Airport electric ground service vehicle scheduling method integrating MAMBA and neural network
By integrating MAMBA and neural network methods, the scheduling of electric ground vehicles is optimized, the problem of inefficient traditional manual scheduling is solved, efficient and correct scheduling plan generation is achieved, and vehicle utilization and emergency response capabilities are improved.
Patent Information
- Application Number
- CN202510196843.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-21
- Publication Date
- 2025-06-06
AI Technical Summary
Traditional manual scheduling methods are inefficient and it is difficult to effectively optimize the scheduling of electric ground vehicles, resulting in low vehicle utilization, difficulty in emergency response and increased risk of flight delays.
Using the method of integrating MAMBA and neural networks, a hybrid integer planning model is established, and combining the Mamba model, LSTM neural network and attention mechanism is used to optimize the scheduling scheme of electric ground trains.
It has achieved efficient and correct handling of scheduling problems of electric ground vehicles, generated high-quality scheduling solutions, improved vehicle utilization and emergency response capabilities, and reduced flight delay risks.
Smart Images
Figure CN120106487A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of ground vehicle control, and in particular to an airport electric ground vehicle dispatching method integrating MAMBA and neural network. Background Art
[0002] Efficient airport operations rely heavily on the support of ground service vehicles, which play a key role in ensuring flight safety, improving operational efficiency, and meeting passenger needs. Based on the study of actual airport operations, the main processes of ground support services can be summarized as follows: Figure 1 As environmental awareness and sustainable development concepts increase, many airports are transitioning to electric ground handling vehicles. However, this transition brings significant scheduling challenges, as the scheduling of electric vehicles has not been fully studied, and traditional models cannot cope with the complexity of modern airport operations. Therefore, optimizing the scheduling of electric ground handling vehicles has become a key issue for airport management.
[0003] To ensure smooth airport operations, it is essential to develop scheduling solutions for different flight types. However, most airports still rely on manual operations for scheduling single-vehicle, single-flight services, which is inefficient, especially considering the complex needs of electric vehicles. Manual scheduling leads to low vehicle utilization, difficulty in emergency handling, and increased risk of flight delays.
[0004] Therefore, there is an urgent need for an intelligent and efficient scheduling method that can not only give full play to the advantages of electric vehicles but also solve the inefficiency of the current manual scheduling method. Summary of the invention
[0005] To solve at least one of the above problems, the present invention proposes an airport electric ground vehicle dispatching method integrating MAMBA and neural network.
[0006] The technical solution of the present invention is: a method for dispatching electric ground vehicles for airports integrating MAMBA and neural network, comprising the following steps: S1. Establish a mixed integer programming model: Taking the minimization of the driving distance of the electric ground handling vehicle as the objective function, consider factors including service order, driving distance, charging, remaining power of the electric ground handling vehicle and the number of electric ground handling vehicles to establish constraints; S2. Collect dispatch characteristics, including the location of the charging station, the aircraft landing location, the initial location of the electric ground handling vehicle, the service duration of the electric ground handling vehicle, the service start time of the electric ground handling vehicle, and the travel distance of the electric ground handling vehicle; S3, based on the Mamba model, encode these scheduling features to obtain feature representation; S4, based on LSTM neural network, combined with attention mechanism and feature representation of S3, solves the mixed integer programming model and outputs the scheduling plan.
[0007] Beneficial effects: The method of the present invention establishes a mixed integer programming model by considering factors such as the remaining power of the vehicle and the charging requirements, and at the same time, uses the Mamba model, LSTM neural network and attention mechanism to solve the model. This enables the present invention to efficiently and correctly handle the scheduling problem of electric ground vehicles and generate high-quality scheduling solutions. BRIEF DESCRIPTION OF THE DRAWINGS
[0008] Figure 1 is an algorithm flow chart of an embodiment of the present invention; Figure 2 Heatmaps for computing time analysis of different algorithms at different numbers of flights; Figure 3a Comparison chart of total driving distance and performance gap between different algorithms under 20 flights; Figure 3b Comparison chart of total driving distance and performance gap between different algorithms under 50 flights; Figure 3c The comparison chart of total driving distance and performance gap between different algorithms under the number of 100 flights; Figure 3d The comparison chart of total driving distance and performance gap between different algorithms under the number of 200 flights; Figure 4a The probability distribution comparison diagram of total driving distance and performance gap between different algorithms under Gaussian distribution; Figure 4b The probability distribution comparison diagram of total driving distance and performance gap between different algorithms under Poisson distribution; Figure 5 Heat map of probability distribution analysis of the corresponding computation time for different algorithms. DETAILED DESCRIPTION
[0009] The specific implementation modes of the present invention will be described clearly and completely below with reference to examples. Obviously, the examples described are only some embodiments of the present invention, rather than all embodiments.
