Rail transit-oriented digital scheduling generation method based on general-purpose algorithm

Abstract

The application belongs to the technical field of intelligent scheduling, and particularly relates to a digital scheduling generation method for rail transit based on a general algorithm device, which comprises the following steps: extracting the time period density index, online driver position distribution entropy, task load degree and vehicle turnover deviation degree of a current time window, and standardizing to form an operation state characteristic vector; dynamically adjusting the weight according to the time period density index to retrieve a historical scheduling segment by weighted similarity; screening and mapping the fatigue index based on the fusion of continuous driving saturation, cumulative load degree and rest quality coefficient to the secondary qualification to an initial template; identifying conflicts along the driver task chain and the vehicle connection chain, constructing a dependency level, adaptively adjusting the beam width to eliminate fatigue over-limit operation layer by layer, and excluding fatigue over-limit operation until a scheduling scheme meeting the hard constraint is generated or a time limit is reached; and the scheduling scheme can meet the requirements of both transport capacity and active fatigue safety protection.

Description

Technical Field

[0001] This invention relates to the field of intelligent scheduling technology. More specifically, this invention relates to a digital scheduling generation method based on a general algorithm for rail transit. Background Technology

[0002] In rail transit operations, when unexpected events such as additional trains, line interruptions, or large-scale deviations from the planned schedule occur, extremely high demands are placed on the real-time generation of driver and conductor scheduling plans.

[0003] In existing technologies, dispatchers typically rely on manual adjustments or global optimization algorithms for scheduling. Manual adjustments are limited by the dispatcher's ability to coordinate multi-dimensional constraints, resulting in low efficiency and a high risk of oversights. On the other hand, global optimization methods such as genetic algorithms and integer programming require searching within a vast combination space consisting of driver task chains and vehicle connection chains, leading to exponentially increasing computation time and making it difficult to meet the real-time decision-making requirements within minutes.

[0004] To accelerate schedule generation, the industry has introduced template reuse methods based on matching historical schedule records. These methods calculate the similarity between the current operational status and historical schedule segments to retrieve reusable initial templates, thereby lowering the search threshold.

[0005] However, in similarity measurement, existing methods generally assign pre-set fixed weights to features representing different dimensions such as task density, driver spatial distribution, capacity supply and demand, and vehicle turnover stability, without considering the differences in the impact of each feature on the scheduling structure under different operating scenarios.

[0006] Especially during high-density periods, even small changes in task intensity can have a decisive impact on scheduling schemes. Fixed-weight similarity calculations will result in insufficient matching accuracy between retrieved historical templates and the current real-time status, introducing a large number of unnecessary hard constraint conflicts and increasing the complexity and time consumption of subsequent corrections.

[0007] In addition, existing scheduling methods for driver fatigue management mostly rely on single-dimensional regulatory thresholds such as the maximum continuous driving time for compliance verification, lacking the integration and quantification of multi-dimensional fatigue factors such as the cumulative driving load and the effect of rest and recovery.

[0008] This results in a scheduling scheme that, while formally meeting legal requirements, fails to effectively mitigate the risk of drivers actually being in a state of high fatigue.

[0009] At the same time, when historical templates are directly mapped to currently online driver and vehicle resources, a large number of interdependent hard constraint conflicts will arise along the newly constructed driver continuous task chain and vehicle turnaround connection chain, such as continuous driving timeout, insufficient rest interval, and unqualified qualifications.

[0010] Faced with such high-dimensional coupling constraints, traditional conflict resolution methods based on rule exhaustion or simple reasoning are prone to combinatorial explosion, making it difficult to ensure the production of a feasible scheduling scheme that satisfies all hard constraints within a strict decision-making time limit. Summary of the Invention

[0011] To address the aforementioned technical problems, the present invention provides solutions in the following aspects.

[0012] In a first aspect, the present invention provides a digital scheduling generation method for rail transit based on a general algorithm, employing the following technical solution:

[0013] Extract the time period density index representing the intensity of tasks, the online driver location distribution entropy representing the spatial distribution of drivers, the task load representing the supply and demand of transportation capacity, and the vehicle turnover deviation representing the stability of vehicle operation within the current time window, and form a real-time operation status feature vector after standardization.

[0014] The weighting factors of each feature dimension are dynamically adjusted based on the standardized value of the time period density index in the real-time operation status feature vector. The weighted similarity between the real-time operation status feature vector and the operation status feature vector of the historical scheduling task segment is calculated based on the weighting factors. Candidate scheduling segments that meet the similarity conditions are retrieved from the historical scheduling task segment storage set.

[0015] The candidate scheduling segments are further filtered based on the real-time fatigue dynamic index and skill qualifications of online drivers. The filtered candidate scheduling segments are mapped to the initial scheduling template in the current scenario. The real-time fatigue dynamic index is obtained by weighted fusion of continuous driving saturation, daily cumulative load and rest quality coefficient.

[0016] The system identifies hard constraint conflict nodes in the initial scheduling template along the driver task chain and vehicle connection chain. Based on the causal dependencies between conflict nodes and the risk impact of the real-time fatigue dynamic index on task succession between nodes, a conflict dependency hierarchy is constructed. In each expansion step, the bundle width is dynamically adjusted based on the remaining computation time, the standardized value of the time period density index, the complexity of the current conflict dependency hierarchy, and the average driver fatigue. Candidate operations are executed layer by layer along the conflict dependency hierarchy to resolve conflicts. During the evaluation of candidate operations, operations that cause the real-time fatigue dynamic index to exceed the safety threshold are prohibited or given extremely high costs. The comprehensive operation cost of the current step is fed back to adjust the weight factor until a scheduling scheme that satisfies all hard constraints is output or the computation time limit is reached.

