Highway service area vehicle flow scheduling method and system based on intelligent algorithm

Abstract

The application provides a highway service area vehicle flow scheduling method and system based on an intelligent algorithm, which effectively utilizes a to-be-learned template set containing multiple template scheduling learning data through a deep knowledge learning algorithm. In the target training stage, the method dynamically evaluates the effective annotation training value and invalid annotation global training value of the corresponding training vehicle flow scheduling network relative to the to-be-called vehicle flow scheduling network for each template scheduling learning data, thereby ensuring the pertinence and efficiency of the training process. Through continuous optimization learning, when the training error no longer continues to decline, the deep knowledge learning is terminated, and finally the target vehicle flow scheduling network capable of accurately determining the target vehicle flow dynamic data scheduling knowledge description is obtained. The method improves the intelligent level of vehicle flow scheduling, optimizes the vehicle flow scheduling effect of the highway service area, helps to relieve traffic congestion in the service area, and improves the operation efficiency and user experience of the service area.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and more specifically, to a method and system for traffic flow scheduling in highway service areas based on intelligent algorithms. Background Technology

[0002] With the continuous expansion of the expressway network and the sustained growth of traffic flow, the traffic flow scheduling problem of expressway service areas, as important nodes on expressways, is becoming increasingly prominent. Traditional traffic flow scheduling methods often rely on manual experience and fixed rules, which are difficult to adapt to complex and ever-changing traffic conditions, and can easily lead to problems such as service area congestion and uneven resource utilization.

[0003] Intelligent algorithms have achieved remarkable results in various fields, providing new ideas for traffic flow scheduling in highway service areas. However, existing traffic flow scheduling methods based on intelligent algorithms still have some shortcomings. On the one hand, these methods often lack effective learning mechanisms and cannot fully utilize the scheduling knowledge in historical traffic flow data, resulting in limited scheduling effects. On the other hand, existing methods are prone to problems such as low training efficiency, overfitting, or underfitting when processing large-scale traffic flow data, affecting the accuracy and stability of the scheduling network. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for traffic flow scheduling in highway service areas based on intelligent algorithms, the method comprising:

[0005] Obtain the traffic flow scheduling network to be invoked and the set of templates to be learned, which contains multiple template scheduling learning data; the template scheduling learning data includes the dynamic data of the traffic flow to be scheduled in the highway service area, the invalid labeled scheduling knowledge description of the dynamic data of the traffic flow to be scheduled, and the valid labeled scheduling knowledge description of the dynamic data of the traffic flow to be scheduled.

[0006] For each template scheduling learning data in the set of templates to be learned, during the deep knowledge learning operation based on the template scheduling learning data in the target training phase, the effective labeled training value of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be called is determined; the training traffic flow scheduling network corresponding to the template scheduling learning data is obtained by deep knowledge learning based on the forward training allocation group of the training allocation group to which the template scheduling learning data belongs; in the first deep knowledge learning phase, the network learning target of the initial training allocation group is the traffic flow scheduling network to be called.

[0007] Based on the template scheduling learning data and the deep knowledge learning stage preceding the template scheduling learning data, determine the global training value of the invalid annotations of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be invoked;

[0008] When the training error in the target training phase no longer decreases, based on the effective labeled training value and the invalid labeled global training value corresponding to each template scheduling learning data, deep knowledge learning is terminated, and the target traffic flow scheduling network corresponding to the traffic flow scheduling network to be invoked is obtained; the target traffic flow scheduling network is used to determine the scheduling knowledge description of the target traffic flow dynamic data.

[0009] In one possible implementation of the first aspect, determining the invalid labeled global training value of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be invoked, based on the template scheduling learning data and the deep knowledge learning stage preceding the template scheduling learning data, includes:

[0010] Determine the first sample performance index of the template scheduling learning data, and the second sample performance index of the previous template scheduling learning data;

[0011] In the deep knowledge learning stage based on the previous template scheduling learning data, the global training value of the training traffic flow scheduling network corresponding to the previous template scheduling learning data relative to the traffic flow scheduling network to be called is obtained from the previous invalid annotations.

[0012] In the deep knowledge learning stage based on the template scheduling learning data, the invalid labeled training value of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be invoked is determined.

[0013] Based on the first sample performance index and the second sample performance index, the invalid label training value and the past invalid label global training value are fused and calculated to determine the invalid label global training value of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be called.

[0014] In one possible implementation of the first aspect, determining the invalid labeled training value of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be invoked during the deep knowledge learning stage based on the template scheduling learning data includes:

[0015] In the deep knowledge learning operation based on the template scheduling learning data during the target training phase, the second training confidence of the training traffic flow scheduling network corresponding to the template scheduling learning data that obtains invalid labeled scheduling knowledge description, and the second label confidence of the traffic flow scheduling network to be invoked that obtains invalid labeled scheduling knowledge description;

[0016] Based on the error between the second training confidence and the second label confidence, the invalid label training value of the template scheduling learning data corresponding to the training traffic flow scheduling network relative to the traffic flow scheduling network to be called is determined.

[0017] In one possible implementation of the first aspect, determining the invalid labeled global training value of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be invoked, based on the template scheduling learning data and the deep knowledge learning stage preceding the template scheduling learning data, includes:

[0018] Determine the target sample performance index of the training allocation group where the template scheduling learning data is located, and the past sample performance index of the previous training allocation groups before the training allocation group.

[0019] In the deep knowledge learning phase based on the past training allocation group, the invalid labeled past training value of the training traffic flow scheduling network corresponding to the past training allocation group relative to the traffic flow scheduling network to be called is obtained during the target training phase.

[0020] In the target training phase, based on the deep knowledge learning phase of the training allocation group, the invalid labeled target training value of the training traffic flow scheduling network corresponding to the training allocation group relative to the traffic flow scheduling network to be called is determined.

