A two-stage energy consumption prediction path planning method based on deep learning

Abstract

The application discloses a two-stage energy consumption prediction path planning method based on deep learning, which utilizes deep learning technology to predict and optimize the energy consumption of electric vehicles. It not only considers various factors such as road slope, traffic conditions, and vehicle load, but also combines historical driving data and real-time traffic data to achieve accurate prediction of energy consumption for specific routes. In addition, it uses a two-stage solution, where the initial stage focuses on pre-trip prediction, and the subsequent stage focuses more on post-trip prediction during the journey. This allows for a rough estimate of energy consumption before the journey begins, and dynamic adjustments based on real-time traffic data during the journey, resulting in more accurate energy consumption prediction.

Description

Technical Field

[0001] The present invention belongs to the field of logistics route planning, and specifically belongs to a method for two-stage energy consumption prediction route planning based on deep learning. Background Art

[0002] Logistics route planning is a key component in optimizing the transportation of goods in a wide range of fields, including freight and express delivery. It aims to reduce costs and significantly improve transportation efficiency through meticulous planning. This field utilizes a comprehensive range of classic algorithms—including the Dijkstra and A* algorithms for finding the shortest path, as well as the Traveling Salesman Problem (TSP) and Vehicle Routing Problem (VRP) that consider more complex constraints—that form the foundation of logistics route planning by solving problems through both precise and heuristic methods.

[0003] With the development of technology, modern methods such as heuristic algorithms and deep learning techniques are widely used in real-time route planning in dynamic environments, making logistics route planning more flexible and adaptable to complex and changing transportation needs. Furthermore, logistics route planning focuses on multi-objective optimization, not only pursuing cost minimization but also considering factors such as time efficiency and energy consumption, striving to find the optimal balance between these multiple objectives.

[0004] Leveraging IoT technology, real-time traffic data, weather information, and vehicle status data can be effectively utilized. Dynamic planning methods can be used to adjust routing plans in real time to accommodate changing circumstances. These technologies are supported by a variety of software and platforms that integrate these algorithms and methods, providing user-friendly interfaces and enabling businesses and individuals to efficiently and conveniently manage logistics routing issues.

[0005] Logistics routing planning continuously adapts to and solves logistics challenges through the integration of interdisciplinary technologies. Its methods and tools are constantly evolving with technological advancements, aiming to provide more efficient, reliable, and economical logistics solutions to meet growing logistics demands.

[0006] In recent years, the research on logistics path planning mainly includes:

[0007] Chinese patent "A Logistics Route Planning Method Based on Deep Learning" (CN110929149A): This patent discloses a method for logistics route planning using deep learning technology. This method calculates the optimal route for logistics vehicles by acquiring real-time traffic information and factoring in the time cost of vehicles arriving at intersections. This method can effectively improve the efficiency and accuracy of logistics transportation.

[0008] Chinese Patent "A Method for Implementing Path Planning in Real-Time Logistics Scenarios" (CN109241413A): This patent discloses a path planning method suitable for real-time logistics scenarios. This method uses a two-tiered path selection mechanism to match riders with the path with the lowest delivery cost and predict the rider's arrival time at each node and any timeouts, thereby improving delivery efficiency and service quality.

[0009] Chinese Patent "Route Planning and Distribution Patent Applied by JD Logistics" (CN110929149A): This patent discloses a route planning and distribution method applied for by JD Logistics. This method obtains historical routes for the current station and a preset time period, determines the overlap between the stations covered by the historical routes and the current station, and then selects a template route for the current station. This method can effectively improve the efficiency and accuracy of logistics distribution.

[0010] The above inventions have all played a certain role in promoting the improvement of logistics distribution efficiency by considering the impact of real-time traffic information or historical traffic trajectories on logistics, but there are still deficiencies in adopting an efficient mechanism to integrate historical and real-time traffic information: 1) The accuracy of real-time traffic information is crucial for logistics route planning. However, obtaining accurate real-time traffic information may face many challenges, such as the update frequency of traffic information, the coverage of information, and the reliability of information. 2) Current patented technologies may lack an effective mechanism to integrate historical and real-time traffic information. For example, if a route has often experienced traffic congestion in the past, but the traffic conditions are good at the current point in time, how should these two types of information be weighed to make the optimal route planning. Summary of the Invention

[0011] The purpose of this invention is to provide a two-stage energy consumption prediction path planning method based on deep learning, so as to improve the distribution efficiency of cold chain electric vehicles while reducing the energy consumption of electric vehicles.

