Intelligent transportation route planning method based on deep learning

Abstract

The invention relates to the technical field of path planning, in particular to an intelligent transportation path planning method based on deep learning, which comprises the following steps: constructing a joint network model based on a convolutional neural network and a long short-term memory network, and training the joint network model by using historical traffic data; in the current transportation task, the real-time traffic data of the non-driving road sections in the current path are analyzed based on the joint network model, and the traffic change trend of each non-driving road section at the next moment is predicted; if the undriven road section of the current path has the sudden traffic event, or one or more undriven road sections in the current path are predicted to have traffic jam at the next moment, re-planning the path; and when the path is re-planned, dynamically optimizing the current path through the reinforcement learning model under the constraint of the multi-objective optimization function. According to the method, the accuracy of the future traffic condition can be improved, and the adaptive optimization of the transportation path can be realized.

Description

Technical Field

[0001] The present invention relates to the field of path planning technology, and more specifically to an intelligent transportation route planning method based on deep learning. Background Art

[0002] In recent years, with the acceleration of urbanization and the increasing complexity of economic activities, the transportation network has become larger and larger, and the transportation demand has also increased. In this context, traditional route planning and transportation management methods have been unable to adapt to the ever-changing and complex transportation environment. In the transportation and logistics industry, how to achieve efficient, accurate and low-cost transportation route planning has become a research hotspot.

[0003] At present, the mainstream technologies in the field of path planning include the shortest path method based on graph search, the path optimization method based on real-time data drive, and the intelligent path planning method based on artificial intelligence and machine learning. Each technology has its own strengths, but there are still deficiencies in real-time, accuracy and personalization.

[0004] Among them, the shortest path algorithm based on graph search shows great limitations when traffic conditions change dynamically. For example, when the road is blocked, an accident occurs, or an emergency causes the road conditions to change, the original path selection may no longer be optimal, and these algorithms are difficult to adjust quickly.

[0005] Path optimization methods driven by real-time data usually rely on local updates and lack forward-looking predictions of global traffic conditions. Therefore, in complex situations such as rush hour and bad weather, path optimization technologies based only on real-time data cannot provide a comprehensive optimal path.

[0006] Intelligent path planning methods based on artificial intelligence and machine learning are mostly static training models, which are difficult to cope with the rapid dynamic changes in the traffic environment; in addition, these technologies are still not perfect in handling uncertain factors in complex environments, and it is difficult to respond efficiently to sudden traffic events (such as accidents and weather changes); existing intelligent path planning models have limited support for different user needs and lack multi-objective optimization functions, such as differentiated planning capabilities for transportation needs of different priorities (such as cargo priority, time priority, and cost priority).

[0007] Therefore, how to provide a path planning method that can quickly adjust the path in the event of sudden traffic changes and has multi-objective optimization functions has become a technical problem that technical personnel in this field urgently need to solve. Summary of the invention

[0008] In view of this, the present invention provides an intelligent transportation route planning method based on deep learning, which can not only improve the accuracy of future traffic conditions, but also realize adaptive optimization of transportation paths.

[0009] In order to achieve the above object, the present invention adopts the following technical solution:

[0010] A deep learning-based intelligent transportation route planning method comprises the following steps:

[0011] A joint network model is constructed based on convolutional neural network and long short-term memory network, and the joint network model is trained using historical traffic data;

[0012] In the current transportation task, the real-time traffic data of the untraveled sections in the current route are analyzed based on the joint network model to predict the traffic change trend of each untraveled section at the next moment;

[0013] If a sudden traffic incident occurs on an unused section of the current route, or if it is predicted that one or more unused sections of the current route will be congested at the next moment, the route will be replanned;

[0014] When replanning the path, the reinforcement learning model is used to dynamically optimize the current path under the constraints of the multi-objective optimization function.

[0015] Furthermore, in the joint network model, spatial features in the real-time traffic data of the untraveled road section in the current path are extracted based on the convolutional neural network to obtain a feature matrix with road section features as an input of the long short-term memory network;

[0016] Based on the long short-term memory network, the feature matrix is ​​recursively processed time by time step to predict the traffic change trend of the unused road section at the next moment.

