CN105761488B - Real-time extreme learning machine Short-time Traffic Flow Forecasting Methods based on fusion - Google Patents

Abstract

The invention discloses a fusion-based real-time extreme learning machine short-term traffic flow prediction method, aiming at the prediction technology in the short-term non-stationary traffic flow scene, based on the three characteristics of short-term traffic flow real-time, accuracy and reliability and Fused real-time extreme learning machine to predict short-term traffic flow. The short-term traffic flow prediction method proposed by the present invention is based on a simplified single hidden layer feed-forward neural network structure, which can quickly train historical data during the peak period of traffic flow and incrementally update the arriving data, while ensuring a certain prediction accuracy Also save study time. In addition, the adoption of fusion mechanism ensures the stability and robustness of short-term traffic flow prediction. Reconstruction is performed during periods of missing data and severe fluctuations, and the training phase takes a short time, and the root mean square error and standard error percentage of the prediction results are all within the confidence area.

Description

Fusion-based real-time extreme learning machine short-term traffic flow prediction method

technical field

The invention mainly relates to the fields of intelligent transportation systems such as machine learning and traffic flow forecasting, especially a fusion-based real-time extreme learning machine short-term traffic flow forecasting method.

Background technique

With the development of the global economy and the progress of social urbanization, the development of the transportation industry has become more and more important. As an important material basis for the progress of human society, the transportation industry is the lifeblood of the entire national economic development. However, in recent years, the gradual increase of road vehicles has led to a deteriorating traffic operation efficiency. Traffic congestion, traffic exhaust pollution, complicated and inefficient traffic operations, and frequent traffic accidents have brought troubles to people's travel and other social activities. In order to alleviate traffic problems, a series of advanced management and control systems such as Intelligent Transportation System (ITS) have developed rapidly and been widely used, which can solve some road traffic problems to a certain extent. Traffic flow forecasting plays an important role in supporting the demand forecasting function of traffic management systems. Traffic flow data (also called traffic parameters) can directly reflect the macroscopic traffic status and is the basic data of traffic business. At the same time, traffic flow data is also the most easily collected data in traffic, which can be detected by induction coils, microwaves, and video , Global Positioning System (Global Position System, referred to as GPS), social media equipment and other methods to obtain, including traffic flow, traffic speed, traffic proportion, travel time and other information. Traffic flow prediction is essentially the prediction of these basic parameters of traffic flow. According to the length of the forecast period, traffic flow forecasting can be divided into two categories: short-term forecasting and medium- and long-term forecasting. The distribution of traffic flow data presents two peaks and two troughs, similar to the Gaussian distribution, peak prediction, and improving the real-time and high precision of short-term traffic flow prediction can not only inform drivers and passengers of traffic information in time, but also design and implement mobile infrastructure. Short-term traffic flow forecasting requires real-time performance and is difficult to predict. Thanks to the continuous improvement of vehicle and road sensing facilities and the support of ITS and traffic control systems, short-term traffic forecasting technology is also developing continuously.

In the early short-term traffic flow forecasting methods, most of the traffic flow forecasting methods were forecasted under the assumption of simple and balanced traffic flow data, and there were many restrictions on the data set, such as equal interval sampling, moderate interval length, appropriate number of historical data samples, The sample data has no noise, etc., but in the real traffic scene, due to the failure of the equipment itself or the interference of external factors (such as bad weather, abnormal roads, etc.), the traffic flow data is prone to loss and mutation during the collection and transmission process; During holidays, the traffic flow tends to surge and reach the peak value. These situations can lead to non-stationary and nonlinear heterogeneity in the traffic flow data. Here, the heterogeneity value data distribution is uneven and complex, coupled with the consideration of equipment cost factors , many data monitoring, collection, processing, transmission and other equipment cannot cover the entire traffic network, which increases the heterogeneity of traffic flow data. All these factors add difficulty to the modeling of traffic flow prediction models. Performance, accuracy, and stability need to be improved.

In order to strengthen the scalability and robustness of the short-term traffic flow forecasting method, so that it can still meet certain real-time and accuracy requirements in the case of unstable and heterogeneous traffic flow data, and expand the applicability of existing forecasting models, the Research on forecasting methods of real-time heterogeneous time-series traffic flow is necessary.

Contents of the invention

The invention proposes a fusion-based real-time extreme learning machine short-term traffic flow prediction method to improve the prediction accuracy and reliability of short-term traffic flow data, and is suitable for real-time traffic flow prediction.

