Single-line bus operation dynamic optimization method based on multi-source data

Abstract

The invention relates to the technical field of intelligent public transport systems, in particular to a single-line bus operation dynamic optimization method based on multi-source data. The method comprises the following steps: establishing a short-time public transport passenger flow prediction model based on a bidirectional long-short-time memory network, and outputting a passenger flow prediction result; establishing a single-line bus operation dynamic optimization model according to passenger flow distribution spatial-temporal characteristics of the passenger flow prediction result; and solving the single-line bus operation dynamic optimization model by adopting a hybrid algorithm, and outputting an optimal scheduling scheme of the current bus. Compared with the prior art, the method has good precision of short-time passenger flow prediction of the line level and even the station level, can effectively improve the line operation efficiency on the premise of ensuring the service level, and has excellent universal performance.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent public transportation systems, and in particular to a single-line public transportation operation dynamic optimization method based on multi-source data. Background Art

[0002] The multi-source data passenger flow prediction method is the core of the dynamic optimization of bus operation. The passenger flow prediction method of the traditional parameter model has shown fatigue when dealing with the increasing amount and increasingly complex characteristics of bus passenger flow data. Although the high-complexity hybrid prediction model has excellent prediction accuracy, the model structure and process are complex, and it is difficult to implement it in the actual operation process of the bus, and its general performance is limited. Traditional bus operation optimization is simply aimed at the control behavior of two adjacent buses, without considering that the changes in the operating status of adjacent buses will affect the overall operation of the bus line, which makes the final scheduling plan have a large optimization space from the global optimal solution.

[0003] This invention aims to analyze and integrate the multi-source data generated and related to the operation of public transportation, and propose a short-term public transportation passenger flow prediction method with reliable accuracy and strong practicality. On this basis, a single-line public transportation operation dynamic optimization model based on a combination strategy is established, and a feasible evolutionary algorithm and solution are explored. While improving the travel experience of the majority of passengers, it also provides a certain reference for solving urban public transportation problems and refining high-level operations in the future. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a single-line bus operation dynamic optimization method based on multi-source data.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] The present invention provides a single-line bus operation dynamic optimization method based on multi-source data, comprising the following steps:

[0007] S1. Establish a short-term public transport passenger flow prediction model based on a bidirectional long short-term memory network, input multidimensional passenger flow data into the short-term public transport passenger flow prediction model, and output the passenger flow prediction result;

[0008] S2. Establishing a single-line bus operation dynamic optimization model based on the temporal and spatial characteristics of passenger flow distribution of the passenger flow prediction results;

[0009] S3. A hybrid algorithm is used to solve the single-line bus operation dynamic optimization model. The solution process includes two stages:

[0010] The first stage: taking the minimum total cost of bus operation cost and passenger travel cost as the objective function of the single-line bus operation dynamic optimization model, and calculating the optimal departure interval in each period of the line;

[0011] The second stage: in the set time period before the current vehicle departure time, the short-term bus passenger flow prediction model is used to predict the passenger flow at the line station that has completed the departure time, and the prediction results and road traffic information data are input into the hybrid algorithm to output the optimal scheduling plan for the current vehicle.

[0012] The multi-dimensional passenger flow data is first pre-processed and then expanded in multiple dimensions to obtain multi-dimensional passenger flow data including coupling time characteristics, weather condition information, station information, real-time vehicle GPS, and historical passenger flow factors.

[0013] The short-term bus passenger flow prediction model is obtained by the following steps: using the holdout method to construct a data set from the original passenger flow data, and dividing it into two mutually exclusive sets, one of which is used as a training set and the other is used as a test set; using the training set to train the short-term bus passenger flow prediction model, and using the test set to verify the short-term bus passenger flow prediction model; selecting the short-term bus passenger flow prediction model with the best prediction accuracy and effect according to the average relative error, the average absolute percentage error and the relative root mean square error.

[0014] The public transportation operating cost is expressed by the following calculation formula:

[0015]

[0016] Among them, μ1 is the variable operating cost control coefficient of the bus; L is the total length of the bus route; T k is the length of the kth time period; K is the total number of scheduling time periods; x k is the departure interval of the first stop of the bus line in the kth period.

[0017] The sum of the passenger waiting cost and the passenger's time cost in transit is taken as the passenger travel cost.

[0018] The passenger travel cost is expressed by the following calculation formula:

[0019] W2=W a +W b

[0020] Among them, W a The travel cost for passengers, μ2 is the general waiting time cost control coefficient for passengers, μ3 is the waiting time cost control coefficient for passengers including satisfaction penalty, n is the total number of buses on the line, m is the total number of bus stops on the line, is the number of passengers boarding bus i when it arrives at station j, y i，j is a 0-1 decision variable, controlling whether bus i will skip the stop at station j; W b is the cost of the passenger's travel time, μ4 is the passenger travel time cost control coefficient, N i，j is the number of passengers carried in the bus compartment in real time when bus i departs from station j; l i，j is the travel path length of bus i between station j and station j+1, v i，j is the average speed of bus i between station j and station j+1.

[0021] Constraints are set for the objective function, including: vehicle overstepping constraint, station overstepping protection constraint, vehicle running speed constraint, vehicle capacity constraint, departure interval constraint and passenger service rate constraint.

[0022] The genetic algorithm is combined with the simulated annealing algorithm as a hybrid algorithm.

[0023] The hybrid algorithm solution process specifically includes: encoding the vehicle path in an integer coding manner, taking the minimum total cost of bus operating cost and passenger travel cost as the overall optimization goal, taking the inverse of the objective function as the fitness function, and terminating the hybrid algorithm when the iterative calculation reaches a set number of thresholds, and outputting the optimal scheduling plan for the current vehicle.