[0017] Furthermore, the spatial features extracted by the convolutional neural network include at least: road location, traffic density, congestion level and road construction conditions.

[0018] Furthermore, the multi-objective optimization function aims to minimize the comprehensive cost, and its expression is:

[0019] minimizef(x)=αT(x)+βC(x)+γF(x)

[0020] Among them, f(x) represents the comprehensive cost of path x; T(x) represents the total transportation time of path x; C(x) represents the transportation cost of path x; F(x) represents the energy consumption of path x; α, β, and γ represent the weights of transportation time, transportation cost, and energy consumption, respectively.

[0021] Furthermore, the reinforcement learning model performs path planning by setting the state space, action space and reward function.

[0022] Furthermore, the state space includes the current location of the vehicle, the location of the target location, and the surrounding traffic state vector, wherein the surrounding traffic state vector includes the traffic volume, average speed, and road construction information of the road section;

[0023] The action space is the selectable paths at each intersection.

[0024] Furthermore, the reward function is expressed as:

[0025] R(s,a)=-(α·T(s,a)+β·C(s,a)+γ·F(s,a))

[0026] Among them, R(s,a) represents the reward value of taking action a in state s, T(s,a) represents the estimated transportation time, C(s,a) represents the transportation cost, F(s,a) represents the energy consumption; α, β, and γ represent weights respectively.

[0027] Furthermore, the method further comprises:

[0028] After the current transportation task is completed, the actual transportation route, actual transportation time, energy consumption and cost of the transportation process are recorded and saved in the experience replay pool.

[0029] It can be seen from the above technical solutions that, compared with the prior art, the present invention has the following beneficial effects:

[0030] 1. The present invention uses a joint network model of convolutional neural network (CNN) and long short-term memory network (LSTM) to accurately predict traffic flow and congestion trends, and can dynamically adapt to sudden changes in traffic conditions (such as traffic accidents or temporary road closures), thereby achieving higher response speed and path adjustment efficiency, and effectively avoiding time delays caused by traffic jams.

[0031] 2. The present invention uses a multi-objective optimization function to comprehensively consider factors such as transportation time, cost and energy consumption, and can flexibly generate multiple path optimization solutions according to the specific needs of users. Users can choose a path that prioritizes time, cost or energy consumption according to actual conditions to meet different transportation task requirements, thereby making transportation planning more personalized and diverse.

[0032] 3. By introducing a reinforcement learning model, the present invention can continuously improve the decision-making strategy according to real-time feedback in path optimization, so that the path planning process has adaptive capabilities, and can continuously learn and improve its own planning accuracy in the process of repeated path selection and optimization, thereby enhancing adaptability in complex environments and meeting the precise requirements of long-term, large-scale transportation tasks.

[0033] 4. The present invention systematically optimizes the time cost and fuel consumption in the transportation process through a multi-objective optimization function to achieve the most economical transportation solution selection. The system comprehensively considers the distance, congestion level and expected fuel consumption of the path, effectively reduces the fuel cost and vehicle wear in the transportation process, achieves the environmental protection effect of reducing carbon emissions and saving energy, and is particularly suitable for green logistics and intelligent transportation systems.

[0034] 5. The present invention can not only perform route planning, but also monitor and adjust various parameters in the transportation process in real time to ensure that the transportation task can be successfully completed when an emergency occurs. Through the automatic adjustment mechanism, the route planning of the present invention is more stable and can improve the reliability of the entire transportation process. It is suitable for multiple scenarios such as logistics distribution, public transportation management, and emergency dispatch.

