Sage Journals: Discover world-class research

Abstract

In view of the complex road conditions in today’s cities, the traditional prediction methods for road conditions are not so accurate, and the optimization algorithm for the logistics distribution path is not sensitive to changes in the road conditions so that its application in an actual logistics distribution system is not effective. This article proposes a road condition prediction and logistics distribution path optimization algorithm based on traffic big data. First, it analyses the characteristics of the road condition information of traffic big data. By combining the powerful feature extraction and self-learning ability of a deep belief network, it establishes a road condition prediction model based on a deep belief network and completes the model training and verification through the learning of traffic big data. Then, it combines the road condition prediction (result) information, traffic network information, and logistics distribution information to construct the time-share weighted traffic network. It then modifies the access set and pheromone variables of the ant algorithm based on the time-share traffic network to establish the road condition prediction and logistics distribution path optimization algorithm based on traffic big data. Finally, it conducts comparative experiments with other logistics distribution path optimization algorithms. The experimental results show that the proposed algorithm is superior to other logistics distribution optimization algorithms. Therefore, this algorithm is an effective method for optimizing logistics distribution.

Keywords

Traffic big data deep belief network road condition prediction logistics distribution path optimization algorithm

Introduction

Path optimization for logistics distribution is comparatively complex, and it mainly involves path selection in logistics distribution. The transport fleet, time limit, transport cost, and other conditions should be considered in the logistics distribution path optimization process. In actual logistics distribution activities, the proportion of these factors changes; therefore, many aspects of logistics distribution path optimization are researched and various algorithms are also adopted. The problems encountered in the actual distribution process should be fully considered when selecting an algorithm by incorporating the distribution capability, time, and other cost restrictions; then, an excellent distribution path solution could be obtained by the algorithm.

With the thriving development of China’s logistics industry, the logistics industry urgently needs the support of the relevant theory and technology. In recent years, China’s cities have experienced very severe traffic congestion, and the logistic distribution path optimization has not only involved simple combinatorial optimization. Since the complexity and changes of road conditions have a great impact on the logistic distribution path optimization and the path travel time would have larger fluctuations in different time ranges, the road condition changes should be considered in the path optimization research for logistics distribution. Accordingly, the research on path optimization for logistics distribution has important application value and research significance for the development of the logistics industry.

Relevant research

In the 1960s, Dantzig and Ramser proposed path optimization¹ for the first time, and its essence is embodied in the development and change of traveling salesman. In the early, the road traffic had small pressure, the impact factor was less, the calculated quantity of road net was not huge, and the relevant theory just started its development. At such period, the optimization algorithm of logistics distribution path mainly refers to the accurate algorithm in combination with the mathematical theory, including the cutting plane algorithm, branch-and-bound method and Dijkstra algorithm.² These algorithms mainly apply the combinatorial optimization ideas in mathematics to find out the best solution as far as possible. Later, with the development of traffic scale and category, the road net becomes huge, and the traditional path optimization algorithm has not solved the complex road net optimization. At this time, with the prosperous development of natural calculation theory, many scholars raise the heuristic algorithm by combining the natural calculation theory and path optimization. Although such kind of algorithm guarantees the best solution, it approximates the best solution to the maximum extent, and its computation efficiency boosts greatly upon the large-scale solution. Common heuristic algorithm includes ant algorithm, genetic algorithm, simulated annealing algorithm, and particle swarm optimization.^3–6 These heuristic algorithms are the natural calculation of simulating the phenomenon principle in the natural world, which seeks for the best solution or similar best solution rapidly based on the priori knowledge, and the best solution is sought within the acceptable time. Recently, there are more algorithms about the neural network, especially, the rising of deep learning provides a kind of new support in theory for solving the logistics distribution path optimization.

Most algorithms adopted by the scholars of China’s colleges and universities for path optimization in logistics distribution apply the ant algorithm. Zhang Wenguang⁷ applies the ant algorithm to path optimization, and the slope and congestion degree are converted into energy to consume the equivalent flat journal so that the route solution better improves the efficiency of logistics distribution. Chen Yuquan⁸ proposes an ant algorithm based on an improved Pareto approach, and Li Jing⁹ proposes an ant transfer strategy and a pheromone updating method to address the slow convergence rate of the traditional ant algorithm and the best partial problems to improve the convergence speed and computational time. Chen Jianjun¹⁰ applies the ant algorithm based on the optimized mathematical model of the logistics distribution path. In the simulation experiment, such an algorithm model has a strong optimizing ability overall, with a rapid search speed. Xu Xing¹¹ analyses the advantages and disadvantages of the ant algorithm, introduces a genetic manipulation, and amends the pheromone updating method to solve the logistics distribution path problem; this approach effectively determines the optimal route or a near optimal route. Nie Jingjing¹² introduces the partial best strategy and the nearest-neighbor algorithm to the ant algorithm, and addresses the low optimization speed and partial convergence.

