Optimal Transshipment Route Planning Method Based on Deep Learning for Multimodal Transport Scenarios

Abstract

The optimal path problem is an important topic in the current geographic information system (GIS) and computer science fields. The Dijkstra algorithm is a commonly used method to find the shortest path, which is usually used to find the least cost path from a single source. Based on the analysis and research of the traditional Dijkstra algorithm, this paper points out the problems of the Dijkstra algorithm and optimizes it to improve its storage capacity and operation efficiency. Then, combined with the traffic elements, a new network-based optimal path planning method is established. However, the existing network is far from actual operation in terms of the expansion of the transportation network, the uncertainty of the transportation environment, and the differences in the transportation area. Therefore, this paper proposes an optimal transshipment path planning method based on deep learning, which is oriented to multimodal transportation scenarios. This paper mainly introduces the intelligent transportation system and intelligent navigation system, and then conducts in-depth research on optimal path planning. This paper also uses the deep neural network algorithm to optimize the calculation, and finally analyzes its use and application. Simulation experiments were also performed to analyze the relationship between energy consumption, emissions, speed, load cost, and other factors under the optimal path. The final experimental results show that within the range of the emission limit of [100,200], the emission is 50%, the emission is less than 100%, but the emission is higher than 75%. In [100,200], 75% of the loading rate emits no less than 100%. In [200,300], the 50% and 100% emissions are the same. This also means that the emissions are the same but the paths are not necessarily the same.

1. Introduction

With the continuous expansion of the scale of cities, the continuous improvement of the degree of urbanization, the large number of vehicles, and the increasingly frequent exchanges between cities, the traffic problem in Chinese cities has increasingly become a problem that cannot be ignored. In modern traffic, traffic accidents occur frequently, the traffic environment deteriorates, and traffic congestion increases. These problems not only waste a lot of resources, but also seriously affect people’s travel safety. In the field of intelligent transportation, concepts such as vehicle navigation technology (VNS) and virtual navigation systems have also emerged.

Through reasonable path planning, urban traffic congestion can be effectively reduced and urban traffic resources can be reduced, thereby shortening the time required to reach the destination. This can not only save a lot of transportation resources, but also reduce vehicle exhaust emissions and reduce environmental pollution. Therefore, it is of great practical value to optimize the path problem.

Uncertainty in optimal path planning is essential for many applications. Subramani et al. proposed and applied stochastic level set partial differential equations to control stochastic time-optimal reachability fronts and time-optimal paths for vehicles navigating in dynamic flow fields [1]. The per-flow or per-path state on the switch can be eliminated by encoding the desired path of each packet in its header. A key component of this approach is the efficient encoding of paths through the network. Hari et al. introduced the mathematical formulation of this optimal path encoding problem [2]. For stealth unmanned aerial vehicle (UAV), path safety and penetration path search efficiency are the two most important factors for mission execution. Zhang et al. studied an optimal penetration path planning method considering kinematics principle, dynamic radar cross section of stealth UAV and network radar system at the same time [3]. Ramasudha introduced the optimal path finding algorithm to identify the optimal path for a distribution system during restoration using Dijkstra’s method. The goal is to reduce power losses and obtain an effective recovery plan after a large-scale outage in the system [4]. Algorithm research on optimal path planning is imminent, and deep learning combines underlying features to generate higher-level abstract attributes to discover distributed data. It thus discovers the characteristics of path planning and calculates its optimal route.

At present, the common detection techniques of steganographic communication mainly include three aspects: residual calculation, feature extraction, and binary classification. Jian et al. proposed a new method for image hiding analysis based on a convolutional neural network. This method can replicate and optimize each key link under a unified framework, and it can directly obtain hierarchical representation from the original image [5]. Over the past two decades, there have been many studies on the classification of hyperspectral data. However, most algorithms are unable to extract it hierarchically. Chen et al. first proposed the concept of deep learning in the classification of hyperspectral data [6]. Deep learning is rapidly becoming a state-of-the-art technology, thereby improving the performance of various medical applications. Shen et al. mainly expounded the basic theory based on deep learning and gave an overview of its applications in image registration, anatomy, cell structure, tissue segmentation, computer-aided diagnosis, and prognosis [7]. Morsali et al. applied fractional order PI and D controller to the design of series capacitor damping controller based on thyristor control and cooperates with the secondary integral controller as the automatic generation control loop [8]. They all introduced an optimal route and deep learning algorithm, but they did not apply it in real life, and did not discits advantages.