[0010] like Figure 1 As shown, a method for dispatching electric ground vehicles for airports integrating MAMBA and neural network includes the following steps: S1. Establish a mixed integer programming model: Taking the minimization of the driving distance of the electric ground handling vehicle as the objective function, consider factors including service order, driving distance, charging, remaining power and the number of electric ground handling vehicles to establish constraints; in, , where f represents the driving distance; i represents the number of the electric ground handling vehicle, and i=1,2,…,n v , n v represents the maximum number of electric ground handling vehicles; j represents the flight number, and j=1,2,…,n f , n f Indicates the maximum flight number; x ij An indicator variable indicating whether electric ground handling vehicle i provides service for flight j, x ij =1 means that electric ground handling vehicle i can provide service for flight j, otherwise x ij =0; TD ij is the distance travelled by electric ground handling vehicle i during the period of providing service for flight j; The constraints are as follows: Service order constraint: BT ij +ST ij ≤BT ij' , BT ij +ST ij ≤BT i'j , where BT ij is the time point when electric ground handling vehicle i starts to be occupied by flight j; BT ij' is the time point when the electric ground handling vehicle i starts to be occupied by flight j'; ST ij The total time that electric ground handling vehicle i provides service for flight j; BT ij' is the time point when the electric ground handling vehicle i starts to be occupied by flight j'; BT i'j is the time point when electric ground handling vehicle i' begins to be occupied by flight j; electric ground handling vehicle i is ranked before electric ground handling vehicle i'. After electric ground handling vehicle i completes the service for flight j, electric ground handling vehicle i' can start the service. At the same time, electric ground handling vehicle i has priority to serve flight j. Flight j' is the next task of electric ground handling vehicle i after it completes the service for flight j. Travel distance constraint: TD ij =(TT ij )×S i +C i ×D jp , where TT ij The total time of the electric ground handling vehicle is ST ij S is the flight time of flight j; i is the average speed of electric ground vehicle i; C i Indicator variable indicating whether electric ground handling vehicle i is charging, C i =1 means that the electric ground handling vehicle i is charging, otherwise C i =0;D jp is the distance from flight j to warehouse p; Charging limit: VD ij −Djp ≥0, VD ij −TD ij' ×S i −D j'p ≤0, CR i =VD i m / CT i , CT i c =C i ×(VD i m −VD ij −D jp ) / CR i , FT i =BT ij' −BT ij −ST ij −C i ×D jp / S i , FT i ≥CT i c 、VD ij ≤0.2×VD i m , where VD ij D is the distance that electric ground handling vehicle i travels with the remaining power when it is about to serve flight j; jp is the distance from flight j to warehouse p; TD ij' D is the distance travelled by electric ground handling vehicle i during the period of providing service for flight j'; j'p is the distance from flight j' to warehouse p; CR i VD is the distance traveled by electric ground vehicle i per unit charging time; i m CT is the distance traveled by the electric ground handling vehicle i when fully charged; i m The time required to charge the remaining power of the electric ground handling vehicle from 0 to full power for the electric ground handling vehicle i; CT i c The time required to charge the electric ground handling vehicle i from the current remaining power of the electric ground handling vehicle to full power; FT i is the idle time of electric ground handling vehicle i between serving flights j and j', BT ij' is the time point when electric ground handling vehicle i starts to be occupied by flight j'; Remaining power limit: D j ≤Q i , where D j is the demand of flight j within the service time; Q i is the battery capacity of the electric ground handling vehicle i; Electric ground handling vehicle quantity limit: n vs ≤n v , where n vs The number of electric ground handling vehicles in service; n v is the total number of electric ground vehicles.
[0011] The mixed integer programming model of this embodiment aims to minimize the driving distance and takes into account issues such as the charging constraints of the electric ground vehicle, making the model more in line with current actual conditions.
[0012] S2. Collect dispatch characteristics, including the location of the charging station, the aircraft landing location, the initial location of the electric ground handling vehicle, the service duration of the electric ground handling vehicle, the service start time of the electric ground handling vehicle, and the travel distance of the electric ground handling vehicle; These characteristics can be easily obtained from the daily operation records of electric ground handling vehicles and the flight records of airports.