[0012] The purpose of the present invention is to achieve the following technical means: a two-stage energy consumption prediction path planning method based on deep learning, comprising the following steps:

[0013] Step 1: Data collection: Collect data on the actual operation of cold chain distribution vehicles in the past, including historical driving data, vehicle parameter data and real-time traffic data;

[0014] Step 2: Data preprocessing, normalizing and standardizing the data, and performing feature engineering to select and construct features that are informative for energy consumption prediction;

[0015] Step 3: Energy consumption model construction: formulas for calculating driving energy consumption and cooling energy consumption are constructed, and feature sets are extracted using a convolutional neural network model and the collected data.

[0016] Input the feature set into the long short-term memory network (LSTM) to generate an energy consumption prediction model;

[0017] Step 4: Long-term energy consumption prediction. Based on the historical data of vehicles on similar delivery tasks, we extract the characteristics related to the current task that affect energy consumption. These characteristics are input into the energy consumption prediction model to predict the energy consumption of possible future road sections. The driving energy consumption and cooling energy consumption between each two nodes are predicted.

[0018] Step 5: Short-term energy consumption prediction: During the vehicle's driving process, data affecting energy consumption is collected in real time and input into the energy consumption prediction model. The driving energy consumption and cooling energy consumption between every two nodes of the current and future road sections are predicted, and the long-term energy consumption prediction is revised.

[0019] Step 6: Path planning and optimization. Path planning is performed with the goal of minimizing total energy consumption and time constraints, that is, the planned route time does not exceed the delivery time limit. The non-dominated sorting genetic algorithm (NSGA-II) is used to optimize the delivery path. During the subsequent driving process, the obtained delivery path is optimized and corrected based on real-time data and new short-term energy consumption forecasts.

[0020] The data is data that affects delivery time and refrigeration energy consumption, and the historical driving data includes the actual driving speed, driving route, time and date, external environmental conditions and road slope of the vehicle on each road section;

[0021] Vehicle parameter information refers to the physical and mechanical parameters of the vehicle, including fully loaded and unloaded weight, battery type and capacity, and cooling system specifications;

[0022] Real-time traffic data includes current traffic flow, road information, traffic accident data, and road topology.

[0023] In the step 2, the features that are informative for energy consumption prediction are selected and constructed, including speed change rate, slope characteristics and environmental characteristics.

[0024] The driving energy consumption calculation formula is:

[0025] n is the mass of the vehicle, g is the acceleration of gravity, θ is the road slope, C d is the air resistance coefficient, ρ is the air density, K is the shape factor of the vehicle, v is the vehicle speed, Cr is the rolling resistance coefficient, P tract is the traction power of the electric vehicle;

[0026] Average speed and energy consumption during comprehensive driving μ eled Energy conversion efficiency for electric vehicles;

[0027] The calculation formula for cooling energy consumption is E cool =γ·d. Where γ is the cooling energy consumption per unit distance and d is the driving distance.

[0028] In step three, the convolutional neural network model and the collected data are used to extract a feature set. Specifically, the preprocessed historical driving data is input into the CNN network. The data format is a multi-dimensional matrix. The feature map of the data is extracted through multiple convolutional layers to reveal the correlation and spatial pattern between different features. The convolution feature map is reduced in dimension using a pooling layer to reduce the amount of calculation while retaining key features. The final output feature set retains the key information of the data and provides input for the subsequent LSTM network.

[0029] In step 6, the total energy consumption minimization objective function is:

[0030]

[0031] Among them, E drive (i,i+1) represents the energy consumption of the vehicle traveling from node i to node i+1, E cool (i, i+1) represents the cooling energy consumption of the same section, N is the total number of distribution nodes in the path, E total is the total energy consumption;

[0032] The time constraint objective function is Among them, t i Indicates that the vehicle is at node i and from node i

[0033] The travel time to node i+1, T is the total time limit of the delivery task.

[0034] The specific steps of using the non-dominated sorting genetic algorithm NSGA-II to optimize the delivery path are:

[0035] Initial population generation: Each individual in the initial population represents a possible delivery path. The path is represented by a set of node sequences. For the delivery points P1, P2, ..., P N , the path can be encoded as a sequence [P11,P2,P3,…,P N ],

[0036] By randomly disrupting the order of delivery nodes, multiple different paths are generated. Each path corresponds to an individual, and each individual represents a possible delivery plan in the decision variable space.

[0037] Fitness evaluation: For each individual, that is, path, the total energy consumption is calculated through the energy consumption prediction model, including the driving energy consumption Edrive and cooling energy consumption E cool , use the objective function to calculate the fitness of the individual,

[0038]

[0039] Calculate the total delivery time for each route and use it as the second objective function:

[0040]

[0041] If f2(P)>T, the path does not meet the time constraint and the fitness value will be penalized and adjusted;

[0042] Non-dominated sorting and selection: In the population, each individual is sorted according to energy consumption f1(P) and time f2(P), and the crowding distance of the individual is calculated in each non-dominated layer;

[0043] Selection operation: Use the binary tournament selection method based on non-dominated sorting and crowding distance to select individuals from the current population to generate the next generation population.