[0024] The optimal scheduling solution includes the current vehicle running path and vehicle overtaking control position information.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] (1) The present invention conducts a multi-level analysis of the factors affecting the short-term passenger flow of public buses. In addition to conventional factors, the public health event processing information is taken as a key factor and incorporated into the quantification process of the passenger flow prediction model.

[0027] (2) Based on the results of equivalent learning and comparative modeling, a method for predicting bus short-term passenger flow demand based on the BiLSTM network was proposed after evaluation and comparison, which achieved short-term passenger flow prediction at the route and even station level with good accuracy.

[0028] (3) Compared with the macro passenger flow prediction at the traditional line network level, the prediction results of this patent can be directly used to guide the scheduling optimization process of bus routes.

[0029] (4) In addition to conventional static scheduling optimization methods, this patent discusses in detail the application scenarios and implementation methods of dynamic bus scheduling methods represented by vehicle skipping control and local path optimization. Based on the organic combination of multiple control strategies, a single-line bus operation dynamic optimization model is constructed. While making up for the shortcomings of general scheduling methods, it also further realizes the dynamic optimization adjustment of the local reasonable range of the vehicle operation path on the basis of traditional skipping methods. The results of numerical simulation experiments show that the model can effectively improve the line operation efficiency while ensuring the service level, and has excellent general performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a schematic diagram of the steps of the present invention;

[0031] Figure 2 This is a schematic diagram of chromosome encoding;

[0032] Figure 3 (a) is a schematic diagram of random insertion perturbation. Figure 3 (b) is a schematic diagram of cross-interchange perturbation;

[0033] Figure 4 Select schematic for roulette

[0034] Figure 5 It is a schematic diagram of two-point crossover operation;

[0035] Figure 6 This is the route map of bus No. 115;

[0036] Figure 7 This is the convergence curve of the hybrid algorithm.

[0037] Figure 8 It is the convergence curve diagram of the conventional genetic algorithm;

[0038] Fig. 9 This is a distribution map of the main roads and bus stops around Longdu Shopping Mall. DETAILED DESCRIPTION

[0039] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0040] Example

[0041] This embodiment provides a single-line bus operation dynamic optimization method based on multi-source data. The method takes a single-line conventional bus as the research object, starts from the prediction process of bus passenger flow, proposes a short-term prediction method for bus passenger flow, and then proposes a combined control means for vehicle operation control based on the prediction results, establishes a single-line bus operation dynamic optimization model and designs a corresponding solution algorithm.

[0042] like Figure 1 As shown, the specific steps include:

[0043] Step 1) Establish a short-term bus passenger flow prediction model. Short-term bus station passenger flow prediction refers to the process of predicting the passenger flow of a specified bus line or a specified station in a certain period of time in the future based on historical passenger flow data, real-time traffic information, vehicle operation status, etc., with a smaller time granularity (usually days, hours, minutes, etc.). The present invention selects a recurrent neural network with good information mining capabilities for data with sequence characteristics, and completes the establishment of a short-term passenger flow prediction model and data set construction based on the historical passenger flow of the line and other related data. Through a multi-dimensional analysis of the factors affecting bus passenger flow, the construction of the training set and the test set is completed after selecting the benchmark feature parameters, and an evaluation index system for evaluating the prediction effect of the model is determined. Subsequently, the two trained short-term bus passenger flow prediction network models are used to perform passenger flow prediction under the same conditions, and the various indicators of the results are calculated to comprehensively evaluate the prediction effect. Finally, a network model with better prediction accuracy and effect is selected.

[0044] Step 1) The method for establishing the short- and medium-term bus passenger flow prediction model is as follows:

[0045] First, establish the LSTM network model, and its basic operation process can be described as:

[0046] f(t)=σ(a f h t-1 +b f x t +d f ) (1)

[0047] i(t)=σ(a i h t-1 +b i x t +d i ) (2)

[0048] u(t)=tanh(a u h t-1 +b u x t +d u ) (3)

[0049] o(t)=σ(a oh t-1 +b o x t +d o ) (4)

[0050] In the above formula, a f 、b f d f They represent h in the forget gate respectively. t-1 The weight coefficient, x t The weight coefficient and bias value of i 、b i d i They represent the update gate h t-1 The weight coefficient, x t The weight coefficient and bias value of u 、b u d u They represent the h in the feature extraction process. t-1 The weight coefficient, x t The weight coefficient and bias value of o 、b o d o They represent h in the output gate respectively. t-1 The weight coefficient, x t The weight coefficient and bias value. In addition, tanh is the standard hyperbolic tangent function, σ is the activation function Sigmoid, and the calculation method is:

[0051]

[0052] At time t, the unit state c transmitted by the system t and the hidden state h t The calculation method is:

[0053] c t =c t-1 ⊙f(t)+i(t)⊙u(t) (7)

[0054] h t =o(t)⊙tanh(c t ) (8)

[0055] Where ⊙ is the Hadamard product.

[0056] LSTM network model input layer construction:

[0057] The input data of the model is pre-normalized time series multi-source data. The original passenger flow data only contains the mapping relationship between time and vehicle, which makes it difficult to quantify the relationship between passenger flow and multi-dimensional traffic factors. After data preprocessing, the original passenger flow data is dimensionally expanded to obtain multi-dimensional passenger flow data including coupling time characteristics, weather conditions, station information, real-time vehicle GPS, historical passenger flow and other factors.

[0058] LSTM network model output layer construction:

[0059] Responsible for the output process of the network model prediction results. The prediction value of this model is the passenger flow data of the bus stops along the route at a certain time. The data dimension is 1, so the number of neurons in the output layer is set to 1.