[0017] Further, the standardization process includes: the time period density index being the ratio of the planned number of departures to the maximum capacity of the line within the current time window; the online driver location distribution entropy being the information entropy calculated using the discrete probability distribution of the current station where the online driver is located; the task load being the ratio of the total remaining driving time of online drivers to the total driving time required for all tasks within the current time window; and the vehicle turnaround deviation being the average absolute deviation between the actual vehicle turnaround time and the planned turnaround time. Then, the time period density index, the online driver location distribution entropy, the task load, and the vehicle turnaround deviation are mapped to the interval between 0 and 1 using a standardization method, thus completing the standardization process.

[0018] Furthermore, the dynamic adjustment of the weight factors of each feature dimension includes: setting the time period density weight factor corresponding to the standardized value of the time period density index to 1 and the sum of the standardized value of the time period density index multiplied by a first preset ratio; allocating the remaining weights of the online driver location distribution entropy weight factor, task load weight factor, and vehicle turnover deviation weight factor according to the original ratio, wherein the remaining weight is 1 minus the time period density weight factor; and the weight factors are periodically fine-tuned according to the accumulated comprehensive operation cost during the conflict resolution process: when the average comprehensive operation cost of the most recent few steps is higher than a preset threshold, the time period density weight factor is increased and the other weight factors are decreased accordingly.

[0019] Furthermore, the weighted similarity is a weighted Euclidean distance, which includes: calculating the difference between each feature value of the real-time operation status feature vector and the corresponding feature value of the historical scheduling task segment and squaring it, multiplying each squared difference by the corresponding weight factor and summing them, and then taking the square root of the summation result.

[0020] Furthermore, the real-time fatigue dynamic index is calculated in the following ways: the continuous driving saturation is the ratio of the driver's current continuous driving time to the prescribed maximum continuous driving time; the daily cumulative load is the ratio of the driver's cumulative driving time from the start of the workday to the current time to the maximum allowed cumulative driving time for the day; the rest quality coefficient is determined by a preset fuzzy inference rule table based on the ratio of the most recent rest time to the prescribed minimum rest time and the continuous driving time before the rest, and is used to obtain the degree of fatigue recovery from the previous rest; the real-time fatigue dynamic index is equal to the continuous driving saturation weight multiplied by the continuous driving saturation, plus the daily cumulative load weight multiplied by the daily cumulative load, plus the rest quality supplement weight multiplied by 1 minus the rest quality coefficient.

[0021] Furthermore, the secondary screening includes: excluding online drivers whose real-time fatigue dynamic index exceeds a preset safety threshold; for the remaining online drivers, verifying the matching degree between the qualification requirements of the driver's assigned position in the historical scheduling task segment and the legal qualifications held by the online driver, and assessing the adaptability of the online driver's current real-time fatigue dynamic index to the required increase in driving load for the position; selecting the historical scheduling task segment with the highest overall score, and when the highest score corresponds to multiple segments, selecting the segment with the lowest weighted similarity as the most similar segment, and mapping the driver assignment structure of the most similar segment to the initial scheduling template in the current scenario; during the mapping process, replacing the historical driver identifier with the identifier of the currently online driver whose skills best match the minimum qualification deviation principle.

[0022] Furthermore, the construction of the conflict dependency hierarchy includes: starting from the newly added sudden task node, traversing the driver's continuous task chain forward and backward, and expanding along the vehicle connection chain, checking the hard constraints of each node, and marking nodes that do not meet any hard constraints as conflict nodes; the hard constraints include the upper limit of continuous driving time, the minimum rest time, the matching of driver qualifications with tasks, and the minimum turnaround time for vehicles; and, based on the causal dependency relationship between conflict nodes and the task succession risk relationship based on the real-time fatigue dynamic index, a directed edge is established: if the resolution of the first conflict node depends on the priority adjustment of the second conflict node, or if the fatigue index of the driver of the first conflict node exceeds a threshold, which will force the driver of the second conflict node to take over, thus generating a new conflict, then a directed edge is established from the second conflict node to the first conflict node, forming a directed acyclic conflict dependency hierarchy structure after traversing all nodes; in the conflict dependency hierarchy structure, the closer the node is to the newly added sudden task node and the higher the fatigue risk impact, the lower the hierarchy level and the higher the adjustment priority.

[0023] Further, the joint adjustment of the beamwidth includes: determining the beamwidth value for the current step according to the following rules based on the current remaining available computing time, the total allowed computing time, the standardized value of the time period density index, the average number of directed edges in the current conflict dependency level, and the driver's average fatigue dynamics: multiply the preset maximum beamwidth by the ratio of the remaining available computing time to the total allowed computing time, then multiply by 1 by the difference obtained by subtracting the product of the standardized value of the time period density index and the driver's average fatigue dynamics, then divide by the sum obtained by the sum of 1 and the normalized value of the average number of directed edges, round the result down and compare it with a value to take the maximum value as the beamwidth for the current step.