[0044] Crossover operation: Use sequential crossover or partial mapping crossover methods to cross selected individuals and generate new path individuals. The crossover operation helps explore new path combinations in order to find a better path.

[0045] Mutation operation: Use point mutation or neighbor exchange method to randomly change the order of one or more nodes in the path to increase the diversity of the population and avoid local optimal solutions;

[0046] The parent and child populations are merged to form a new population, and then the non-dominated sorting and crowding distance calculation are performed on the merged population again. The individuals with the best fitness value are selected from the merged population to form a new population. These individuals will enter the next generation of evolution, that is, iteration;

[0047] Repeat the above process of selection, crossover, mutation and production of new populations until the predetermined number of iterations is reached or the population fitness converges. At this time, the optimal solution is the best delivery path under the current conditions.

[0048] The present invention utilizes deep learning technology, specifically a CNN-LSTM model, to predict and optimize the energy consumption of electric vehicles. This approach is unique in that it not only considers multiple factors, such as road slope, traffic conditions, and vehicle load, but also incorporates historical driving data and real-time traffic data to accurately predict energy consumption for specific routes.

[0049] In addition, a two-stage solution is also adopted, in which the initial stage focuses on pre-departure predictions, while the subsequent stage focuses more on post-departure predictions during the journey. This allows for a rough prediction of energy consumption before the journey begins, and dynamic adjustments can be made during the journey based on real-time traffic data, thereby achieving more accurate energy consumption predictions.

[0050] Finally, the NSGA-II algorithm is used to optimize the route to further reduce the electric vehicle's energy consumption and extend its driving range. This optimization not only takes into account the current traffic conditions but also predicts possible future traffic changes, allowing the electric vehicle to choose the optimal route under any traffic conditions.

[0051] Ultimately, it can effectively reduce the energy consumption of electric vehicles, extend the driving range of electric vehicles, and provide important reference information for cold chain distribution, thereby improving the efficiency and quality of cold chain distribution. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A flow chart of a two-stage energy consumption prediction path planning method based on deep learning;

[0053] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. DETAILED DESCRIPTION

[0054] [Example 1]

[0055] A two-stage energy consumption prediction path planning method based on deep learning includes the following steps:

[0056] Step 1: Data collection: Collect data on the actual operation of cold chain distribution vehicles in the past, including historical driving data, vehicle parameter data and real-time traffic data;

[0057] The data is data that affects delivery time and refrigeration energy consumption, and the historical driving data includes the actual driving speed, driving route, time and date, external environmental conditions and road slope of the vehicle on each road section;

[0058] Vehicle parameter information refers to the physical and mechanical parameters of the vehicle, including fully loaded and unloaded weight, battery type and capacity, and cooling system specifications;

[0059] Real-time traffic data includes current traffic flow, road information, traffic accident data, and road topology.

[0060] First, we actively collected extensive historical driving data, covering everything from vehicle speed to road slope, to build a historical data framework to support model training and improve prediction accuracy. This data provided the model with a solid historical foundation, helping to improve the accuracy and reliability of predictions.

[0061] Advanced sensing technologies (on-board sensors, traffic cameras, and online traffic information services, etc.) are used to monitor and collect detailed data on road conditions and traffic in real time, ensuring that the information obtained is both real-time and comprehensive, providing data support for immediate route adjustments, thereby playing a key role in real-time decision-making, namely short-term energy consumption forecasting.

[0062] Specifically, historical driving data includes delivery time, delivery location, customer address, and delivery route information. This data provides a record of vehicle operation at different times and under different road conditions, providing a foundation for the subsequent training of energy consumption prediction models.

[0063] Including but not limited to the following:

[0064] Driving speed: record the actual driving speed of the vehicle on each road section.

[0065] Driving route: Track the vehicle's driving path to obtain detailed information on different road sections.

[0066] Time and date: Mark the specific time period and date corresponding to the data, considering the impact of time on energy consumption.

[0067] External environmental conditions: including weather, temperature, humidity, road conditions, etc., which may significantly affect energy consumption.

[0068] Traffic network data: This includes road topology, road slope, speed limit information, etc. This information is used to accurately calculate vehicle energy consumption on different road sections and provide data support for route planning.

[0069] Vehicle parameter data, which collects key physical and mechanical parameters of the vehicle, such as:

[0070] Vehicle weight: includes both fully loaded and unloaded weight.

[0071] Battery type and capacity: The type of battery used in the vehicle and its capacity, which determines the vehicle's range.

[0072] Refrigeration system specifications: The power consumption parameters and operating characteristics of the refrigeration equipment, which directly affect the energy consumption in cold chain transportation.