The design idea of the present invention is to predict the short-term traffic flow based on the integrated real-time extreme learning machine. The short-term traffic flow data presents the characteristics of periodicity, real-time, accuracy and reliability, and the real-time traffic flow data is randomly selected, and the sequence learning idea It is applied to the ELM algorithm and a real-time sequence ELM algorithm is proposed. The real-time sequence ELM algorithm is a new algorithm proposed based on the original ELM algorithm and using the online learning mode. In this algorithm, data can be added to the network one by one or piece by piece, and the original data learning is completed. Then it will be discarded and no longer used.

The technical scheme adopted in the present invention is: a kind of fusion-based real-time extreme learning machine short-term traffic flow prediction method, comprising the following steps:

S1. Randomly select detectors on the road to collect short-term traffic flow data according to the preset time period;

S2, preprocessing and normalizing the obtained traffic flow data, and judging whether the traffic scene is in a stable situation or a non-stationary situation;

S3. If it is a non-stationary traffic scene, initialize the short-term traffic flow prediction model;

S4, establishing the real-time sequence learning part of the short-term traffic flow prediction model;

S5. Complete the prediction module in the short-term traffic flow prediction model;

S6. Denormalize and evaluate the prediction results.

S7. If it is a stable scene, the prediction can be performed directly according to steps S4 to S6.

Further, initializing the short-term traffic flow prediction model in step S3 includes the following steps:

S31. The initial data set is Randomly assign the input parameters of the prediction model, including the weight vector wi and the threshold bi between the input node and the hidden node, and randomly select the input weight ai and threshold bi of the hidden node, where i=1,2,..., L;

S32. Calculate the hidden layer output matrix H ₀ :

S33. Calculate the initial output weight β ⁽⁰⁾ . In order to ensure that the extreme learning machine can maintain the same learning performance, it is assumed that the rank of H is Have And it has been proved that β=H ⁺ T and H+=(H ^T H) ^-1 H ^T , then:

Where P ₀ =(H ₀ ^T H ₀ ) ^-1 , M ₀ =H ₀ ^T H ₀ =P ₀ ^-1 ;

where β, H and T refer to the sets {β ⁽⁰⁾ , β ⁽¹⁾ , β ⁽²⁾ , ..., β ⁽ⁿ⁾ }, {H ₀ , H ₁ , H ₂ , ..., Hn}, {T ₀ , T ₁ , T ₂ ,..., Tn};

S34. Set the arrived data block sequence k=0.

Further, the real-time sequence learning part of establishing the short-term traffic flow prediction model in step S4 includes the following steps:

S41. Calculate the hidden layer output matrix H _k+1 of newly added data:

S42. Order

and Then the output weight can be calculated:

S43, obtained according to the formula

S44. Set the arriving data block sequence k=k+1, indicating that the sliding window moves forward by one position, that is, the size of the sliding window is 1, and return to step S41.

Further, completing the prediction module in the short-term traffic flow prediction model in step S5 includes the following steps:

S51. Whenever a new data block k+1 arrives, each real-time sequence learning machine trains β ^(k+1) to calculate f _k+2 , where f _k+2 represents the traffic flow value predicted at k+2 moment;

S52, put f _k+2 into the test set to predict the traffic flow value at the next moment;

S53, as long as there are new data blocks arriving, return to step S51;

S54, according to the formula Calculate the weighted average to realize the weighted fusion mechanism;

Among them, let the number of single real-time extreme learning networks be L, and each network has the same number of hidden layer nodes and activation functions.

Further, the evaluation index of the step S6 is:

Assume that the observed actual traffic flow data sequence is fi or Yp(t), and the predicted traffic flow value result is ti or Yr(t)

(1) Absolute percentage error

(2) Mean absolute percentage error

(3) root mean square error

(4) Average relative error

The average relative error reflects the degree of deviation of the traffic flow prediction value from the real value, and the smaller the value, the better the prediction effect;

(5) Mean absolute error

The average relative error reflects the absolute value of the error between the traffic flow prediction value and the real value, and the smaller the value, the better the prediction effect;

(6) mean square error

This index not only reflects the magnitude of the traffic flow forecast error, but also reflects the discrete distribution of the error. The smaller the value, the smaller the error dispersion and the better the prediction effect;

(7) Fitting degree

The fitting degree reflects whether the traffic flow prediction curve fits the change trend of the actual observation curve from the aspect of the geometric characteristics of traffic flow. The larger the value, the closer the predicted value of traffic flow is to the actual observed value, and the better the prediction effect.