[0013] S3, based on the Mamba model, encode these scheduling features to obtain feature representation; In this step, considering the complexity of the airport electric ground vehicle scheduling problem, it involves multi-dimensional features such as location information, service requirements and time windows. It is difficult to process the above scheduling features using conventional methods. Even if the unprocessed scheduling features are processed by machine learning, it is difficult to contribute and guide actual production. Therefore, after a large number of experiments, the inventor found that the Mamba model can be used to process these multi-dimensional features and accurately capture the complex relationships between nodes. This step mainly includes the following sub-steps: S31, combining the learnable weight matrix and the bias term, projecting the scheduling features into the embedding space and constructing the initial features; Among them, the initial features are expressed as follows: , where represents the initial feature; e 0 is the node information of the parking lot, which is used to initialize the feature representation of the parking lot node; e j The node corresponding to flight j is based on D jp W is the learnable weight matrix that projects the scheduling features into the embedding space; b is a bias term; T ij The total time that electric ground handling vehicle i provides service for flight j; BT ij is the time point when electric ground handling vehicle i starts to be occupied by flight j; TD ij is the distance traveled by electric ground handling vehicle i during the period of providing service for flight j.
[0014] In this step, a node is a basic unit in Mamba, which is used to represent a key entity or position in a scheduling problem. This is common knowledge in the art and will not be described in detail here.
[0015] S32, processing the initial features through Mamba Block and generating new node features; in this step, the initial features are processed through N=3 Mamba Blocks, and each Mamba Block uses state space modeling to process node information, thereby generating new node features; Among them, the new node feature is generated by the following formula: j' (l) =LayerNorm (l) (h j (l-1) +MambaBlock (l) (h j (l-1) ;h 0 (l-1) ,...,h nf (l-1) )), where h j (l) is the feature representation of the node corresponding to flight j in layer l, LayerNorm (l) is the layer normalization operation of the lth layer; h j (l-1) is the feature representation of the node corresponding to flight j in the l-1th layer, MambaBlock (l) is the Mamba module of the lth layer, used for feature update, h 0 (l-1) ,....,h nf (l-1) is the feature representation of all nodes in the l-1th layer.
[0016] S33, repeat S32 until the operation of all nodes is completed; S34, based on the node features of S33, combined with Mamba Block, generates feature representation; The characteristic representation is as follows: h j (l) =LayerNorm (l) (h j' (l) ), where h 0 (N) is the feature aggregation of all nodes; h j is the feature representation of node j; h 0 ,...,h nf is the feature representation of all nodes, including the current node and all flight nodes.
[0017] S4, based on LSTM neural network, combined with attention mechanism and feature representation of S3, solves the mixed integer programming model and outputs the scheduling plan.
[0018] The inventors found in actual use that the LSTM neural network based on the attention mechanism builds a powerful context representation through the multi-head attention mechanism, combined with the optimization of the reinforcement learning strategy, and performs well in dealing with the problems of the embodiments of the present invention. While considering the charging conditions, the electric ground vehicle selects the flights to be served, and optimizes the scheduling strategy through reinforcement learning. This design enables the algorithm to efficiently and accurately handle the scheduling problem of the electric ground vehicle and automatically generate high-quality scheduling solutions.
[0019] This step includes the following sub-steps: S3.1. Integrate the unified pooling embedding h from the Mamba encoder 0 (N) , combined with the electric ground vehicle logo e i 、Previous service flights embedded h prev (N) , Remaining driving distance VD ij and current service start time BT ij Generate comprehensive context embedding h c (N) ;
[0020] The context embedding formula is as follows: c (N) =[ph 0 (N) ,e i ,h prev (N) ,VD ij ,BT ij ],∀j∈V,∀j∈F; where hc (N) Embedded for context; ph 0 (N) represents the embedding h from the Mamba encoder 0 (N) Unified pooling of i Indicates the current electric ground handling vehicle i, used to distinguish it from other electric ground handling vehicles; h prev (N) Embed for previously serviced flights; VD ij Indicates the remaining driving distance of the electric ground vehicle i; BT ij Indicates the current service start time of the electric ground handling vehicle i.