Road condition prediction model based on deep belief network

Traffic big data and their characteristics

Traffic big data mainly include the driving position, license plate, traffic flow, road conditions, surrounding environment and other information collected by road traffic vehicle cameras, global positioning systems, road cameras, speedometers, radio-frequency identification technology, and other sensors; it may also include information from transportation and express delivery services regarding distribution vehicles, routes, and delivery times of logistics companies; weather information, such as sun, rain, temperature, humidity, and wind may be provided by the Meteorological Bureau; and road construction and closure information may be sourced from road construction companies. The characteristics of traffic big data are as follows. Traffic big data consist of many types of data from a wide number of sources, and the amount of data is massive. The amount of data for a medium-sized city is approximately 500 PB per year and there is a lack of data in some areas, and there may also be errors, redundancy, and other data issues. The data have multidimensional features, such as time, space, which contains great value, and the data stream is rapidly time-varying. Therefore, the accuracy of predicting the road traffic conditions depends on the data from motor vehicles, vehicle owners, roadside facilities, expressway toll stations, traffic police, meteorological bureaus, road construction companies, etc., and more data will enable a more accurate prediction.^13–17

Deep belief network

Deep learning is used in popular technical models such as the deep belief network (DBN) and convolutional neural network.¹³ The DBN model is characterized by a rapid training speed and strong learning ability. This paper applies a DBN to achieve the learning of traffic data, and the traffic forecast model is constructed based on the DBN. Since the DBN model is a directed acyclic graph overlaid by many restricted Boltzmann machines (RBMs), its input data are located at the bottom unit and are visible, and its parameter values are set according to a priori knowledge, so it could also be called a probability model. The DBN could also be understood as a network model or multilayer RBM based on Bayesian probability with the following training processes:

The unsupervised training is conducted in the RBM by a self-encoded network method layer by layer to ensure that the feature vector maps to different feature spaces, the feature information is reserved as much as possible, and the pre-training of the model is completed in this process.

On the last layer of the DBN, the RBM input is regarded as the input of the classifier, and the supervised training classifier conducts the back propagation (BP) of the error from top to down. The entire DBN network is slightly adjusted to complete the training process.^18-20

Road condition prediction model

Combined with the characteristics of DBN technology, to solve the accuracy problem of road condition prediction, this article proposes a road condition prediction model based on a DBN, as shown in Figure 1.

Figure 1.

Flowchart of traffic forecast.

The data pre-processing module is mainly the combination of traffic data pre-processing and road condition information (weather, time, etc.). The data pre-processing uses data repair, data replacement, data denoising, and other techniques to pre-process the abnormal data, recording errors, data null values, data distortion, etc. to improve the quality of the data, which will facilitate the model learning and improve the training speed and learning effect.

DBN training module: Cv indicates the average speed of the vehicle, Cn indicates the traffic flow, J indicates the holiday, the input vector also includes the characteristic data such as the weather value and the weekend value, Cv' and Cn' indicate the reconstructed feature value.

The traffic condition prediction of DBN needs to consider the weather, road construction, holidays, and commuting peaks, and the traffic flow data include the traffic volume Cn of road segment in unit time, the average traffic speed Cv, and the travel time Rt. The weather information is represented by Tq, which, according to the new version of rainstorm warning signal level standards, is divided into four levels of blue, yellow, orange, and red (IV, III, II, I), respectively, representing the general, heavier, severe, and particularly severe. Therefore, the weather information Tq = 1, 2, 3, 4, respectively, indicates that the rainstorm level is general, heavier, severe, and particularly severe. The time period t_i – t_i₊₁, the holiday value J (0, 1). The merged data set is the input to the DBN model.

Model construction

Considering the scale of urban traffic data, the number of layers for DBN model is set as five layers, and the structure diagram is shown in Figure 2.

Figure 2.

DBN model structure.

The five-layer DBN structure has met the requirements for traffic data learning very well, in which the first layer is the input layer, the last layer is the output layer, and the middle layer is the hidden layer. Based on the experience of setting the number of hidden layer neurons artificial neural networks,¹⁴ the number of neurons on the second layer is two-thirds of that on the first layer; the number of neurons on the third layer is two-thirds of that on the second layer; and the fourth and fifth layers are the inverse of the first and second layers, respectively.