[0060] Then, based on the established LSTM model, a layer of reverse LSTM units is added to the hidden layer. A bidirectional long short-term memory network (BiLSTM) is constructed. It is a network structure obtained by compounding two standard LSTMs in opposite directions. Based on the sequence direction of the original data, the one that processes the input sequence in sequence is called the forward LSTM, and the one that processes it in reverse order is called the reverse LSTM. The final output result of the BiLSTM is obtained by splicing the two groups of sub-result vectors together through linear fusion.

[0061] At time t, one of the LSTM networks is defined as a forward network, and the unit state of the network is recorded as The hidden state is recorded as The computational process within the network is the same as that of the standard LSTM network, expressed as:

[0062]

[0063]

[0064] Similarly, another LSTM network is defined as a reverse network. At the same time, the unit state of the network is recorded as The hidden state is recorded as It is expressed as:

[0065]

[0066] The final output of the BiLSTM network is to combine the operation results of the forward and reverse sub-networks through linear fusion and other methods to obtain the unit state c of the entire network. t and the hidden state h t .

[0067]

[0068] G(x) is a combinatorial function.

[0069] The bidirectional loop structure design of the BiLSTM network optimizes the learning weakness of ordinary LSTM and can better mine the correlation features within the data sequence.

[0070] BiLSTM network model input layer construction:

[0071] The input data is the same as that of the LSTM model, which is multi-dimensional passenger flow data coupled with time characteristics, weather conditions, station information, real-time vehicle GPS and other factors.

[0072] BiLSTM network model output layer construction:

[0073] The prediction value of this model is also the passenger flow data of the bus stops along the route at a certain moment. Before using the established network model for passenger flow prediction, the network needs to be trained and learned. This embodiment uses the holdout method to divide the original data set into two mutually exclusive sets, one of which is used as a training set and the other as a test set. Considering that the prediction result of a single use of the holdout method may be insufficient in stability and reliability, methods such as adjusting the division ratio and multiple averages are further used to ensure the training effect.

[0074] Finally, in order to more intuitively and accurately verify the prediction effects of different models, relevant performance indicators are selected for calculation and evaluation. Considering the convenience, practicality, comprehensiveness and objectivity of the calculation, the mean relative error (MAE%), mean absolute percentage error (MAPE), and relative root mean square error (RMSE%) are selected for result evaluation.

[0075] 1) MAE%

[0076] Mean absolute error (MAE) is a common indicator to measure the absolute deviation between the predicted value and the actual value, and is used to judge the accuracy of the model prediction. Due to the large volatility of passenger flow values, in order to ensure the consistency and intuitiveness of error measurement, the ratio of MAE to average passenger flow demand MAE% is used as the evaluation indicator. The calculation method is:

[0077]

[0078] Where yi is the predicted value and y is the true value.

[0079] 2) Mean Absolute Percentage Error (MAPE)

[0080] MAPE is another common key performance indicator used to evaluate forecast accuracy. MAPE is calculated in a way that it does not change due to global scaling of the target variable, which is particularly suitable for dealing with problems where variables of different scales exist at the same time. Also due to this calculation method, the indicator is very sensitive to higher errors when the true value is low. The calculation method is:

[0081]

[0082] 3) Relative root mean square error (RMSE%)

[0083] Root mean square error (RMSE), also known as standard error, can measure the degree of dispersion of the predicted value. Similar to MAE, the ratio of RMSE to average passenger flow demand, RMSE%, is also used as an evaluation indicator. The calculation method is:

[0084]

[0085] Step 2) Based on the sequential passenger flow prediction obtained in step 1, the predicted passenger volume (persons / times) is obtained according to the prediction results, and the spatiotemporal characteristics of passenger flow distribution (including the overall trend of passenger flow, peak station locations and peak time) are further obtained to establish a single-line bus operation dynamic optimization model with combined scheduling means including vehicle overtaking control and vehicle operation path optimization.

[0086] The established model first stipulates that the following assumptions are true:

[0087] (1) During the driving process, buses will not experience abnormal speed reduction, stop or other emergencies due to the influence of other vehicles or non-motor vehicles on the route, but the average driving speed will be reduced due to congestion on the driving section.

[0088] (2) Before buses depart from their first stop, they can receive dispatch plans from the dispatch center. Drivers are familiar with road conditions and can accurately understand and implement vehicle skipping and route adjustment strategies.

[0089] (3) Before a bus arrives at a bus stop, the stop is vacant and no bus from other routes is occupying it.

[0090] (4) When a bus passes through a road intersection, the average waiting time of the node in different time periods is used to replace the passing process, without considering the specific effects of traffic intersection, signal phase, etc.

[0091] (5) Passenger arrivals at bus stops approximately follow a Poisson distribution.

[0092] (6) Public buses will depart on time according to the dispatch plan, and the departure frequency will remain consistent during the same operating period.

[0093] (7) During the vehicle path optimization process, the urban roads corresponding to the generated feasible paths all allow buses to pass through without any width, height, or speed limit requirements.

[0094] In the process of establishing the single-line bus operation dynamic optimization model, the relevant variable parameters and their meanings are shown in Table 1.

[0095] Table 1 Parameter definition table

[0096]

[0097]

[0098] The proposal and establishment of a single-line bus operation dynamic optimization model aims to balance the operating costs of bus operators and the travel costs of passengers through a reasonable bus dispatching plan, improve the line operation efficiency while ensuring the passenger riding experience, and find a balance of interests between the two parties. Therefore, the objective function of the single-line bus operation dynamic optimization model is to minimize the bus operation cost and the passenger travel cost. Both are used as optimization functions for dual-objective optimization modeling and solution. The decision variables of the model include bus departure intervals, bus passing station locations, and operation path optimization interval distribution.