Compared with the prior art, the present invention has the advantages of:

Based on the simplified single hidden layer feed-forward neural network structure, it can quickly train historical data and incrementally update the arriving data during the peak period of traffic flow, saving learning time while ensuring a certain prediction accuracy. In addition, the adoption of fusion mechanism ensures the stability and robustness of short-term traffic flow prediction. The invention performs reconstruction in the period of data absence and severe fluctuation, and the time consumption of the training stage is short, and the root mean square error and standard error percentage of the prediction result are all within the confidence region.

Description of drawings

Fig. 1 is the flow chart of short-term traffic flow prediction of the present invention;

Fig. 2 is the realization flowchart of the real-time extreme learning machine prediction algorithm based on fusion of the present invention;

Fig. 3 is a comparison diagram of the actual traffic flow and the predicted traffic flow value when the detection point US101N is in the peak period 5:00AM-10:00AM in the first example scene of the present invention without missing data;

Fig. 4 is a comparison chart of the actual traffic flow and the predicted traffic flow value under the condition that the detection point US101N is in the peak period 5:00PM-10:00PM in the example scene one of the present invention without missing data;

Fig. 5 is a comparison chart of APE values under the condition that the detection point US101N is in the case of no missing data during the two peak periods in the example scenario 1 of the present invention;

Fig. 6 is a comparison diagram of the actual traffic flow and the predicted traffic flow value under the condition that the detection point SR120E is in the peak period 5:00AM-10:00AM in the second example scene of the present invention when there is missing data;

Fig. 7 is a comparison diagram of the actual traffic flow and the predicted traffic flow value under the condition that the detection point SR120E is in the peak period of 5:00PM-10:00PM in the second example scene of the present invention when there is missing data;

Fig. 8 is a comparison chart of APE values in the case of missing data during the two peak periods at the detection point SR120E in the second example scenario of the present invention.

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings and specific preferred embodiments, but the protection scope of the present invention is not limited thereby.

Referring to Figure 1 and Figure 2, the fusion-based real-time sequence extreme learning machine short-term traffic flow prediction method mainly includes the following steps:

Step 1. Collect short-term traffic flow data. The present invention uses the PeMS system to randomly select the detection points of four highways in California, USA to obtain the traffic flow historical data and perform prediction analysis. And randomly select the traffic flow data from 2014-11-24 to 2014-12-1, the data includes the traffic flow data of normal working days and holidays, which can represent the scenes of smooth and non-stationary periods. Among them, the data of the first 7 days is used as the training set, and the data of the last day is used as the test set. These data are used to predict the traffic flow changes in the two cases of complete data and missing data during the peak traffic flow period of 5:00-10:00AM and 5:00-10:00PM in a day. The invention divides the collected data into two parts: Scenario 1, using the collected data set to learn and predict the complete data of the traffic flow peak period in a day; Scenario 2, in the case of lack of traffic flow data, use the improved The real-time extreme learning machine framework is used for data reconstruction prediction, and the fusion mechanism is added and verified to make the prediction result more stable under the premise of ensuring a certain prediction accuracy.

Step 2. Import the collected short-term traffic flow data. If it succeeds, continue to judge the traffic scene, if it fails, exit and re-import the data.

Step 3. Preprocessing the collected short-term traffic flow data. As the basis of short-term traffic flow forecasting, data quality plays an important role in the effectiveness of short-term forecasting models. To judge whether the traffic flow data is abnormal, it needs to be judged: the driving of vehicles on the road satisfies certain rules, so the collected traffic flow data must belong to one of the following two situations: (1) If the traffic flow flow>0, speed> 0 and occupancy>0; (2) if flow=0, then speed=0. Traffic flow data that does not meet any of the above conditions is considered to be obviously abnormal data. After preprocessing, the data is normalized. Normalization is to speed up the prediction calculation speed and reduce time consumption.