[0073] Real-time traffic data: The traffic management system acquires and processes data on current traffic flow, road conditions, traffic accidents, and other data in real time. The introduction of real-time traffic data can significantly improve the real-time and accuracy of energy consumption forecasts, especially in urban road networks with complex and changing traffic conditions.

[0074] Traffic flow: The density and flow of vehicles on the road.

[0075] Road information: current road conditions, such as closures, construction, accidents, etc.

[0076] Traffic accident data: sudden traffic incidents that may affect driving speed and energy consumption.

[0077] Step 2: Data preprocessing, normalizing and standardizing the data, and performing feature engineering to select and construct features that are informative for energy consumption prediction;

[0078] In the step 2, the features that are informative for energy consumption prediction are selected and constructed, including speed change rate, slope characteristics and environmental characteristics.

[0079] Utilize sophisticated data cleaning algorithms to precisely remove noise and redundancy from unstructured data sets, ensuring high-quality data input for analysis.

[0080] Apply a series of statistical standardization techniques and machine learning methods to normalize the data and optimize the data structure to ensure that data from different sources can be seamlessly connected and processed in the model, laying a solid foundation for subsequent in-depth analysis.

[0081] Specifically, remove duplicate values: delete duplicate records to ensure data uniqueness.

[0082] Handling missing values: Process missing data through interpolation, mean filling, etc. to avoid model errors caused by incomplete data.

[0083] Outlier removal: Identify and remove obvious anomalous data, such as extreme speed or energy consumption values.

[0084] In order to improve the efficiency and accuracy of model training, the data needs to be standardized or normalized:

[0085] Standardization: Transform the data into a standard normal distribution with mean 0 and variance 1.

[0086] Normalization: Adjust the data values ​​to the range of 0 to 1 to eliminate the dimensional differences between different features.

[0087] Select and construct the most informative features for energy consumption prediction:

[0088] Speed ​​change rate: Based on historical speed data, the rate of change of vehicle speed is calculated to capture the characteristics of acceleration or deceleration during driving.

[0089] Slope characteristics: Utilize road network information to calculate the slope changes of each road section, which has a direct impact on driving energy consumption.

[0090] Environmental characteristics: Based on weather and road condition data, relevant features such as humidity and temperature are extracted to understand their impact on vehicle driving and cooling system energy consumption.

[0091] Step 3: Energy consumption model construction: formulas for calculating driving energy consumption and cooling energy consumption are constructed, and feature sets are extracted using a convolutional neural network model and the collected data.

[0092] Input the feature set into the long short-term memory network (LSTM) to generate an energy consumption prediction model;

[0093] The energy consumption model is the core of the system, covering the calculation of driving energy consumption and cooling energy consumption.

[0094] Driving energy consumption refers to the energy consumed by a vehicle when overcoming rolling resistance, air resistance and slope. The driving energy consumption calculation formula is: m is the mass of the vehicle (kg), g is the acceleration due to gravity (9.81m / s 2 ), θ is the road slope (degrees), C d is the air resistance coefficient, ρ is the air density (kg / m 3 ), K is the shape factor of the vehicle, v is the vehicle speed (m / s), Cr is the rolling resistance coefficient, P tract is the traction power of the electric vehicle (kw);

[0095] Average speed and energy consumption during comprehensive driving μ eled Energy conversion efficiency for electric vehicles;

[0096] Refrigeration energy consumption is the electrical energy consumed to maintain the required low temperature environment for the cargo. The calculation formula is simplified to multiply the energy consumption per kilometer by the distance traveled.

[0097] The calculation formula for cooling energy consumption is E cool =γ·d. Where γ is the cooling energy consumption per unit distance and d is the driving distance.

[0098] After completing data collection and preprocessing, the next step is to train the energy consumption prediction model through deep learning methods.

[0099] In step three, the convolutional neural network model and the collected data are used to extract a feature set. Specifically, the preprocessed historical driving data is input into the CNN network. The data format is a multi-dimensional matrix. The feature map of the data is extracted through multiple convolutional layers to reveal the correlation and spatial pattern between different features. The convolution feature map is reduced in dimension using a pooling layer to reduce the amount of calculation while retaining key features. The final output feature set retains the key information of the data and provides input for the subsequent LSTM network.

[0100] Apply convolutional neural networks (CNNs) to conduct in-depth analysis of complex road images, leveraging their powerful feature extraction capabilities to accurately capture and analyze road condition details;

[0101] Use long short-term memory (LSTM) networks to explore and learn time series patterns of distance, load, speed, and energy consumption to predict future conditions and respond in real time;

[0102] Integrate the analytical results of CNN and LSTM, build a comprehensive integrated prediction model through precise algorithm tuning, and train and rigorously verify large amounts of data on high-performance servers;

[0103] Convolutional Neural Network (CNN) is used to extract spatial features of input data. The specific process is as follows:

[0104] Input data formatting

[0105] The pre-processed historical driving data is input into the CNN. The data format includes multi-dimensional matrix form, such as vehicle speed, environmental conditions, road slope, etc.

