CN111311905A - Particle swarm optimization wavelet neural network-based expressway travel time prediction method - Google Patents

Abstract

The invention relates to a particle swarm optimization-based travel time prediction method for a wavelet neural network. The particle swarm optimization algorithm optimizes parameters of the wavelet neural network through continuous iteration, so that the defects of a wavelet neural network model are overcome, including slow convergence speed, easiness in falling into a local minimum value and easiness in generating an oscillation effect. Through comparison experiment analysis, the particle swarm optimization wavelet neural network model can accurately predict the change trend of the travel time and the fluctuation condition of the travel time, and the advantages of high convergence speed, high prediction precision and strong adaptability are proved.

Description

A Highway Travel Time Prediction Based on Particle Swarm Optimization Wavelet Neural Network method

technical field

The invention designs a practicability model, which belongs to the category of intelligent transportation, and specifically is a method for predicting highway travel time based on particle swarm optimization wavelet neural network, which can provide travelers with high real-time and reliable travel information.

Background technique

Accurate prediction of travel time provides a basis for traffic managers to make management decisions and travelers to rationally arrange travel plans. At present, the methods of travel time prediction include: time series model, multiple regression model, Kalman filter prediction and artificial neural network. However, time series models do not take into account other factors that affect travel times, but predict future travel times by analyzing similar trends in past data. Multiple regression models study the data of the target road segment and adjacent road segments, and the prediction results are usually unsatisfactory. The Kalman filter model needs to constantly modify the weights in the prediction process, which makes the calculation time is too large and the real-time prediction is poor. Artificial neural network is a commonly used method for travel time prediction, but in the process of application, it is easy to fall into local optimal solution, easy to produce oscillation effect and slow convergence.

Therefore, when the present invention predicts the travel time, the advantages and disadvantages of the existing methods are comprehensively considered, and a new expressway travel time prediction model is established.

SUMMARY OF THE INVENTION

In view of the deficiencies and research trends of the prior art, the purpose of the present invention is to provide a method model of expressway travel time prediction based on particle swarm optimization wavelet neural network, which can accurately predict the travel time, so as to provide a good solution for traffic management and optimization by relevant departments. theoretical support and decision-making basis.

In order to realize the purpose of the present invention, the technical scheme adopted is:

Research on the prediction method of expressway travel time. The main steps of the research method include:

A. Process the highway toll data and extract the receipt fields required for research;

B. Perform data processing on the extracted charging data fields;

C. On the basis of data processing, use the wavelet neural network model to predict the travel time of the expressway;

D. On the basis of data processing, use particle swarm optimization wavelet neural network model to predict highway travel time;

E. Using mean absolute error (MAE), mean relative error (MAPE) and mean square error (MSE) three evaluation indicators to compare and analyze the accuracy of the two prediction models. By comparing the results, it can be seen that the wavelet neural network model of particle swarm optimization can more accurately predict the travel time of expressways, and show better adaptability in actual data verification.

The research on the expressway travel time prediction method, wherein, the specific analysis process of the step A is:

A1. my country's highway toll collection adopts an information system that fully covers the toll collection process, so a large amount of toll data can be collected. The data fields required for the study include INSTATIONID (entry toll station code), INTIME (entry time), EXITSTATION (exit toll station code), EXITTIME (exit time), as shown in Table 1.

Table 1 Charge data field description table

The research on the expressway travel time prediction method, wherein, the specific analysis process of the step B is:

B1. The charging data processing is carried out in two steps. The first is to eliminate abnormal data, including missing data and wrong data. , whose interval is [t _min ,t _max ].