[0021] S3.2. Use multi-head attention mechanism to process context embedding h c (N) ; S3.3. Evaluate all possible service decisions and update the state variable h (N) ;
[0022] Among them, the variable update formula is: , BT ij' =BT ij +ST ij +C i ×CT i c +TT ij' ; where, TD ij represents the travel distance of service flight j; CT i c Indicates the time it takes for the electric ground handling vehicle i to reach full charge, ST ij represents the service time of electric ground handling vehicle i occupied by flight j, BT ij' represents the starting time point of the next potential flight j'.
[0023] S3.4. Calculate the selection score V using an LSTM neural network based on an attention mechanism and nonlinear transformation t (j) ; The selection score calculation formula is as follows: , where v t (j) is the selection score for flight j at time step t; ReLU is the activation function used to introduce nonlinearity; W q and W k is the learnable weight matrix, W q Used to embed context into h c (N+1) Transformed into a query vector, W k Used to convert flight characteristics h j (N) Transform to Key Vector ， h c (N+1) is the context embedding, the embedding updated by the multi-head attention mechanism at layer N+1; h j (N) is the feature representation of flight j at layer N; d is the dimension of the feature, which is used to scale the dot product to prevent the gradient from disappearing.
[0024] In the decoding framework of the LSTM neural network, its strategy includes one of the greedy strategy and the sampling strategy. Finally, different scheduling schemes are generated according to different strategies, and the optimal scheduling scheme is selected as the output scheme.
[0025] S3.5. Check the satisfaction of the operational constraints and assign a negative infinite score -∞ to the infeasible flights; S3.6. Use the Softmax function to convert the scores into probability distribution, form a complete scheduling strategy distribution, and select the optimal flight allocation plan;
[0026] Among them, the probability distribution conversion formula is as follows: π θ (a i |d i ,a i 1:t-1 )=Softmax(v t (a i )); where π θ (a i |d i ,a i 1:t-1 ) is the probability distribution of the dispatching plan of electric ground vehicle i, which is determined by the parameter θ; a i is the dispatching action of electric ground handling vehicle i, i.e., electric ground handling vehicle i selects the flight to serve; a i 1:t-1 is the current partial solution (route) constructed before the tth iteration; d i is the dispatching problem of electric ground vehicle i, v t (a i ) is the vector used to calculate the selection score. This formula explicitly considers the distance-related operating costs, while implicitly considering the constraints of the remaining power and charging time of the electric ground vehicle through the decision-making process of the policy network.
[0027] The scheduling formula is as follows: ; In the formula, d i For the dispatching problem of electric ground handling vehicle i, the solution is π θ (a i |d i ) is parameterized by θ; T represents the total time steps or total decision steps of the scheduling scheme; a i t represents the dispatching action of electric ground handling vehicle i at the tth time step, that is, the flight that electric ground handling vehicle i chooses to serve at the tth time step; a i 1:t-1 For the tth th The current partial solution that was built iterations ago.
[0028] The scheme evaluation formula is as follows: L(θ|d i ) = Eπ θ (a i ∣d i )[L(a i )]; where L(θ|d i) is the evaluation index, and the constraints of the remaining power and charging time of the electric ground vehicle are implicitly considered through the decision-making process of the policy network; E is the expected value, that is, for all possible scheduling schemes a i The probability distribution of π θ (a i |d i ) is the probability distribution of the dispatching plan of electric ground vehicle i; L(a i ) is the travel distance of electric ground vehicle i at the tth time step. Through this formula, each scheduling scheme is evaluated and the optimal scheduling scheme is selected.
[0029] S3.7. Use BS with rolling baseline (d i )’s reinforcement algorithm trains the policy network, optimizes the network parameters θ through the gradient update formula, and uses the verification mechanism X val Ensure continuous improvement of solution quality.
[0030] The gradient update formula is as follows: ; where θ is the model parameter used to parameterize the probability distribution of the scheduling scheme, L(a i ) is the evaluation index of the scheme, such as the operating cost related to the distance. ∇L(θ|d) is the gradient update formula. z is a randomly generated sample used to generate a scheduling scheme; A set of B random samples, each of which is used to generate a scheduling solution; BS(d i ) is the baseline function, used to stabilize the gradient update; ∇logπ θ (a i ∣d i ) is the gradient of the log probability distribution.
[0031] For a specific flow chart of the algorithm of the embodiment of the present invention, see Figure 1 .
[0032] In order to further illustrate the advantages of the embodiments of the present invention, specific examples are given below.