The training model

This paper adopts the greedy unsupervised learning algorithm layer by layer based on the self-encode model as the pre-training process of BP algorithm. Each layer of deep learning model applies the stacked self-coded method, and the input on the rear layer is from the output of the front layer. The weight parameter on the first layer of training depth network model corresponding to the first automatic encoder is W1, as well, the weight parameter on the second layer of training depth learning model for the second automatic encoder is W2. The automatic encoder applies the model of RBM, and uses the parameter weight in the training model with specific dispersion. The preliminary pre-training of model refers to the weight learning and unsupervised learning by reconstructing the data. At the classification stage of model, the label is set manually for the supervised learning. Finally, the BP algorithm is adopted to slightly adjust the weight parameter, as shown in Figure 3.

Figure 3.

Training of prediction model.

At the classification stage, the learning sample data with manually set labels are provided to the classifier for learning. The classifier conducts the learning classification according to the characteristic quantity set after the reconstruction according to the labeled data and conducts the classification operation according to the characteristic quantity set after the learning process.¹⁵ First, build the training set. The vector V for the traffic information characteristics and traffic impact factor will be the new vector X after the learning of the DNN model. It is predicted that the model outputs 10 traffic rank values (0–9), in which level 0 indicates the failure of a passage, and level 9 indicates the rare vehicle at the traffic road section, which passes unimpeded. In accordance with the rank value set manually, set the class label, y⁽ ⁱ ⁾∈{0,1,2, … 9}, to each vector to indicate the 10 rank values predicated by the classifier. At this time, the expression of vector set with labels is as follows $L = {(x^{1}, y^{1}), (x^{2}, y^{2}) \dots (x^{m}, y^{m})}$ (1)

Then, the classifier predictions are conducted with the solution calculation method. The SoftMax model is promoted by and changed from the logistic regression model in the multi-classification research. The training set is input with the label vector set, L, and a method with more repetitions could be selected. The setting method of the hypothesis function is used to evaluate the probability value, p = (y = j|x), of each class j. In this way, the hypothesis function, h(x), can be expressed as $h_{θ}^{(x^{(i)})} = [\begin{matrix} p^{(y^{(i)} = 0 | x^{(i)} : θ)} \\ p^{(y^{(i)} = 1 | x^{(i)} : θ)} \\ ⋮ \\ p^{(y^{(i)} = 9 | x^{(i)} : θ)} \end{matrix}] = \frac{1}{\sum_{j = 1}^{10} e^{θ_{j}^{T_{x (i)}}}} [\begin{matrix} e^{θ_{1}^{T_{x (i)}}} \\ e^{θ_{2}^{T_{x (i)}}} \\ ⋮ \\ e^{θ_{k}^{T_{x (i)}}} \end{matrix}]$ (2)where θ indicates the parameter value of the classifier that must be calculated. In this way, only the cost function shall be set. Determine the parameter θ_k after the iterative process by selecting the classifier model parameter with the minimum number of iterations, thus obtaining the final prediction classifier. The cost function, j(θ), is expressed as $j (θ) = - \frac{1}{m} [\sum_{i = 1}^{m} \sum_{j = 1}^{k} 1 {y^{i} = j} \log \frac{e^{θ_{j}^{T_{x (i)}}}}{\sum_{j = 1}^{10} e^{θ_{j}^{T_{x (i)}}}}]$ (3)where y^(t) indicates the output eigenvector with the traffic rank value of 10, and the kth 1 (not 0) indicates the traffic rank value of K–1, for example, y⁽ⁱ⁾=[0,0,0,0,0,1,0,0,0,0] ^T indicates that the traffic rank value is 5, and the traffic is not blocked and smooth in the time period. The traffic rank value, Y, indicates the traffic capacity, the Y value is the label data of the DBN algorithm for supervision and training, and the solution formula for the Y value is as follows $Y = \frac{v_{t}}{v_{\max}} (1 - \frac{Cn}{C n_{\max}}) \times 9 φ$ (4)

In the above formula, the v_t, v_max, Cn, $C n_{\max}$ and φ indicate the average speed at the current time period, speed limit of the road, traffic flow within the time period, peak value of the traffic flow, and the adjustable parameter of the rank value, respectively. It is more accurate to reflect the traffic capacity of the road sections based on the speed and traffic flow.