[0106] Convolution operation

[0107] The feature maps of the data are extracted through multiple convolutional layers, revealing the correlation and spatial patterns between different features.

[0108] Feature Mapping and Pooling

[0109] The pooling layer is used to reduce the dimension of the convolution feature map, reducing the amount of computation while retaining key features.

[0110] Feature Output

[0111] The final output feature set retains the key information of the data and provides input for the subsequent LSTM network.

[0112] Step 3.2: LSTM model training

[0113] Long short-term memory (LSTM) networks are used to process time series data and capture temporal dependencies in the data.

[0114] Input time series data

[0115] The feature set extracted by CNN is converted into a time series and input into the LSTM model.

[0116] Construction of LSTM layer

[0117] The LSTM network captures the temporal dynamics of input data through multiple memory units and learns the complex relationship between energy consumption and time.

[0118] Fully connected layer and output layer

[0119] After the LSTM layer, fully connected layers are added to map the learned features to the energy consumption prediction values, and finally the prediction results are output.

[0120] Model validation

[0121] The model is tested using a validation dataset to calculate the prediction accuracy and errors of the model. If necessary, the model performance is optimized by adjusting the network parameters.

[0122] Step four, long-term energy consumption prediction, according to the historical data of the vehicle in similar distribution tasks, extract the features related to the current task that affect energy consumption, input the energy consumption prediction model, and predict the energy consumption of the future possible road section, get the driving energy consumption and refrigeration energy consumption prediction between each two nodes;

[0123] Step five, short-term energy consumption prediction, in the process of vehicle driving, real-time collection of data affecting energy consumption, input energy consumption prediction model, driving energy consumption and refrigeration energy consumption prediction between each two nodes for current and future road section, correct long-term energy consumption prediction;

[0124] Energy consumption prediction is the core part of this method, through two-stage energy consumption prediction method, energy consumption evaluation is carried out for each possible distribution path. Energy consumption prediction includes two stages: long-term prediction and short-term prediction.

[0125] First stage: long-term energy consumption prediction

[0126] In this stage, the system uses historical logistics distribution data to predict the energy consumption required by the vehicle to complete a specific distribution route. Long-term prediction aims to provide a preliminary energy consumption estimate based on historical data, helping to identify high energy consumption routes.

[0127] Second stage: short-term energy consumption prediction

[0128] Short-term energy consumption prediction is started during the actual driving of the vehicle, and the system incorporates real-time distribution data into the energy consumption prediction model to dynamically predict the remaining journey energy consumption of the vehicle. This short-term prediction combines real-time traffic conditions and the current state of the vehicle, ensuring the accuracy and timeliness of energy consumption prediction.

[0129] Through the above two-stage energy consumption prediction, the system can perform multi-dimensional analysis of energy consumption at different time scales, thereby improving the accuracy of prediction and optimizing path selection.

[0130] Long-term energy consumption prediction is a preliminary energy consumption evaluation based on the historical distribution data of the vehicle for the entire distribution task, aiming to provide a general energy consumption trend to guide the preliminary selection of path planning.

[0131] Historical data analysis

[0132] The system first analyzes the historical data of the vehicle in similar distribution tasks, extracting features related to the current task, such as the vehicle's driving speed on different road sections, average energy consumption, and the impact of environmental temperature on the energy consumption of the refrigeration system.

[0133] These features are fed into an energy consumption prediction model to infer the overall energy consumption trend of the current task. Especially for long-distance delivery tasks, this prediction based on historical data can provide a preliminary energy consumption assessment.

[0134] Long-term energy consumption trend forecast

[0135] After analyzing historical data, the system uses deep learning models (including CNN and LSTM) to predict long-term energy consumption trends. This allows the system to identify road sections that are likely to consume more energy and avoid these high-energy sections in advance.

[0136] For example, vehicles tend to consume more energy on roads with steep slopes or heavy traffic. By analyzing historical data trends, the system can identify and mark these sections early in planning, providing data support for subsequent route optimization.

[0137] Short-term energy consumption forecast

[0138] Short-term energy consumption forecasts are conducted after a vehicle begins its actual delivery mission. They refine and adjust long-term forecasts and rely primarily on real-time data input. The goal of short-term forecasts is to dynamically respond to changes in the delivery process, such as real-time traffic conditions and emergencies.

[0139] Real-time data collection and input

[0140] When a vehicle begins a delivery mission, the system continuously collects real-time data, including current traffic conditions (such as traffic flow and congestion levels), road information (such as road closures and construction information), and vehicle status (such as current battery charge, temperature, and refrigeration system operating status).