[0021] Based on the target sample performance index and the past sample performance index, the invalid labeled target training value and invalid labeled past training value are fused and calculated to determine the invalid labeled global training value of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be called.

[0022] In one possible implementation of the first aspect, determining the past sample performance metrics of previous training assignment groups prior to the training assignment group includes:

[0023] Determine the group spacing between previous training assignment groups and the current training assignment group;

[0024] Based on the group spacing, the past sample performance indicators of the past training allocation groups are determined; except for the initial training allocation group, the past sample performance indicators of the remaining past training allocation groups are negatively correlated with the group spacing.

[0025] In one possible implementation of the first aspect, the step of obtaining a set of templates to be learned that contains multiple template scheduling learning data includes:

[0026] Acquire traffic flow scheduling tasks and dynamic data of multiple traffic flows to be scheduled;

[0027] For each piece of dynamic traffic flow data to be scheduled, determine scheduling session information that matches the traffic flow scheduling task; the scheduling session information includes the dynamic traffic flow data to be scheduled.

[0028] The scheduling session information is loaded into the previously generated traffic flow scheduling knowledge network, and the obtained traffic flow scheduling knowledge network is used as an invalid annotation scheduling knowledge description of the dynamic data of the traffic flow to be scheduled.

[0029] Based on the adjustment instruction generated from the invalid labeled scheduling knowledge description of the dynamic data of the traffic flow to be scheduled, the valid labeled scheduling knowledge description of the dynamic data of the traffic flow to be scheduled is obtained.

[0030] Generate template scheduling learning data that includes the dynamic data of the traffic flow to be scheduled, the effective labeled scheduling knowledge description of the dynamic data of the traffic flow to be scheduled, and the invalid labeled scheduling knowledge description of the dynamic data of the traffic flow to be scheduled, to obtain a set of templates to be learned that contains multiple template scheduling learning data.

[0031] In one possible implementation of the first aspect, the step of obtaining the traffic flow scheduling network to be invoked includes:

[0032] Obtain the previously generated deep learning network;

[0033] Based on the set of templates to be learned, an initial set of templates to be learned is generated, which contains multiple initial template scheduling learning data; each initial template scheduling learning data includes dynamic traffic flow data to be scheduled, and either a valid labeled scheduling knowledge description of the dynamic traffic flow data to be scheduled or an invalid labeled scheduling knowledge description of the dynamic traffic flow data to be scheduled.

[0034] The deep learning network is trained on the network parameters based on the initial set of templates to be learned, and the traffic flow scheduling network to be called is obtained.

[0035] In one possible implementation of the first aspect, in the operation of deep knowledge learning based on the template scheduling learning data during the target training phase, determining the effective labeled training value of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be invoked includes:

[0036] In the deep knowledge learning operation based on the template scheduling learning data during the target training phase, the first training confidence of the training traffic flow scheduling network corresponding to the template scheduling learning data to obtain an effective labeled scheduling knowledge description, and the first labeling confidence of the traffic flow scheduling network to be invoked to obtain an effective labeled scheduling knowledge description;

[0037] Based on the error between the first training confidence and the first annotation confidence, the effective annotation training value of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be called is determined.

[0038] In one possible implementation of the first aspect, the method further includes:

[0039] For each template scheduling learning data, during the deep knowledge learning operation based on the template scheduling learning data in the target training phase, the invalid labeled training value of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be called is determined;

[0040] When the training error in the target training stage no longer decreases after determining, based on the effective labeled training value and the invalid labeled global training value corresponding to each template scheduling learning data, the deep knowledge learning is terminated, and the target traffic flow scheduling network corresponding to the traffic flow scheduling network to be invoked is obtained, including:

[0041] When the training error in the target training stage no longer decreases, based on the effective labeled training value, invalid labeled training value, and invalid labeled global training value corresponding to each template scheduling learning data, deep knowledge learning is terminated, and the target traffic flow scheduling network corresponding to the traffic flow scheduling network to be called is obtained.

[0042] In another aspect, embodiments of the present invention also provide a highway service area traffic flow scheduling system based on intelligent algorithms, including a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or code, and the processor is used to execute the programs, instructions or code in the machine-readable storage medium to implement the above-described method.

[0043] Based on the above, this application embodiment effectively utilizes a set of learning templates containing multiple template scheduling learning data through a deep knowledge learning algorithm. During the target training phase, for each template scheduling learning data, the method dynamically evaluates the effective labeled training value and invalid labeled global training value of its corresponding training traffic flow scheduling network relative to the traffic flow scheduling network to be invoked, thereby ensuring the targeting and efficiency of the training process. Through continuous optimization, when the training error no longer decreases, deep knowledge learning is terminated, ultimately obtaining a target traffic flow scheduling network capable of accurately determining the dynamic data scheduling knowledge description of the target traffic flow. This method improves the intelligence level of traffic flow scheduling, optimizes the traffic flow scheduling effect of highway service areas, helps alleviate traffic congestion in service areas, and improves the operational efficiency and user experience of service areas. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the execution flow of the highway service area traffic flow scheduling method based on intelligent algorithms provided in an embodiment of the present invention.

[0045] Figure 2 This is a schematic diagram of the hardware architecture of a highway service area traffic flow scheduling system based on intelligent algorithms provided in an embodiment of the present invention. Detailed Implementation

[0046] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a traffic flow scheduling method for highway service areas based on intelligent algorithms, provided in one embodiment of the present invention. The following is a detailed description of this traffic flow scheduling method for highway service areas based on intelligent algorithms.

[0047] Step S110: Obtain the traffic flow scheduling network to be invoked and a set of templates to be learned, which contains multiple template scheduling learning data. The template scheduling learning data includes dynamic traffic flow data to be scheduled from highway service areas, invalid labeled scheduling knowledge descriptions of the dynamic traffic flow data to be scheduled, and valid labeled scheduling knowledge descriptions of the dynamic traffic flow data to be scheduled.