This paper focuses on the intelligent transportation system and intelligent navigation system. On this basis, the optimal route planning problem is deeply studied, the method based on deep neural network is used to solve it, and a simulation test is carried out. The innovation of this paper is that the paper optimizes the design of the algorithm while researching the algorithm and performs a simulation experiment in the experimental part. This research data is more accurate and convincing, and also has innovative value. The advantages and disadvantages of intelligent transportation system, intelligent navigation system, and optimal route planning are introduced.

2. Optimal Transshipment Path Planning Method Based on Deep Learning; Algorithms for Multimodal Transport Scenarios

2.1. Intelligent Transportation System

Intelligent transportation is an intelligent technology system that uses modern science and technology to improve traffic operations and relieve traffic pressure. Intelligent transportation system refers to the comprehensive application of various high technologies in the field of transportation. Intelligent transportation is a comprehensive transportation system aimed at improving urban road traffic conditions, increasing road traffic utilization, reducing motor vehicle emissions, reducing traffic accidents, facilitating travel, and reducing environmental pollution. Its network topology is shown in Figure 1.

2.2. Intelligent Navigation System

The car navigation system allows the accurate display and monitoring of the real-time traffic system in a certain area through GIS, GPS, communication technology, and other technologies. Its intelligent monitoring system is shown in Figure 2 [9,10]. The car navigation system can determine the current position of the car through the electronic map and it can provide the driver with the best running route according to the set requirements [11,12,13]. In addition, according to the precise route information that is provided by the navigation system to reach the area that has not been passed, it can improve the road conditions of the driver. By making full use of road resources, it selects the best path, thereby reducing traffic congestion, optimizing the road network, and realizing the optimal configuration of the road network.

The vehicle positioning system is a comprehensive application of advanced technologies such as intelligent vehicle positioning, communication, geographic information, and computers. It makes the use of vehicle positioning and navigation systems more convenient and reliable, which can provide users with different information and services [14,15].

(1) Vehicle positioning: The geographic information of the surrounding environment can be accurately positioned to the current position of the vehicle and displayed clearly on the electronic map.

(2) Route design: According to the driver’s route planning, the starting point and end point are automatically determined.

(3) Route guidance: It can provide the driver with relevant voice or image navigation instructions, so that the driver can pay attention to the traffic lights, speed limit, and other information on the referenced driving route, so that they can reach the destination safely.

(4) Integrated information: It can provide other peripheral search or query services for drivers. For example, according to the driver’s needs, the required information and data on features such as hotels, restaurants, scenic spots, and parking places can be displayed on the electronic map [16,17].

(5) Wireless communication: The car that is driven by the driver can receive real-time traffic broadcasts, so that the driver can grasp the road conditions in real-time. At the same time, the driver’s vehicle status is reported to the traffic radio or traffic control center to realize real-time traffic information.

2.3. Optimal Route Planning

The optimal path planning can make people plan the route ahead of time or at any time during the travel process. Now it is a basic application of the vehicle navigation system in the urban road network. In contemporary times, traffic congestion has become a very serious social, economic, and environmental problem. The R&D and development of ITS effectively controls traffic jams. ITS can monitor traffic conditions in real-time through various types of sensors or other traffic information acquisition devices. Through the collected monitoring data, the optimal road can be obtained, thereby providing the user with the best route. It thus achieves the maximum satisfaction of the needs of users [18,19,20]. Path planning has become a great help for people’s travel. It allows people to plan their own itinerary during travel, save people time and money, and reduce vehicle losses. At present, the optimal path algorithm has become a research hotspot in many fields, and the optimal path algorithm is being more and more widely used in today’s world. In addition, there is also more research of optimal path algorithms [21,22,23].