[0035] In general, the present invention not only significantly improves the efficiency and accuracy of transportation route planning, but also has the ability of adaptive optimization and multi-objective planning. It has significant beneficial effects in improving transportation efficiency, saving energy, enhancing the stability of route planning and urban traffic management capabilities, and provides important technical support for the realization of smart transportation and intelligent transportation systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0037] Figure 1 A flowchart of the deep learning-based intelligent transportation route planning method provided by the present invention. DETAILED DESCRIPTION

[0038] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] like Figure 1 As shown, an embodiment of the present invention discloses a method for intelligent transportation route planning based on deep learning, comprising the following steps:

[0040] S1. Construct a joint network model based on convolutional neural network and long short-term memory network, and use historical traffic data to train the joint network model;

[0041] S2. In the current transportation task, the real-time traffic data of the untraveled sections in the current route are analyzed based on the joint network model to predict the traffic change trend of each untraveled section at the next moment;

[0042] S3. If a sudden traffic incident occurs on an unused section of the current route, or if it is predicted that one or more unused sections of the current route will be congested at the next moment, the route is replanned;

[0043] When replanning the path, the reinforcement learning model is used to dynamically optimize the current path under the constraints of the multi-objective optimization function.

[0044] The above steps are further described below.

[0045] S1. Construction and training of joint network model.

[0046] The present invention builds a joint network model based on convolutional neural network (CNN) and long short-term memory network (LSTM); wherein CNN is used to extract spatial features of road traffic, and LSTM is used to capture time series features. This combined application utilizes the spatial feature extraction capability of CNN and the time series processing advantage of LSTM, and can more comprehensively process and analyze real-time traffic data. Through the joint action of CNN and LSTM, key features can be extracted in the spatial dimension of traffic data, and the prediction of future traffic status can be realized in the time dimension, providing reliable path optimization decisions for road transportation planning.

[0047] Specifically, CNN includes convolutional layer, activation layer and pooling layer.

[0048] The convolution layer is the core of CNN, which is used to extract local spatial features from the input traffic data. In practical applications, after the traffic data is input into the convolution layer, it is scanned and processed step by step through the convolution kernel (i.e., a small filter matrix). The convolution kernel performs a dot product operation when sliding in the feature matrix of each road section, and accumulates and generates a feature map. The basic formula of the convolution operation is:

[0049] f(x)=W*x+b

[0050] W is the convolution kernel; * represents the convolution operation; x is the feature matrix of the input traffic data (such as the real-time traffic conditions of each road section); b is the bias term. Through layer-by-layer convolution, features can be gradually extracted from low dimensions to high dimensions, such as identifying high congestion areas and traffic bottlenecks.

[0051] Activation layer: The output after each convolution operation is processed by a nonlinear activation function (ReLU function) to further enhance the nonlinear fitting ability of the model. The activation layer enables the model to capture more complex traffic feature patterns. The ReLU function is defined as:

[0052] ReLU(x)=W*x+b

[0053] Through ReLU activation, the model can ignore non-important features, allowing the network to focus more on extracting congestion and traffic features with higher relevance.

[0054] The pooling layer is used to reduce the dimension of the feature map while retaining the key features. Usually, the maximum pooling or average pooling is used on the feature map generated by each convolution. The pooling operation not only compresses the amount of data, but also enhances the model's adaptability to position changes. For example, the maximum pooling takes the maximum value in each area, and the calculation formula is:

[0055] P(x)=max(x i )

[0056] Among them, x i Represents the pixel values ​​in the region of the convolution feature map. Through pooling, the system is able to retain key traffic features while reducing the amount of computation.

[0057] The key structure of LSTM is the gate unit, including input gate, forget gate and output gate. These gate structures control the transmission of information between time steps so as to selectively retain or forget historical traffic data. In road transportation planning, traffic data has temporal continuity and dependency. LSTM is designed specifically for processing time series data. Its unique gating mechanism can effectively capture the temporal dependency in traffic data, thereby realizing the prediction of future traffic flow and congestion.

[0058] The calculation formula for each gate is as follows:

[0059] Input gate: controls the input of new information:

[0060] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0061] Forget gate: decide whether to forget historical information:

[0062] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0063] Output gate: Outputs the hidden state of the current time step:

[0064] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0065] Among them, σ represents the sigmoid activation function, W is the weight matrix, b is the bias term, and h t-1 is the hidden state of the previous time step, x t Through these gating structures, LSTM can effectively handle long-term and short-term dependencies in time series, enabling the model to comprehensively consider past and current traffic conditions for prediction.