[0048] In this embodiment, a highway traffic management scenario is described, where a crisscrossing highway network connects multiple cities and regions, and service areas are densely distributed. The traffic flow scheduling network to be invoked is a deep learning network model, pre-trained using a large amount of basic traffic data and preliminary traffic flow scheduling strategies. This network model has multiple hidden layers, each containing numerous neurons, and the connection weights between neurons are adjusted through a series of initialization and pre-training processes.

[0049] Taking a large highway service area as an example, the template scheduling learning data in the template set to be learned includes detailed information from multiple aspects. First, from the perspective of vehicle origin, it includes vehicle information from different cities and different road access points. For example, vehicles may come from main roads in surrounding cities, branches of other highways, or the ring road of the local city. These vehicles originate from places with different traffic flow characteristics, economic activity levels, and travel demand patterns.

[0050] In terms of vehicle type, it not only distinguishes between small cars, large buses, and large trucks, but also further subdivides them. Small cars include ordinary family sedans, high-performance sports cars, and small commercial vehicles, etc. Different types of small cars may have different driving habits, parking needs, and purposes for stopping at service areas. Large buses are divided into tour buses, long-distance passenger buses, and company commuter buses, etc. Tour buses may have concentrated travel peaks during specific tourist seasons, long-distance passenger buses have fixed schedules and stop times, and company commuter buses frequently travel during specific hours on weekdays. Large trucks are further classified according to factors such as cargo type (e.g., dangerous goods, fresh food, general cargo), load capacity, and whether they are used for cold chain transportation. Different types of trucks have very different operational needs at service areas. For example, dangerous goods trucks require special safety inspection and parking areas, while cold chain trucks need to be close to power facilities to maintain the low temperature of the cargo.

[0051] Within a specific timeframe, such as a full workday (from midnight to midnight), vehicle traffic entering and leaving service areas exhibits complex and dynamic changes. For example, in the early morning, long-haul freight vehicles dominate, requiring refueling, driver rest, or simple vehicle checks. As the morning commute begins, the flow of minivans and some family cars gradually increases, perhaps only making brief stops to buy breakfast or use the restroom. In the morning, tour buses may begin arriving at the service area in large numbers, requiring ample parking space, passenger boarding and alighting guidance, and coordinated arrangements for catering and rest facilities. Around noon, the demand for meals from all types of vehicles peaks, necessitating the efficient management of restaurant capacity and temporary parking lot turnover. The afternoon is a period of transition between returning cars and continuing long-haul buses, requiring efforts to avoid congestion at vehicle entry and exit. In the evening, long-haul freight traffic increases again, while the rest needs of some fatigued minivan drivers must also be considered.

[0052] Invalid labeling of scheduling knowledge descriptions (negative labels) is a scheduling strategy based on simple rules that fails to adequately consider various complex factors. For example, assigning parking positions solely based on the order in which vehicles enter a service area, without considering vehicle type, purpose of stop, or subsequent travel arrangements, can lead to tour buses being parked far from dining and restroom facilities, forcing many tourists to travel long distances to reach their destinations, causing chaos and inconvenience. For trucks, assigning parking without distinguishing between hazardous materials could pose safety hazards. Furthermore, this simplistic scheduling method can cause unnecessary intersections and congestion when vehicles leave the service area; for example, small cars may be blocked in narrow passageways while waiting for large buses and trucks to start and leave.

[0053] Effective labeling of dispatch knowledge descriptions (positive labels) is a strategy that comprehensively considers numerous factors, including the vehicle's origin, type, arrival time, purpose of stay, expected departure time, and the current availability of facilities in the service area. For example, large parking areas near catering and restroom facilities are reserved in advance based on the expected arrival time of tour buses, and staff are arranged to guide tourists to disembark and board in an orderly manner, avoiding conflicts between pedestrian and vehicle traffic. For long-distance freight vehicles, different parking areas are arranged according to cargo type and destination; hazardous materials trucks are parked in designated safe areas, maintaining a safe distance from other vehicles, while ordinary trucks are arranged in areas easily accessible according to their expected departure time. For cars, they are guided to different parking areas based on their purpose of stay (e.g., short rest, meal, refueling) and expected stay duration, and their routes are completely separated from those of large vehicles to ensure safety and efficiency. Simultaneously, when vehicles leave the service area, traffic lights and guidance signs are dynamically adjusted to guide them in an orderly manner according to vehicle type and expected direction of travel, avoiding cross-traffic congestion.

[0054] By collecting dynamic data of traffic flow to be dispatched from different service areas at different time periods, as well as the corresponding invalid and valid labeled dispatch knowledge descriptions, a set of templates to be learned containing multiple template dispatch learning data is formed.

[0055] Step S120: For each template scheduling learning data in the set of templates to be learned, during the deep knowledge learning operation based on the template scheduling learning data in the target training phase, the effective labeled training value of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be called is determined. The training traffic flow scheduling network corresponding to the template scheduling learning data is obtained through deep knowledge learning based on the forward training allocation group of the training allocation group to which the template scheduling learning data belongs. In the first deep knowledge learning phase, the network learning target of the initial training allocation group is the traffic flow scheduling network to be called.

[0056] Considering the scenario of the aforementioned highway service area, during the target training phase, deep knowledge learning is performed on a specific template scheduling learning data. The training traffic flow scheduling network corresponding to this template scheduling learning data is gradually constructed through a series of training allocation groups.