Road condition prediction algorithm

The unsupervised learning process of the model is the core process of the road condition prediction deep belief network traffic forecast algorithm (DBNTFA), and its specific steps are as follows:

Input: Training sample data queue ((x(1), y(1)),(x(2), y(2)), … (x(M), y(M)))

Output: Weight coefficient [W¹, W², … .W^L]

Step 1: Select an initial value of the weight coefficient matrix, W, and a learning rate value, μ;

Step 2: Use the unlabeled training data {x⁽¹⁾, x⁽²⁾, … , x⁽ ^M ⁾} as the input layer of the DBN model;

Step 3: Determine the labeled training data queue (x⁽ ⁱ ⁾, y⁽ ^j ⁾) according to Formula (4);

Step 4: Use the output layer $a_{n}^{5}$ of the DBN model as the input of the SoftMax classifier and complete the SoftMax classifier solution according to Formulas (2) and (3), and the labeled data queue (x⁽ ⁱ ⁾, y⁽ ^j ⁾) is the output;

Step 5: Globally optimize the weights [W¹, W², … .W^L] by back propagation algorithm of the error;

Step 6: Output the final weight matrix [W¹, W², … .W^L].

Logistics distribution path optimization algorithm based on road condition prediction

Path optimization concept

The DBNTFA algorithm and logistics distribution path optimization are combined to realize the DBNTFA-based logistics distribution path optimization. Its basic concept is to solve the time-share traffic class value, Y^t, of the road through the DBNTFA algorithm and introduce Y^t into the traffic network to construct a time-share weighted traffic network; then, in the time-share weighted traffic network, the traffic road condition information, DBNTFA forecast information, and the departure time and distribution point of the logistics distribution are combined to determine the optimal path of logistics distribution by the improved ant algorithm. Its workflow is shown in Figure 4.

Figure 4.

Flowchart of logistics distribution path optimization based on DBNTFA.

Time-share weighted traffic network

The time-share weighted traffic network can be obtained by introducing the time-share weight into the topology of the traffic network, and the time-share weight could be solved on the basis of the traffic rank values and road section length. The weight could be understood as the combination of the passing time and risk. The expression of the time-share weight, $W_{l}^{t}$ , can be defined as follows $W_{l}^{t} = S \times \frac{C}{Y^{t}}$ (5)

In formula (5), S indicates the road section length; the longer the road section length, the longer is the passing time, and the larger is the weight accordingly. Y^t is the traffic rank value predicted at time period t. When the traffic level is 0, the road section is impassable due to a fault. The weight decreases when the weight is infinite and the traffic level is high. The constant C is used to adjust the weight scope to a range that is convenient for calculation.

A road section is the basic unit in the time-share weighted traffic network and is defined as follows in the traffic network: a road section is the minimum unit of a road, and there is no intersection leading to other road sections in the middle of a road section. The time period is determined by the updating cycle of the time-share weighted traffic network, and it is set according to the collection interval for the traffic data. The traffic data are collected every 5 min. The updating cycle of the rank value for the corresponding traffic network is 5 min. The weight in formula (5) is substituted into the traffic network to solve the time-share weighted traffic network.

Improvement of ant algorithm

Logistics distribution path optimization is solved by comparing the time-share weighted traffic network and the traditional ant algorithm with the following distinction aspects: in the traditional ant algorithm, the ant could directly visit other cities when starting from one city, and could only visit the conjoint road section in the traffic network; the path length of traditional ant algorithm is fixed, and the weight of the time-share traffic network is updated; the starting point in traditional ant algorithm includes all urban nodes, and the starting point in the time-share traffic network is the distribution point. The city and path length in ant algorithm are corresponding to the intersection node and road section weight in the time-share traffic network, respectively.

Formula (6) for the degree of expectation is defined as follows, and formula (7) is logistics distribution path formula. Substitute formulas (6) and (7) into formula (8) of the moving probability,¹⁶ to obtain formula (9) that ant moves from one intersection node to another intersection node $η_{ij} = \frac{1}{d_{ij} \cdot w_{l}}$ (6) $d_{ij}^{t} = S \times W_{l}^{t}$ (7) $p_{ij}^{k} (t) = {\begin{array}{l} \frac{{[τ_{ij} (t)]}^{α} \cdot {[η_{ij}]}^{β}}{\sum_{allowe d_{k}} {[τ_{ij} (t)]}^{α} \cdot {[η_{ij}]}^{β}} & if j \in allowe d_{k} \\ 0 & otherwise \end{array}$ (8)where allowed_k={0,1, … , n–1} indicates that the ant k can access the delivery point in the next step. $η_{ij}$ denotes the degree of expectation, α denotes the pheromone accumulated in the movement, and β denotes the different influence of the heuristic factor in the path selection. During the access process, the pheromone will continue to decay until it disappears. The pheromone disappears with the parameter 1−ρ, and the pheromone on the line is updated after the ant algorithm accesses the distribution point (with time n). The variables and parameters in formulas (6), (7), and (8) can be found in Ortega et al¹⁶ $p_{ij}^{k} (t) = {\begin{array}{l} \frac{{[τ_{ij} (t)]}^{α} \cdot {[S \cdot W_{l}^{t}]}^{- 2 β}}{\sum_{allowe d_{k}} {[τ_{ij} (t)]}^{α} \cdot {[S \cdot W_{l}^{t}]}^{- 2 β}} if j \in allowe d_{k} \\ 0 otherwise \end{array}$ (9)where allowed_k in the formula indicates the next intersection node that the ant could visit, and could be obtained by the time-share traffic network. When there is no visitorial intersection node, the ant stops moving. In case all distribution points are completed in the visit, the pheromone is updated. In case there is a distribution point which is not visited, the visit fails, and such visit information is abandoned.