B2. Abnormal data mainly includes the following four types:

Lack of entry/exit toll booth or entry/exit time information: When the toll system software is abnormal or there is an error in the line transmission, the toll system cannot input complete toll data information;

The same data entering and leaving the toll booth: When the driver exchanges the toll card in the service area to avoid the toll collection, it will cause the vehicle recorded in the toll collection system to have the same time at the entrance toll booth and the exit toll booth;

Abnormal time data record. Although the toll collection system is becoming more and more perfect, there is also the disadvantage that the systems of different toll stations are not synchronized, which leads to the deviation of the time of different toll stations, resulting in the phenomenon that the entrance time is later than the exit time. Usually this happens in the early hours of the morning when the charging system is updated;

Abnormal time data record. Due to the particularity of expressways, service areas and parking areas are generally set up on the road sections, resulting in the phenomenon of long-distance vehicles staying for a long time.

B3. On the basis of removing erroneous data, screen valid data based on quartile method. The formula for calculating the lower limit and upper limit of the valid data set of travel time is:

t _min =t _25% -1.5×(t _75% -t _25% )

t _max =t _75% +1.5×(t _75% -t _25% )

t _25% --25% quantile

t _75% --- 75% quantile

t _min --- lower limit value of valid data set of travel time

t _max ---The upper limit value of the valid data set of travel time.

The research on the expressway travel time prediction method, wherein, the specific analysis process of the step C is:

C1. Network construction. Construct a three-layer wavelet neural network with input layer, hidden layer and output layer. As shown in Figure 1, where X ₁ , X ₂ ,...,X _m (m represents the number of input neurons) represents the input of the wavelet neural network, Y ₁ , Y ₂ ,..., Y _n (n represents the input of the wavelet neural network) The number of output neurons), J represents the number of hidden layer nodes, ω _ij and ω _jk respectively refer to the weight between the input layer and the hidden layer, the hidden layer and the output layer, where i=1, 2, .. .,m; j=1,2,...,J; k=1,2,...,n.

C2. Network initialization. Randomly initialize the scaling factor a _j of the wavelet function, the translation factor b _j , the connection weights ω _ij (input layer and hidden layer) and ω _jk (hidden layer and output layer). Set the learning rates η ₁ and η ₂ of the wavelet neural network.

The adjustment formula of the scaling factor is as follows:

---调整前伸缩因子；

--- Adjust the front scaling factor;

--调整后伸缩因子；

-- Adjusted scaling factor;

η ₂ --- the learning rate of wavelet parameters, which can be taken as η ₂ =0.01.

The adjustment formula of the translation factor is as follows:

---调整前平移因子值；

---Translation factor value before adjustment;

---调整后平移因子值；

--- Adjusted translation factor value;

η ₂ --- the learning rate of wavelet parameters, which can be taken as η ₂ =0.01.

The adjustment between the input layer and the hidden layer can be adjusted by the following formula:

---调整之前输入层与隐含层之间的权值；

--- Adjust the weights between the input layer and the hidden layer before;

---调整之后输入层与隐含之间的权值；

---The weight between the input layer and the hidden layer after adjustment;

η ₁ --- the learning rate of the network weight parameter, which can be η ₁ =0.01.

The adjustment between the hidden layer and the output layer can be adjusted by the following formula:

---调整之前隐含层与输出层之间的权值；

--- Adjust the weights between the hidden layer and the output layer before;

---调整之后隐含层与输出层之间的权值；

---After adjustment, the weights between the hidden layer and the output layer;

η ₁ --- the learning rate of the network weight parameter, which can be η ₁ =0.01.

C3. Predict output. The training samples are fed into the network to calculate the predicted output of the network and calculate the error e between the network output and the expected output.

C4. Error calculation. Calculate the error e between the predicted output of the network and the expected output. The specific calculation formula is:

y _p,k (k)---actual output value;

t _p,k --- ideal output value.

C5. Weight correction. In order to make the error meet the requirements, the weights and parameters of the wavelet neural network are modified by the gradient descent method.

C6. If the number of training times is greater than 1000 or e meets the prediction accuracy requirements (e<0.0001, the error accuracy is set to 0.0001. If the error accuracy is set too high, the network convergence speed will be slow; if the error accuracy is set too low, the prediction result will be affected. accuracy), end the training and return the prediction result, otherwise continue learning and training.