[0057] Suppose that the template scheduling learning data is about the traffic flow dynamics of a specific service area on a peak weekend tourist day. Effective labeled scheduling knowledge description involves comprehensive scheduling arrangements. For example, due to the large number of tour buses and their concentrated arrival times on weekends, temporary guidance signs are set up a few kilometers from the highway entrance to avoid congestion at the service area entrance, and initial diversion is carried out based on the tour buses' destinations and expected stop times. For buses with short stop times (only for passenger rest and simple meals), they are guided to the service area's fast service area, which is equipped with efficient catering facilities and temporary vehicle maintenance points, and the parking area is designed for quick entry and exit. For buses with longer stop times (including transit stops for sightseeing), they are guided to a dedicated large tourist vehicle parking area, which is close to the service area's tourist information center and specialty product sales area, facilitating tourists' sightseeing and shopping during their stop.

[0058] Meanwhile, for small cars, given the high proportion of family trips on weekends, cars from the same region or heading to the same tourist attraction are guided to adjacent parking areas based on their origin and destination information, facilitating communication and carpooling among tourists. Furthermore, based on tourist flow forecasts for various attractions within the service area (such as viewing platforms and themed gardens), the signage from the parking lot to these attractions is dynamically adjusted to disperse tourist flow and prevent overcrowding at any particular attraction.

[0059] Both the training network and the waiting network predict the validity of this labeled scheduling knowledge description. When processing this data, the training network calculates the first training confidence level of the labeled scheduling knowledge description using its internal neural network. For example, for a scheduling decision to direct a specific batch of tour buses to a rapid service area, the training network calculates a confidence level of 0.65. This means that based on the network's current weights and structure, it considers the scheduling decision to be 65% likely to be correct.

[0060] The traffic flow scheduling network to be invoked obtains a first label confidence score of 0.55 for the same valid labeled scheduling knowledge description. This indicates that the traffic flow scheduling network to be invoked has a 55% acceptance rate of this scheduling decision.

[0061] The effective labeled training value is determined by the error between these two confidence levels: 0.65 - 0.55 = 0.1. This effective labeled training value reflects the improvement of the trained traffic flow scheduling network in learning effective labeled scheduling knowledge descriptions compared to the traffic flow scheduling network to be invoked on this specific template scheduling learning data. In subsequent training, a larger value indicates that the trained traffic flow scheduling network has greater room for improvement in learning this scheduling knowledge, and the network weights and structure can be adjusted based on this value to better learn this effective scheduling knowledge.

[0062] Step S130: Based on the template scheduling learning data and the deep knowledge learning stage preceding the template scheduling learning data, determine the invalid annotation global training value of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be invoked.

[0063] Continuing with the example of a highway service area, we will elaborate on the process of determining the global training value of invalid annotations.

[0064] Suppose the current template scheduling learning data is about the dynamic traffic flow of a certain service area during holidays, a period characterized by high traffic volume and complex vehicle types. First, determine the first sample performance index for this template scheduling learning data. The calculation of this index involves several factors, such as the fluctuation range of traffic volume, the diversity of vehicle types, the complexity of interactions between different vehicle types, and the degree of correlation with service area facilities.

[0065] In this example, traffic flow fluctuates greatly during holidays, with a continuous increase from early morning to noon, peaking at noon and then gradually decreasing in the afternoon. This significant fluctuation in traffic flow results in a relatively high performance index for the first sample, assumed to be 0.8. Furthermore, the vehicle types include tour buses, long-distance coaches, various types of cars (including family trips and self-driving tour groups), and a large number of freight vehicles (including trucks transporting holiday supplies). This diversity increases the complexity of the sample, further improving the index. Moreover, the interactions between different types of vehicles are very complex, such as tour buses and cars entering and exiting parking lots, and freight vehicles queuing with other vehicles at refueling stations. These factors combined contribute to the first sample's performance index reaching 0.8.

[0066] Then, the performance index of the second sample from the preceding template scheduling learning data is determined. The preceding template scheduling learning data is about the traffic flow dynamics of the service area before holidays. During this period, traffic flow is relatively low compared to holidays, and the vehicle types are relatively homogeneous, mainly consisting of small cars traveling in advance and a small number of freight vehicles. The fluctuation range of traffic flow is small, and the interaction between vehicles is relatively simple. Therefore, the performance index of the second sample is low, assumed to be 0.6.

[0067] Next, in the deep knowledge learning stage based on the previous template scheduling learning data, the global training value of the training traffic flow scheduling network corresponding to the previous template scheduling learning data compared to the calling traffic flow scheduling network is assumed to be 0.3.

[0068] In the deep knowledge learning phase based on the current template scheduling learning data, the training value of invalid labeled traffic flow scheduling networks corresponding to the template scheduling learning data is determined relative to the invalid labeled traffic flow scheduling networks to be called. Invalid labeled scheduling knowledge descriptions are simple but potentially chaotic scheduling strategies, such as allocating vehicles according to fixed parking areas without considering the special changes in vehicle traffic flow and type during holidays.

[0069] For training the traffic flow scheduling network, when processing the current template scheduling learning data, the second training confidence score of the invalid labeled scheduling knowledge description is 0.45. However, the second label confidence score of the invalid labeled scheduling knowledge description obtained by the traffic flow scheduling network to be invoked is 0.4. The training value of the invalid labels is determined based on the difference between the two scores: 0.45 - 0.4 = 0.05.

[0070] Finally, based on the performance metrics of the first and second samples, the training value of invalid labels and the global training value of past invalid labels are fused and calculated. Using a weighted average calculation method, assuming the weight of the first sample performance metric is 0.7 and the weight of the second sample performance metric is 0.3, the calculation process is: (0.8 × 0.05 + 0.6 × 0.3) ÷ (0.8 + 0.6) = (0.04 + 0.18) ÷ 1.4 = 0.157. This 0.157 represents the global training value of the trainable traffic flow scheduling network corresponding to the template scheduling learning data, relative to the traffic flow scheduling network to be called.

