive growth of subway passenger flow has caused a series of problems in daily operations, such as passenger crowding, ineffective service schedules, and insufficient capacity of subway facilities, and has caused an imbalance in urban public management. One solution to this problem is to provide forward-looking passenger flow information to provide operators with decision-making guidance at the station level in terms of passenger flow control, transfer station management, and passenger-oriented service design.

[0004] Therefore, accurate passenger flow prediction is crucial for effective traffic planning and the formulation of efficient passenger flow management strategies. Early passenger flow prediction methods mainly follow parameter models based on statistical principles to deduce future passenger flow from data, including autoregression, seasonal autoregression integrated moving average, autoregression integrated moving average and other prediction methods. Although they can predict passenger flow to a certain extent, these models usually require good stability and are applicable to linear data. They cannot handle complex nonlinear features in the data, so they are not very versatile. Unlike statistical theoretical models, artificial intelligence models can better handle nonlinear data. At present, machine learning models and deep learning models have been widely used. Many researchers have developed traditional parameter models, machine learning models such as random forests and support vector regression models, and later deep learning models such as recursive neural networks (RNNs) and convolutional neural networks (CNNs).

[0005] As an improvement of CNN, the temporal convolutional network (TCN) consists of causal convolution, dilated convolution and residual connection. Dilated causal convolution can effectively avoid future information leakage. By adjusting the convolution kernel size, number of layers, dilation coefficient and other parameters to expand the receptive field, it can capture the dependency between future moments and longer historical moments. Therefore, the temporal convolutional network is selected to predict the passenger flow in the future.

[0006] However, faced with different types of complex data sets, neural network models are easily disturbed by complex data, which reduces the accuracy of prediction. In order to improve the generalization ability and prediction accuracy of the model, an effective processing strategy is to decompose the original data into multiple simple modes with obvious characteristics, and use neural networks to predict and aggregate the decomposed components to solve the problem of complex characteristics of the data set. The most classic technique is empirical mode decomposition (EMD), followed by its related derivative techniques such as integrated EMD, complementary EEMD, and other decomposition techniques, including empirical wavelet transform (EWT) and variational mode decomposition (VMD).

[0007] However, wavelet transform and variational mode decomposition require setting of hyperparameters of decomposition, which are difficult to preset without prior knowledge. Empirical mode decomposition belongs to adaptive iterative decomposition, but there are frequency aliasing and mode mixing problems when processing signals, which affects the continuity of frequencies between adjacent scales. Summary of the invention

[0008] To solve the above technical problems, the present invention provides a passenger flow prediction method (I-ICEEMDAN-BiTCN) based on optimized improved adaptive noise complete ensemble empirical mode decomposition (Improved ICEEMDAN) and bidirectional temporal convolutional network (BiTCN). It can better predict the future trend of subway passenger flow and improve the accuracy of passenger flow prediction.

[0009] The technical solution adopted by the present invention is:

[0010] The subway passenger flow prediction method based on the I-ICEEMDAN-BiTCN model includes the following steps:

[0011] Step 1: Collect passenger flow information, process missing values, decompose the data using the optimized ICEEMDAN, and output the decomposed data with the minimum residual.

[0012] Step 2: After normalizing the decomposed data in step 1, input it into the BiTCN model for training, perform feature extraction, and output the prediction results.

[0013] Step 3: Add up the data obtained in step 2 and output it as the subway passenger flow prediction result.

[0014] In the step 1, the passenger flow data of the subway station is collected, and the missing values ​​are supplemented by using median interpolation.

[0015] In the step 1, the processed data is input into the optimized improved adaptive noise complete set empirical mode decomposition model for multi-frequency decomposition. ICEEMDAN is an algorithm improved based on EMD. ICEEMDAN improves the frequency aliasing and mode mixing problems that occur when EMD processes signals by introducing an adaptive white noise mechanism, and achieves the continuity of frequencies between adjacent scales. Different from traditional Fourier transform and wavelet transform, ICEEMDAN adopts an iterative decomposition strategy without pre-selecting basis functions, which is more suitable for nonlinear signal analysis. The specific process is as follows:

[0016] 1) Construct a new sequence by combining the original sequence (subway passenger flow) with the white noise sequence

[0017]

[0018] Where: is the original signal, is the i-group white noise w added to the original signal i The first-order modal component after EMD decomposition, α0 is the noise intensity adjustment coefficient.