[0099] (1) Bus operating costs

[0100] The operating costs of bus operators include fuel consumption, wage expenditures, asset depreciation, etc., which are closely related to the departure interval and operating mileage during bus operation. As an intuitive feedback method for evaluating the effectiveness of the implementation of the scheduling plan, it can be used as an objective function.

[0101] The calculation method of bus operator's operating cost can be expressed as:

[0102]

[0103] In the above formula:

[0104] μ1: variable operating cost control coefficient of public transport vehicles;

[0105] L: the total length of the bus route;

[0106] T k : The length of the kth period;

[0107] K: total number of scheduling time periods;

[0108] x k : The first stop departure interval of the bus line in the kth period.

[0109] (2) Passenger travel costs

[0110] The processes directly affected by the change of vehicle scheduling scheme include the waiting process after arriving at the station and the vehicle operation process after boarding. Therefore, the model of this embodiment focuses on the passenger waiting time cost and the on-the-way time cost generated in the above two processes.

[0111] The cost of passenger waiting time is linearly correlated with the time interval between the time when the passenger arrives at the station and the time when the next vehicle arrives. The longer the departure interval of the line, the longer the average waiting time of passengers at the station, and the higher the waiting cost. Especially when the vehicle implements the over-station control strategy, the passengers at the over-station must continue to wait for the arrival of the next vehicle. The model assumes that the arrival process of passengers at each station follows the Poisson distribution, and the passenger waiting cost can be expressed as:

[0112]

[0113] In the above formula: μ2: passenger general waiting time cost control coefficient;

[0114] μ3: passenger waiting time cost control coefficient including satisfaction penalty;

[0115] n: total number of buses on the route;

[0116] m: total number of bus stops on the route;

[0117] The number of passengers boarding bus i when it arrives at stop j;

[0118] y i，j : 0-1 decision variable, controlling whether bus i will skip the stop at station j.

[0119] The cost of passenger travel time refers to the travel cost from the time the passenger gets on the bus to the time the passenger gets off the bus. The higher the average speed of the bus, the shorter the time the passengers spend on the bus, and the lower the corresponding time cost. When implementing the over-stop control strategy, since the bus does not need to slow down and stop and start the acceleration process at some stations, it can maintain the driving speed and pass directly. For the passengers on the bus, the travel time is shorter than the conventional uncontrolled time, and the time cost will also be reduced accordingly.

[0120] The calculation method of passenger travel time cost is:

[0121]

[0122] In the above formula: μ4: passenger travel time cost control coefficient;

[0123] N i，j : The number of passengers carried in the bus compartment in real time when bus i departs from station j;

[0124] l i，j : The travel path length of bus i between station j and station j+1;

[0125] v i，j : The average speed of bus i between station j and station j+1.

[0126] In summary, the total cost of passenger travel can be expressed as:

[0127]

[0128] In order to be close to the actual operation scenario, ensure the normal and efficient operation of the line vehicles while avoiding excessive skipping of stations which may cause the normal riding needs of passengers to be unable to be guaranteed, the constraints are set as follows:

[0129] (1) Vehicle overstepping restrictions

[0130] For any bus i, it is not allowed to skip two bus stops consecutively during driving, and the total number of skipped stops in a single operating trip shall not exceed 5 times.

[0131]

[0132] (2) Site Cross-site Protection Constraints

[0133] For any bus stop j, two consecutive arriving vehicles are not allowed to pass through the stop.

[0134]

[0135] (3) Vehicle speed constraints

[0136] During the entire operation process of the bus, the vehicle's driving speed must meet the speed limit requirements of the road section along the route.

[0137]

[0138] (4) Vehicle capacity constraints

[0139] Public transport vehicles must not be overloaded, that is, at any time, the number of passengers in the vehicle must not exceed the vehicle's approved passenger capacity N. max .

[0140]

[0141] (5) Departure interval constraints

[0142] The bus operator has a limited number of vehicles and drivers, so the bus departure interval cannot be too small to avoid unavailability. At the same time, in order to reflect the public service attributes, the departure interval should not be set too large just for cost reasons, causing passengers to wait for a long time during off-peak hours.

[0143] x min ≤x k ≤x max (27)

[0144] (6) Passenger service rate constraints

[0145] The proportion of passengers who are unable to get on or off the bus normally due to over-scheduling of any bus i shall not exceed the prescribed upper limit.

[0146]

[0147] in, is the number of waiting passengers for bus i at stop j.

[0148] Step 3) According to the model characteristics and complexity, taking into account both the speed of solution and the quality of solution, a hybrid genetic algorithm integrating the idea of ​​simulated annealing is proposed and a specific solution is designed.

[0149] Specifically, according to the basic principles of genetic algorithm and simulated annealing algorithm, the hybrid algorithm is designed for the scenario-oriented and bus operation dynamic optimization model of this embodiment to solve the model. The whole solution process is divided into two stages: in the first stage, the characteristic information of the operation day (i.e., multi-dimensional passenger flow data) is input into the trained BiLSTM network to predict the total passenger flow of the line during the operation period, and the sum of the operating cost and the passenger waiting cost is minimized as the stage objective function to calculate the optimal departure interval of each period of the line; in the second stage, 10 minutes before the departure time of each vehicle, the passenger flow prediction model is used to complete the line station passenger flow prediction at the departure time, and the prediction results and road traffic information data are input into the hybrid algorithm. The vehicle path processed by integerization is encoded as the chromosome gene, the passenger travel cost and the operator's operating cost are minimized as the total optimization goal, and the inverse of the objective function is used as the fitness function to iteratively calculate the feasible solution set. After reaching the maximum number of iterations, the maximum iteration of this embodiment is 300 times. Output the optimal scheduling plan including the current vehicle travel path, vehicle crossing control position and other information.