[0071] This calculation process reflects a comprehensive evaluation of the overall value of the trained traffic flow scheduling network relative to the traffic flow scheduling network to be called in terms of learning invalid labeled scheduling knowledge, based on the current template scheduling learning data and previous learning. This evaluation of value helps to more accurately adjust network parameters throughout the training process and improve the network's learning performance.

[0072] Step S140: When it is determined that the training error in the target training stage no longer decreases based on the effective labeled training value and the invalid labeled global training value corresponding to each template scheduling learning data, deep knowledge learning is terminated, and the target traffic flow scheduling network corresponding to the traffic flow scheduling network to be invoked is obtained. The target traffic flow scheduling network is used to determine the scheduling knowledge description of the target traffic flow dynamic data.

[0073] Throughout the training process, the effective labeled training value and the invalid labeled global training value are continuously calculated for each template scheduling learning data. Data from multiple highway service areas at different time periods (including weekdays, holidays, peak tourist seasons, and off-seasons) are used as template scheduling learning data for training.

[0074] In the early stages of training, the training error is relatively large. For example, during initial training, because the initial values ​​of the network weights are random or based on simple pre-training settings, the ability to understand and predict complex traffic flow scheduling knowledge is limited. As the traffic flow scheduling network is continuously adjusted and trained based on template scheduling learning data, the training error gradually decreases.

[0075] For example, the initial training error might reach 0.6. As training progresses, the network weights are gradually optimized by considering the effective labeled training value and the invalid labeled global training value of different template scheduling learning data. For instance, for some high-value template scheduling learning data (data with significant effective labeled training value and invalid labeled global training value), the network will focus on adjusting the weights related to these data.

[0076] As training progresses, the training error gradually decreases. When the training error drops to around 0.1, the training situation is further observed. At this point, although continued training may cause fluctuations in the value calculation of some template scheduling learning data, the overall training error no longer decreases. For example, in the subsequent training rounds, the training error remains around 0.1, without any significant further decreasing trend.

[0077] In this case, deep knowledge learning is terminated. The network obtained through this training process is the target traffic flow scheduling network corresponding to the traffic flow scheduling network to be invoked.

[0078] This target traffic flow scheduling network can be used to process scheduling knowledge descriptions of new target traffic flow dynamic data. For example, when faced with traffic conditions at a new highway service area, the target traffic flow scheduling network can generate a comprehensive scheduling knowledge description based on complex factors such as the origin of vehicles (which may be a newly opened city or road access point), vehicle type (new types of transport vehicles or special-purpose vehicles may be present), arrival time (which may be during a special event or under weather conditions), purpose of stay (there may be new tourist attractions or commercial activities near the service area), and the current facility status of the service area (new facilities may be in use or some facilities may be under maintenance). This scheduling knowledge description may include complex scheduling decisions such as planning guidance routes for vehicles entering the service area, precise allocation of parking areas, planning driving routes for different vehicle types within the service area, the order of leaving the service area, and the setting of guidance signs, to ensure the efficient, safe, and orderly operation of the entire highway service area and surrounding traffic flow.

[0079] Based on the above steps, this embodiment of the application effectively utilizes a set of learning templates containing multiple template scheduling learning data through a deep knowledge learning algorithm. During the target training phase, for each template scheduling learning data, the method dynamically evaluates the effective labeled training value and invalid labeled global training value of its corresponding training traffic flow scheduling network relative to the traffic flow scheduling network to be invoked, thereby ensuring the targeting and efficiency of the training process. Through continuous optimization, when the training error no longer decreases, deep knowledge learning is terminated, ultimately obtaining a target traffic flow scheduling network capable of accurately determining the target traffic flow dynamic data scheduling knowledge description. This method improves the intelligence level of traffic flow scheduling, optimizes the traffic flow scheduling effect of highway service areas, helps alleviate traffic congestion in service areas, and improves the operational efficiency and user experience of service areas.

[0080] In one possible implementation, step S130 may include:

[0081] Step A110: Determine the first sample performance index of the template scheduling learning data and the second sample performance index of the previous template scheduling learning data.

[0082] Step A120: In the deep knowledge learning stage based on the previous template scheduling learning data, obtain the global training value of the training traffic flow scheduling network corresponding to the previous template scheduling learning data relative to the traffic flow scheduling network to be called.

[0083] Step A130: In the deep knowledge learning stage based on the template scheduling learning data, determine the invalid labeled training value of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be called.

[0084] Step A140: Based on the first sample performance index and the second sample performance index, the invalid label training value and the past invalid label global training value are fused and calculated to determine the invalid label global training value of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be called.

[0085] In one possible implementation, step A130 includes:

[0086] Step A131: In the deep knowledge learning operation based on the template scheduling learning data during the target training phase, the second training confidence of the training traffic flow scheduling network corresponding to the template scheduling learning data and the second label confidence of the traffic flow scheduling network to be invoked and the invalid label scheduling knowledge description are determined.

[0087] Step A132: Based on the error between the second training confidence and the second label confidence, determine the invalid label training value of the template scheduling learning data corresponding to the training traffic flow scheduling network relative to the traffic flow scheduling network to be called.

[0088] In this embodiment, for a specific highway service area, traffic flow and vehicle types are complex and variable, and the normal operation of the service area depends on efficient and reasonable traffic flow scheduling. This embodiment has a series of template scheduling learning data, which contains rich information, such as the dynamic data of the traffic flow to be scheduled in the service area, the invalid labeled scheduling knowledge description of the dynamic data of the traffic flow to be scheduled, and the valid labeled scheduling knowledge description of the dynamic data of the traffic flow to be scheduled.