The test flow of travel time prediction based on wavelet neural network is shown in Figure 2.

The research on the method for predicting the travel time of the expressway, wherein, the specific analysis process of the step D is:

D1. Network construction: construct a three-layer POS-WNN neural network including input layer, hidden layer and output layer

D2. Parameter initialization processing: the system randomly generates S particles, and uses the position vectors of these particles to represent the scaling factors and translation factors a _j , b _j in the wavelet basis function; and the weight value ω _ij (the ith input and the first The weight value between the j hidden layers) and ω _j (the weight value between the jth hidden layer and the output layer), the calculation formula is as follows. At the same time, set the maximum and minimum speed of particles, the learning rate and the maximum number of iterations.

x _s =(x _s,1 ,...,x _s,d ,...,x _s,D )=(a ₁ ,...,a _J ,b ₁ ,...,b _J ,ω ₁₁ ,...,ω _1J ,ω ₂₁ ,...,ω _2J ,...,ω ₂₉₁ ,...,ω _29J ,ω ₁ ,...,ω _J )

D=m×J+J×(m-n)+(m-1)×J

m---number of input layer nodes

J---the number of hidden layer nodes

n---the number of output layer nodes

D3. Network training. Based on the error between the actual output value and the ideal output value, the particle fitness in each iteration process is calculated, and the calculation formula is as follows.

Q---Total number of training samples

n---the number of output neurons of the network

y _p,k (k)---actual output value

t _p,k --- ideal output value.

D4. Use the particle fitness as an indicator to determine whether the set error requirement is met. If the set error requirement can be met, that is, the fitness value remains basically unchanged, then the training is completed and go to step F; if it still fails to meet the set error requirement To the error requirement, proceed to the next step.

D5. Determine whether the number of training has reached the set maximum number of iterations, if so, jump out of the loop, go to step F, and stop training; otherwise, update the speed and position of the particle according to the following formula, and return to step C to continue training.

v _s,d (i+1)=ω×v _s,d (i)+c ₁ r ₁ [p _s,d -x _s,d (i)]+c ₂ r ₂ [p _g,d -x _s,d (i)]

x _s,d (i+1)=x _s,d (i)+v _s,d (i+1)

s=1,2,..,S; d=1,2,...,D

c ₁ ,c ₂ --- acceleration factor

i---current iteration number

ω---Inertia factor

r ₁ ,r ₂ --- Random numbers uniformly distributed between 0 and 1.

D6. Network debugging: Select the sample value of the training sample in a specific period as the input value, and the actual value of another specific period in the training sample as the ideal output value, and calculate the error according to the fitness calculation formula of step C. If the error accuracy setting is satisfied , then end the operation; if the set error accuracy cannot be met, go to step C.

The experimental flow of the travel time prediction model based on particle swarm optimization wavelet neural network is shown in Figure 3.

The research on the method for predicting the travel time of the expressway, wherein, the specific analysis process of the step E is:

E1. Using mean absolute error (MAE), mean relative error (MAPE) and mean square error (MSE) three evaluation indicators to compare and analyze the accuracy of the two prediction models.

E2. Let t _p,k represent the actual value of travel time, y _p,k (k) represent the predicted value of travel time, and l represent the number of predicted time periods.

E3. Calculate the mean absolute error evaluation index, and the specific calculation expression is:

Mean Absolute Error (MAE) - A measure used to assess the difference between actual and predicted values.

E4. Calculate the average relative error evaluation index, and the specific calculation expression is:

Mean Relative Error (MAPE)—In the process of travel time prediction, it is used to evaluate the accuracy of travel time prediction results.

E5. Calculate the mean square error evaluation index, and the specific calculation expression is:

Mean Squared Error (MSE)----Comprehensive evaluation of the degree of change in the data.