[0019] 2) For the newly constructed sequence Apply the local mean operation and calculate the first set of residuals:

[0020]

[0021] Where: M(·) represents the local mean calculated by applying the EMD algorithm to the sequence.

[0022] 3) Subtract the first set of residuals from the original signal to obtain the first mode function (IMF1).

[0023]

[0024] 4) Iteratively calculate the residuals and modal components. Repeat the above process, add white noise to the residual signal of each round, calculate new residuals and modal components, until all decompositions are completed:

[0025]

[0026] Where: R k is the kth group of residuals, is the kth modal component. The noise intensity adjustment coefficient α in the decomposition process k It is used to control the proportion of white noise added, which depends on the characteristics of the signal and can be expressed as:

[0027]

[0028] Where: std is the standard deviation; ε0 is the inverse of the signal-to-noise ratio between the first added noise and the analyzed signal.

[0029] 5) After decomposition, the original sequence can be equivalent to:

[0030]

[0031] Where: t is the time variable; is the remaining decomposition residual.

[0032] 6) Eliminate the residual signal from the decomposed data and retain the decomposed components as input for subsequent models.

[0033] In step 1, although the ICEEMDAN used is an adaptive decomposition, the initial threshold for extracting noise needs to be set manually. The quality and time of the decomposition are related. If it is too small, it will fail all the time and then approach the maximum number of iterations. If it is too large, the decomposition will be incomplete and the residual will contain more effective information.

[0034] In order to improve the decomposition accuracy, this paper sets the minimum residual as the objective function and optimizes the noise threshold through the subtraction average optimization algorithm (SABO). The specific process is as follows:

[0035] 1) Set the objective function, variables, and constraints. The variable is set to the noise intensity adjustment coefficient α k , then α k

[0036] The corresponding constraint condition can be set as an interval according to the size of the sequence. The present invention uses ICEEMDAN for effective decomposition and sets the minimum absolute error (residual) L loss is the objective function. The corresponding objective function formula is:

[0037]

[0038] Where: S is the sequence length.

[0039] 2) Set the SABO population size N and the number of iterations T.

[0040] 3) Randomly generate the initial proxy matrix:

[0041]

[0042] x i,d =lb d +r i,d ·(ub d -lb d ), i=1,...,N, d=1,...,m (10)

[0043] Where: is the population matrix, x i is the i-th population member, x i,d is its d-th dimension decision variable in the search space, N is the number of search agents, m is the number of decision variables, r i,d is a random number in the interval [0,1], ub d and lb d are the lower and upper bounds of the d-th decision variable. Each search agent is

[0044] Candidate solutions to a problem.

[0045] 4) Calculate the objective function based on the initial search agent.

[0046] 5) For t = 1 to T; for i = 1 to N, calculate the new recommendation position of the i-th SABO search agent

[0047] Place

[0048]

[0049] Where: is the i-th search agent X i is the new suggested location of , N is the total number of search agents, is a vector of dimension m whose components follow a normal distribution in the interval [0,1]. If this new proposed position leads to an improvement in the objective function value, i.e., a smaller value than the previous one, it is replaced, i.e., the new X i .

[0050] 6) Update X i :

[0051]

[0052] Where: F i and F i new They are search agent X i and The objective function value of .

[0053] 7) End. Save the optimal solution. End.

[0054] 8) Output the best quasi-optimal solution obtained using SABO. End

[0055] The step 2 comprises the following steps:

[0056] S2.1. Perform batch normalization on the decomposed data in step 1 to eliminate the impact of non-standard data on model accuracy. The expression is:

[0057]

[0058] Where: x k for A single value of x min for The minimum value in x max for The maximum value in . x* is the normalized value of the error data. The batch normalized data is divided into a test set and a validation set in a ratio of 7:3. The test set is used for offline model training, and the validation set is used to verify the final prediction effect of the model.