The variable quantity of pheromone is Q/Z, the total length Z of the path could be calculated by formula (7). By combining with formulas (10) and (11), the updating formula of pheromone could be worked out $τ_{ij} (t + n) = ρ \cdot τ_{ij} (t) + Δ τ_{ij}$ (10)

The pheromone on the line should be updated after the ant algorithm accesses the delivery point once (with time n) $Δ τ_{ij} = \sum_{k = 1}^{m} Δ τ_{ij}^{k}$ (11)

In the above formula, $Δ τ_{ij}^{k}$ denotes the pheromone left by the kth in this iteration between distribution points i and j, which reflects the size of the path length, among them, m denotes have m ants between distribution points i and j. From formula (10) and formula (11), pheromone updating formula (12) can be obtained. $τ_{ij} (t + n) = ρ \cdot τ_{ij} (t) + \sum_{k = 1}^{m} \frac{Q}{d_{ij}^{t}}$ (12)

Through the above improvements, in the time-share traffic network, the following improved ant colony optimization algorithm is obtained as follows:

Algorithm 1: Improved Ant Path Optimization Algorithm

Input: Accessible node set {Rn}, delivery point [p]

Output: Optimal path S

Step1: Initialization time T = 0, cycle number N_c = 0, set the maximum number of cycles, put m ants randomly to n delivery points, the initial pheromone of all road segments is the constant C, the pheromone update $Δ τ_{ij}^{k}$ is 0, and put the initial intersection node in the taboo table;

Step2: Obtain the current intersection accessible node set {Rn} from the traffic network, and move it to the next intersection node according to Formulas (11) and {Rn};

Step3: Update the taboo table, and place the intersection node moved in Step (2) into the taboo table;

Step4: Step 2 and step 3 are executed cyclically, and this iteration ends when the ant stops moving;

Step5: Calculate the weight and Wa of walking route of each ant that successfully accesses all delivery points, and update the road segment pheromone and the optimal path S according to Formula (7).

Logistics distribution path optimization algorithm

The road condition prediction and logistics distribution path optimization algorithm based on DBN may be referred to as the Deep belief network traffic forecast path optimization (DBNTFPO) algorithm. This algorithm is a hybrid new algorithm, which combines the advantages of both time-share weighted traffic network and improved ant colony algorithm to make the systematic predictions of urban road conditions, so as to give the optimal solution for the logistics distribution path optimization problem under complex road conditions. The DBNTFPO algorithm can be described in two parts: the construction of time-share weighted traffic network [N'] and the solution of distribution path.

Algorithm 2: DBNTFPO Algorithm

Inputs: traffic feature data [D], traffic network logistics structure [N]; distribution time t, and distribution site n

Output: time-share weighted traffic network [N']; distribution route S

Construction of time-share traffic network

Use the traffic characteristic data [D] as the input of DBNTF algorithm, and obtain the time-share traffic class value Y^t;

Substitute the time-share traffic rank value Y^t into the formula (5), and the time-share weight $W_{l}^{t}$ of road section is obtained;

The weight of logical structure [N] is set in accordance with the road section time-share weight $W_{l}^{t}$ ;

Output the time-share weighted traffic network [N'].

Solution of distribution route

(v) Initialize the time-share weighted traffic network [N'] according to the delivery time t, that is, solving the initial weight $W_{}^{t}$ of road segment in the network;

(vi) Put m ants into n distribution points to conduct the solution through the Algorithm 1;

(vii) Output the optimal route S.

The DBNTFPO algorithm combines the DBNTF algorithm and the improved ant algorithm. Although the process is more complex, after one-time solution, the route solution could be obtained for the future solution after the completion of Algorithm 1.

Algorithm analysis

The DBNTFPO algorithm consists of two parts: the construction of time-sharing weighted traffic network [N'] and the solution of distribution route. Among them, the core algorithm for the construction of time-sharing weighted traffic network [N'] is DBN, and the time complexity of the algorithm is O(n²). For the solution part of distribution route, we assume that there are m ants placed in n distribution points, and the time complexity of the algorithm is O(n² m log m). Therefore, the time complexity of DBNTFPO algorithm is O(n² m log m).