[0141] This real-time data is immediately fed into the energy consumption forecasting model to revise the energy consumption estimates generated during the long-term forecast phase. The frequency and accuracy of real-time data collection directly impact the accuracy of short-term energy consumption forecasts.

[0142] Dynamic energy consumption calculation and correction

[0143] Integrating real-time data, the system dynamically calculates energy consumption for the remaining delivery routes. By capturing short-term changes in time series data through an LSTM network, the system can update energy consumption forecasts in real time, ensuring that the forecasts closely match actual conditions.

[0144] For example, if the system detects a traffic accident on the road ahead, which may cause the vehicle to travel at a low speed for a long time on that road section, the system will immediately re-evaluate the energy consumption of that road section and adjust the path planning to avoid excessive energy consumption.

[0145] Real-time energy consumption feedback and adjustment

[0146] During the short-term energy consumption forecasting process, the system not only continuously corrects the energy consumption calculation, but also compares and analyzes the actual energy consumption with the predicted energy consumption. If any deviation is found, the system will adaptively adjust the model parameters to improve the forecast accuracy.

[0147] The real-time feedback mechanism allows the system to gradually optimize energy consumption prediction and route planning during driving, ensuring that each adjustment can more accurately reflect the current delivery environment.

[0148] Step 6: Path planning and optimization. Path planning is performed with the goal of minimizing total energy consumption and time constraints, that is, the planned route time does not exceed the delivery time limit. The non-dominated sorting genetic algorithm (NSGA-II) is used to optimize the delivery path. During the subsequent driving process, the obtained delivery path is optimized and corrected based on real-time data and new short-term energy consumption forecasts.

[0149] After completing the energy consumption forecast, the delivery route needs to be optimized based on the energy consumption forecast results. The core goal of route planning is to select the route with the lowest energy consumption while meeting the delivery time requirements.

[0150] The path planning problem can be formalized as an optimization problem with multiple objective constraints. The goal is to start from a distribution center ψ, pass through multiple distribution nodes, and return to ψ within a given urban transportation network Ω, so that the delivery vehicle completes the task within the specified time T while minimizing energy consumption. The following is the specific construction of the mathematical model.

[0151] In step 6, the total energy consumption minimization objective function is:

[0152]

[0153] Among them, E drive (i,i+1) represents the energy consumption of the vehicle traveling from node i to node i+1, E cool (i, i+1) represents the cooling energy consumption of the same section, and N is the total number of delivery nodes in the route;

[0154] The time constraint objective function is Among them, t i represents the driving time of the vehicle at node i and from node i to node i+1, and T is the total time limit of the delivery task.

[0155] NSGA-II is a multi-objective optimization algorithm suitable for solving problems such as path planning that involve multiple conflicting objectives.

[0156] The specific steps of using the non-dominated sorting genetic algorithm NSGA-II to optimize the delivery path are:

[0157] Initial population generation:

[0158] Path encoding,

[0159] Each individual in the initial population represents a possible delivery path, which is represented by a set of node sequences. N , the path can be encoded as a sequence [P11,P2,P3,…,P N ],

[0160] Randomly generate the initial population:

[0161] By randomly shuffling the order of delivery nodes, multiple different paths are generated, each corresponding to an individual. Each individual represents a possible delivery solution in the decision variable space.

[0162] Fitness evaluation:

[0163] Energy consumption calculation:

[0164] For each individual, i.e., path, the total energy consumption is calculated through the energy consumption prediction model, including the driving energy consumption E drive and cooling energy consumption E cool , use the objective function to calculate the fitness of the individual,

[0165]

[0166] Time Assessment:

[0167] Calculate the total delivery time for each route and use it as the second objective function:

[0168]

[0169] If f2(P)>T, the path does not meet the time constraint and the fitness value will be penalized and adjusted;

[0170] Non-dominated sorting and selection:

[0171] Non-dominated sorting:

[0172] In the population, each individual is sorted according to energy consumption f1(P) and time f2(P); path P a Dominant path P b The conditions are:

[0173] f1(P a )≤f1(P b ) and f2(P a )≤f2(P b )

[0174] And at least one inequality sign is a strict inequality.

[0175] Crowding distance calculation, in order to maintain the diversity of the population, the crowding distance of individuals is calculated in each non-dominated layer. The crowding distance indicates the degree of isolation of individuals in the target space, and the calculation formula is:

[0176]

[0177] Among them, f k (i+1) and f k (i-1) are the neighboring individual values ​​of individual i on the kth target.

[0178] Selection, crossover, and mutation

[0179] Selection operation: Use the binary tournament selection method based on non-dominated sorting and crowding distance to select individuals from the current population to generate the next generation population, giving priority to individuals with lower non-dominated levels or larger crowding distances.

[0180] Crossover operation: Use the sequential crossover (OX) or partial mapping crossover (PMX) method to crossover the selected individuals to generate new path individuals. The crossover operation helps explore new path combinations in order to find a better path.