[0024] Further, the candidate operations include: driver replacement operation, task order fine-tuning operation, and standby insertion operation; during the evaluation of candidate operations, the comprehensive operation cost of the candidate operation is calculated, which is equal to the safety cost weight multiplied by the weighted sum of the changes in the real-time fatigue dynamic index of the relevant drivers after the operation, plus the operating cost weight multiplied by the empty driving distance caused by the operation, plus the conflict consequence weight multiplied by the number of newly generated conflict nodes after the operation; and for any candidate operation, a short-term forward simulation is performed to predict the changes in the fatigue dynamic index within the next two task steps. If the fatigue dynamic index of any driver exceeds the safety threshold in the simulation, a prohibition cost is directly assigned; wherein, the driver replacement operation is to reassign the task corresponding to the conflict node to another qualified and available driver, the task order fine-tuning operation is to shift the execution time of the task within a preset time tolerance window, and the standby insertion operation is to use a standby driver to execute the conflict node task; for the standby insertion operation, its comprehensive operation cost is directly assigned a preset fixed cost value, without using the weighted calculation of the safety cost, operating cost, and conflict consequence.

[0025] Beneficial effects

[0026] The present invention has at least the following beneficial effects:

[0027] 1. By extracting time-period density index, online driver location distribution entropy, task load, and vehicle turnover deviation, a four-dimensional operational status feature is constructed. Based on the standardized value of real-time time-period density, the feature weights in the weighted Euclidean distance are adaptively adjusted, automatically strengthening the contribution of density features during high-density periods. This allows for the accurate retrieval of scheduling segments that highly match the current status from historical scheduling records. After secondary screening based on driver qualifications and fatigue, the number of hard constraint conflicts within the initial template is reduced, and the optimized search space is significantly compressed, thereby improving the real-time response speed and initial solution quality for sudden scheduling generation.

[0028] 2. By constructing a conflict-dependent directed acyclic graph along the driver task chain and vehicle connection chain starting from a new task, and performing incremental correction layer by layer according to the hierarchy of dependencies, structured local search replaces global enumeration, thus avoiding combinatorial explosion in terms of mechanism; the bundle width is adaptively adjusted according to the remaining computing time, time period density index and average driver fatigue, automatically shrinking when the time limit is approaching to ensure rapid convergence, and appropriately expanding when the time limit is generous to explore better solutions, thereby stably outputting a scheduling scheme that satisfies all hard constraints or the suboptimal solution with the minimum cost within the strict decision time limit, achieving an adaptive balance between computing efficiency and scheme quality.

[0029] 3. By weighted fusion of continuous driving saturation, daily cumulative load, and rest quality coefficient based on fuzzy inference, a multi-dimensional dynamic index of driver fatigue is constructed, breaking through the limitations of the traditional single indicator that only relies on continuous driving time, and comprehensively reflecting the risks of continuous and cumulative fatigue. This index is set as a rigid safety threshold and embedded with the operation evaluation of conflict resolution, which automatically blocks candidate operations that cause fatigue to exceed limits, reduces scheduling that is compliant in form but high in reality, and enables the final solution to have proactive fatigue safety protection capabilities while meeting the capacity demand. Attached Figure Description

[0030] Figure 1 The flowchart illustrates the steps of the digital scheduling generation method based on a general algorithm for rail transit in this invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] The following is in conjunction with the appendix Figure 1 The specific embodiments of the present invention will be described in detail below.

[0033] S1: Extract the time period density index representing the intensity of tasks, the online driver location distribution entropy representing the spatial distribution of drivers, the task load representing the supply and demand of transportation capacity, and the vehicle turnover deviation representing the stability of vehicle operation within the current time window.

[0034] When scheduling needs are triggered in non-stationary scenarios, a 30-minute analysis time window is defined based on the current time. From the operation chart, real-time driver location data, and vehicle arrival and departure records within this time window, the values ​​required to calculate the time period density index, online driver location distribution entropy, task load, and vehicle turnover deviation are extracted respectively. The duration of the analysis time window is an empirical value and can be adjusted by the implementer according to the specific implementation scenario.

[0035] Time Period Density Index The number of scheduled departures within this time window With the maximum throughput capacity of the line The ratio, according to Calculation, where Read directly from the runtime graph. The minimum headway and turnaround capability of the line signal are determined and stored as fixed constants in a preset storage location. This ratio is dimensionless, and the closer the value is to 1, the more intensive the tasks are in the current time window.

[0036] Online driver location distribution entropy The information entropy is calculated using the discrete probability distribution of online drivers at each station, as shown in the formula: ,in The total number of stations on the line. For the first The percentage of online drivers at each station out of the total number of online drivers.

[0037] Drivers operating within a section are assigned to the nearest station based on the principle of proximity.

[0038] The unit of online driver location distribution entropy is bits. The value is higher when the drivers are more evenly distributed, and tends to 0 when all drivers are concentrated at the same station.

[0039] Task load This represents the total remaining driving time for all online drivers within the current time window. Total driving time required for all missions during the same period The ratio, i.e. , and The units are all minutes. This ratio is dimensionless. A value less than 1 indicates a capacity shortage, while a value greater than 1 indicates relatively abundant capacity.

[0040] Vehicle turnover deviation Using the completed vehicle turnaround events within the current time window as the statistical object, calculate the absolute value of the difference between the actual turnaround time and the planned turnaround time for each trip, and then average the difference across all turnaround trips.