[0033] The configuration of electric ground handling vehicles is specifically based on the statistical data of Xiamen Gaoqi Airport, involving 10 different ground handling vehicle services in 4 stages. Regarding the service time of various ground handling vehicles, it takes 2 minutes to dock the passenger elevator, 6 minutes for passengers to get off the plane, 4 minutes for catering service, 4 minutes for cleaning service, 8 minutes for refueling service, 2 minutes for unloading luggage, 4 minutes for loading luggage, 8 minutes for passengers to board the plane, 8 minutes for sewage treatment, and 2 minutes for disconnecting the passenger elevator. Xiamen Gaoqi Airport was selected as the research object. The airport includes 2 terminals and 89 parking spaces. Through the field investigation of the airport, the airport plan was loaded into SUMO to obtain the locations of the parking lot, terminal, and parking space (the parking lot is the charging location and also the initial location of the electric ground handling vehicle; the parking space is the aircraft parking location, and the service starts after the aircraft is docked). The driving distance is automatically calculated using SUMO. At the same time, combined with the existing research data, the service start time is determined to be the flight arrival time plus the turnaround time (30 minutes) plus the latest time when all services in the previous stage end. Based on the flight data of Xiamen Gaoqi Airport, experimental cases with 20, 50, 100, and 200 flights were created. In order to comprehensively evaluate the performance of the algorithm and verify its effectiveness, various scheduling cases were designed and compared with the CPLEX solver, the latest heuristic algorithm, and the self-built heuristic algorithm. In addition, a situation involving real-time flight arrivals was simulated to test the algorithm's ability to handle dynamic flight changes. For each flight, the flight demand was set to [1,…,12]. Then, for 20, 50, 100, and 200 flights, the number of electric ground handling vehicles was set to 30, 40, 50, and 60, respectively, to ensure that the actual vehicle demand of the airport was met. The vehicle speed was set to 30 km / h, and the cruising range of the electric ground handling vehicle was set to 80 km. The policy network was trained 200 times. In order to match the actual situation, 12,800 instances were generated per epoch based on the flight information of Xiamen Gaoqi Airport. Then, 1,000 cases with the same scale and random seed were randomly generated at the end of each epoch as a validation set to evaluate the model performance. During the testing phase, 1,000 cases with 20, 50, 100, and 200 flights were generated using different random seeds. Finally, during the testing process, comparisons were made with CPLEX, ILS, ACS-SA, LNS-SA, GA-ACO, CP-LNS, Random Insertion, Nearest Insertion, Farthest Insertion, and Nearest Neighbor algorithms.
[0034] Among them, two strategies are set in the decoding framework of the LSTM neural network: 1) Greedy strategy, at each time step, the algorithm selects the flight with the highest probability based on the prediction of the policy network; 2) Sampling strategy, first sampling from the probability distribution of the policy network to generate 1,000 different candidate solutions, and then selecting the solution with the best performance as the output.
[0035] For CPLEX and heuristic algorithms, a time limit of 60 minutes was set for the cases with solution scales of 20, 50, 100, and 200. For the heuristic algorithm, the population size was uniformly set to 200, the crossover probability was set to 0.8, and the mutation probability was set to 0.3. For the algorithm designed in this patent, the policy network was trained for 200 epochs using the flight data of Xiamen Gaoqi Airport. Each training epoch consisted of 200 batches with a batch size of 64 instances, and at the end of each epoch, 1000 instances were evaluated. In the neural structure, the Mamba model used 3 layers in the encoder and 8 heads in each multi-head attention layer in the decoder. The embedding dimension of the algorithm was set to 128. The Adam optimizer updated the parameters with a constant learning rate of 10−4. To verify the performance of the algorithm, the significance level of the t-test was set to 5%.
[0036] In addition, the average processing time for each case of each algorithm is recorded, which depends on the specific method used. The experiments were performed on a machine equipped with a GeForce RTX-4060 GPU and an Intel i7-12650H CPU running at 4.70 GHz.
[0037] Test Example 1 This example uses airport electric ground vehicle dispatching as an example to evaluate the algorithm performance. Based on the flight data of Xiamen Gaoqi Airport, experimental cases with 20, 50, 100, and 200 flights were created and compared with the CPLEX solver, the latest heuristic algorithm, and the self-built heuristic algorithm. In addition, a situation involving real-time flight arrivals was simulated to test the algorithm's ability to handle dynamic flight changes.