[0089] First, we determine the first-sample performance index of the template scheduling learning data and the second-sample performance index of the previous template scheduling learning data. Taking a large highway service area as an example, we assume the current template scheduling learning data pertains to the dynamic traffic flow of this service area during major holidays. During this period, the service area faces immense traffic pressure, with vehicles originating from a wide range of sources, including vehicles from multiple surrounding cities and long-distance travelers. The vehicle types are diverse, including not only common cars, buses, and trucks, but also special-purpose vehicles such as RVs. Different types of vehicles exhibit vastly different driving habits and stopping needs, and traffic flow shows a complex fluctuation pattern throughout the day, gradually increasing from early morning, peaking at noon, and then gradually declining in the afternoon. The interaction between various facilities within the service area, such as refueling areas, catering areas, rest areas, and parking lots, and vehicles becomes highly complex. Based on these factors, the first-sample performance index of this template scheduling learning data is calculated to be 0.8.

[0090] Looking at the previous template's scheduling learning data, it concerns the traffic flow dynamics of the service area before a holiday. During this period, traffic volume is relatively low compared to the holiday season. While vehicle types are diverse, their proportions differ, with cars and some long-distance buses dominating, large trucks being relatively few, and special-purpose vehicles almost nonexistent. Traffic flow fluctuations are small, placing less pressure on service area facilities, and the interactions between vehicles are relatively simple. Considering these factors, the performance index for the second sample is calculated to be 0.6.

[0091] Next, in the deep knowledge learning stage based on the previous template scheduling learning data, the global training value of the past invalid annotations of the training traffic flow scheduling network corresponding to the previous template scheduling learning data is obtained. Assume that in the previous training process, this global training value of the past invalid annotations was determined to be 0.3.

[0092] Then, in the deep knowledge learning phase based on the current template scheduling learning data, the invalid labeled training value of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be called is determined. In this deep knowledge learning phase, for the traffic flow dynamic data of this service area during major holidays, both the training traffic flow scheduling network and the traffic flow scheduling network to be called process the invalid labeled scheduling knowledge descriptions. Invalid labeled scheduling knowledge descriptions may be scheduling strategies based on simple rules that do not fully consider the complexities of actual situations, such as allocating vehicles to parking lots in a fixed order without considering factors such as vehicle type, arrival time, and dwell time requirements.

[0093] In this process, the second training confidence of the training traffic flow scheduling network corresponding to the template scheduling learning data for obtaining invalid labeled scheduling knowledge descriptions, and the second label confidence of the traffic flow scheduling network to be invoked for obtaining invalid labeled scheduling knowledge descriptions, are determined. For example, based on its current structure and weights, the training traffic flow scheduling network obtains a second training confidence of 0.45 for the invalid labeled scheduling knowledge description of allocating vehicles to parking lots in a fixed order. This means that based on the calculations of the training traffic flow scheduling network, it considers the reliability of scheduling according to this simple rule to be 45%. The traffic flow scheduling network to be invoked, however, obtains a second label confidence of 0.4 for the same invalid labeled scheduling knowledge description.

[0094] The training value of invalid labels is determined based on the error between the second training confidence and the second label confidence. In this example, the error is 0.45 - 0.4 = 0.05. This 0.05 is the training value of invalid labels for the trainable traffic flow scheduling network corresponding to this template scheduling learning data compared to the traffic flow scheduling network to be called.

[0095] Finally, based on the performance metrics of the first and second samples, the training value of invalid labels and the global training value of past invalid labels are fused and calculated to determine the global training value of the trainable traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be called. Using a reasonable weighted calculation method, assuming the weight of the first sample performance metric is 0.7 and the weight of the second sample performance metric is 0.3, the calculation process is: (0.8 × 0.05 + 0.6 × 0.3) ÷ (0.8 + 0.6) = (0.04 + 0.18) ÷ 1.4 = 0.157. This 0.157 is the global training value of the trainable traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be called. This calculation method comprehensively considers the characteristics of the current template scheduling learning data, the characteristics of the previous template scheduling learning data, and the previous training results, thereby comprehensively and accurately evaluating the overall value of the trained traffic flow scheduling network relative to the traffic flow scheduling network to be called in terms of learning invalid labeled scheduling knowledge. This helps to more accurately adjust the network parameters during the training process of the traffic flow scheduling network, so as to improve the network's learning and processing capabilities for traffic flow scheduling.

[0096] In another possible implementation, the specific operation for determining the invalid labeled training value of the training traffic flow scheduling network corresponding to the template scheduling learning data in the deep knowledge learning stage based on the template scheduling learning data is as follows.

[0097] Taking traffic flow scheduling at highway service areas as an example, during the deep knowledge learning process based on template scheduling learning data in the target training phase, both the training traffic flow scheduling network and the traffic flow scheduling network to be called process invalid labeled scheduling knowledge descriptions for specific template scheduling learning data. For example, invalid labeled scheduling knowledge descriptions are simple and insufficiently optimized traffic flow scheduling strategies, such as allocating parking areas only based on the approximate size of vehicles (small, large), without considering the specific type of vehicle (such as the different needs of sports cars, family cars, and commercial vehicles among small cars, and the special needs of tour buses and long-distance passenger buses among large buses), the arrival time of vehicles (different handling methods for vehicles during peak and off-peak periods), and the purpose of vehicle stay (reasonable arrangements corresponding to different needs such as refueling, rest, and dining).

[0098] In this deep knowledge learning operation, the second training confidence of the training traffic flow scheduling network corresponding to the template scheduling learning data and the second label confidence of the traffic flow scheduling network to be invoked for the invalid labeled scheduling knowledge description are determined. Assuming that for the traffic flow dynamic data of a certain service area, the training traffic flow scheduling network, based on its internal network structure and weights, calculates a second training confidence of 0.4 for the invalid labeled scheduling knowledge description of allocating parking areas according to the approximate size of vehicles. This indicates that the training traffic flow scheduling network considers the scheduling based on this simple strategy to have a 40% reliability. Meanwhile, the traffic flow scheduling network to be invoked, for the same invalid labeled scheduling knowledge description, calculates a second label confidence of 0.35 through its own calculations.