The mean absolute error (MAE), mean relative error (MAPE) and mean square error (MSE) were used to compare and analyze the accuracy of the two prediction models. The experimental results show that the mean absolute error, mean relative error and mean square error of the prediction results of the wavelet neural network travel time prediction model of particle swarm optimization are reduced by 83.36%, 82.20% and 98.15% respectively compared with the wavelet neural network model. The wavelet neural network travel time prediction model of particle swarm optimization not only has high prediction accuracy, but also can accurately predict the trend and fluctuation of travel time. It also shows certain advantages in convergence speed and has good adaptability.

The invention discloses a travel time prediction method based on particle swarm optimization wavelet neural network. The particle swarm optimization algorithm solves the defects of the wavelet neural network model through continuous iterative optimization of the parameters of the wavelet neural network, including slow convergence speed, easy to fall into local minimum and easy to produce oscillation effect. Through comparative experimental analysis, the particle swarm optimization wavelet neural network model can not only accurately predict the change trend of travel time, but also more accurately predict the fluctuation of travel time. The expressway travel time prediction experiment based on the particle swarm optimization wavelet neural network model proves that the particle swarm optimization wavelet neural network has the advantages of fast convergence, high prediction accuracy and strong adaptability.

It should be understood that the present invention is not limited to the above-mentioned best embodiment, and anyone can draw other various forms of products under the inspiration of the present invention, but no matter if any changes are made in its shape or structure, any The same or similar technical solutions of the present application all fall within the protection scope of the present invention.

Description of drawings

Fig.1 Structure of compact wavelet neural network

Figure 2 Flow chart of wavelet neural network

Fig. 3 The experimental flow chart of the particle swarm optimization wavelet neural network travel time prediction model

Research on the prediction method of expressway travel time. The main steps of the research method include:

A. Process the highway toll data and extract the receipt fields required for research;

B. Perform data processing on the extracted charging data fields;

C. On the basis of data processing, use the wavelet neural network model to predict the travel time of the expressway;

D. On the basis of data processing, use particle swarm optimization wavelet neural network model to predict highway travel time;

E. Using mean absolute error (MAE), mean relative error (MAPE) and mean square error (MSE) three evaluation indicators to compare and analyze the accuracy of the two prediction models. By comparing the results, it can be seen that the wavelet neural network model of particle swarm optimization can more accurately predict the travel time of expressways, and show better adaptability in actual data verification.

The research on the expressway travel time prediction method, wherein, the specific analysis process of the step A is:

A1. my country's highway toll collection adopts an information system that fully covers the toll collection process, so a large amount of toll data can be collected. The data fields required for the study include INSTATIONID (entry toll station code), INTIME (entry time), EXITSTATION (exit toll station code), EXITTIME (exit time), as shown in Table 1.

Table 1 Charge data field description table

The research on the expressway travel time prediction method, wherein, the specific analysis process of the step B is:

B1. The charging data processing is carried out in two steps. The first is to eliminate abnormal data, including missing data and wrong data. , whose interval is [t _min ,t _max ].

B2. Abnormal data mainly includes the following four types:

Lack of entry/exit toll booth or entry/exit time information: When the toll system software is abnormal or there is an error in the line transmission, the toll system cannot input complete toll data information;

The same data entering and leaving the toll booth: When the driver exchanges the toll card in the service area to avoid the toll collection, it will cause the vehicle recorded in the toll collection system to have the same time at the entrance toll booth and the exit toll booth;

Abnormal time data record. Although the toll collection system is becoming more and more perfect, there is also the disadvantage that the systems of different toll stations are not synchronized, which leads to the deviation of the time of different toll stations, resulting in the phenomenon that the entrance time is later than the exit time. Usually this happens in the early hours of the morning when the charging system is updated;

Abnormal time data record. Due to the particularity of expressways, service areas and parking areas are generally set up on the road sections, resulting in the phenomenon of long-distance vehicles staying for a long time.