(1) Concepts related to optimal path planning

Definition of graph:

$$H(U,Z)$$

(1)

Path expression:

$$u=u_{j,0},u_{j,1,\dots },u_{j,n}=u^{\prime }$$

(2)

Definition of sparse graph:

$$z<m\log m$$

(3)

Adjacency matrix:

$$arc\left[j\right]\left[i\right]=\{\begin{array}{ll}{S}_{j,i},& if (u_j,u_i)\in E or <u_j,u_i>\in E\\ 0& if j=i\\ \infty & other\end{array}$$

(4)

An adjacency table is a concatenated memory of a graph. If the graph H is a weighted network, then the construction of the adjacent table is shown in Figure 3.

Using weights to represent the distance and power consumption between two vertices. The starting point, end point, intersection, etc., of the network are taken as a node, and the length of the road segment, travel time, etc., are used as the weight of the road segment. This works by abstracting the network into a weighted directed graph and applying the methods of graph theory to solve related problems [24].

The expression of the path weight sum is as follows:

$$DS(Q)={\displaystyle \sum _{r=0}^{m-1}t(u_r,u_{r+1})}$$

(5)

If the path weight is the length of a curve, then the distance from the start point to the end point of the route is the route. In a given weighted directed graph, a minimum path problem is acquired from the initial point to the sum of the final weights. If the weight is regarded as the road usage fee, then in a specific road network, finding the optimal route with the smallest travel cost is the optimal route planning problem. Therefore, the optimal path planning problem is closely related to the optimal path problem [25].

(2) Discussion on the optimal path calculation method

In the process of solving, this paper uses the minimum heap in the algorithm structure and considers the sorting and the selection of the optimal path node at the same time. First, the nodes to be classified are put into a group or a group of linked lists in order, and divide them into the smallest group, and then treat them as a binary tree. The tree data structure has two characteristics: first, each node in the tree has at most two child node; second, the child node is equal to or greater than its parent node. The second feature refers to the root node of the stack, which is the minimum value that is contained in the stack. The following is the basic operating procedure [26,27].

Step 1: The expression of the initial value of the optimal path length is as follows.

$$dist\left[j\right]=Adj(Locate(H,u_0), j) u_j\in U$$

(6)

Step 2: The minimum heap is initialized and the elements are inserted in the graph one by one into the heap. If $dist\left[j\right]$ represents the value of the heap element numbered $j$, it is shown in the following formula.

$$Heap\left[j+1\right]\leftarrow (j,dist\left[j\right]), 0\le j\le m$$

(7)

Step 3: The starting point $u_0$ is deleted in the heap as shown in the following formula.

$$Heap\left[m+1\right]\to u_0$$

(8)

$$dist\left[j\right]+Adj(u_j,u_r)<dist\left[r\right]$$

(9)

Which is updated to:

$$dist\left[r\right]=dist\left[j\right]+Adj(u_j,u_r)$$

(10)

2.4. Optimal Path Based on Deep Learning

The route planning problem refers to a series of connections that are marked with the beginning and the end on the connectivity graph, that is, the route. At present, mechanical algorithms and corresponding improved algorithms are commonly used. One of the biggest problems is that the input must be precise boundaries and nodes, which must be processed manually or by computer vision. This process is very cumbersome and time-consuming. Secondly, in the new path planning problem, it is necessary to start from scratch and perform iterative calculations based on CPU instructions, so there is no intelligent awareness and cannot be extended when solving the path planning problem. With the development of deep learning technology, the path planning of the original image can be realized, and an end-to-end model with learning generalization ability can be constructed. In the past, due to the lack of in-depth research on neural networks, neural networks have always been a “black box”, making it difficult for researchers to realize their potential [28]. Currently, research on deep learning-based road planning problems is still in its infancy, and it must be combined with reinforcement learning. On the theoretical basis of reinforcement learning, path planning is regarded as a series of single decision-making processes, and the way of deep learning of its decision-related functions is called deep reinforcement learning.