[0066] In each time step, LSTM gradually predicts the future traffic status based on historical traffic data. By recursively transferring the traffic flow changes at each time step, LSTM can gradually generate the congestion trend and traffic flow prediction results for the future time period, thereby providing reliable data support for subsequent path selection. For example, during peak hours, LSTM can judge whether the future traffic conditions will be further congested through historical patterns, providing early warning for the system's planning decisions.

[0067] After the joint network model is built, a large amount of historical traffic data is obtained to train and test the model, and the parameters of CNN and LSTM are adjusted to achieve the best performance of the joint network model. During training, the gradient descent method and back propagation algorithm are used to optimize the weights to ensure that CNN can accurately extract spatial features and LSTM can effectively capture time dependencies. Through continuous iterative training, the model gradually optimizes its prediction effect in actual traffic data, providing more accurate support for path planning.

[0068] S2. In the current transportation task, the real-time traffic data of the untraveled sections in the current path are analyzed based on the trained and tested joint network model to predict the traffic change trend of each untraveled section at the next moment.

[0069] The combined application of CNN and LSTM can achieve comprehensive integration from spatial features to time series features:

[0070] Traffic data is obtained through real-time data collection equipment such as GPS, traffic monitoring cameras, and traffic flow sensors. After cleaning, normalization, and feature extraction, the collected data is stored in the data storage module to provide input for the model. The real-time traffic data of each road section is first extracted through CNN to extract spatial features (such as road location, traffic density, congestion level, and road construction conditions). After convolution and pooling operations, these features form a feature matrix with road section characteristics. These feature matrices serve as inputs to LSTM, providing LSTM with initial state information of road traffic, so that LSTM can better capture the temporal changes in traffic status.

[0071] LSTM receives the feature vector processed by CNN, recursively processes it step by step, and predicts the traffic change trend (such as congestion probability) of each road section at the next moment. In this way, dynamic update of traffic data can be achieved in time series, providing more accurate future predictions for path selection at each moment.

[0072] S3. Path dynamic optimization based on reinforcement learning model and multi-objective function.

[0073] The present invention introduces a multi-objective optimization function, combining multi-dimensional objectives such as path distance, transportation time, and fuel consumption to evaluate each path. This optimization problem is expressed as follows:

[0074] minimizef(x)=αT(x)+βC(x)+γF(x)

[0075] Among them, f(x) represents the comprehensive cost of path x; T(x) represents the total transportation time of path x; C(x) represents the transportation cost of path x; F(x) represents the energy consumption of path x; α, β, and γ represent the weights of transportation time, transportation cost, and energy consumption, respectively.

[0076] The reinforcement learning model performs multi-objective path planning by setting the state space, action space and reward function.

[0077] State space S: The state space contains the current location of the vehicle, the location of the target location, and the surrounding traffic state vector. The surrounding traffic state vector includes the traffic volume, average speed, and road construction information of the road section to ensure that the reinforcement learning model can fully understand the current traffic environment.

[0078] Action space A: The action space is the path that can be selected at each intersection, such as continuing along the current path or choosing an alternative path. Through action selection, it is possible to dynamically switch paths in different states and adapt to environmental changes.

[0079] The reward function defines the goal of the system. Based on the above multi-objective function, the path planning can meet the optimization of time, cost, energy consumption and other aspects. Specifically, the expression of the reward function is:

[0080] R(s,a)=-(α·T(s,a)+β·C(s,a)+γ·F(s,a))

[0081] Among them, R(s,a) represents the reward value of taking action a in state s, T(s,a) represents the estimated transportation time, C(s,a) represents the transportation cost, and F(s,a) represents the energy consumption. α, β, and γ represent weights respectively. By adjusting the weight parameters, it is possible to generate a personalized path with optimal time, optimal cost, or optimal energy consumption according to actual needs.

[0082] The path planning results for different requirements based on the reinforcement learning model are not necessarily the optimal results. The path selection strategy needs to be continuously optimized to continuously approach the optimal strategy. In the process of optimizing the path selection strategy, the deep Q network (DQN) algorithm is used to continuously update the optimal strategy of each state-action pair by learning the state-action value (Q value) to find the path with the lowest comprehensive cost.