B3. On the basis of removing erroneous data, screen valid data based on quartile method. The formula for calculating the lower limit and upper limit of the valid data set of travel time is:

t _min =t _25% -1.5×(t _75% -t _25% )

t _max =t _75% +1.5×(t _75% -t _25% )

t _25% --25% quantile

t _75% --- 75% quantile

t _min --- lower limit value of valid data set of travel time

t _max ---The upper limit value of the valid data set of travel time.

The research on the expressway travel time prediction method, wherein, the specific analysis process of the step C is:

C1. Network construction. Construct a three-layer wavelet neural network with input layer, hidden layer and output layer. As shown in Figure 1, where X ₁ , X ₂ ,...,X _m (m represents the number of input neurons) represents the input of the wavelet neural network, Y ₁ , Y ₂ ,..., Y _n (n represents the input of the wavelet neural network) The number of output neurons), J represents the number of hidden layer nodes, ω _ij and ω _jk respectively refer to the weight between the input layer and the hidden layer, the hidden layer and the output layer, where i=1, 2, .. .,m; j=1,2,...,J; k=1,2,...,n.

C2. Network initialization. Randomly initialize the scaling factor a _j of the wavelet function, the translation factor b _j , the connection weights ω _ij (input layer and hidden layer) and ω _jk (hidden layer and output layer). Set the learning rates η ₁ and η ₂ of the wavelet neural network.

The adjustment formula of the scaling factor is as follows:

---调整前伸缩因子；

--- Adjust the front scaling factor;

---调整后伸缩因子；

--- Adjusted scaling factor;

η ₂ --- the learning rate of wavelet parameters, which can be taken as η ₂ =0.01.

The adjustment formula of the translation factor is as follows:

---调整前平移因子值；

---Translation factor value before adjustment;

---调整后平移因子值；

--- Adjusted translation factor value;

η ₂ --- the learning rate of wavelet parameters, which can be taken as η ₂ =0.01.

The adjustment between the input layer and the hidden layer can be adjusted by the following formula:

---调整之前输入层与隐含层之间的权值；

--- Adjust the weights between the input layer and the hidden layer before;

---调整之后输入层与隐含之间的权值；

---The weight between the input layer and the hidden layer after adjustment;

η ₁ --- the learning rate of the network weight parameter, which can be η ₁ =0.01.

The adjustment between the hidden layer and the output layer can be adjusted by the following formula:

---调整之前隐含层与输出层之间的权值；

--- Adjust the weights between the hidden layer and the output layer before;

---调整之后隐含层与输出层之间的权值；

---After adjustment, the weights between the hidden layer and the output layer;

η ₁ --- the learning rate of the network weight parameter, which can be η ₁ =0.01.

C3. Predict output. The training samples are fed into the network to calculate the predicted output of the network and calculate the error e between the network output and the expected output.

C4. Error calculation. Calculate the error e between the predicted output of the network and the expected output. The specific calculation formula is:

y _p,k (k)---actual output value;

t _p,k --- ideal output value.

C5. Weight correction. In order to make the error meet the requirements, the weights and parameters of the wavelet neural network are modified by the gradient descent method.

C6. If the number of training times is greater than 1000 or e meets the prediction accuracy requirements (e<0.0001, the error accuracy is set to 0.0001. If the error accuracy is set too high, the network convergence speed will be slow; if the error accuracy is set too low, the prediction result will be affected. accuracy), end the training and return the prediction result, otherwise continue learning and training.

The test flow of travel time prediction based on wavelet neural network is shown in Figure 2.