At present, the neural network based on reflection strategy is mainly used in path planning, and it does not have the ability to deal with long-term planning problems. Secondly, in order to improve students’ learning ability, it is sometimes difficult to obtain, and it takes a lot of time and money. Therefore, both in theory and in practice, there is an urgent need for an intelligent algorithm with “end-to-end” and “planning” functions, but its implementation and improvement still have certain difficulties.

Deep learning is a new field in the field of machine learning, and its core purpose is to build a neural network that is similar to the human brain. In addition, according to its working principle, it is interpreted and analyzed accordingly, and it is graded and explained in different conversion stages. To understand neural networks, one must first understand the simplest neurons. The neuron model is shown in Figure 4.

The output of the neuron can be expressed as:

$$g_{s,y}(a)=g(S^{D}a)=g\left({\displaystyle \sum _{j=1}^m{s}_j{a}_j+y}\right)$$

(11)

A neural network model is simply a number of simple neurons that are connected according to certain rules. Figure 5 shows a neural network model.

Such as:

$$e_j^{(2)}={\displaystyle \sum _{i=1}^m{S}_{ji}^{(1)}{a}_i+y_j^{(1)}}$$

(12)

Then

$$x_j^{(o)}=g\left(e_j^{(o)}\right)$$

(13)

After defining the symbols of each network, the computational steps of the neural network are as follows:

$$x_1^{(2)}=g\left(S_{11}^{(1)}{a}_1+S_{12}^{(1)}{a}_2+S_{13}^{(1)}{a}_3+y_1^{(1)}\right)$$

(14)

If the activation function $g(•)$ is extended to be represented by a vector:

$$g\left(\left[e_1,e_2,e_3\right]\right)=\left[g(e_1),g(e_2),g(e_3)\right]$$

(15)

Then, the above equation can be expressed as:

$$\begin{array}{l}{e}^{(2)}=S^{(1)}a+y^{(1)}\\ x^{(2)}=g\left(e^{(2)}\right)\\ e^{(3)}=S^{(2)}{x}^{(2)}+y^{(2)}\\ f_{S,y}(a)=x^{(3)}=g\left(e^{(3)}\right)\end{array}$$

(16)

The above calculation process is the forward propagation of the neural network. If the activation value of the zth layer of the network is $x^{(z)}$, then the activation value of the z + 1 th layer can be obtained as follows:

$$\begin{array}{l}{e}^{(z+1)}=S^{(z)}{x}^{(z)}+y^{(z)}\\ x^{(z+1)}=g\left(e^{(z+1)}\right)\end{array}$$

(17)

The forward propagation of the neural network will get the output value. A cost function is used to evaluate the difference between the output value of the neural network and the target value. The cost function of a single sample $(a,b)$ is expressed as:

$$J(S,y,a,b)=\frac{1}{2}{\Vert f_{S,y}(a)-b\Vert }^2$$

(18)

The above formula is the cost function of one sample data, then the overall cost function of n sample data is:

$$J(S,y)=\frac{1}{n}{\displaystyle \sum _{j=1}^{n}J(S,y,a^{(j)},b^{(j)})}$$

(19)

$J(S,y)$ is a mean square error term.

After getting the error term, we can get:

$$\begin{array}{l}\frac{\varphi }{\varphi s_{ji}^{(z)}}J(S,y,a,b)=x_i^{(z)}{\eta }_j^{(z+1)}\\ \frac{\varphi }{\varphi y_j^{(z)}}J(S,y,a,b)={\eta }_j^{(z+1)}\end{array}$$

(20)

After the gradient value is obtained, the value of the error term $J(S,y)$ can be reduced again and again through the gradient descent method, thereby training the entire neural network.