[0083] In a more complex traffic environment, DQN uses deep neural networks to process high-dimensional state space, input traffic status and output Q values ​​for all possible paths. The network is trained through experience replay and target network technology, thereby reducing the deviation caused by non-stationary data and making the learning process more stable.

[0084] DQN updates the Q value through the Bellman equation. The update formula is as follows:

[0085]

[0086] Among them, Q(s,a) is the value estimate of the current state and action, is the learning rate, γ is the discount factor; Q(s',a') is the sum of the immediate reward and the subsequent expected return of taking action a' in state s''. This formula ensures that the value estimate of each state is updated after each path selection, and constantly approaches the optimal strategy.

[0087] The present invention introduces a dynamic adjustment mechanism of reinforcement learning. During the transportation process, the model will monitor the traffic conditions of the current path in real time. When major changes are detected (such as an accident ahead or severe congestion on the road), the path selection will be recalculated based on the reward value, and the optimal action will be selected (i.e. switching to other sections or continuing the current section) to achieve dynamic optimization of the path.

[0088] In other embodiments, the method of the present invention further comprises:

[0089] After the current transportation task is completed, the actual transportation route, actual transportation time, energy consumption and cost of the transportation process will be recorded and saved in the experience replay pool to improve the accuracy and adaptability of the next path planning and realize continuous learning and optimization of the reinforcement learning model.

[0090] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0091] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for intelligent transportation route planning based on deep learning, characterized in that: The following steps are involved: A joint network model is constructed based on convolutional neural network and long short-term memory network, and the joint network model is trained using historical traffic data; In the current transportation task, the real-time traffic data of the untraveled sections in the current route are analyzed based on the joint network model to predict the traffic change trend of each untraveled section at the next moment; If a sudden traffic incident occurs on an unused section of the current route, or if it is predicted that one or more unused sections of the current route will be congested at the next moment, the route will be replanned; When replanning the path, the reinforcement learning model is used to dynamically optimize the current path under the constraints of the multi-objective optimization function.

2. The method for intelligent transportation route planning based on deep learning according to claim 1, characterized in that: In the joint network model, spatial features in real-time traffic data of untraveled sections in the current path are extracted based on the convolutional neural network to obtain a feature matrix with section features as input to the long short-term memory network; Based on the long short-term memory network, the feature matrix is ​​recursively processed time by time step to predict the traffic change trend of the unused road section at the next moment.

3. The deep learning-based intelligent transportation route planning method according to claim 2 is characterized in that: The spatial features extracted by the convolutional neural network include at least: road location, traffic density, congestion level and road construction conditions.

4. The deep learning-based intelligent transportation route planning method according to claim 1, characterized in that: The multi-objective optimization function aims to minimize the comprehensive cost, and its expression is: minimizef(x)=αT(x)+βC(x)+γF(x) Among them, f(x) represents the comprehensive cost of path x; T(x) represents the total transportation time of path x; C(x) represents the transportation cost of path x; F(x) represents the energy consumption of path x; α, β, and γ represent the weights of transportation time, transportation cost, and energy consumption, respectively.

5. The deep learning-based intelligent transportation route planning method according to claim 1, characterized in that: The reinforcement learning model performs path planning by setting the state space, action space and reward function.

6. The deep learning-based intelligent transportation route planning method according to claim 5 is characterized in that: The state space includes the current location of the vehicle, the location of the target location, and the surrounding traffic state vector, wherein the surrounding traffic state vector includes the traffic flow, average speed, and road construction information of the road section; The action space is the selectable paths at each intersection.

7. The deep learning-based intelligent transportation route planning method according to claim 5, characterized in that: The expression of the reward function is: R(s,a)=-(α·T(s,a)+β·C(s,a)+γ·F(s,a)) Among them, R(s,a) represents the reward value of taking action a in state s, T(s,a) represents the estimated transportation time, C(s,a) represents the transportation cost, F(s,a) represents the energy consumption; α, β, and γ represent weights respectively.

8. The method for intelligent transportation route planning based on deep learning according to claim 1, characterized in that: Also includes: After the current transportation task is completed, the actual transportation route, actual transportation time, energy consumption and cost of the transportation process are recorded and saved in the experience replay pool.

Images