The research on the method for predicting the travel time of the expressway, wherein, the specific analysis process of the step D is:

D1. Network construction: construct a three-layer POS-WNN neural network including input layer, hidden layer and output layer

D2. Parameter initialization processing: the system randomly generates S particles, and uses the position vectors of these particles to represent the scaling factors and translation factors a _j , b _j in the wavelet basis function; and the weight value ω _ij (the ith input and the first The weight value between the j hidden layers) and ω _j (the weight value between the jth hidden layer and the output layer), the calculation formula is as follows. At the same time, set the maximum and minimum speed of particles, the learning rate and the maximum number of iterations.

x _s =(x _s,1 ,...,x _s,d ,...,x _s,D )=(a ₁ ,...,a _J ,b ₁ ,...,b _J ,ω ₁₁ ,...,ω _1J ,ω ₂₁ ,...,ω _2J ,...,ω ₂₉₁ ,...,ω _29J ,ω ₁ ,...,ω _J )

D=m×J+J×(m-n)+(m-1)×J

m---number of input layer nodes

J---the number of hidden layer nodes

n---the number of output layer nodes

D3. Network training. Based on the error between the actual output value and the ideal output value, the particle fitness in each iteration process is calculated, and the calculation formula is as follows.

Q---Total number of training samples

n---the number of output neurons of the network

y _p,k (k)---actual output value

t _p,k --- ideal output value.

D4. Use the particle fitness as an indicator to determine whether the set error requirement is met. If the set error requirement can be met, that is, the fitness value remains basically unchanged, then the training is completed and go to step F; if it still fails to meet the set error requirement To the error requirement, proceed to the next step.

D5. Determine whether the number of training has reached the set maximum number of iterations, if so, jump out of the loop, go to step F, and stop training; otherwise, update the speed and position of the particle according to the following formula, and return to step C to continue training.

v _s,d (i+1)=ω×v _s,d (i)+c ₁ r ₁ [p _s,d -x _s,d (i)]+c ₂ r ₂ [p _g,d -x _s,d (i)]

x _s,d (i+1)=x _s,d (i)+v _s,d (i+1)

s=1,2,..,S; d=1,2,...,D

c ₁ ,c ₂ --- acceleration factor

i---current iteration number

ω---Inertia factor

r ₁ ,r ₂ --- Random numbers uniformly distributed between 0 and 1.

D6. Network debugging: Select the sample value of the training sample in a specific period as the input value, and the actual value of another specific period in the training sample as the ideal output value, and calculate the error according to the fitness calculation formula in step C. If the error accuracy setting is satisfied , then end the operation; if the set error accuracy cannot be met, go to step C.

The experimental flow of the travel time prediction model based on particle swarm optimization wavelet neural network is shown in Figure 3.

The research on the method for predicting the travel time of the expressway, wherein, the specific analysis process of the step E is:

E1. Using mean absolute error (MAE), mean relative error (MAPE) and mean square error (MSE) three evaluation indicators to compare and analyze the accuracy of the two prediction models.

E2. Let t _p,k represent the actual value of travel time, y _p,k (k) represent the predicted value of travel time, and l represent the number of predicted time periods.

E3. Calculate the mean absolute error evaluation index, and the specific calculation expression is:

Mean Absolute Error (MAE) - A measure used to assess the difference between actual and predicted values.

E4. Calculate the average relative error evaluation index, and the specific calculation expression is:

Mean Relative Error (MAPE)—In the process of travel time prediction, it is used to evaluate the accuracy of travel time prediction results.

E5. Calculate the mean square error evaluation index, and the specific calculation expression is:

Mean Squared Error (MSE)----Comprehensive evaluation of the degree of change in the data.

The mean absolute error (MAE), mean relative error (MAPE) and mean square error (MSE) were used to compare and analyze the accuracy of the two prediction models. The experimental results show that the mean absolute error, mean relative error and mean square error of the prediction results of the wavelet neural network travel time prediction model of particle swarm optimization are reduced by 83.36%, 82.20% and 98.15% respectively compared with the wavelet neural network model. The wavelet neural network travel time prediction model of particle swarm optimization not only has high prediction accuracy, but also can accurately predict the trend and fluctuation of travel time. It also shows certain advantages in convergence speed and has good adaptability.