3. Optimal Transshipment Route Planning Method Based on Deep Learning; Experiment on Multimodal Transportation Scenarios

3.1. Simulation Experiment

Based on 12 randomly distributed cities, including four modes of transportation, road, water, rail, and air, the solid line shows that there is at least one mode of transportation between the two cities as shown in Figure 6.

It is assumed that the parameters are set as in

Table 1. Parameter assumptions for different modes of transport.

Shipping Method	Cost	Speed	Fuel Consumption	Power	Fuel Emissions	Power Emissions
Highway	35	90	1	NULL	500	NULL
Waterway	18	40	0.6	1	NULL	10
Railway	20	75	0.6	1	NULL	10
Aviation	145	600	10	10	NULL	10

. The transit process adopts road transportation. The freight railway station and port are usually located in the suburbs of the city, and the airport usually has a high speed. Therefore, in

Table 2. Average speed of transit between different modes of transport (km/h).

Shipping Method	Highway	Waterway	Railway	Aviation
Highway	NULL	40	35	50
Waterway	40	NULL	50	50
Railway	35	50	NULL	50
Aviation	50	50	50	NULL

, the transit speed of the traffic mode is specifically set. It starts from the railway station of the fifth city and goes to the wharf of the tenth city. In order to facilitate the analysis of the simulation results, the standard experimental data is assumed as follows: the load ratio is 75%, the emission trading price is assumed to be 0.01 RMB/kg, and the fuel price is 7.8 RMB/L.

The 12 cities are expanded into 12 “bottoms”, and the transfer points of each city are set according to the coordinates (x, y) of cars, ports, railway stations, and airports as shown in

Table 3. Transit point coordinates of the simulated city (a).

City	Bus Station	Port	Train Station	Airport
1	16.88	9.83	19.82	17.98
2	42.98	38.93	44.92	32.96
3	68.96	60.97	70.96	58.94
4	96.78	93.75	98.70	97.87
5	8.70	3.62	39.61	9.80
6	36.68	28.61	39.62	36.76

(a) are the coordinates of bus station, ports, train station and airport for each simulation city.

and

Table 4. Transit point coordinates of the simulated city (b).

City	Bus Station	Port	Train Station	Airport
7	69.58	62.55	65.50	68.67
8	24.38	20.37	36.37	24.48
9	65.37	60.34	66.35	63.46
10	96.38	93.45	98.30	86.39
11	44.33	40.18	46.20	44.30
12	84.14	80.4	86.5	85.23

(b) are the coordinates of bus station, ports, train station and airport for each simulation city.

. The car transfer station is placed at the middle because the flexibility of car transportation makes it easier to transport with various other transfer stations in the “underside”. In addition, in reality air transportation between two cities is generally direct.

From Figure 7, it can be concluded that the emission caps are divided into seven different regions according to their impact on the total cost. From a slope point of view, in the range [0,100], the change in the emission limit has the greatest impact on the total cost. In the three regions [100,200], [200,300], and [300,400], the impact of changes in emission caps on the total cost is relatively modest. The impact of changes in emissions caps on the total cost is slightly similar between [100,200], but there are still differences, with the latter having a slightly larger impact than the former. In [300,400], the emission limit is lowered and the total cost remains the same. That said, changes to the emission limit will not have any effect on the total cost, so the upper end of this range can be infinite.

3.2. Parameter

(1) Load rate

The load rate of 50% and 100% was tested, respectively, and the load rate in the standard experimental data was 75%. It can be seen from Figure 8 that when the load rate remains unchanged, the relationship between the emission limit and the comprehensive cost is the same as the curve under the standard data. Under the same emission limit, the overall cost is proportional to the load factor. The change of load rate did not affect the curve of the emission limit and comprehensive cost. However, the higher the load factor, the larger the emission limit point of the inflection point of the impact of the emission limit on the overall cost.