Step 4. Initialize the short-term traffic flow prediction model. The initial dataset is Randomly assign the input parameters of the prediction model, including the weight vector wi and the threshold bi between the input node and the hidden node, and randomly select the input weight ai and threshold bi of the hidden node, where i=1,2,..., L;

Step 5, establishing a real-time sequence learning mechanism. In this algorithm, a sliding window is used to dynamically slide according to the spatio-temporal relationship of the trained and newly arrived traffic flow data. The data can be added to the network one by one or piece by piece. As the data blocks added in step k+1 continue to arrive, where N _k+1 represents the number of data added in step k+1, calculate the hidden layer output matrix H _k+1 of the newly added data:

Reorder Calculate output weights:

Set the arriving data block sequence as k=k+1, and return to step 4.

Step 6. Add an adaptive discarding mechanism. For the data that has been trained, it has little effect on the newly added data. Different weights are assigned according to the sequence of distance from the target time and the characteristics of the data itself, and the part is discarded adaptively, and then predicted together with the new training samples to reflect the difference in data. Qualitative while ensuring the accuracy of the forecast.

Step 7. Add a weighted average fusion mechanism. The weighted average fusion mechanism takes into account the influence of multiple real-time sequence learning machines with the same structure, and performs weighted average of multiple predicted results according to the following formula:

Step 8. The experimental results are denormalized, and then evaluated and analyzed. In order to intuitively reflect the predicted value and actual observed value of short-term traffic flow data in the peak period, the present invention respectively selects the observation data and Forecast data, compare different prediction algorithms, and draw images as shown in Figure 3 to Figure 8. Among them, the black dotted line represents the actual observation value, and the rest represent the traffic flow predicted by the multi-layer perceptual neural network, the predicted value of the wavelet neural network, the traffic flow predicted by the extreme learning machine, and the traffic flow predicted by the fusion of real-time extreme learning machine. In Scenario 1, the predicted value of the traffic flow at the detection point US101N is compared with the actual value. The prediction interval is 5 minutes, and the predicted value integrated with the real-time extreme learning machine is closest to the trend of the actual observed value. However, the multi-layer perceptual neural network fluctuates obviously when the traffic flow data fluctuates greatly, and the prediction error is relatively large. It can be seen from Figure 3 that the predicted value calculated by ERS-ELM is the closest to the real traffic flow trend, even in the time period of 6:10-7:00 AM with relatively large fluctuations, so most of the APE error is the smallest. Extreme learning machine is the second best algorithm, but the error value during 6:00-7:05AM is still very large. During 6:00-7:30AM, the predicted value obtained by the wavelet neural network deviates the farthest from the actual value. However, the fluctuation of traffic flow data during 5:00-10:00PM is jagged, which brings great difficulties to prediction. Figure 4 shows that the predicted value calculated by ERS-ELM is the closest to the actual traffic flow trend, while the predicted values obtained by the other three algorithms are not as good as ERS-ELM. Likewise, the APE value is also the smallest.

There is corrupted traffic flow data during the peak traffic flow in Scenario 2. Every time the algorithm is run, the algorithm needs to predict the corrupted data by reconstructing the traffic flow trend curve. It can be seen from the experimental results that the proposed algorithm can learn historical data faster, and obtain the next-moment forecast data by sliding the window and incrementally adding the previous-moment forecast data, and the prediction accuracy is better. It can be seen from the figure that there is damage in the three time periods of 5:20-6:05AM, 6:55-7:10PM, and 8:45-9:00PM. In order to better compare the actual traffic flow value with the predicted The error between the traffic flow values, this paper draws all the curves, including the missing part of the traffic flow data due to damage. The diagram is shown with a box icon. It can be seen from the figure that in the time period when the trend is stable, the four prediction methods can be well matched, but in the damage stage, ERS-ELM is the algorithm that matches the traffic flow trend best. For example, in Figure 7, in the time period of 6:55-7:10PM and 8:45-9:00PM, only ERS-ELM predicts a corner trend, while the other three algorithms simply follow the existing trend. The prediction performance of ELM is second only to the ERS-ELM algorithm, while the wavelet neural network and multi-layer perceptual neural network do not adapt to fluctuation points.

We can also draw the following conclusion that the smoother the traffic flow trend, the higher the prediction accuracy. But the dataset is not the only factor that affects predictive performance. In non-stationary scenarios, such as corrupted data, the algorithm needs to effectively learn the relationship between historical data and analyze the trend of data changes. ERS-ELM can learn historical data well and quickly, and maintain high prediction accuracy through variable-sized sliding windows.

The data used in the embodiment of the present invention comes from the open performance evaluation system platform PeMS 14.0.