Claims (5)

1. a kind of expressway travel time prediction method based on particle swarm optimization wavelet neural network, is characterized in that, comprises the following steps: A. Process the highway toll data and extract the receipt fields required for research; B. Perform data processing on the extracted charging data fields; C. On the basis of data processing, use the wavelet neural network model to predict the travel time of the expressway; D. On the basis of data processing, use the particle swarm optimization wavelet neural network model to predict the travel time of the expressway. 2. a kind of expressway travel time prediction method based on particle swarm optimization wavelet neural network according to claim 1, is characterized in that, the concrete analysis process of described step A is: A1. my country's highway toll collection adopts an information system that fully covers the toll collection process, so a large amount of toll data can be collected; the data fields required for the study include INSTATIONID (entry toll station code), INTIME (entry time), EXITSTATION (exit toll station code) ), EXITTIME (exit time). 3. a kind of expressway travel time prediction method based on particle swarm optimization wavelet neural network according to claim 1, is characterized in that, wherein, the concrete analysis process of described step B is: B1. The charging data processing is carried out in two steps. The first is to eliminate abnormal data, including missing data, wrong data, etc.; the second is to filter valid data according to the quartile method. The interval is [t _min ,t _max ]; B2. Abnormal data includes the following four types: Lack of entry/exit toll booth or entry/exit time information: When the toll system software is abnormal or there is an error in the line transmission, the toll system cannot input complete toll data information; The same data entering and leaving the toll booth: When the driver exchanges the toll card in the service area to avoid the toll collection, it will cause the vehicle recorded in the toll collection system to have the same time at the entrance toll booth and the exit toll booth; Abnormal time data recording; although the toll collection system is becoming more and more perfect, it also has the disadvantage that the systems of different toll stations are not synchronized, which leads to the deviation of the time of different toll stations, resulting in the phenomenon that the entrance time is later than the exit time; Abnormal time data recording; due to the particularity of expressways, service areas and parking areas are set up on the road sections, resulting in the phenomenon of long-distance vehicles staying for a long time; B3. On the basis of removing the erroneous data, the valid data is screened based on the quartile method; the calculation formula of the lower limit and upper limit of the valid data set of travel time is: t _min =t _25% -1.5×(t _75% -t _25% ) t _max =t _75% +1.5×(t _75% -t _25% ) t _25% --25% quantile t _75% --- 75% quantile t _min --- lower limit value of valid data set of travel time t _max ---The upper limit value of the valid data set of travel time. 4. a kind of expressway travel time prediction method based on particle swarm optimization wavelet neural network according to claim 1, is characterized in that, the concrete analysis process of described step C is: C1. Network construction; construct a three-layer wavelet neural network including input layer, hidden layer and output layer; as shown in Figure 1, where X ₁ , X ₂ ,..., X _m , represent the input of the wavelet neural network, Where m represents the number of input neurons; Y ₁ , Y ₂ ,...,Y _n , where n represents the number of output neurons, J represents the number of hidden layer nodes, ω _ij and ω _jk refer to the input layer and the Hidden layer, weight between hidden layer and output layer, where i=1,2,...,m; j=1,2,...,J; k=1,2,..., n; C2. Network initialization; randomly initialize the scaling factor a _j of the wavelet function, the translation factor b _j , the connection weight ω _ij between the input layer and the hidden layer and the connection weight ω _jk between the hidden layer and the output layer; set the wavelet the learning rates η ₁ and η ₂ of the neural network; The adjustment formula of the scaling factor is as follows:

--- Adjust the front scaling factor;

--- Adjusted scaling factor; η ₂ --- the learning rate of wavelet parameters, preferably η ₂ =0.01; The adjustment formula of the translation factor is as follows:

---Translation factor value before adjustment;