Test results and analyses

Application examples

For an intelligent distribution platform, if a large-scale logistics company stored approximately 3.9 PB of big data such as the information for the vehicle, cargo, customer, finance, road traffic, weather, and road condition, we could extract the road condition prediction-related parameters and distribution-related information from approximately 1 PB of data over the past 5 years, including the road segment number, time period, average speed, traffic volume, weather level, wind speed, holidays, weekends, vehicle number, delivery point, intersection node, and other information. Using Hadoop’s big data processing platform and Map Reduce, we extracted the traffic information, weather information, and road condition information of 1012 roads from traffic big data. Among them, the information of 712 road lines is used as the training set, with 128 road lines used as the evaluation set and 172 road lines as the test set.

Experimental environment and parameter settings

The experimental hardware consists of 6 Blade servers (256 CPUs), 2 T K40 GPU graphics accelerator cards, and the cloud platform with 2 PB capacity. The software includes Linux (CentOS 7_x64), hadoop-2.7.2, jdk1.8.45, Eclipse 4.3.2 and MATLAB 2012 R.

One training period is one-half year (180 days), and the traffic data of 5–22 points per day is selected as the sample time range with a data sample collection interval of 5 min. The conversion formula for time (Time) and the time period, n, is Time = 5(h) + 5n(min), thus 5:05 a.m. is recorded as time period 1, 5:10 is recorded as time period 2, and there are 204 samples per road per day. In terms of the traffic flow statistics, the average motor vehicle is set as 1, the bus is set as 3, and the electric bike is set as 0.2. The blockage condition is divided into grades 1–3, of which grade 3 is a severe blockage, and the total number of training samples for 5 years is approximately 265,368,096. Table 1 shows the training sample data of 48 time periods for the 10 most congested roads at 7–9 a.m. and 5–7 p.m. The learning set of the classifier needs to set the traffic class value label for the training sample; the iteration number of the depth model is 100, and the number of iterations for the prediction classifier is 1000.

Table 1.

Some training sample data.

Road section number	Time period	Average speed	Vehicle flow	Weather grade	Wind speed	Holiday	Weekend	Blocking
1	48	12	283	3	3	11	5	3
2	48	15	236	2	2	5	8	3
3	48	22	258	2	4	6	13	3
4	48	27	198	1	7	10	14	2
5	48	30	215	1	1	8	7	2
6	48	15	266	2	1	3	5	3
7	48	24	150	1	2	4	4	3
8	48	17	178	4	5	7	8	3
9	48	18	196	2	3	9	13	3
10	48	21	201	1	2	7	55	3

Road condition prediction results and analysis

In this paper, the DBNTFPO algorithm is compared with the historical average method and support vector machine (SVM) algorithm for the road condition prediction experiments. Figure 5 shows the comparison chart of the three algorithms in terms of the traffic class value prediction accuracy.

Figure 5.

Prediction accuracy comparison chart.

Figure 6.

Accuracy under different road conditions.

Figure 7.

Different traffic class value prediction accuracy.

Figure 8.

Distribution time figure in three methods within 1 week.

As shown in Figure 5, when the learning period is short, the road condition prediction accuracy of the proposed algorithm in this paper is relatively poor, and the road condition prediction based on the SVM is more accurate. However, with the increase of the learning period, the prediction accuracy of this algorithm is greatly improved and finally stabilizes at approximately 86%. The accuracy of the SVM decreases with the increasing training scale, which is approximately 71%. The calculation of the historical average method is simple and convenient, and the accuracy does not change much with the increase of the data, and the accuracy is low. The reason is that the scale of the study set is small at the beginning stage, and the road condition characteristic factor effect learning is not enough. In addition, the depth model is under-fitting, and the prediction effect is not ideal. When the training set is increased, the deep learning model has a high degree of training, and the prediction accuracy is correspondingly improved.

As shown in Figure 6, the accuracy for the normal working day is relatively stable, but its accuracy is lower than the other two cases, and the accuracy is the highest in rainy weather. The reason is that, under normal circumstances, the road traffic conditions have fewer influencing factors and the road conditions are more random, so the prediction accuracy is relatively low. In rainy weather, the road traffic is obviously affected, and various factors have a greater impact on the road conditions. Nanning is in a rainy area, so the data training is relatively adequate, and the prediction accuracy is relatively high.