[0099] The error between the second training confidence score and the second annotation confidence score is used to determine the invalid annotation training value of the trainable traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be called. In this example, the error is 0.4 - 0.35 = 0.05, and this 0.05 is the invalid annotation training value of the trainable traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be called. Through this calculation, the difference between the trainable traffic flow scheduling network and the traffic flow scheduling network to be called in terms of invalid annotation scheduling knowledge description can be accurately measured. This provides a basis for subsequent overall evaluation and network parameter adjustment, helping to improve the traffic flow scheduling network's ability to cope with complex traffic flow situations and optimize traffic flow scheduling strategies.

[0100] In one possible implementation, step S130 may further include:

[0101] Step B110: Determine the target sample performance index of the training allocation group where the template scheduling learning data is located, and the past sample performance index of the previous training allocation groups before the training allocation group.

[0102] Step B120: Obtain the invalid labeled past training value of the training traffic flow scheduling network corresponding to the past training allocation group relative to the traffic flow scheduling network to be called during the deep knowledge learning stage based on the past training allocation group in the target training stage.

[0103] Step B130: In the target training phase, based on the deep knowledge learning phase of the training allocation group, determine the invalid labeled target training value of the training traffic flow scheduling network corresponding to the training allocation group relative to the traffic flow scheduling network to be called.

[0104] Step B140: Based on the target sample performance index and the past sample performance index, perform a fusion calculation on the invalid labeled target training value and the invalid labeled past training value to determine the invalid labeled global training value of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be called.

[0105] In one possible implementation, step B110 includes:

[0106] Step B111: Determine the group spacing between the previous training assignment groups and the training assignment group.

[0107] Step B112: Based on the group spacing, determine the past sample performance indicators of the past training allocation groups. Except for the initial training allocation group, the past sample performance indicators of the remaining past training allocation groups have a negative correlation with the group spacing.

[0108] In this embodiment, service area traffic flow scheduling in a large-scale highway network is taken as an example. In this scenario, traffic flow scheduling needs to comprehensively consider numerous factors to ensure efficient and safe traffic. The learning process of traffic flow scheduling can be divided into different training assignment groups, each of which makes a unique contribution to the training of the traffic flow scheduling network.

[0109] First, the performance index of the target sample in the training allocation group containing the template scheduling learning data, as well as the performance index of previous samples from previous training allocation groups, are determined. For a specific highway service area, traffic flow exhibits complex variations across different time periods and dates. Assume that the training allocation group containing the template scheduling learning data being studied corresponds to the traffic flow dynamics of this service area during a specific time period (e.g., 2 PM to 6 PM on a weekend) during the peak tourist season. During this period, vehicles come from a wide range of sources, including self-driving tour vehicles, tour buses, and long-distance freight vehicles from surrounding cities. Vehicle types are diverse; small cars are further subdivided into family sedans, sports cars, and small commercial vehicles; large buses have different needs for tour buses and long-distance passenger buses; and large trucks have their own characteristics depending on the type of cargo and destination. The usage of facilities within the service area is complex, with significant challenges in coordinating refueling areas, catering areas, rest areas, and parking lots. Considering these factors, and after detailed calculations and evaluations, the performance index of the target sample for this training allocation group is determined to be 0.85.

[0110] Let's examine the previous training allocation groups before the current training allocation group, taking the closest previous training allocation group as an example. This previous training allocation group corresponds to the traffic flow dynamics data of the service area during the peak tourist season from 10:00 AM to 2:00 PM. Although the traffic flow is also large during this period, the composition ratio of vehicles is different compared to the afternoon. For example, the proportion of tour buses is relatively high in the morning, while the proportion of long-distance freight vehicles is relatively low. The interaction between vehicles also differs. For instance, in the morning, more tour buses arrive in concentrated numbers, and the flow of tourists after disembarking and the parking demand of vehicles are different from those in the afternoon. The group spacing between this previous training allocation group and the current training allocation group is determined to be 1 (this group spacing can be determined according to a pre-set order or logical relationship, such as sequential numbering according to time, with an adjacent group spacing of 1). Based on this group spacing, the past sample performance indicators of the previous training allocation group are determined. Since, apart from the initial training assignment group, the past sample performance metrics of each of the remaining past training assignment groups have a negative correlation with the group spacing, the past sample performance metrics will decrease as the group spacing increases. Assuming that based on this negative correlation, the calculated past sample performance metric for this past training assignment group is 0.75.

[0111] Next, during the deep knowledge learning phase based on the past training allocation groups, the invalid labeled past training value of the training traffic flow scheduling network corresponding to the past training allocation groups relative to the traffic flow scheduling network to be called is obtained. In the previous deep knowledge learning phase targeting the past training allocation groups (i.e., the groups corresponding to the traffic flow dynamic data from 10 AM to 2 PM), both the training traffic flow scheduling network and the traffic flow scheduling network to be called were processing invalid labeled scheduling knowledge descriptions. Invalid labeled scheduling knowledge descriptions may be simplistic scheduling strategies that do not fully consider the complexities of actual situations, such as allocating parking positions according to the order in which vehicles enter the service area, without considering factors such as vehicle type and purpose of stay. After calculation and analysis, the invalid labeled past training value of the training traffic flow scheduling network corresponding to the past training allocation groups relative to the traffic flow scheduling network to be called is determined to be 0.3.