The above are only preferred embodiments of the present invention, and do not limit the present invention in any form. Although the present invention has been disclosed above with preferred embodiments, it is not intended to limit the present invention. Any person familiar with the art, without departing from the scope of the technical solution of the present invention, can use the technical content disclosed above to make many possible changes and modifications to the technical solution of the present invention, or modify it into an equivalent implementation of equivalent changes example. Therefore, any simple modifications, equivalent changes and modifications made to the above embodiments according to the technical essence of the present invention shall fall within the protection scope of the technical solution of the present invention.

Claims (3)

1. A fusion-based real-time extreme learning machine short-term traffic flow prediction method is characterized in that, comprising the following steps: S1. Randomly select detectors on the road to collect short-term traffic flow data according to the preset time period; S2, preprocessing and normalizing the obtained traffic flow data, and judging whether the traffic scene is in a stable situation or a non-stationary situation; S3. If it is a non-stationary traffic scene, initialize the short-term traffic flow prediction model; S4, establishing the real-time sequence learning part of the short-term traffic flow prediction model; S5. Complete the prediction module in the short-term traffic flow prediction model; S6. Denormalizing and evaluating the prediction results; S7. If it is a stable scene, it can directly predict according to the steps from S4 to S6; S31. The initial data set is Randomly assign the input parameters of the prediction model, including the weight vector wi and the threshold bi between the input node and the hidden node, and randomly select the input weight ai and threshold bi of the hidden node, where i=1,2,..., L; S32. Calculate the hidden layer output matrix H ₀ : S33. Calculate the initial output weight β ⁽⁰⁾ . In order to ensure that the extreme learning machine can maintain the same learning performance, it is assumed that the rank of H is Have And it has been proved that β=H ⁺ T and H+=(H ^T H) ^-1 H ^T , then: Where P ₀ =(H ₀ ^T H ₀ ) ^-1 , M ₀ =H ₀ ^T H ₀ =P ₀ ^-1 ; where β, H and T refer to the sets {β ⁽⁰⁾ , β ⁽¹⁾ , β ⁽²⁾ , ..., β ⁽ⁿ⁾ }, {H ₀ , H ₁ , H ₂ , ..., Hn}, {T ₀ , T ₁ , T ₂ ,..., Tn}; S34. Set the arrived data block sequence k=0; The real-time sequence learning part of setting up short-term traffic flow prediction model in step S4 comprises the following steps: S41. Calculate the hidden layer output matrix H _k+1 of newly added data: S42. Order and Then the output weight can be calculated: S43, obtained according to the formula S44. Set the arriving data block sequence k=k+1, indicating that the sliding window moves forward by one position, that is, the size of the sliding window is 1, and return to step S41. 2. a kind of real-time extreme learning machine short-term traffic flow prediction method based on fusion according to claim 1, is characterized in that, in the step S5, completes the prediction module in the short-term traffic flow prediction model and comprises the following steps: S51. Whenever a new data block k+1 arrives, each real-time sequence learning machine trains β ^(k+1) to calculate f _k+2 , where f _k+2 represents the traffic flow value predicted at k+2 moment; S52, put f _k+2 into the test set to predict the traffic flow value at the next moment; S53, as long as there are new data blocks arriving, return to step S51; S54, according to the formula Calculate the weighted average to realize the weighted fusion mechanism; Among them, let the number of single real-time extreme learning networks be L, and each network has the same number of hidden layer nodes and activation functions. 3. a kind of real-time extreme learning machine short-term traffic flow prediction method based on fusion according to claim 1, is characterized in that, the evaluation index of described step S6 is: Assume that the observed actual traffic flow data sequence is f _i or Y _p (t), and the predicted traffic flow value is t _i or Y _r (t) (1) Absolute percentage error (2) Mean absolute percentage error (3) root mean square error (4) Average relative error The average relative error reflects the degree of deviation of the traffic flow prediction value from the real value, and the smaller the value, the better the prediction effect; (5) Mean absolute error The average relative error reflects the absolute value of the error between the traffic flow prediction value and the real value, and the smaller the value, the better the prediction effect; (6) mean square error This index not only reflects the size of the traffic flow forecast error, but also reflects the discrete distribution of the error. The smaller the value, the smaller the dispersion of the error and the better the forecast effect; (7) Fitting degree The fitting degree reflects whether the traffic flow prediction curve fits the actual observation curve from the geometric characteristics of traffic flow; the larger the value, the closer the predicted value of traffic flow is to the actual observed value, and the better the prediction effect is.