As seen from Figure 7, the prediction model can realize accurate results in the faulty road section and can achieve accurate predictions in the congested road section and the smooth traffic road. However, when the road conditions are irregular or the features are not obvious, the prediction accuracy is correspondingly reduced. Table 2 shows a comparison of the performance for three different algorithms. The prediction accuracy values of the algorithm proposed in this paper are the highest in the various road conditions and are all above 90%, and the historical average method has the worst prediction effect.

Table 2.

Performance comparison of several road condition processing methods.

Traffic condition forecast algorithms	Mean square error			Mean absolute percentage error
Traffic condition forecast algorithms	Workday	Rainy day	Weekend	Workday	Rainy day	Weekend
Historic average	0.5839	0.5236	0.5791	23.1625	23.8376	22.4851
SVM	0.1531	0.1529	0.1628	11.2617	12.121	12.3150
DBNTFPO	0.0664	0.0582	0.0607	8.8972	6.5411	7.2175

DBNTFPO: deep belief network traffic forecast path optimization; SVM: support vector machine.

Distribution path optimization test results and analysis

Now, take 1 week (7 days) for the distribution optimization test. There are 30 delivery vehicles every day to deliver the goods to 27,000 delivery points; the delivery time is from 7:00 a.m. to 19:00 p.m., and the distribution logistics path optimization algorithm proposed in this paper, ant algorithm, and driver experience methods are used to determine the distribution time, as shown in Table 3.

Table 3.

Number of solutions for each method in a week.

Algorithm	The number of solutions	Average number of road sections
Ant algorithm	5	189
Driver experience	35	178
DBNTFPO algorithm	59	175

DBNTFPO: deep belief network traffic forecast path optimization.

Table 4.

Average schedule.

Ant algorithm	144.5
Driver experience	121.0
DBNTFPO algorithm	108.5

DBNTFPO: deep belief network traffic forecast path optimization.

The path length of traditional ant algorithm is fixed and is not updated after one-time solution, and the driver could select the path depending on the driving experience. The DBNTFPO algorithm generates plenty of route resolutions by depending on the abundant traffic data advantage. The average quantity value of road section for such three algorithms has little difference, because different road sections have a very large difference in road condition, the road section contains many road sections, and the time may be short uncertainly. These route solutions are evaluated on the basis of the distribution time.

The distribution time data are recorded and collected by the distribution driver. Based on Figure 8, the algorithm in such paper has a relatively obvious time advantage, and the distribution time is relatively stable. Traditional ant algorithm includes the single route, and has larger change and fluctuation in time, with poor anti-risk capability. The distribution time is also unstable in the driver experience method, because the drivers mainly select the path according to individual experience, with the relatively stable distribution, but the path is not the optimal.

As seen from Table 4, the traditional ant algorithm has become not suitable for the path optimization problem under the complex road condition, and the algorithm in this paper could effectively reduce the distribution time compared with the distribution experience of drivers. In the actual logistics distribution, the algorithm in this paper has excellent applicability and effectiveness, and the time complexity is consistent with the theoretical analysis.

Conclusion

The innovation point of this paper is to introduce the big data and deep learning method into the path optimization problem. All traditional path optimization algorithms are mostly based on the simulation theories, which disjoint with the actual application. This paper uses the deep learning method to study the traffic big data, extract the feature factors of road condition, and accurately predict the road condition in combination with the classifier. This paper solves the poor practical application problem of traditional algorithm, and provides a new thought for solving the logistics distribution path optimization problem. The development of deep learning brings new solution to many fields. Thus, in future, as the technical problems of deep learning is solved, more efficient method will be brought to the logistics distribution path optimization, and it will change the model of logistic distribution path optimization.

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research,authorship,and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research,authorship,and/or publication of this article: This study is supported by the National Natural Science Foundation of China under Grant Nos. 61866006,61741203;Guangxi Natural Science Foundation (2016GXNSFAA380243);Guangxi innovation-driven development of special funds project (Gui Ke AA17204091);Guangxi Nanning Science and Technology Development Planning Project (20181015–5).

References

Dantzig

Ramser

RH.

The truck dispatching problem. Manage Sci 1959; 6(1): 80–91.

Neves-Silva

Tshirintzis

GA.

An extended Dijkstra’s algorithm for calculating alternative routes for evacuee agents in disaster simulation. Front Artif Intell Appl 2014; 262: 318–323.

Daneshzand

The vehicle-routing problem. Amsterdam: Elsevier, 2011.

Polat

A parallel variable neighborhood search for the vehicle routing problem with divisible deliveries and pickups. Comput Oper Res 2017; 85: 71–86.

Okulewicz M and Mańdziuk J. The impact of particular components of the PSO-based algorithm solving the Dynamic Vehicle Routing Problem. Appl Soft Comput 2017; 58: 586–604.