[0059] S2.2. Input the test set divided by S2.1 into the input layer of the time series convolutional network model. Each input layer uses the subway passenger flow data corresponding to the time series as the input of the time series convolutional network model. According to the number of components k, a total of k sub-models are established.

[0060] S2.3, the temporal convolutional network consists of residual modules stacked in sequence and propagated backward. Each residual module consists of causal convolution, dilated convolution and residual connection. The input layer data is passed to the residual block of the time series convolutional network model for high-dimensional feature extraction.

[0061] S2.4. Feature extraction is performed through causal convolution and dilated convolution, and batch normalization is performed after each convolution calculation; after activation using the ReLU activation function, it is propagated backward.

[0062] S2.5, causal convolution allows TCN to use only data before the current moment at each step, avoiding the leakage of future information. An identity mapping is performed for every two convolution calculations in S2.4, and each convolution calculation is a causal convolution with a hole. After the convolution calculation, in order to improve the classification accuracy, batch normalization is performed, and then the obtained value is activated using the ReLU activation function. TCN uses dilated causal convolution to improve the receptive field, and skips the input sequence to perform convolution operations. The expression is:

[0063]

[0064] Where: c it is the convolution result of the tth element in the sequence; w(i) is the convolution kernel; n is the dilation factor; * is the convolution operator; k er represents the convolution kernel size; X t-n·iCorresponding to the input sequence value after interval sampling. And after each dilated convolution, the Dropout method is introduced for regularization, randomly discarding a certain proportion of neurons to reduce overfitting and improve the generalization ability of the model. In order to be able to add smoothly, the value dimension of the convolution calculation result and the identity mapping must be the same, so a 1×1 convolution calculation is used on the path of the identity mapping to eliminate the shape tensor difference before the residual block is connected, so that the dimensions of the two are consistent.

[0065] S2.6. Output the residual module result of the S2.5 time series convolutional network model. The output h is as follows:

[0066] h=ReLu(W1c it +b1) (15)

[0067] Where W1 and b1 are the weight matrix and bias matrix. ReLu represents the activation function. This step is a process of linear change, mapping through the activation function and outputting the result. After the input data undergoes such multiple convolution operations, the abstract features contained in the subway passenger flow data are effectively extracted, and the in-line nature of various types of data can still be maintained. After multiple layers of convolution operations, the output matrix with the same dimension as the input is obtained through the full convolution layer in the time series convolution network.

[0068] S2.7. Further, changing the temporal convolutional network to a bidirectional processing path can expand the amount of data and prevent underfitting. First, based on the forward processing, reverse the input sequence x, denoted as x′. h is the feature set output by the processing path, denoted as Use the temporal convolutional network to simultaneously process the forward and backward feature sets for feature extraction and prediction. Therefore, the final output formula is as follows:

[0069]

[0070] Where: is the feature set output by the backward processing path; W2 and b2 represent the corresponding weight matrix and bias matrix. This paper uses BiTCN to extract features from the input matrix, processes the input tensor through the Flatten layer, and then connects to the fully connected layer to output the prediction result; Stands for addition.

[0071] S2.8. Accumulate the output results of the bidirectional time series convolutional network model of S2.7 to obtain the predicted subway passenger flow data.

[0072] In the step three, the validation set in S2.1 is used as input to the trained bidirectional temporal convolutional network model to predict the decomposed passenger flow components; the component prediction results are accumulated to obtain the final subway passenger flow prediction result.

[0073] In order to comprehensively evaluate the effect of the model, the mean absolute error (MAE), root mean square error (RMSE), and determination coefficient (R-Square, R 2 ) Three evaluation indicators are used to judge the prediction effect of the model. The judgment formula is as follows:

[0074]

[0075] Where: y(t) represents the true output at the current moment; y^(t) is the predicted output of the model; S represents the sequence length, which is a finite number.