[0150] In the second stage of solving, the main steps and processes of the hybrid algorithm include:

[0151] (1) Chromosome encoding

[0152] The vehicle path is encoded using integer coding. Each gene position corresponds to each station on the bus line, and each chromosome contains the complete path of the vehicle in the corresponding time period. Figure 2 As shown in Figure 2, when the vehicle runs in one direction along the initial route and does not pass any bus stop, the path is represented by When the vehicle passes through one of the stations (such as the 3rd station), the gene bit value of the station is 0, and the path can be expressed as The chromosome encoding of the initial population can be obtained as Where n is the total number of chromosomes in the initial population.

[0153] (2) Population initialization

[0154] The population initialization of the genetic algorithm can be performed in many ways. The common practice is to use random generation to generate random points of the same size as the population. Due to the complexity of the optimization model, the distribution of the initial points obtained by this random generation method in the search space has great uncertainty, which can easily cause the algorithm to fall into the local optimum too early. When the quality of the initial solution is average, it will also greatly extend the solution time. Therefore, the Tent chaos mapping method is used to homogenize the distribution of initial values.

[0155] Tent chaos mapping is a one-dimensional discrete segmented mapping method based on chaos theory. The function graph presents a "tent" shape, with a relatively uniform distribution function and good correlation. The expression of this method is:

[0156]

[0157] Generally, 0<α<1, but when α<0.5, the system periodic state will converge quickly, resulting in unsatisfactory results. Therefore, α=0.75 is taken, and the chaotic components of each optimization parameter are calculated according to the above formula to initialize the population.

[0158] (3) Fitness function

[0159] The fitness function is the key to evaluating the quality of each solution, and directly affects which solutions in each generation can be inherited to the next generation. Since the purpose of this model is to find the scheduling solution with the lowest total cost in the set of feasible solutions, the larger the objective function value W(x) of a solution, the smaller the probability of the solution being selected should be, that is, the lower the fitness should be. Therefore, the reciprocal of the objective function is used as the fitness function Fit(x). For inferior solutions that directly violate the constraints, the penalty function P(x) is set to a sufficiently large value, which greatly increases the probability of the solution being eliminated in the next round of inheritance; otherwise P(x) = 0.

[0160] Fit(x) is calculated as:

[0161]

[0162] (4) Simulated annealing perturbation

[0163] The perturbation operation in the simulated annealing process includes the insertion of new cross-site sites and the cross-exchange of cross-site sites in the original path. Figure 3 (a) represents the perturbation process of random insertion, selecting any site in the parent chromosome whose gene bit value is not 0 and assigning it a value of 0; Figure 3 (b) represents the perturbation process of crossover, where two adjacent gene positions in the parent chromosome are selected for coding exchange. Two perturbation operations are performed at the same time, and the other gene positions that are not perturbed are retained in the offspring according to the coding order of the original parent chromosome.

[0164] (5) Selection, crossover, and mutation

[0165] The roulette wheel method based on the elite selection strategy is used for operator selection. The roulette wheel strategy converts the fitness value of an individual into a selection probability, and then selects the individual based on this selection probability. The probability of each individual being selected is the individual's fitness value divided by the sum of the fitness values ​​of all individuals. The principle is as follows: Figure 4 As shown. Then, a random number in the range of [0, 1] is generated and the corresponding individuals are selected according to the corresponding range. The operation is repeated until the number of selected individuals reaches the preset value.

[0166] If the population size is n, the fitness of individual i is Fit(x i ), then the probability of the individual being selected can be expressed as:

[0167]

[0168] Two-point crossover is used to achieve the exchange of parent chromosome segments, such as Figure 5 As shown in the figure. The crossover operation enables the algorithm to search a wider range in the search space, which helps to find a better solution. First, two crossover gene sites are randomly selected, and the two parent individuals are cut at these two crossover points. Then, the two cut fragments are interchanged and fused to obtain new offspring individuals. The fragment length and cutting position are determined by the selected site, so they are also random, which increases the evolutionary direction as much as possible while retaining the advantages of the parent generation.

[0169] Common genetic mutation operations include point mutation, insertion, deletion, exchange, inversion, etc. This study uses a mutation operation based on adaptive mutation rate. Select a certain number of offspring individuals from the new population generated by crossover, and calculate the mutation probability of each individual according to the method in formula (32). Suppose the fitness of individual i in the offspring is fit(i), and the average fitness of all individuals is fit ave , the fitness of the optimal individual is fit max , then the probability of adaptive mutation is:

[0170]

[0171] ω1 is the probability control coefficient to avoid excessive variation.

[0172] For individuals with confirmed mutations, two gene positions in the fragment are randomly selected to generate mutations. The mutation rule is: when the original gene position value is 0, the gene position value after mutation is 1; when the original gene position value is 1, the value after mutation is 0.

[0173] (6) Algorithm termination condition

[0174] When the hybrid algorithm reaches the upper limit of the specified number of iterations, the algorithm terminates.

[0175] In order to verify the effect of the method provided in this embodiment, a calculation example is designed to verify the model of the single-line bus operation dynamic optimization method based on multi-source data and its solution algorithm and analyze the effectiveness of the algorithm, including the following stages:

[0176] 1) Example description

[0177] According to the technical solution of the present invention, and in combination with Figure 6 The bus route of Route 115 is shown in the figure, and a bus optimization example based on multivariate data is designed. The basic data used in the example comes from the actual vehicle operation data, passenger flow data, on-board real-time GPS data, driving logs, and urban road network structure and coordinates of Route 115 provided by Ganzhou Public Transport Corporation. The urban road network structure and coordinate data are combined with the positioning information of Amap for composite calibration to finally obtain the GPS coordinates of the bus station, the route coordinates, and the network coordinates of the surrounding main road nodes.