--- Adjusted translation factor value; η ₂ --- the learning rate of wavelet parameters, preferably η ₂ =0.01; The adjustment between the input layer and the hidden layer is adjusted by the following formula:

--- Adjust the weights between the input layer and the hidden layer before;

---The weight between the input layer and the hidden layer after adjustment; η ₁ --- the learning rate of the network weight parameter, preferably η ₁ =0.01; The adjustment between the hidden layer and the output layer can be adjusted by the following formula:

--- Adjust the weights between the hidden layer and the output layer before;

---After adjustment, the weights between the hidden layer and the output layer; η ₁ --- the learning rate of the network weight parameter, preferably η ₁ =0.01; C3. Predict the output; input the training samples into the network to calculate the predicted output of the network and calculate the error e between the network output and the expected output; C4. Error calculation; calculate the error e between the predicted output and the expected output of the network; the specific calculation formula is:

y _p,k (k)---actual output value; t _p,k --- ideal output value; C5. Weight correction; use the gradient descent method to correct the weights and parameters of the wavelet neural network; C6. If the number of training times is greater than 1000 or e meets the prediction accuracy requirements; e<0.0001, the error accuracy is set to 0.0001, and the prediction result is returned after the training, otherwise the learning and training continue. 5. a kind of expressway travel time prediction method based on particle swarm optimization wavelet neural network according to claim 1, is characterized in that, The specific analysis process of the step D is: D1. Network construction: construct a three-layer POS-WNN neural network including input layer, hidden layer and output layer D2. Parameter initialization processing: the system randomly generates S particles, and uses the position vectors of these particles to represent the scaling factors and translation factors a _j , b _j in the wavelet basis function; and between the ith input and the jth hidden layer The weight value ω _ij and the weight value ω _j between the jth hidden layer and the output layer, the calculation formula is as follows; at the same time, set the maximum and minimum speed of the particle, the learning rate and the maximum number of iterations; x _s =(x _s,1 ,...,x _s,d ,...,x _s,D )=(a ₁ ,...,a _J ,b ₁ ,...,b _J ,ω ₁₁ ,...,ω _1J ,ω ₂₁ ,...,ω _2J ,...,ω ₂₉₁ ,...,ω _29J ,ω ₁ ,...,ω _J ) D=m×J+J×(m-n)+(m-1)×J m---number of input layer nodes J---the number of hidden layer nodes n---the number of output layer nodes D3. Network training; based on the error between the actual output value and the ideal output value, calculate the particle fitness in each iteration process, and the calculation formula is as follows;

Q---Total number of training samples n---the number of output neurons of the network y _p,k (k)---actual output value t _p,k --- ideal output value; D4. Use the particle fitness as an indicator to determine whether the set error requirement is met. If the set error requirement can be met, that is, the fitness value remains basically unchanged, then the training is completed and go to step F; if it still fails to meet the set error requirement To the error requirement, proceed to the next step; D5. Determine whether the number of training has reached the set maximum number of iterations, if so, jump out of the loop, go to step F, and stop training; otherwise, update the speed and position of the particle according to the following formula, and return to step C to continue training; v _s,d (i+1)=ω×v _s,d (i)+c ₁ r ₁ [p _s,d -x _s,d (i)]+c ₂ r ₂ [p _g,d -x _s,d (i)] x _s,d (i+1)=x _s,d (i)+v _s,d (i+1) s=1,2,..,S; d=1,2,...,D c ₁ ,c ₂ --- acceleration factor i---current iteration number ω---Inertia factor r ₁ , r ₂ --- random numbers uniformly distributed between 0 and 1; D6. Network debugging: Select the sample value of the training sample in a specific period as the input value, and the actual value of another specific period in the training sample as the ideal output value, and calculate the error according to the fitness calculation formula of step C. If the error accuracy setting is satisfied , then end the operation; if the set error accuracy cannot be met, go to step C.

Images