[0076] The present invention is based on the subway passenger flow prediction method of optimizing the ICEEMDAN-BiTCN model, and the technical effects are as follows:

[0077] 1) The present invention is based on the subway passenger flow prediction method of optimizing the ICEEMDAN-BiTCN model, and realizes the prediction of future subway passenger flow through subway passenger flow data. The present invention has good adaptability and can realize the effective trend of future changes in subway passenger flow under different data conditions.

[0078] 2) The multi-layer architecture design of the temporal convolutional network provided by the present invention enables each layer to be stacked with residual blocks of different expansion coefficients, thereby enhancing the model's modeling ability for long sequence data and retaining key original information. This causal convolution method allows the temporal convolutional network to use only data before the current moment at each step, avoiding the leakage of future information. The application of dilated causal convolution enables the time series convolutional network to have a larger field of view with fewer layers, improving the ability of the time series convolutional network to receive historical data. The introduction of the residual network "connects" network layers separated by one or more layers, effectively avoiding the gradient vanishing problem in complex models. The time series convolutional network can input sequences of any length, and output sequences of equal length in the output layer to be passed on to the next unit. Its architecture can be adjusted to any length as needed, and has strong applicability.

[0079] 3) The present invention improves the interactivity of context by modifying the temporal convolutional network into a bidirectional processing structure, thereby enhancing the realization of multi-level feature mining from local to global, avoiding the situation in which the traditional model may have a large difference in context values ​​when there is an abnormal error, and improving the prediction accuracy. At the same time, the length of the data is further expanded to avoid falling into the local optimum.

[0080] 4) The present invention provides a decomposition method to solve the problem of difficulty in decomposing parameters of complex data models. The present invention provides a decomposition method. ICEEMDAN is used to pre-process the data before prediction, and the learning ability of the model is enhanced by decomposition. The pre-processing increases the prediction stability of the model for different sequences.

[0081] 5) The present invention optimizes the selection of decomposition threshold based on the decomposition of ICEEMDAN, so that the model can adaptively decompose according to different sequence features, avoiding the interpretation errors, under-decomposition, or over-decomposition caused by prior assumptions, and improving the decomposition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 This is a comparison chart of the decomposition residuals of ICEEMDAN after optimization (the range of the standard deviation is set to [0.2, 300], N = 20, T = 100, and the optimization is performed. The standard deviation after optimization is set to 8.19; the standard deviation is set to 0.2; the standard deviation is set to 300).

[0083] Figure 2 Figure 2 is a graph of the components after using optimized-ICEEMDAN on the original data (standard deviation = 8.19).

[0084] Figure 3 Figure 2. The decomposition structure diagram of the Bidirectional Temporal Convolutional Network (BiTCN) model.

[0085] Figure 4 It is a flow chart of the present invention.

[0086] Figure 5 The following is a comparison chart of passenger flow prediction results based on the optimized ICEEMDAN-Bidirectional Temporal Convolutional Network (I-ICEEMDAN-BiTCN) (the horizontal axis 0 to 20 steps are 4:00-23:00 on Tuesday, 21 to 40 steps are 4:00-23:00 on Wednesday, and so on, for a total of one week).

[0087] Figure 6 This is a comparison chart showing the passenger flow prediction results for weekdays (4:00 to 23:00) (one day is randomly selected from Monday to Friday, this chart is Tuesday).

[0088] Figure 7 This is a comparison chart showing the passenger flow prediction results for weekends (4:00 to 23:00) (one day is randomly selected from Saturday to Sunday, this chart is Sunday).

[0089] Figure 8 Forecast of total passenger volume per day. DETAILED DESCRIPTION

[0090] A passenger flow prediction method (I-ICEEMDAN-BiTCN) based on the optimized improved adaptive noise complete ensemble empirical mode decomposition (Improved-ICEEMDAN) and bidirectional temporal convolutional network (BiTCN). After adaptive processing of passenger flow data, a deep learning model is built. The present invention can effectively follow the future trend of subway passenger flow, and the prediction effect of the algorithm is closer to the actual situation. It has good adaptability and stability, and can better predict different types of passenger flow conditions.