[0178] 2) Test data

[0179] The road network traffic condition information and key traffic parameters are determined from the full-day traffic condition monitoring data of Ganzhou City’s urban roads on March 28, 2022 provided by the Ganzhou City Comprehensive Transportation Service Center.

[0180] Table 2 Coordinates of bus stops and the distance between bus stops on route 115

[0181]

[0182]

[0183] 3) Parameter settings

[0184] Combining the existing research results of scholars in the field and the preliminary experimental results of test cases, in order to take into account the algorithm solution speed, stability and feasibility of the searched optimal solution, the key parameter settings of the hybrid algorithm proposed in this paper are shown in Table 3.

[0185] Table 3 Algorithm parameter settings

[0186]

[0187] The parameters involved in the establishment and solution of other models are set as shown in Table 4, based on relevant literature and actual investigation and research data on public transportation in Ganzhou City.

[0188] Table 4 Other related parameter settings

[0189]

[0190]

[0191] 4) Analysis of example results

[0192] Use Python 3.9 environment to program the hybrid algorithm and solve the bus operation dynamic optimization model. The convergence curve of the fitness function is as follows: Figure 7 As shown. A conventional genetic algorithm is also used for comparison and solution, and the encoding method and related parameters are consistent with the hybrid algorithm. Compared with the hybrid algorithm, the conventional genetic algorithm does not perform operations such as simulated annealing perturbation and probability acceptance of solutions. The convergence curve of its fitness function is shown in Figure 8 shown.

[0193] Comparing the two figures, it can be seen that within the maximum number of iterations, both algorithms can iterate to obtain the optimal solution of the problem. From the perspective of iteration speed, the initial convergence speed of the hybrid algorithm is significantly improved compared with the conventional genetic algorithm. The objective function value decreases rapidly within 30 generations, and then the convergence rate decreases. It converges to the optimal value at about 200 generations; the conventional genetic algorithm converges to the optimal value at about 180 generations, and the convergence speed is relatively stable throughout the process. From the perspective of solution quality, the optimal solution obtained by the conventional genetic algorithm is 8.81% higher than that of the hybrid algorithm, indicating that the simulated annealing perturbation operation can actively accept some inferior solutions, and after crossover mutation, the algorithm can get rid of the interference of the local optimal solution to a certain extent, thereby obtaining a better evolutionary direction. At 50 generations, the solution obtained by the hybrid algorithm is better than the final optimization result of the conventional genetic algorithm, proving that the algorithm has good practicality in solving such problems. In addition, although the perturbation operation of simulated annealing helps the algorithm to decrease rapidly in the early stage, it will also slow down the search speed of the optimal solution in the later stage due to the generation and addition of inferior solutions, resulting in a certain lag in the convergence point of the optimal solution.

[0194] In the optimal dispatching scheme calculated by the hybrid algorithm, the line has 65 daily departures and a total cost of 58,144.2 yuan. Taking the uplink direction as an example, the final calculated vehicle time-sharing departure interval results and the optimal vehicle dispatching scheme output by the hybrid algorithm are converted into the forms shown in Tables 5 and 6. The service rate is defined as: the proportion of passengers who can travel normally according to the original expected travel time, boarding stations and alighting stations during the full operation of a certain bus, and who are unable to board the bus immediately or get off at the station due to the vehicle crossing the station, to the total number of passengers. This indicator can intuitively reflect the availability of the dispatching scheme, and specific restrictions have been imposed in the constraints of the model to avoid vehicles abandoning a large number of passengers in order to achieve lower costs.

[0195] Table 5 Vehicle departure intervals before and after optimization

[0196]

[0197] As can be seen from the table above, the average departure interval of the line is about 13 minutes throughout the day, the average departure interval during the morning peak period (7:00-9:00) is about 7 minutes, and the average departure interval during the evening peak period (17:00-19:00) is about 8 minutes. Compared with the original departure interval, the optimized departure interval is more consistent with the time distribution of passenger flow on this line. The departure interval during the peak passenger flow period is further shortened, which helps to improve the peak capacity and quickly relieve the passenger flow waiting for the platform. The appropriate extension of the departure interval during the off-peak period can reduce the cost of vehicle idle driving, save capacity resources, and is also conducive to the handover work such as crew shift change during the off-peak period.

[0198] In addition, since the calculation results of the time-sharing intervals at this stage are only directly related to the total daily passenger flow of the line, they are not affected by the feedback of the subsequent vehicle operation optimization results. The passenger flow level of bus lines generally has a stable level and similar characteristics over a long period of time, so the time-sharing departure interval scheme is less sensitive to a specific date. After taking the passenger flow forecast data of other dates from March 25 to March 30, 2022 for calculation, the results also show that the time-sharing departure intervals have differences in performance between different groups of forecast day attributes (weekdays and weekends), while the performance within the group is basically the same, which proves that the scheme has good stability and redundancy. This means that in actual operation scenarios, the operator only needs to formulate two different departure interval arrangements for weekdays and weekends, and make monthly or even quarterly adjustments based on operating conditions, historical passenger flow and model prediction results. For passengers, it can also effectively avoid the inconvenience of waiting for actual travel caused by the short time scale of dynamic adjustment of departure times.