[0041] Assume there is a total The second turnaround event, the first The actual turnaround time for this time is The planned turnaround time is ,but .

[0042] If there are no rewind events within the current time window, then The value is 0.

[0043] The actual turnaround time and the planned turnaround time are both measured in minutes. The greater the deviation, the more serious the deviation between the actual turnaround and the plan, and the lower the stability of the vehicle operation chain.

[0044] The four original feature values ​​mentioned above are uniformly standardized using range, and each feature... Standardized value according to Calculation, where and These are the minimum and maximum values ​​of this feature, pre-statistically calculated and fixed in historical operational data. After standardization, each feature value is mapped to... If the real-time value exceeds the historical extreme value range, the standardized value is treated as the boundary value 0 or 1.

[0045] This results in a standardized value derived from the time period density index. Standardized value of online driver location distribution entropy Standardized values ​​of task load Standardized value of vehicle turnover deviation The real-time operational status feature vector is composed of these features.

[0046] S2: Dynamically adjust the weight factors of each feature dimension based on the standardized value of the time period density index in the real-time operation status feature vector.

[0047] The sensitivity of each feature dimension in the real-time operation status feature vector to the scheduling structure varies. The task density within a time period directly determines the number of drivers required, and even a small change in this density can lead to a significant shift in the scheduling plan. However, the changes in driver location distribution entropy and vehicle turnover deviation have a relatively small impact on the scheduling structure. Therefore, when comparing the similarity between the real-time vector and the historical vector, it is necessary to apply differentiated weights to the feature dimensions based on the real-time task density to strengthen the dominant role of density features in similarity measurement.

[0048] Standardized value of density index for the current period Based on this, let the time period density weighting factor be... ,in, This is the first preset ratio, an empirical value that can be adjusted by the implementer according to the specific implementation scenario.

[0049] Online driver location distribution entropy weighting factor Task load weighting factor and vehicle turnover deviation weighting factor Based on the original proportions Allocate remaining weights ,Right now:

[0050]

[0051] The aforementioned original proportional constant is a preset parameter that can be adjusted by the implementer according to the specific implementation scenario.

[0052] The weighted similarity between the real-time operational status feature vector and the operational status feature vector of any historical scheduling task segment is measured by a weighted Euclidean distance: the differences of four standardized features, namely the time period density index, the online driver location distribution entropy, the task load degree, and the vehicle turnover deviation degree, are calculated respectively. The squares of each difference are multiplied by the corresponding weight factor and summed. The square root of the summation result is then taken.

[0053] This distance is dimensionless, and the smaller the value, the more similar the real-time state is to the historical segment.

[0054] Iterate through the historical task scheduling fragment storage set, and for each historical fragment, substitute it into the real-time vector to calculate its relationship. Value, will all Values ​​are sorted in ascending order, and the first few are taken. One historical segment was selected as a candidate scheduling segment that met the similarity criteria. The empirical value of 3 is set as a preset parameter, which can be adjusted by the implementer according to the specific implementation scenario. This outputs candidate scheduling segments obtained by weighted similarity filtering.

[0055] Weight feedback fine-tuning: In the subsequent conflict resolution process, the above weight factors are periodically fine-tuned based on the accumulated comprehensive operational costs.

[0056] After resolving each conflict node, the overall operational cost of that step is recorded. When the average overall operational cost of the most recent five consecutive steps exceeds a preset threshold of 15.0, the time-period density weighting factor is adjusted. Increase by 0.05 but not exceeding 0.8, while simultaneously decreasing proportionally according to the original ratio. , and This ensures that the sum of all factors is 1. The preset threshold, 0.05, and 0.8 are empirical values ​​that can be adjusted by the implementer according to the specific implementation scenario.

[0057] The fine-tuned weight factors will be immediately used for subsequent supplementary similarity searches that may be triggered in the same batch, as well as for the initial weight calculation of the next batch.

[0058] S3: Perform secondary screening on candidate scheduling segments and map them to the initial scheduling template.

[0059] A secondary screening is performed based on the real-time fatigue dynamic index and skill qualifications of online drivers. The selected candidate scheduling segments are mapped to the initial scheduling template in the current scenario. The real-time fatigue dynamic index is obtained by weighted fusion of continuous driving saturation, daily cumulative load and rest quality coefficient.

[0060] In one embodiment, the fuzzy inference rule table takes the rest duration ratio and the normalized values ​​of the continuous driving time before the rest as inputs.

[0061] The rest duration ratio is the ratio of the most recent rest duration to the minimum prescribed rest duration. The normalized value of the continuous driving time before the rest is the ratio of the continuous driving time before the rest begins to the upper limit of the continuous driving time. Both inputs are divided into three fuzzy sets: low, medium, and high. The membership function adopts a triangular or trapezoidal function.

[0062] High rest quality is output when the rest duration ratio is high and the normalized value of the continuous driving time before the rest is low; low rest quality is output when the rest duration ratio is low or the normalized value of the continuous driving time before the rest is high; and medium rest quality is output for the remaining combinations.

[0063] During defuzzing, low, medium, and high rest quality are mapped to 0.4, 0.7, and 1.0, respectively, and the rest quality coefficient is obtained by weighted averaging according to the activation intensity of each inference rule. Among them, 0.4, 0.7, and 1.0 are empirical values ​​that can be adjusted by the implementer according to the specific implementation scenario.