[0181] Mutation operation: Use point mutation or neighbor exchange method to randomly change the order of one or more nodes in the path to increase the diversity of the population and avoid local optimal solutions;

[0182] Environmental selection and generation of new populations,

[0183] The parent and offspring populations are merged to form a new population. The non-dominated sorting and crowding distance calculations are then performed on the merged population again. The individuals with the best fitness values ​​(usually the top N) are selected from the merged population to form a new population. These individuals will enter the next generation of evolution, i.e. iteration.

[0184] Repeat the above process of selection, crossover, mutation and production of new populations until the predetermined number of iterations is reached or the population fitness converges. At this time, the optimal solution is the best delivery path under the current conditions.

[0185] The algorithm is considered to have converged when the optimal solution of the population no longer improves significantly over several generations. At this point, the optimal solution is the best delivery route under the current conditions.

[0186] Before the start of distribution, the distribution path is optimized by the non-dominated sorting genetic algorithm NSGA-II based on the results of long-term energy consumption prediction to obtain a preliminary distribution path;

[0187] During driving, short-term energy consumption predictions are made based on the collected real-time data. Energy consumption predictions are also made for subsequent road sections between nodes to obtain the energy consumption of each road section that may be involved (i.e., between two nodes). The delivery route is then optimized using the non-dominated sorting genetic algorithm (NSGA-II) to obtain the optimized delivery route.

[0188] During subsequent driving, new real-time data is collected and short-term energy consumption forecasts are repeated, thereby continuously optimizing the previous delivery route.

[0189] After the path planning is completed and the optimal path is selected, the system also needs to process the data and provide feedback after the task is completed.

[0190] After completing a delivery mission, actual driving data (such as actual energy consumption, driving time, changes in environmental conditions, etc.) is fed back to the database. This data is used to evaluate the accuracy and effectiveness of the current energy consumption prediction model and path planning algorithm.

[0191] Based on feedback data, the energy consumption prediction model parameters are automatically adjusted, specifically fine-tuning the weights in the CNN and LSTM networks to adapt to different delivery environments and task requirements. The updated model will be applied to future delivery tasks to improve prediction accuracy and planning efficiency.

[0192] The total energy consumption of each delivery task is evaluated and summarized, and the main components of energy consumption (the ratio of driving energy consumption to cooling energy consumption) are analyzed.

[0193] The pathfinding algorithm is optimized and iteratively learned through feedback. Specifically, based on actual delivery data, key parameters in the NSGA-II algorithm, such as the crossover rate, mutation rate, and population size, are optimized and adjusted to improve the algorithm's applicability and convergence speed for different delivery tasks.

[0194] Through the execution and feedback of multiple delivery tasks, the performance of the model and algorithm is gradually improved to make it more adaptable to the complex and changing urban road network environment and the special requirements of cold chain transportation.

[0195] In summary, if Figure 1 As shown, in specific use, data is first collected and processed, and the processed data is input into the energy consumption prediction model to obtain the energy consumption data of each node;

[0196] When the destination is reached and the delivery is completed, the route optimization ends. If the delivery is not completed, the data from this trip is fed back to the entire system to optimize the energy consumption prediction model and the non-dominated sorting genetic algorithm NSGA-II. The process from collecting real-time data to route planning is repeated until the delivery is completed.

[0197] When planning a route, a long-term prediction is first performed to plan a long-term route. During driving, the long-term route is adjusted based on the real-time data collected, and the remaining mileage route is planned. For example, based on traffic information, if an accident occurs on a road in the long-term route, the energy consumption of this section will increase. When subsequently planning the route, due to the excessive energy consumption here, the section will be bypassed and replanned to obtain a new route. The above steps are repeated until the delivery is completed.