As can be seen from Figure 8, the relationship between the emission limit and emissions at different load rates is divided into four segments, each segment corresponding to a corresponding path. Emissions are limited to the [100,200] range, and their emissions are 50%. In [100,200], the emission at 75% loading rate is not less than 100%; in [195,221], the emission at 50% and 100% is the same. In [200,300], a 75% loading rate is indistinguishable from 100% discharge. When the load rate is higher, the emission limit change has less effect on the route, for example, at the first route change, the load factor is 100%, 75%, and 50%.Similar to the standard data, the load rate is constant, and each discrete point corresponds to an inflection point of the total cost curve at the corresponding load rate, but the inflection point does not necessarily correspond to a discrete point. Emissions are directly or indirectly related to the load rate when the emission constraints remain unchanged. Under different load rates, the total emissions can be the same, but the changes in the overall cost are not the same. This shows that the same emissions do not mean the same roads.

(2) Cost factor

Taking the road cost factor as an example. When the test road cost coefficient is 30 and 80, in the standard test data, the road cost coefficient is 40. When the road cost factor is the same, the relationship between the emission limit and the total cost is basically the same as the curve under the standard data. Under the same emission limit, the comprehensive cost is proportional to the road cost factor. Under different road cost factors, the impact of the emission limit on the total cost has the same turning point, that is, the same turning point at the same level.

As can be seen from Figure 9, under different road cost factors, the relationship between emission limits and emissions is divided into four stages, each stage corresponding to a corresponding road. With an emissions limit of 165, changes to the road cost factor will not have any effect on total emissions. In [100,200], the emission amount of 40 is equal to 30, and if the emission limit is 165, the emission amount is greater than 100. In [100,200], the emission of 100 is equal to 40, and the total emission is 30; between [200,300], the emission of these three substances is the same. In [200,300], the emissions of 40 and 100 are the same, and the number is less than 30. In [300,400], the 40 and 30 emissions are the same, and above 100. The road cost coefficient remains unchanged, and each discrete point corresponds to the inflection point of the total cost curve under the road cost coefficient that is shown in Figure 9, but the inflection point does not necessarily correspond to a discrete point. When the emission limit is unchanged, the emission of the road cost factor has a certain proportional relationship with the road cost factor. Under different road cost factors, the total emissions of the road can be the same, but the changes in the overall cost are not the same. It also shows that the emissions are the same, but it does not mean that they are the same.

(3) Speed

Taking the average speed as an example, the relationship between emission limits and total cost and total emissions was carried out on railways with an average speed of 140 and 60, respectively. In standard test data, the average speed is 80. All three curves are consistent, that is, the changes in the average speed will not have any effect on the total cost. In [0,100], the total cost at average speeds of 80 and 140 varies with the emission limit to the same extent and regularity, and it is less than the average speed of 60 in total. Between [0,100], the average speed and total cost are inversely proportional.

As shown in Figure 10, when the average train speed changes, the impact of the emission limit on the emission of the entire route has the same abscissa. That is to say, at the same emission limit at different average train speeds, there will be changes in the path. At each scatter, the lower the average velocity, the greater the difference in emissions from each scatter. The higher the average velocity, the greater the difference in emissions from each dispersion point. As shown in Figure 10, the average track speed remains constant, and each discrete point corresponds to an inflection point of the total cost curve in Figure 10 that corresponds to the average speed, although the inflection points do not necessarily correspond to discrete points. When the emission limit is the same, the emission is inversely proportional to the full train speed. In [300,400] the total emissions and combined cost of the route are the same at an average speed.

4. Conclusions

The intelligent transportation and car navigation technology are deeply analyzed, and the key problem of optimal path planning is discussed in detail. The optimal criterion that affects the route planning is analyzed, and the three main factors that affect the route planning are: motor vehicle speed, blocking delay, and one-way route. A detailed study is carried out, the method is studied, and a network-based optimal path planning method is presented. This paper briefly describes the development environment of the system and describes its design and implementation in detail. Through the discussion of different path planning criteria, no matter which criterion is chosen, the best path selection can be achieved. On this basis, a set of optimization path query simulation systems is established, which aims to provide users with the shortest path and the smallest path.