[0038] The objective function values, gaps, and number of solutions of each algorithm under different conditions are shown in Table 1. The heat map of the number of calculations is shown in Figure 2 As shown, the objective function value and gap result obtained by the algorithm are as follows Figure 3a , as shown in 3b, 3c, and 3d. Figure 2The solution time of the algorithm was further analyzed, and the results showed that the algorithm developed in this study had the fastest computation time, outperforming all other algorithms except the nearest neighbor algorithm. Even in cases where the number of flights is large, the difference in computation time between the algorithm developed in this study and the nearest neighbor algorithm is still small, and still much less than other algorithms, which typically take several minutes or even an hour to complete. This is because the algorithm designed in this study directly constructs the solution to the case without the need for additional iterative improvement steps, making it faster than heuristic methods. Therefore, it can be concluded that the algorithm designed in this study produces an optimal solution in a reasonable amount of time, proving its superiority over other optimal algorithms. From Tables 1 and Figure 3a , 3b, 3c, 3d, it can be seen that the algorithm designed by the present invention has the shortest total distance in all cases, the average gap of the sampling strategy is close to 0, and the total distance and gap of the greedy strategy are also short. Compared with algorithms such as ACS-SA, LNS-SA, and CP-LNS, the total distance of the sampling strategy and the greedy strategy is significantly shortened, proving the superiority of the algorithm. The specific data are as follows: Table 1 Comparison of results of different algorithms
[0039] For the case of 20 flights, the ACS-SA algorithm performs best among the other compared algorithms and produces the shortest total distance. Compared with ACS-SA, the proposed method performs better, where the total distance of the sampling strategy is reduced by 9.19% and the total distance of the greedy strategy is reduced by 4.18%.
[0040] When 50 flights are considered, the LNS-SA algorithm has the shortest total distance compared to other algorithms. In this case, the total distance of the sampling strategy is reduced by 16.76% compared to LNS-SA, and the total distance of the greedy strategy is reduced by 11.32% compared to LNS-SA.
[0041] In the case of 100 flights, the CP-LNS algorithm is the best scheduling solution among the other compared algorithms. Here, the sampling strategy outperforms CP-LNS, reducing the total distance by 20.92%, and the greedy strategy also outperforms CP-LNS, reducing the total distance by 16.28%.
[0042] Finally, for 200 flights, CP-LNS again provides the best solution among the other compared algorithms. Compared with CP-LNS, the total distance of the sampling strategy is reduced by 28.50%, and the total distance of the greedy strategy is reduced by 22.39%. The sampling strategy of the present invention and the greedy strategy still have a shorter total driving distance than the best method among the other methods.
[0043] Comparing the calculation time of the algorithms, the algorithm developed by the present invention has the fastest calculation time, and is superior to all other algorithms except the nearest neighbor algorithm. Although the difference in calculation time with the nearest neighbor algorithm is small, the objective function value is much lower than it, and when the number of flights is large, the calculation time is still much less than that of other algorithms, usually taking several minutes or even an hour to complete. Because this algorithm directly constructs case solutions without additional iterative improvement steps, it is faster than the heuristic method.
[0044] Test Example 2 To verify the performance of the algorithm in various unknown scenarios, the generalization ability of the model is evaluated by sampling cases from different distributions. Flight demand is sampled from a Gaussian distribution with a mean of 6 and a variance of 3 and a Poisson distribution with an expected number of events of 6, and flight arrival times are sampled according to the hourly distribution probability of Xiamen Gaoqi Airport.
[0045] The objective function values and gap results obtained by the algorithm are shown in Table 2. Figure 4a and 4b is the objective function value and gap result graph obtained by the algorithm, Figure 5 It is the calculation frequency graph. It can be seen from the three figures that our algorithm reaches the lowest target value and gap in a shorter time. Combined with Table 2, under the Gaussian distribution and Poisson distribution, compared with the CP-LNS algorithm, the total distance of the sampling strategy and greedy strategy of the present invention is significantly reduced. The experiment shows that the algorithm has good generalization ability for both distributions.
[0046] Table 2 Variable parameter generalization research table
[0047] Under the Gaussian distribution, the total distance produced by the CP-LNS algorithm is the shortest among the other compared algorithms. Compared with CP-LNS, the sampling strategy in the present invention reduces the total distance by 27.49%, while the greedy strategy reduces the total distance by 20.41%. Similarly, under the Poisson distribution, the total distance produced by the CP-LNS algorithm is also the shortest among the other compared algorithms. Compared with CP-LNS, the total distance of the sampling strategy in the present invention is reduced by 16.32%, while the greedy strategy reduces the total distance by 20.56%.