[0091] The subway passenger flow prediction method based on the I-ICEEMDAN-BiTCN model includes the following steps:

[0092] Step 1: Collect passenger flow information, process missing values, and use the optimized ICEEMDAN to decompose the data and output the decomposed data with the minimum residual.

[0093] Step 2: After normalizing the decomposed data in step 1, input it into the BiTCN model for training, perform feature extraction, and output the prediction results.

[0094] Step 3: Add up the data obtained in step 2 and output it as the subway passenger flow prediction result.

[0095] In the step 1, the passenger flow data of the subway station is collected, and the missing values ​​are supplemented by using median interpolation.

[0096] In the step 1, the processed data is input into the optimized improved adaptive noise complete set empirical mode decomposition model for multi-frequency decomposition. ICEEMDAN is an algorithm improved based on EMD. ICEEMDAN improves the frequency aliasing and mode mixing problems that occur when EMD processes signals by introducing an adaptive white noise mechanism, and achieves the continuity of frequencies between adjacent scales. Different from traditional Fourier transform and wavelet transform, ICEEMDAN adopts an iterative decomposition strategy without pre-selecting basis functions, which is more suitable for nonlinear signal analysis. The specific process is as follows:

[0097] 1) Construct a new sequence by combining the original sequence (subway passenger flow) with the white noise sequence

[0098]

[0099] Where: is the original signal, is the i-group white noise w added to the original signal i The first-order modal component after EMD decomposition, α0 is the noise intensity adjustment coefficient.

[0100] 2) For the newly constructed sequence Apply the local mean operation and calculate the first set of residuals:

[0101]

[0102] Where: M(·) represents the local mean calculated by applying the EMD algorithm to the sequence.

[0103] 3) Subtract the first set of residuals from the original signal to obtain the first mode function (IMF1).

[0104]

[0105] 4) Iteratively calculate the residuals and modal components. Repeat the above process, add white noise to the residual signal of each round, calculate new residuals and modal components, until all decompositions are completed:

[0106]

[0107] Where: R k is the kth group of residuals, is the kth modal component. The noise intensity adjustment coefficient α in the decomposition process k It is used to control the proportion of white noise added, which depends on the characteristics of the signal and can be expressed as:

[0108]

[0109] Where: std is the standard deviation; ε0 is the inverse of the signal-to-noise ratio between the first added noise and the analyzed signal.

[0110] 5) After decomposition, the original sequence can be equivalent to:

[0111]

[0112] Where: t is the time variable; is the remaining decomposition residual.

[0113] 6) Eliminate the residual signal from the decomposed data and retain the decomposed components as input for subsequent models.

[0114] In step 1, although the ICEEMDAN used is an adaptive decomposition, the initial threshold for extracting noise needs to be set manually. The quality and time of the decomposition are related. If it is too small, it will fail all the time and then approach the maximum number of iterations. If it is too large, the decomposition will be incomplete, and the residual will contain more effective information. In order to improve the decomposition accuracy, this paper sets the minimum residual as the objective function and optimizes the noise threshold through the subtraction average optimization algorithm (SABO). The specific process is as follows:

[0115] 1) Set the objective function, variables, and constraints. The variable is set to the noise intensity adjustment coefficient α k , then α k The corresponding constraint condition can be set as an interval according to the size of the sequence. The present invention uses ICEEMDAN for effective decomposition and sets the minimum absolute error (residual) L loss is the objective function. Then the corresponding objective function is

[0116] The formula is:

[0117]

[0118] Where: S is the sequence length.

[0119] 2) Set the SABO population size N and the number of iterations T.