Alam Khan N and Shaikh

A smart amalgamation of spectral neural algorithm for nonlinear Lane-Emden equations with simulated annealing. J Artif Intell Soft Comput Res 2017; 7: 215–224.

Zhang

A research on E-business logistics terminal distribution mode and path optimization based on ant colony algorithm. Guizhou Normal University, China, 2016.

Chen

Logistics distribution vehicle routing problem based on Ant algorithm mode. Sci Technol Econ Guide 2016; (35): 152.

A unidirectional logistics distribution path optimization based on improved ant algorithm. Electron Des Eng 2016; 10: 68–70.

10.

Chen

Research on Ant algorithm in logistics distribution routing optimization. Comput Simul 2011; 2: 268–271.

11.

Research on logistics distribution routing optimization. Zhejiang University, China, 2006.

12.

Nie

Research on logistics distribution optimal path based on Ant algorithm. Automat Instrum 2016; 5 : 3–5.

13.

Fan

Fei

Shen

Accelerating deep learning with a parallel mechanism using CPU+MIC. Int J Parallel Prog. 2017; 46: 1–14.

14.

Singh

Guntuku

Thakur

, et al. Big data analytics framework for peer-to-peer botnet detection using random forests. Inform Sci 2014; 278: 488–497.

15.

Suchanek

Weikum

Knowledge bases in the age of big data analytics. Proc VLDB Endow. 2014; 7: 1713–1714.

16.

Ortega

NAA

Ruiz-Vanoye

, et al. A-means: Improving the cluster assignment phase of K-means for big data. Int J Combinatorial Optimization Prob Inform 2018; 9: 3–10.

17.

Zhang

Yang

Chen

ZK.

A survey on deep learning for big data. Inform Fusion 2018; 42: 146–157.

18.

NIeson2012. Machine learning - Detailed analysis of DBN deep belief network [EB/OL].[2017-03-15], http://m.blog.csdn.net/ article/details?id=52184189 (accessed 20 January 2019).

19.

Wang

Determination of number of hidden neurons in artificial neural network. Chongqing University, China, 2012.

20.

Tan

Wang

Research on traffic jam prediction model based on deep learning. Res Comput Appl 2015; 10: 2951–2954.

Road section number	Time period	Average speed	Vehicle flow	Weather grade	Wind speed	Holiday	Weekend	Blocking
1	48	12	283	3	3	11	5	3
2	48	15	236	2	2	5	8	3
3	48	22	258	2	4	6	13	3
4	48	27	198	1	7	10	14	2
5	48	30	215	1	1	8	7	2
6	48	15	266	2	1	3	5	3
7	48	24	150	1	2	4	4	3
8	48	17	178	4	5	7	8	3
9	48	18	196	2	3	9	13	3
10	48	21	201	1	2	7	55	3

Road section number	Time period	Average speed	Vehicle flow	Weather grade	Wind speed	Holiday	Weekend	Blocking
1	48	12	283	3	3	11	5	3
2	48	15	236	2	2	5	8	3
3	48	22	258	2	4	6	13	3
4	48	27	198	1	7	10	14	2
5	48	30	215	1	1	8	7	2
6	48	15	266	2	1	3	5	3
7	48	24	150	1	2	4	4	3
8	48	17	178	4	5	7	8	3
9	48	18	196	2	3	9	13	3
10	48	21	201	1	2	7	55	3

Road condition prediction and logistics distribution path optimization algorithm based on traffic big data

Abstract

Keywords

Introduction

Relevant research

Road condition prediction model based on deep belief network

Traffic big data and their characteristics

Deep belief network

Road condition prediction model

Model construction

The training model

Road condition prediction algorithm

Logistics distribution path optimization algorithm based on road condition prediction

Path optimization concept

Time-share weighted traffic network

Improvement of ant algorithm

Logistics distribution path optimization algorithm

Algorithm analysis

Test results and analyses

Application examples

Experimental environment and parameter settings

Road condition prediction results and analysis

Distribution path optimization test results and analysis

Conclusion

Footnotes

Declaration of conflicting interests

Funding

References

Road section number	Time period	Average speed	Vehicle flow	Weather grade	Wind speed	Holiday	Weekend	Blocking
1	48	12	283	3	3	11	5	3
2	48	15	236	2	2	5	8	3
3	48	22	258	2	4	6	13	3
4	48	27	198	1	7	10	14	2
5	48	30	215	1	1	8	7	2
6	48	15	266	2	1	3	5	3
7	48	24	150	1	2	4	4	3
8	48	17	178	4	5	7	8	3
9	48	18	196	2	3	9	13	3
10	48	21	201	1	2	7	55	3