[0048] Test Example 3 To verify the algorithm's ability to handle real-world scenarios, its performance on randomly arriving real-time flights was tested and compared with the fast heuristic algorithm. The algorithm was adjusted to run in reoptimization mode to simulate the real-time dynamics of the airport. 1,000 cases containing 20 flights were generated for testing, and gradually expanded to 200 flights.
[0049] From the results in Table 3, it can be seen that the algorithm proposed in the present invention has significant advantages in real-time scheduling problems. The sampling strategy and the greedy strategy have significantly lower costs than the nearest neighbor algorithm, and the generated scheduling solution is almost real-time, which proves that the algorithm is suitable for solving the real-time scheduling problem of electric ground vehicles.
[0050] Table 3 Real-time electric ground vehicle dispatch results
[0051] Among other algorithms compared for real-time scheduling problems, the nearest neighbor algorithm performs best. However, the algorithm proposed in the present invention is compared with the nearest neighbor algorithm, and the results show that the sampling strategy in the present invention significantly reduces the cost by 36.64%, while the greedy strategy reduces the cost by 34.24%.
[0052] Overall, the algorithm we proposed shows excellent performance in the electric ground handling vehicle scheduling problem. The sampling strategy balances the solution quality and computational speed, and is suitable for dynamic and demanding airport environments. Combined with technologies such as the Mamba model and attention-based neural networks, the algorithm shows excellent adaptability and can efficiently generate real-time solutions, which has wide application potential and great value.
[0053] The above description is only a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as a preferred embodiment as above, it is not intended to limit the present invention. Any technician familiar with the profession can make some changes or modifications to equivalent embodiments of equivalent changes using the technical contents disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of the technical solution of the present invention.
Claims
1. A method for dispatching electric ground vehicles for airports integrating MAMBA and neural network, characterized in that: The following steps are involved: S1. Establish a mixed integer programming model: Taking the minimization of the driving distance of the electric ground handling vehicle as the objective function, consider factors including service order, driving distance, charging, remaining power and the number of electric ground handling vehicles to establish constraints; S2. Collect dispatch characteristics, including the location of the charging station, the aircraft landing location, the initial location of the electric ground handling vehicle, the service duration of the electric ground handling vehicle, the service start time of the electric ground handling vehicle, and the travel distance of the electric ground handling vehicle; S3, based on the Mamba model, encode these scheduling features to obtain feature representation; S4, based on LSTM neural network, combined with attention mechanism and feature representation of S3, solves the mixed integer programming model and outputs the scheduling plan.
2. The method according to claim 1, characterized in that The objective function is as follows , where f represents the driving distance; i represents the number of the electric ground handling vehicle, and i=1,2,…,n v , n v represents the maximum number of electric ground handling vehicles; j represents the flight number, and j=1,2,…,n f , n f Indicates the maximum flight number; x ij An indicator variable indicating whether electric ground handling vehicle i provides service for flight j, x ij =1 means that electric ground handling vehicle i can provide service for flight j, otherwise x ij =0; TD ij is the distance travelled by electric ground handling vehicle i during the period of providing service for flight j; The constraints are as follows: Service order constraint: BT ij +ST ij ≤BT ij' , BT ij +ST ij ≤BT i'j , where BT ij is the time point when electric ground handling vehicle i starts to be occupied by flight j; BT ij' is the time point when the electric ground handling vehicle i starts to be occupied by flight j'; ST ij The total time that electric ground handling vehicle i provides service for flight j; BT ij' is the time point when the electric ground handling vehicle i starts to be occupied by flight j'; BT i'j is the time point when electric ground handling vehicle i' begins to be occupied by flight j; electric ground handling vehicle i is ranked before electric ground handling vehicle i'. After electric ground handling vehicle i completes the service for flight j, electric ground handling vehicle i' can start the service. At the same time, electric ground handling vehicle i has priority to serve flight j. Flight j' is the next task of electric ground handling vehicle i after it completes the service for flight j. Travel distance constraint: TD ij =(TT ij )×S i +C i ×D jp , where TT ij The total time of the electric ground handling vehicle is ST ij S is the flight time of flight j; i is the average speed of electric