[0120] 3) Randomly generate the initial proxy matrix:

[0121]

[0122] x i,d =lb d +r i,d ·(ub d -lb d ), i=1,...,N, d=1,...,m (10)

[0123] Where: is the population matrix, x i is the i-th population member, x i,d is its d-th dimension decision variable in the search space, N is the number of search agents, m is the number of decision variables, r i,d is a random number in the interval [0,1], ub d and lb d are the lower and upper bounds of the d-th decision variable, respectively. Each search agent is

[0124] A candidate solution to a problem.

[0125] 4) Calculate the objective function based on the initial search agent.

[0126] 5) For t = 1 to T; for i = 1 to N, calculate the new recommended position of the i-th SABO search agent

[0127]

[0128] Where: is the i-th search agent X i is the new suggested location of , N is the total number of search agents, Yes

[0129] A vector of dimension m whose components follow a normal distribution in the interval [0,1]. If this new proposed position leads to an improvement in the objective function value, that is, a smaller value than the previous one, it is replaced, that is, the new X i .

[0130] 6) Update X i :

[0131]

[0132] Where: F i and F i new They are search agent X i and The objective function value of .

[0133] 7) End. Save the optimal solution. End.

[0134] 8) Output the best quasi-optimal solution obtained using SABO. End

[0135] The step 2 comprises the following steps:

[0136] S2.1. Perform batch normalization on the decomposed data in step 1 to eliminate the impact of non-standard data on model accuracy. The expression is:

[0137]

[0138] Where: x k for A single value of x min for The minimum value in x max for The maximum value in . x* is the normalized value of the error data. The batch normalized data is divided into a test set and a validation set in a ratio of 7:3. The test set is used for offline model training, and the validation set is used to verify the final prediction effect of the model.

[0139] S2.2. Input the test set divided by S2.1 into the input layer of the time series convolutional network model. Each input layer uses the subway passenger flow data corresponding to the time series as the input of the time series convolutional network model. According to the number of components k, a total of k sub-models are established.

[0140] S2.3, the temporal convolutional network consists of residual modules stacked in sequence and propagated backward. Each residual module consists of causal convolution, dilated convolution and residual connection. The input layer data is passed to the residual block of the sequential convolutional network model for high-dimensional feature extraction.

[0141] S2.4. Feature extraction is performed through causal convolution and dilated convolution, and batch normalization is performed after each convolution calculation; after activation using the ReLU activation function, it is propagated backward.

[0142] S2.5, causal convolution allows TCN to use only data before the current moment at each step, avoiding the leakage of future information. An identity mapping is performed for every two convolution calculations in S2.4, and each convolution calculation is a causal convolution with a hole. After the convolution calculation, in order to improve the classification accuracy, batch normalization is performed, and then the obtained value is activated using the ReLU activation function. TCN uses dilated causal convolution to improve the receptive field, and skips the input sequence to perform convolution operations. The expression is:

[0143]

[0144] Where: c it is the convolution result of the tth element in the sequence; w(i) is the convolution kernel; n is the dilation factor; * is the convolution operator; k er represents the convolution kernel size; X t-n·i Corresponding to the input sequence value after interval sampling. And after each dilated convolution, the Dropout method is introduced for regularization, randomly discarding a certain proportion of neurons to reduce overfitting and improve the generalization ability of the model. In order to be able to add smoothly, the value dimension of the convolution calculation result and the identity mapping must be the same, so a 1×1 convolution calculation is used on the path of the identity mapping to eliminate the shape tensor difference before the residual block is connected, so that the dimensions of the two are consistent.

[0145] S2.6. Output the residual module result of the S2.5 time series convolutional network model. The output h is as follows:

[0146] h=ReLu(W1c it +b1) (15)

[0147] Where W1 and b1 are the weight matrix and bias matrix. ReLu represents the activation function. This step is a process of linear change, mapping through the activation function and outputting the result. After the input data undergoes such multiple convolution operations, the abstract features contained in the subway passenger flow data are effectively extracted, and the in-line nature of various types of data can still be maintained. After multiple layers of convolution operations, the output matrix with the same dimension as the input is obtained through the full convolution layer in the time series convolution network.