[0064] The candidate scheduling segments output by the aforementioned weighted similarity retrieval are selected based on the global statistical characteristics of the operational status, without considering the real-time fatigue risk and legal qualification constraints of individual online drivers.

[0065] Therefore, before mapping the historical driver allocation structure in the candidate segments to the current online driver group, it is necessary to conduct a multi-dimensional applicability evaluation of the online drivers and exclude drivers who are in a state of high fatigue or lack the necessary qualifications. Otherwise, direct mapping will generate a large number of immediate hard constraint conflicts, weakening the feasibility basis of the initial template.

[0066] For each driver in the online driver pool, the real-time fatigue dynamic index is first calculated. .

[0067] This index is based on continuous driving saturation. Daily cumulative load and rest quality coefficient The weighted fusion is obtained.

[0068] Continuous driving saturation Reflects the driver's immediate fatigue stress from continuous driving, according to The calculation involves the driver's continuous driving time from the end of the last rest period to the current moment, which is accumulated in real time by the on-board terminal or attendance records, in minutes; the upper limit of continuous driving time stipulated by regulations is a preset constraint parameter.

[0069] This ratio is dimensionless, and the closer the value is to 1, the closer the continuous driving load is to the limit.

[0070] Daily cumulative load Reflecting the cumulative driving load since the start of the workday, according to The calculation is performed by summing up the driver's total driving time from the start of their shift to the current time, in minutes; the maximum allowed total driving time for the day is a preset constraint parameter.

[0071] This ratio is also dimensionless; the closer the value is to 1, the greater the accumulated driving load for the day.

[0072] Rest quality coefficient The value of is not a fixed constant, but is obtained by calling a preset fuzzy inference rule table based on the driver's most recent rest segment.

[0073] Based on the above three dimensions, the driver fatigue dynamic index is obtained by weighted fusion of continuous driving saturation, daily cumulative load, and rest quality coefficient:

[0074]

[0075] Among them, the weight of continuous driving saturation Daily cumulative load weight Rest quality supplementary weight These are all dimensionless preset parameters, which are empirical values ​​and can be adjusted by the implementer according to the specific implementation scenario.

[0076] The range of values ​​is limited to It is dimensionless, and the higher the value, the greater the real-time fatigue risk faced by the driver.

[0077] After obtaining the real-time dynamic fatigue index of each online driver, perform the first hard-constraint filtering: Exceeding the preset safety threshold Drivers who do not meet the requirements are excluded from the candidate online driver set. The preset safety threshold is an empirical value that can be adjusted by the implementer according to the specific implementation scenario.

[0078] Excluded drivers will not participate in the allocation of any new tasks during the current scheduling cycle, thus eliminating the possibility of drivers at high fatigue risk being included in the initial template from the source of the algorithm.

[0079] Based on the remaining set of online drivers, the driver allocation structure contained in each candidate scheduling segment is read one by one, and the qualification requirements for each historical position are extracted.

[0080] The qualification requirements specifically refer to the legally required certificates or line access permits that must be held to perform the duties of the position, such as traction qualification for line A, signal certification for line B, etc.

[0081] The qualification requirements for each position are compared item by item with the legal qualifications held and valid by each driver in the remaining online driver pool to establish qualification fulfillment relationships.

[0082] For all positions in the candidate scheduling segments, calculate the qualification satisfaction rate for each remaining online driver. It is defined as the ratio of the number of qualifications held by the driver that match the job requirements to the total number of job qualification requirements, with a value of 1 when there is a perfect match.

[0083] Simultaneously, by retrieving driving task details for the corresponding position from historical data, the system estimates the increase in the driver's continuous driving saturation and daily cumulative workload if the driver were to take on the task, thereby calculating the expected dynamic fatigue index after the task is performed. .

[0084] Prioritize the current Lower and expected Still below the safety threshold Drivers should be properly matched to ensure that no new risks of fatigue or exceeding limits are introduced after job placement.

[0085] The overall score for each candidate shift segment is generated by combining the degree of qualification satisfaction and the degree of fatigue suitability.

[0086] One example of how to calculate the overall score is to evaluate all segments within the segment. For each job position, the weighted sum of the average qualification satisfaction and average fatigue adaptability of the best-matched drivers for that position is calculated, with each weighted at 0.5. Fatigue adaptability is... The metric, with each weight set to 0.5, is an empirical value that can be adjusted by the implementer based on the specific implementation scenario.

[0087] The candidate scheduling segment with the highest overall score is selected as the most similar segment; if the highest score corresponds to multiple segments, the segment with the smallest distance among the aforementioned weighted similarities is selected as the most similar segment.

[0088] When mapping the most similar segment to the current scene, the principle of minimum qualification deviation is followed: the driver identifiers in the historical segments are replaced one by one with the best matching current online driver identifiers for each position determined in the previous segment, and after replacement, each driver is verified to have the legal qualifications for the assigned task.

[0089] If a gap occurs in the time and space connection of the original driver task chain due to driver replacement, an available driver within the same station and with the same qualifications will be selected to fill the gap first. If it is not possible to fully maintain the continuity of the original chain, a minimum deviation involving only a single point of connection is allowed. This deviation will be further corrected in the subsequent conflict resolution steps.