ground vehicle i; C i Indicator variable indicating whether electric ground handling vehicle i is charging, C i =1 means that the electric ground handling vehicle i is charging, otherwise C i =0;D jp is the distance from flight j to warehouse p; Charging limit: VD ij −D jp ≥0, VD ij −TD ij' ×S i −D j'p ≤0, CR i =VD i m / CT i m , CT i c =C i ×(VD i m −VD ij −D jp ) / CR i , FT i =BT ij' −BT ij −ST ij −C i ×D jp / S i , FT i ≥CT i c 、VD ij ≤0.2×VD i m , where VD ij is the distance that electric ground handling vehicle i travels with the remaining power when it is about to serve flight j; TD ij’ D is the distance travelled by electric ground handling vehicle i during the period of providing service for flight j'; j'p is the distance from flight j' to warehouse p; CR i VD is the distance traveled by electric ground vehicle i per unit charging time; i m CT is the distance traveled by the electric ground handling vehicle i when fully charged; i m The time required to charge the remaining power of the electric ground handling vehicle from 0 to full power for the electric ground handling vehicle i; CT i c The time required to charge the electric ground handling vehicle i from the current remaining power of the electric ground handling vehicle to full power; FT i is the idle time of electric ground handling vehicle i between serving flights j and j', BT ij' is the time point when electric ground handling vehicle i starts to be occupied by flight j'; Remaining power limit: D j ≤Q i , where D j is the power consumption of flight j during the service time of the ground handling vehicle; Q i is the current remaining power of the electric ground handling vehicle i; Electric ground handling vehicle quantity restrictions: , where n vs The number of electric ground handling vehicles currently in service.
3. The method according to claim 1, characterized in that S3 includes the following steps: S31, combining the learnable weight matrix and the bias term, projecting the scheduling features into the embedding space and constructing the initial features; S32, processing the initial features through Mamba Block and generating new node features; S33, repeat S32 until the operation of all nodes is completed; S34, based on the node features of S33, combined with Mamba Block, generates feature representation.
4. The method according to claim 3, characterized in that The initial features are as follows: , where represents the initial features; e0 is the node information of the parking lot, which is used to initialize the feature representation of the parking lot node; e j The node corresponding to flight j is based on D jp Learnable position embedding of D jp is the distance from flight j to warehouse p; W is the learnable weight matrix that projects the scheduling features into the embedding space; b is the bias term; ST ij The total time that electric ground handling vehicle i provides service for flight j; BT ij is the time point when electric ground handling vehicle i starts to be occupied by flight j; TD ij is the distance travelled by electric ground handling vehicle i during the period of providing service for flight j; Generate new node features by the following formula: j' (l) =LayerNorm (l) (h j (l-1) +MambaBlock (l) (h j (l-1) ;h0 (l -1) ,...,h nf (l-1) )), where h j' (l) The feature representation generated for the node corresponding to flight j in layer l, LayerNorm (l) is the layer normalization operation of the lth layer; h j (l-1) is the feature representation of the final node corresponding to flight j in the l-1th layer, MambaBlock (l) is the Mamba module of the lth layer, used for feature update, h0 (l-1) ,...,h nf (l-1) is the feature representation of all nodes in the l-1th layer; The characteristic representation is as follows: h j (l) =LayerNorm (l) (h j' (l) ), where h j (l) is the final feature representation of the node corresponding to flight j in the lth layer.
5. The method according to claim 1, characterized in that S4 includes the following sub-steps: S4.
1. Integrate the unified pooling embedding h0 from the Mamba encoder (N) , combined with the electric ground vehicle logo e i 、Previous service flights embedded h prev (N) , Remaining driving distance VD ij and the current service start time BT ij Generate comprehensive context embedding h c (N) ; S4.
2. Using multi-head attention mechanism to process context embedding h c (N) ; S4.
3. Evaluate all possible service decisions and update the state variable h (N) ; S4.
4. Calculate the selection score V using an LSTM neural network based on an attention mechanism and nonlinear transformation t (j) ; S4.
5. Check the satisfaction of the operational constraints and assign a negative infinite score -∞ to the infeasible flights; S4.
6. Use the Softmax function to convert the scores into probability distribution, form a complete scheduling strategy distribution, and select the optimal flight allocation plan; S4.
7. Use BS with rolling baseline (d i )’s reinforcement algorithm trains the policy network, optimizes the network parameters θ through the gradient update formula, and uses the verification mechanism X val Ensure continuous improvement of solution quality.
6. The method according to claim 1 or 5, characterized in that: In S4, the decoding framework of the LSTM neural network is based on a greedy strategy or a sampling strategy.