[0148] S2.7. Further, changing the temporal convolutional network to a bidirectional processing path can expand the amount of data and prevent underfitting. First, based on the forward processing, reverse the input sequence x, denoted as x′. h is the feature set output by the processing path, denoted as Use the temporal convolutional network to simultaneously process the forward and backward feature sets for feature extraction and prediction. Therefore, the final output formula is as follows:

[0149]

[0150] Where: is the feature set output by the backward processing path; W2 and b2 represent the corresponding weight matrix and bias matrix. This paper uses BiTCN to extract features from the input matrix, processes the input tensor through the Flatten layer, and then connects to the fully connected layer to output the prediction result; Stands for addition.

[0151] S2.8. Accumulate the output results of the bidirectional time series convolutional network model of S2.7 to obtain the predicted subway passenger flow data.

[0152] In the step three, the validation set in S2.1 is used as input to the trained bidirectional temporal convolutional network model to predict the decomposed passenger flow components; the component prediction results are accumulated to obtain the final subway passenger flow prediction result.

[0153] In order to comprehensively evaluate the effect of the model, the mean absolute error (MAE), root mean square error (RMSE), and determination coefficient (R-Square, R 2 ) Three evaluation indicators are used to judge the prediction effect of the model. The judgment formula is as follows:

[0154]

[0155] Where: y(t) represents the actual output at the current moment; y^(t) is the predicted output of the model; is the average value of the true value; S represents the sequence length, which is a finite number.

[0156] From the data aspect, the decomposition residuals of the optimized ICEEMDAN (I-ICEEMDAN) and the decomposition residuals of the ICEEMDAN with fixed parameters are compared. Figure 1 It can be seen that I-ICEEMDAN has the smallest decomposition residual. The prediction results of CNN algorithm, TCN algorithm, BiTCN algorithm and optimized ICEEMDAN-BiTCN algorithm (I-ICEEMDAN-BiTCN) are compared. Figure 5 , Figure 6 , Figure 7As can be seen in Figure 2, the evaluation indicators of I-ICEEMDAN-BiTCN after the optimized ICEEMDAN perform best.

Claims

1. Passenger flow prediction method based on optimized ICEEMDAN and bidirectional temporal convolutional network, characterized by The following steps are involved: Step 1: Collect passenger flow information, process missing values, decompose the data using the optimized ICEEMDAN, and output the decomposed data with the minimum residual. Step 2: After normalizing the decomposed data in step 1, input it into the BiTCN model for training, perform feature extraction, and output the prediction results. Step 3: Add up the data obtained in step 2 and output it as the subway passenger flow prediction result.

2. The passenger flow prediction method based on optimized ICEEMDAN and bidirectional temporal convolutional network according to claim 1 is characterized by: In the step 1, the passenger flow data of the subway station is collected, and the missing values ​​are supplemented by using median interpolation; the processed data is input into the optimized improved adaptive noise complete set empirical mode decomposition model to perform multi-frequency decomposition.

3. The passenger flow prediction method based on optimized ICEEMDAN and bidirectional temporal convolutional network according to claim 1 is characterized in that: The ICEEMDAN decomposition process in step 1 specifically includes: constructing a new sequence by combining the original sequence (subway passenger flow) with the white noise sequence, calculating the local mean operation, and iteratively calculating the residual and modal components until all decompositions are completed.