[0090] The driver-task allocation relationship generated in this way is the initial scheduling template for the current scenario. This template has embedded hard constraints on fatigue safety and qualification compliance requirements, and can be directly input into the conflict identification and resolution process for processing.

[0091] S4: Identify hard constraint conflict nodes in the initial scheduling template along the driver task chain and vehicle connection chain and construct a conflict dependency hierarchy.

[0092] After the initial scheduling template maps the historical driver allocation structure to the current online drivers, due to differences between the current driver group's real-time fatigue level, specific qualification status, and real-time location distribution and the situation in historical segments, the template still has many hard constraints that are not met, such as continuous driving time exceeding the limit, insufficient rest intervals, and qualifications not matching the job.

[0093] Taking the newly added emergency task node as the origin, traverse the forward and backward directions of each driver's continuous task chain, and simultaneously expand along the vehicle connection chain to verify the hard constraints node by node.

[0094] The verification items include whether the continuous driving time exceeds the 120-minute limit, whether the actual rest time between adjacent tasks is less than the minimum rest time of 10 minutes, whether the driver's legal qualifications match the job requirements, and whether the vehicle turnaround time meets the minimum turnaround time.

[0095] Any node that does not satisfy any constraint is marked as a conflict node.

[0096] After marking all conflict nodes, directed edges are established based on the causal dependencies between conflict nodes and the task succession risk relationship based on the real-time fatigue dynamic index.

[0097] The specific rules are as follows: if the resolution of the first conflict node depends on the priority adjustment of the second conflict node, then a directed edge is established from the second conflict node to the first conflict node; furthermore, if the real-time fatigue dynamic index of the driver at the first conflict node is... If the driver's subsequent tasks are forced to be taken over by the driver of the second conflict node, the replacement behavior may generate a new fatigue over-limit conflict, and a directed edge from the second conflict node to the first conflict node will be established.

[0098] By traversing all conflict nodes, a directed acyclic conflict dependency hierarchy is formed.

[0099] In the hierarchical structure, the closer a node is to a newly added emergency task node and the higher its fatigue risk impact, the lower its level and the higher its adjustment priority. The fatigue risk impact is the sum of the number of directed edges pointing to that node and the number of directed edges originating from that node.

[0100] After entering the conflict resolution iteration, in each expansion step, the remaining computation time is used as a guide. Standardized value of real-time time-segment density index The average number of directed edges at the current conflict dependency level and the dynamic average fatigue level of all online drivers. Joint calculation of adaptive beam width .

[0101] The formula for calculating the beam width is:

[0102]

[0103] Among them, the preset maximum beam width Taking an empirical value of 50, the total allowable calculation time Take 600 seconds, and the remaining calculation time is in seconds. The empirical value can be adjusted by the implementer according to the specific implementation scenario.

[0104] The normalized value of the average number of directed edges is defined as follows: , The maximum number of directed edges is set to 10 in this embodiment, but can be adjusted by the implementer according to the specific implementation scenario. The average number of directed edges... This is the arithmetic mean of the number of directed edges associated with each conflict node in the current conflict dependency level. If the level is empty, it is 1.

[0105] When the remaining computation time decreases, the operational density increases, or the overall driver fatigue increases, the product term... Increase Decrease, at the same time It also decreases, and the more complex the constraint coupling, the more... The larger the denominator, the greater the denominator, leading to Automatic shrinking reduces the number of candidate states retained in each expansion step to adapt to time constraints and security requirements.

[0106] For the conflict node that needs to be processed in the conflict dependency hierarchy graph, three types of candidate operations are generated.

[0107] The first type is driver replacement operation, which reassigns the driving task corresponding to the node to another qualified online driver who is in a task gap.

[0108] The second category is task sequence fine-tuning operations, which shift the start or end time of a task within a preset time tolerance window. The tolerance window width is dynamically determined based on the line operation diagram margin, and is usually in the range of -3 minutes to +3 minutes. The -3 minutes and +3 minutes are empirical values ​​that can be adjusted by the implementer according to the specific implementation scenario.

[0109] The third category is the insertion of standby operations, which involves using standby drivers who are in a ready-to-go state to perform the task.

[0110] For driver replacement and task sequence fine-tuning operations, the comprehensive operational cost is calculated based on changes in fatigue risk, additional vehicle idling distance, and the number of new conflict nodes after the operation. The comprehensive operational cost is obtained by weighted summation of safety costs, operating costs, and conflict consequence costs.

[0111]

[0112] Among them, the safety cost weight Operating cost weight Conflict consequences weighting All three are dimensionless preset parameters that can be adjusted by the implementer according to the specific implementation scenario.

[0113] ΔF is the arithmetic mean of the real-time dynamic index changes of the drivers involved after performing this operation, calculated by substituting the old and new shift segments of each driver into the aforementioned... The calculation method obtains the change in each driver's value, and then takes the arithmetic mean of all drivers involved. If only one driver is involved, then that driver's change in value is the dimensionless value.

[0114] The normalized value of the additional vehicle idling distance caused by the operation is defined as follows: ,in The unit is meters. As a preset benchmark distance, this embodiment uses 1000 meters to eliminate dimensional differences and keep the three costs at similar orders of magnitude.

[0115] This refers to the number of nodes that do not meet hard constraints when the operation is re-checked along the driver's task chain and the vehicle connection chain after execution.

[0116] During the cost assessment process, the real-time dynamic fatigue index of each driver affected by the operation is calculated. .