Claims

1. A two-stage energy consumption prediction path planning method based on deep learning, characterized in that: The following steps are involved: Step 1: Data collection: Collect data on the actual operation of cold chain distribution vehicles in the past, including historical driving data, vehicle parameter data and real-time traffic data; Step 2: Data preprocessing, normalizing and standardizing the data, and performing feature engineering to select and construct features that are informative for energy consumption prediction; Step 3: Energy consumption model construction: construct the driving energy consumption and cooling energy consumption calculation formulas, and extract the feature set through the convolutional neural network model and the collected data; Input the feature set into the long short-term memory network (LSTM) to generate an energy consumption prediction model; Step 4: Long-term energy consumption prediction. Based on the historical data of vehicles on similar delivery tasks, we extract the characteristics related to the current task that affect energy consumption. These characteristics are input into the energy consumption prediction model to predict the energy consumption of possible future road sections. The driving energy consumption and cooling energy consumption between each two nodes are predicted. Step 5: Short-term energy consumption prediction: During the vehicle's driving process, data affecting energy consumption is collected in real time and input into the energy consumption prediction model. The driving energy consumption and cooling energy consumption between every two nodes of the current and future road sections are predicted, and the long-term energy consumption prediction is revised. Step 6: Route planning and optimization. This involves minimizing total energy consumption and ensuring that the planned route time does not exceed the delivery time limit. This involves using the non-dominated sorting genetic algorithm (NSGA-II) to optimize the delivery route. The resulting delivery route is then optimized and revised during subsequent travel based on real-time data and new short-term energy consumption forecasts. The driving energy consumption calculation formula is: m is the mass of the vehicle, g is the acceleration of gravity, θ is the road slope, C d is the air resistance coefficient, ρ is the air density, K is the shape factor of the vehicle, v is the vehicle speed, Cr is the rolling resistance coefficient, P tract is the traction power of the electric vehicle; Average speed and energy consumption during comprehensive driving μ eled is the energy conversion efficiency of electric vehicles; M is the mass of the vehicle; The calculation formula for cooling energy consumption is E cool =γ·d. Where γ is the cooling energy consumption per unit distance and d is the travel distance; In step 6, the total energy consumption minimization objective function is: Among them, E drive (i,i+1) represents the energy consumption of the vehicle traveling from node i to node i+1, E cool (i, i+1) represents the cooling energy consumption of the same section, N is the total number of distribution nodes in the path, E total is the total energy consumption; The time constraint objective function is Among them, t i represents the driving time of the vehicle at node i and from node i to node i+1, and T is the total time limit of the delivery task.

2. The method for two-stage energy consumption prediction path planning based on deep learning according to claim 1, characterized in that: The data is data that affects delivery time and refrigeration energy consumption, and the historical driving data includes the actual driving speed, driving route, time and date, external environmental conditions and road slope of the vehicle on each road section; Vehicle parameter information refers to the physical and mechanical parameters of the vehicle, including fully loaded and unloaded weight, battery type and capacity, and cooling system specifications; Real-time traffic data includes current traffic flow, road information, traffic accident data, and road topology.

3. The method for two-stage energy consumption prediction path planning based on deep learning according to claim 1, characterized in that: In the step 2, the features that are informative for energy consumption prediction are selected and constructed, including speed change rate, slope characteristics and environmental characteristics.

4. The method for two-stage energy consumption prediction path planning based on deep learning according to claim 1, characterized in that: In step three, the convolutional neural network model and the collected data are used to extract a feature set. Specifically, the preprocessed historical driving data is input into the CNN network. The data format is a multi-dimensional matrix. The feature map of the data is extracted through multiple convolutional layers to reveal the correlation and spatial pattern between different features. The convolution feature map is reduced in dimension using a pooling layer to reduce the amount of calculation while retaining key features. The final output feature set retains the key information of the data and provides input for the subsequent LSTM network.

5. The method for two-stage energy consumption prediction path planning based on deep learning according to claim 1, characterized in that: The specific steps of using the non-dominated sorting genetic algorithm NSGA-II to optimize the delivery path are: Initial population generation: Each individual in the initial population represents a possible delivery path. The path is represented by a set of node sequences. For the delivery points P1, P2, ..., P N , the path can be encoded as a sequence [P11,P2,P3,…,P N ], By randomly disrupting the order of delivery nodes, multiple different paths are generated. Each path corresponds to an individual, and each individual represents a possible delivery plan in the decision variable space. Fitness evaluation: For each individual, that is, path, the total energy consumption is calculated through the energy consumption prediction model, including the driving energy consumption E drive and cooling energy consumption E cool , use the objective function to calculate the fitness of the individual, Calculate the total delivery time for each route and use it as the second objective function: If f2(P)>T, the path does not meet the time constraint and the fitness value will be penalized and adjusted; Non-dominated sorting and selection: In the population, each individual is sorted according to energy consumption f1(P) and time f2(P), and the crowding distance of the individual is calculated in each non-dominated layer; Selection operation: Use the binary tournament selection method based on non-dominated sorting and crowding distance to select individuals from the current population to generate the next generation population. Crossover operation: Use sequential crossover or partial mapping crossover methods to cross selected individuals and generate new path individuals. The crossover operation helps explore new path combinations in order to find a better path. Mutation operation: Use point mutation or neighbor exchange method to randomly change the order of one or more nodes in the path to increase the diversity of the population and avoid local optimal solutions; The parent and child populations are merged to form a new population, and then the non-dominated sorting and crowding distance calculation are performed on the merged population again. The individuals with the best fitness value are selected from the merged population to form a new population. These individuals will enter the next generation of evolution, that is, iteration; Repeat the above process of selection, crossover, mutation and production of new populations until the predetermined number of iterations is reached or the population fitness converges. At this time, the optimal solution is the best delivery path under the current conditions.

Images