4. The passenger flow prediction method based on optimized ICEEMDAN and bidirectional temporal convolutional network according to claim 2 is characterized by: The optimization process is as follows: the minimum residual is set as the objective function, and the noise threshold is optimized through the subtraction average optimization algorithm (SABO), thereby improving the decomposition accuracy of ICEEMDAN. The optimization includes the following steps: 1) Set the objective function, variables, and constraints. The variable is set to the noise intensity adjustment coefficient α k , then α k The corresponding constraint condition can be set as an interval according to the size of the sequence. The present invention designs the minimum absolute error L of the residual loss The objective function decomposed by ICEEMDAN. The corresponding objective function formula is: Where: S is the sequence length. 2) Set the SABO population size N and the number of iterations T. 3) Randomly generate the initial proxy matrix: x i,d =lb d +r i,d ·(ub d -lb d ),i=1,…,N,d=1,…,m#(10) Where: is the population matrix, x i is the i-th population member, x i,d is its d-th dimension decision variable in the search space, N is the number of search agents, m is the number of decision variables, r i,d is a random number in the interval [0,1], ub d and lb d are the lower and upper bounds of the dth decision variable, respectively. Each search agent is a candidate solution to the problem. 4) Calculate the objective function based on the initial search agent. 5) For t = 1 to T; for i = 1 to N, calculate the new recommended position of the i-th SABO search agent Where: is the i-th search agent X i is the new suggested location of , N is the total number of search agents, is a vector of dimension m whose components follow a normal distribution in the interval [0,1]. If this new proposed position leads to an improvement in the objective function value, i.e., a smaller value than the previous one, it is replaced, i.e., the new X i . 6) Update X i : Where: F i and F i new They are search agent X i and The objective function value of . 7) End. Save the optimal solution. End. 8) Output the best quasi-optimal solution obtained using SABO. End.

5. The passenger flow prediction method based on optimized ICEEMDAN and bidirectional temporal convolutional network according to claim 1 is characterized by: The step 2 comprises the following steps: S2.

1. Perform batch normalization on the decomposed data in step 1 to eliminate the impact of non-standard data on model accuracy; divide the batch normalized data into a test set and a validation set in a ratio of 7:3; S2.2, input the data processed by S2.1 into the input layer of the time series convolutional network model, and establish multiple sub-models according to the number of components; S2.3, the temporal convolutional network consists of residual modules stacked in sequence and propagated backward. Each residual module consists of causal convolution, dilated convolution and residual connection. The input layer data is transmitted to the residual block of the time series convolutional network model for high-dimensional feature extraction; S2.4, feature extraction is performed through causal convolution and dilated convolution, and batch normalization is performed after each convolution calculation; after activation using the ReLU activation function, backward propagation is performed; S2.5, for every two calculations in S2.4, a residual connection is performed, the result of the convolution calculation is added to the value of the identity mapping, and regularized by the Dropout method; S2.6, output the residual module result of the time series convolutional network model in S2.5 to obtain the prediction result of the decomposition amount; S2.7, change the temporal convolutional network of S2.6 to a bidirectional processing path to obtain the final prediction result of the decomposition amount; S2.

8. Accumulate the output results of the bidirectional time series convolutional network model of S2.7 to obtain the predicted subway passenger flow data.

6. The passenger flow prediction method based on optimized ICEEMDAN and bidirectional temporal convolutional network according to claim 5 is characterized by: The bidirectional temporal convolutional network is a temporal convolutional network that processes bidirectional paths. On the basis of forward processing, by reversing the input sequence, the temporal convolutional network is used to simultaneously process the forward and backward feature sets for feature extraction and prediction.

7. The passenger flow prediction method based on optimized ICEEMDAN and bidirectional temporal convolutional network according to claim 1 is characterized by: In step 3, the validation set in S2.1 is used as input to the trained bidirectional temporal convolutional network model to predict the decomposed passenger flow components; the component prediction results are accumulated to obtain the final subway passenger flow prediction result. 2 )Three evaluation indicators are used to judge the prediction effect of the model.

Citation Information

Patent Citations

  • Urban rail transit short-time passenger flow prediction method based on CEEMDAN-LSTM

    CN115034457A

  • Cutter wear prediction method based on improved subtraction optimizer combined with improved bidirectional long-short term memory neural network

    CN117549139A

  • Heavy overload transformer oil temperature prediction method and system

    CN117688322A

  • Photovoltaic hybrid energy storage system power distribution method based on parameter optimization variational mode decomposition

    CN118199149A

  • Power load prediction method based on decomposition strategy and deep learning

    CN118211844A

Cited By

  • Airport passenger flow multi-factor fusion prediction method based on causal inference

    CN120046771A