[0117] In addition, a short forward simulation is performed on each candidate operation to predict the dynamic exponential change of fatigue in the next two task steps after the operation is executed.

[0118] The simulation method is as follows: assuming the operation takes effect, temporarily update the continuous driving time and cumulative driving time of the affected drivers, and extrapolate up to two tasks based on the subsequent task sequence, calculating the time at the end of each task. value.

[0119] If the dynamic fatigue index of any driver exceeds the safety threshold immediately after the operation. Or any driver within the next two task steps Exceeding the safety threshold If the candidate operation is prohibited, then the prohibition cost is directly assigned; in this embodiment, it is set to... This is far higher than the upper limit of normal operating costs.

[0120] For the insertion of standby operations, the comprehensive operational cost is directly assigned a preset fixed value of 100, without using a weighted calculation of safety costs, operating costs, and conflict consequences, and without performing forward simulation.

[0121] All non-forbidden candidate operations generated in the current step will be processed according to... Sort in ascending order and keep the first few with the lowest cost. Candidate operations (beam width limit).

[0122] From this Among the candidate operations, the one with the smallest cumulative operation cost is selected as the actual operation to be executed in the current step. The corresponding local scheduling state is the current state of the next step, and the cumulative operation cost of this state since the initial template is recorded.

[0123] During each actual state update step, the actual execution operation determined in the current step is selected and implemented, updating the relevant driver task chains and vehicle occupancy time windows, and accumulating the driver's continuous driving time and cumulative driving time for the day. After execution, all hard constraints are re-verified along the driver task chain and vehicle connection chain, and the conflict dependency hierarchy graph is updated. If all hard constraints in the updated scheduling scheme are satisfied, the search is stopped immediately, and the scheduling scheme is output as the final result.

[0124] If the iteration is extended to the total computation time If unresolved conflict nodes remain when the search is exhausted, the search terminates. At this point, the scheduling scheme with the lowest cumulative operation cost is selected from all explored states since the initial template as the suboptimal solution output. The specific unresolved conflict items, driver identifiers, and task nodes involved in this scheme are clearly listed for manual intervention and correction by the dispatcher.

[0125] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make several improvements and substitutions without departing from the technical principles of the present invention, and these improvements and substitutions should also be considered within the scope of protection of the present invention.

Claims

1. A digital scheduling generation method based on a general algorithm for rail transit, characterized in that, When scheduling requirements are triggered in non-stationary scenarios, including: Extract the time period density index representing the intensity of tasks, the online driver location distribution entropy representing the spatial distribution of drivers, the task load representing the supply and demand of transportation capacity, and the vehicle turnover deviation representing the stability of vehicle operation within the current time window, and form a real-time operation status feature vector after standardization. The weighting factors of each feature dimension are dynamically adjusted based on the standardized value of the time period density index in the real-time operation status feature vector. The weighted similarity between the real-time operation status feature vector and the operation status feature vector of the historical scheduling task segment is calculated based on the weighting factors. Candidate scheduling segments that meet the similarity conditions are retrieved from the historical scheduling task segment storage set. The candidate scheduling segments are further filtered based on the real-time fatigue dynamic index and skill qualifications of online drivers. The filtered candidate scheduling segments are mapped to the initial scheduling template in the current scenario. The real-time fatigue dynamic index is obtained by weighted fusion of continuous driving saturation, daily cumulative load and rest quality coefficient. The hard constraint conflict nodes in the initial scheduling template are identified along the driver task chain and vehicle connection chain. Based on the causal dependency relationship between conflict nodes and the risk impact relationship between nodes caused by task handover in the real-time fatigue dynamic index, a conflict dependency hierarchy is constructed. This includes: starting from the newly added emergency task node, traversing the driver continuous task chain forward and backward respectively, and expanding along the vehicle connection chain to check the hard constraints of each node. Nodes that do not meet any hard constraints are marked as conflict nodes. The hard constraints include the maximum continuous driving time, the minimum rest time, the matching of driver qualifications with tasks, and the minimum turnaround time for vehicles. Furthermore, directed edges are established based on the causal dependencies between conflict nodes and the task succession risk relationship based on the real-time fatigue dynamic index: if the resolution of the first conflict node depends on the priority adjustment of the second conflict node, or if the fatigue index of the driver of the first conflict node exceeds the threshold, the driver of the second conflict node will be forced to take over, thus generating a new conflict, then a directed edge from the second conflict node to the first conflict node is established, and a directed acyclic conflict dependency hierarchy structure is formed after traversing all nodes. In the conflict dependency hierarchy, the closer a node is to a newly added emergency task node and the higher its fatigue risk impact, the lower its level and the higher its adjustment priority. In each step of the expansion, the bundle width is dynamically and jointly adjusted based on the remaining computation time, the standardized value of the time period density index, the complexity of the current conflict dependency level, and the driver's average fatigue level. This includes: determining the bundle width value for the current step according to the following rules based on the current remaining available computation time, the total allowed computation time, the standardized value of the time period density index, the average number of directed edges in the current conflict dependency level, and the driver's average fatigue level: multiply the preset maximum bundle width by the ratio of the remaining available computation time to the total allowed computation time, then multiply by 1 minus the difference obtained by the product of the standardized value of the time period density index and the driver's average fatigue level, then divide by the sum obtained by the sum of 1 and the normalized value of the average number of directed edges, round down the result and compare it with a value to take the maximum value as the bundle width for the current step; execute candidate operations layer by layer along the conflict dependency level to resolve conflicts, and during the evaluation of candidate operations, assign prohibition or extremely high cost to operations that cause the real-time fatigue dynamic index to exceed the safety threshold, and feed back the comprehensive operation cost of the current step to adjust the weight factor until a scheduling scheme that satisfies all hard constraints is output or the computation time limit is reached.