[0199] Table 6 The best vehicle dispatching scheme obtained by hybrid algorithm

[0200]

[0201]

[0202]

[0203] Combining the two tables, we can see that:

[0204] From the perspective of the route, the total passenger flow of the route is predicted to be 5,980 per day, and the total passenger flow of the actual service plan is 5,919, with a comprehensive service rate of 98.98% for the whole day, which meets the model constraints and the actual operation scenario requirements. From the perspective of the vehicle, the average speed of the original dispatching plan is 19.7km / h for the whole day, and the average speed during the peak period is 18.1km / h; the average speed of the optimized dispatching plan is 21.6km / h for the whole day, and the average speed during the peak period is 20.3km / h, which are increased by 9.64% and 12.15% respectively, indicating that the operation efficiency of the bus has been improved. From the perspective of passengers, the average waiting time for passengers throughout the day is shortened by about 11.07%, and the average waiting time for passengers during the peak period is shortened by about 22.15%. In addition, due to the increase in the speed of buses, the time for passengers in transit will also be shortened accordingly, further reducing the overall travel time. On the other hand, the implementation of dynamic departure intervals in different time periods will also extend the waiting time for passengers on off-peak routes. For example, in the calculation results, the proportion of affected passengers to the total passengers is 17.64%, and the average waiting time is extended by about 20.85%. This will lead to a decrease in the willingness of passengers to choose to take this line during the off-peak period, resulting in a decrease in passenger flow during this period.

[0205] In addition, under the premise of short departure intervals, the implementation of the stop-over control strategy will cause uneven distribution of the arrival times of adjacent buses, such as the difference of 2 minutes between the final arrival time of buses 13 and 12, and the difference of 3 minutes between the final arrival time of buses 15 and 16, and buses 16 and 17. This shows that the vehicle stop-over will cause a certain degree of decrease in the uniformity of the line headway, which may interfere with the waiting process of passengers at the station in the actual operation scenario, but the degree is limited when the departure interval is small during peak hours.

[0206] In order to further analyze and evaluate the specific effect of the vehicle control strategy proposed in this embodiment, the specific dispatching scheme of some buses that have skipped stations or adjusted routes in Table 5.5 is expanded as shown in Table 7. Among them, "△" means that the vehicle skips the station at this station, passes but does not stop at the station, and the running path in the previous and next intervals remains unchanged; "▲" means that the vehicle adjusts the path at this station, that is, a new running path is selected in the interval between the previous station and the next station, and the station is not passed.

[0207] Table 7 Vehicle operation plan for dynamic dispatching

[0208]

[0209]

[0210] Analyzing the data in the above table, we can draw the following conclusions:

[0211] (1) The trains that implemented the skipping control strategy were concentrated in the morning and evening peak hours, and were not implemented in the off-peak hours. This is mainly because the departure intervals during the off-peak hours are large. If a vehicle skips a station, the passengers at that station will incur high waiting time costs and penalty costs. During peak hours, the routes have dense departures, and after a vehicle skips a station, the next vehicle will arrive soon. The total cost increase is limited and easy for passengers to accept. As mentioned earlier, skipping will increase the average speed of buses and has the ability to avoid serious traffic congestion points on the route. Therefore, for passengers, spending extra waiting time does not necessarily mean that the overall travel time will increase. For example, in the scheduling plan actually implemented on the route that day, the bus that departed at 18:10 departed 6 minutes earlier than the optimization plan proposed in this study, but the final arrival time was delayed by 3 minutes (Table 8).

[0212] Table 8 Comparison of vehicle operating parameters before and after optimization

[0213]

[0214] Although this result is based on model simulation and cannot fully restore the actual scene and complex conditions of vehicle operation, it is not difficult to see that if adjacent trains pass each other continuously and the number of times the vehicles pass each other at stations along the route is sufficient, the reduction in passengers' travel time will be greater than the additional waiting time at the boarding station, which will eventually be reflected in the reduction of overall travel time. Of course, whether it is realistic to implement this "quasi-customized bus" scheduling method in regular bus routes requires more in-depth and specific discussion and research, and it must be used in combination with perfect and efficient multi-channel information release methods such as electronic bus stops and online platforms.

[0215] (2) The stations where vehicles skip the station mainly include the Turbine Factory, Ganzhou Medical School, Longdu Mall, Livestock Farm, Tandong Town Government, No. 4 Middle School, Lighting City, Yunlong Garden, etc. A comprehensive analysis of the station location relationship and passenger flow characteristics shows that these stations can be roughly divided into three categories: the first is stations with sparse overall passenger flow, such as the Livestock Farm Station and Lighting City Station. There are no residential communities and work areas around the stations, and there is almost no commuter flow, and the passenger flow has been maintained at a low level for a long time; the second is stations with acceptable total passenger flow, but the daily distribution of passenger flow is closely related to the time period, showing time-concentrated characteristics, such as the No. 4 Middle School Station, which mainly serves student passenger flow, and passenger flow is mostly generated during school hours; the third is stations with small station spacing and poor traffic conditions, such as the Tandong Town Government Station. This station is only about 360 meters away from the Tandong Farmers' Market Station, and the T-shaped intersection between the two stations has poor traffic conditions during peak hours. Traffic survey results show that many passengers choose to get off at the front station and walk to their destination. Therefore, the dispatch plan generated by this model can not only improve the operating efficiency of public buses, but decision makers can also discover problems in the layout of line stations through the frequency of each station being bypassed in the plan, providing reference for the relocation of original stations and the establishment of new stations.

[0216] (3) The two stations where the vehicle routes are adjusted are Longdu Mall and Yunlong Garden. The situation around Longdu Mall is as follows: Fig. 9 As shown. Taking Longdu Mall Station as an example, on the one hand, the road network near the station is densely developed, and there are many routes to choose from, such as Harmonious Avenue-Rongjiang Fourth Road-Beijing-Macao Line (original route), Harmonious Avenue-Ruilin Road-Beijing-Macao Line, Harmonious Avenue-Jingming Road-Beijing-Macao Line, etc. The two sides of Jingming Road and Ruilin Road are mostly urban areas under construction and residential communities, with small traffic flows and good road and intersection conditions. During peak hours, the original route can avoid congestion time on urban trunk roads (Harmonious Avenue) and large intersections (Harmonious Avenue-Rongjiang Fourth Road, etc.). On the other hand, the original route passes through the exit of the underground garage in the commercial district and a primary school. The parking space at the two places is small. When the traffic volume is large, illegal parking on the roadside often leads to road blockage and poor traffic conditions. When the passenger flow level is low, the vehicle chooses to skip the station and detour to other routes, which will reduce the total cost.

[0217] However, the side door of the original line station of Yunlong Garden Station was temporarily closed due to the needs of community management. Community residents are more willing to choose the bus route through the main gate exit when traveling, and the passenger flow level of the station has dropped significantly. In addition, the original line passes through many intersections of Jingan North Avenue and the waiting phase time is long, so vehicles tend to choose to skip the station during the calculation process and go directly to the next station via Taikang East Avenue-Furong North Avenue-Gannan Avenue. Therefore, the scheduling plan results of the model in this embodiment can also help decision makers identify difficult sections of the line, and can be used as a theoretical basis for bus operators when optimizing existing line routes and laying new lines.

[0218] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A method for dynamic optimization of single-line bus operation based on multi-source data, characterized in that: The following steps are involved: S1. Establish a short-term public transport passenger flow prediction model based on a bidirectional long short-term memory network, input multidimensional passenger flow data into the short-term public transport passenger flow prediction model, and output the passenger flow prediction result; S2. Establishing a single-line bus operation dynamic optimization model based on the temporal and spatial characteristics of passenger flow distribution of the passenger flow prediction results; S3. A hybrid algorithm is used to solve the single-line bus operation dynamic optimization model. The solution process includes two stages: The first stage: taking the minimum total cost of bus operation cost and passenger travel cost as the objective function of the single-line bus operation dynamic optimization model, and calculating the optimal departure interval in each period of the line; The second stage: in the set time period before the current vehicle departure time, the short-term bus passenger flow prediction model is used to predict the passenger flow at the line station that has completed the departure time, and the prediction results and road traffic information data are input into the hybrid algorithm to output the optimal scheduling plan for the current vehicle.

2. According to claim 1, a method for dynamic optimization of single-line bus operation based on multi-source data is characterized in that: The multi-dimensional passenger flow data is first pre-processed and then expanded in multiple dimensions to obtain multi-dimensional passenger flow data including coupling time characteristics, weather condition information, station information, real-time vehicle GPS, and historical passenger flow factors.

3. The method for dynamic optimization of single-line bus operation based on multi-source data according to claim 1, characterized in that: The short-term bus passenger flow prediction model is obtained by the following steps: using the holdout method to construct a data set from the original passenger flow data, and dividing it into two mutually exclusive sets, one of which is used as a training set and the other is used as a test set; using the training set to train the short-term bus passenger flow prediction model, and using the test set to verify the short-term bus passenger flow prediction model; selecting the short-term bus passenger flow prediction model with the best prediction accuracy and effect according to the average relative error, the average absolute percentage error and the relative root mean square error.

4. The method for dynamic optimization of single-line bus operation based on multi-source data according to claim 1, characterized in that: The public transportation operating cost is expressed by the following calculation formula: Among them, μ1 is the variable operating cost control coefficient of the bus; L is the total length of the bus route; T k is the length of the kth time period; K is the total number of scheduling time periods; x k is the departure interval of the first stop of the bus line in the kth period.

5. The method for dynamic optimization of single-line bus operation based on multi-source data according to claim 1, characterized in that: The sum of the passenger waiting cost and the passenger's time cost in transit is taken as the passenger travel cost.

6. The method for dynamic optimization of single-line bus operation based on multi-source data according to claim 5, characterized in that: The passenger travel cost is expressed by the following calculation formula: W2=W a +W b Among them, W a The travel cost for passengers, μ2 is the general waiting time cost control coefficient for passengers, μ3 is the waiting time cost control coefficient for passengers including satisfaction penalty, n is the total number of buses on the line, m is the total number of bus stops on the line, is the number of passengers boarding bus i when it arrives at station j, u i,j is a 0-1 decision variable, controlling whether bus i will skip the stop at station j; W b is the cost of the passenger's travel time, μ4 is the passenger travel time cost control coefficient, N i,j is the number of passengers carried in the bus compartment in real time when bus i departs from station j; l i,j is the travel path length of bus i between station j and station j+1, v i,j is the average speed of bus i between station j and station j+1.

7. The method for dynamic optimization of single-line bus operation based on multi-source data according to claim 1, characterized in that: Constraints are set for the objective function, including: vehicle overstepping constraint, station overstepping protection constraint, vehicle running speed constraint, vehicle capacity constraint, departure interval constraint and passenger service rate constraint.

8. The method for dynamic optimization of single-line bus operation based on multi-source data according to claim 1, characterized in that: The genetic algorithm is combined with the simulated annealing algorithm as a hybrid algorithm.

9. The method for dynamic optimization of single-line bus operation based on multi-source data according to claim 8, characterized in that: The hybrid algorithm solution process specifically includes: encoding the vehicle path in an integer coding manner, taking the minimum total cost of bus operating cost and passenger travel cost as the overall optimization goal, taking the inverse of the objective function as the fitness function, and terminating the hybrid algorithm when the iterative calculation reaches a set number of thresholds, and outputting the optimal scheduling plan for the current vehicle.

10. The method for dynamic optimization of single-line bus operation based on multi-source data according to claim 9, characterized in that: The optimal scheduling solution includes the current vehicle running path and vehicle overtaking control position information.

Images