2. The digital scheduling generation method for rail transit based on a general algorithm, as described in claim 1, is characterized in that... The standardization process includes: The time-period density index is the ratio of the number of planned departures within the current time window to the maximum capacity of the line. The online driver location distribution entropy is the information entropy calculated using the discrete probability distribution of the online driver's current station location; The task load is the ratio of the total remaining driving time of online drivers to the total driving time required for all tasks within the current time window; The vehicle turnaround deviation is the average absolute deviation between the actual vehicle turnaround time and the planned turnaround time. Furthermore, the time period density index, the online driver location distribution entropy, the task load, and the vehicle turnover deviation are mapped to the 0 to 1 range using a standardization method to complete the standardization process.

3. The digital scheduling generation method for rail transit based on a general algorithm, as described in claim 1, is characterized in that... The dynamic adjustment of the weight factors for each feature dimension includes: Set the time period density weight factor corresponding to the standardized value of the time period density index to 1 and the sum of the standardized value of the time period density index multiplied by a first preset ratio. The remaining weights are allocated according to the original proportions of the online driver location distribution entropy weight factor, task load weight factor, and vehicle turnover deviation weight factor, and the remaining weights are 1 minus the time period density weight factor. Furthermore, the weighting factors are periodically fine-tuned based on the accumulated comprehensive operation cost during the conflict resolution process: when the average comprehensive operation cost of the most recent steps is higher than a preset threshold, the time period density weighting factor is increased and the other weighting factors are decreased accordingly.

4. The digital scheduling generation method for rail transit based on a general algorithm, as described in claim 1, is characterized in that... The weighted similarity is a weighted Euclidean distance, including: Calculate the difference between each feature value of the real-time operation status feature vector and the corresponding feature value of the historical scheduling task segment, and square the difference. Multiply the square of each difference by the corresponding weight factor and sum the results. Then take the square root of the summation.

5. The digital scheduling generation method for rail transit based on a general algorithm, as described in claim 1, is characterized in that... The real-time fatigue dynamic index is calculated in the following way: The continuous driving saturation is the ratio of the driver's current continuous driving time to the prescribed maximum continuous driving time. The daily cumulative load is the ratio of the driver's cumulative driving time from the start of the workday to the current time to the maximum allowed cumulative driving time for the day; The rest quality coefficient is determined by a preset fuzzy inference rule table based on the ratio of the most recent rest duration to the minimum prescribed rest duration and the continuous driving time before the rest, and is used to obtain the degree of fatigue recovery from the previous rest. The real-time fatigue dynamic index is equal to the continuous driving saturation weight multiplied by the continuous driving saturation, plus the daily cumulative load weight multiplied by the daily cumulative load, plus the rest quality supplement weight multiplied by 1 minus the rest quality coefficient.

6. The digital scheduling generation method for rail transit based on a general algorithm, as described in claim 1, is characterized in that... The secondary screening includes: Exclude online drivers whose real-time fatigue dynamic index exceeds a preset safety threshold; For the remaining online drivers, verify the matching degree between the qualification requirements of the driver's assigned position in the historical shift task segment and the legal qualifications held by the online driver, and assess the adaptability of the online driver's current real-time fatigue dynamic index to the increase in driving load required for the position. Select the historical scheduling task segment with the highest overall score. When the highest score corresponds to multiple segments, select the segment with the smallest weighted similarity as the most similar segment. Map the driver allocation structure of the most similar segment to the initial scheduling template in the current scenario. During the mapping process, the historical driver identifiers are replaced with the identifiers of currently online drivers whose skills best match the criteria, based on the principle of minimum qualification deviation.

7. The digital scheduling generation method for rail transit based on a general algorithm, as described in claim 1, is characterized in that... The candidate operations include: The candidate operations include driver replacement operation, task sequence fine-tuning operation, and inserting standby operation; During the evaluation of candidate operations, the comprehensive operational cost of the candidate operation is calculated. The comprehensive operational cost is equal to the weighted sum of the safety cost weight multiplied by the change in the dynamic index of the real-time fatigue of the relevant drivers after the operation is executed, plus the operating cost weight multiplied by the empty driving distance caused by the operation, plus the conflict consequence weight multiplied by the number of new conflict nodes generated after the operation. Furthermore, for any candidate operation, a short-term forward simulation is performed to predict the change in the fatigue dynamic index within the next two task steps. If the fatigue dynamic index of any driver exceeds the safety threshold in the simulation, a prohibition cost is directly assigned. Among them, the driver replacement operation is to reassign the task corresponding to the conflict node to another qualified and available driver; the task order fine-tuning operation is to shift the execution time of the task within a preset time tolerance window; and the standby operation is to use a standby driver to execute the conflict node task. For the standby operation, its comprehensive operation cost is directly assigned a preset fixed value, instead of using the weighted calculation of the safety cost, operating cost, and conflict consequences.

Images