[0112] Then, during the deep knowledge learning phase based on the training allocation group (i.e., the group corresponding to the traffic flow dynamic data from 2 PM to 6 PM), the training value of the training traffic flow scheduling network corresponding to the training allocation group relative to the invalid labeled target training of the traffic flow scheduling network to be called is determined. In this deep knowledge learning phase, both the training traffic flow scheduling network and the traffic flow scheduling network to be called process the invalid labeled scheduling knowledge descriptions. The invalid labeled scheduling knowledge descriptions are still simple scheduling strategies that do not fully consider various factors, such as allocating parking areas only based on the vehicle's size (small or large), without considering the specific type of vehicle (e.g., different models of small cars, different uses of large buses), the expected dwell time of the vehicle, and the special needs of the passengers inside. Through analysis of the calculation results of the training traffic flow scheduling network and the traffic flow scheduling network to be called, a certain confidence level of 0.4 is determined for the invalid labeled scheduling knowledge description obtained by the training traffic flow scheduling network, and a corresponding confidence level of 0.35 is determined for the invalid labeled scheduling knowledge description obtained by the traffic flow scheduling network to be called. The difference between the two is 0.4 - 0.35 = 0.05. This 0.05 is the training value of the invalid labeled target of the training traffic flow scheduling network corresponding to the training allocation group compared with the traffic flow scheduling network to be called.

[0113] Finally, based on the performance metrics of the target samples and past samples, the training value of invalid labeled targets and the training value of invalid labeled past samples are fused together to determine the global training value of the trainable traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be called. Using a reasonable weighted calculation method, assuming the weight of the target sample performance metric is 0.6 and the weight of the past sample performance metric is 0.4, the calculation process is: (0.85×0.05 + 0.75×0.3)÷(0.85 + 0.75)= (0.0425 + 0.225)÷1.6 = 0.167. This 0.167 is the global training value of the trainable traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be called. This calculation method comprehensively considers the characteristics of the training allocation group itself (target sample performance indicators), the characteristics of the previous training allocation groups (past sample performance indicators), and the training results at different depths of knowledge learning stages (invalid label target training value and invalid label past training value). This allows for an accurate assessment of the overall value of the trained traffic flow scheduling network relative to the traffic flow scheduling network to be called in terms of invalid label scheduling knowledge learning. This helps to more accurately adjust network parameters during the training process of the traffic flow scheduling network and improve the entire network's learning and processing capabilities for traffic flow scheduling.

[0114] When determining the past performance metrics of previous training allocation groups before the current training allocation group, specific rules must be followed. Taking traffic flow scheduling at a highway service area as an example, when calculating the past performance metrics, the group spacing between the previous training allocation groups and the current training allocation group must first be determined. For example, in the traffic flow scheduling learning process of a certain service area, training allocation groups are configured in chronological order, with each training allocation group corresponding to traffic flow dynamic data for different time periods. Suppose the currently interested training allocation group is the 5th group, which corresponds to the traffic flow dynamic data of the service area on the afternoon of a specific date. The past training allocation group to be determined in this embodiment is the 3rd group, which corresponds to the traffic flow dynamic data of the service area in the morning. Through the pre-set order relationship, the group spacing between these two groups can be determined to be 2 (5th group - 3rd group = 2).

[0115] Based on this group spacing, the past sample performance index of the previous training allocation group is determined. Since the past sample performance index of each of the remaining previous training allocation groups (excluding the initial training allocation group) is negatively correlated with the group spacing, this means that as the group spacing increases, the past sample performance index decreases. Assuming some initial basic parameters and calculation rules, when the group spacing is 1, the initial value of the past sample performance index is 0.8 (this 0.8 is a base value obtained based on a comprehensive evaluation of the initial traffic flow in the service area). With each increase of 1 in the group spacing, the past sample performance index decreases by a certain proportion, for example, by 0.05. Therefore, when the group spacing is 2, the past sample performance index of the previous training allocation group is 0.8 - 0.05 × 2 = 0.7. This method of determining past sample performance indicators takes into account the relationship between different training assignment groups and the fact that the influence of past training assignment groups on the current training gradually weakens as the training progresses. This helps to more accurately evaluate and adjust network parameters during the training process of the entire traffic flow scheduling network, so as to achieve more efficient traffic flow scheduling learning.

[0116] In one possible implementation, step S110 includes:

[0117] Step S111: Obtain the traffic flow scheduling task and dynamic data of multiple traffic flows to be scheduled.

[0118] In this embodiment, the traffic flow scheduling task is set based on the operational needs of the entire highway network. For example, in a highway network covering multiple cities and their surrounding areas, the traffic flow scheduling task may be to ensure smooth traffic flow during peak hours (such as morning and evening rush hours on weekdays) and special periods (such as holidays and large-scale events), while ensuring the normal operation of service areas. The dynamic data of the traffic flow to be scheduled contains a wealth of detailed information. Taking a certain highway service area as an example, this data includes the number of vehicles entering and leaving the service area at different times, vehicle types (such as small cars can be further subdivided into family cars, sports cars, small commercial vehicles, etc.; large buses include tour buses, long-distance passenger buses, etc.; large trucks are further subdivided according to cargo type, load, whether it is cold chain transportation, etc.), the origin of the vehicles (which cities, which highway branches, etc.), destination information, the dwell time of the vehicles in the service area, the behavior patterns of the people in the vehicles (such as whether they are eating, resting, refueling, etc.), and the driving trajectory of the vehicles in the service area, etc. The collection of this dynamic data on traffic flow to be dispatched is accomplished through sensors (such as vehicle counters, license plate recognition systems, speed sensors, etc.) and other related equipment (such as parking management systems that record parking duration) installed at various key locations on highways and service areas.

[0119] Step S112: For each piece of dynamic traffic data to be scheduled, determine scheduling session information that matches the traffic scheduling task. The scheduling session information includes the dynamic traffic data to be scheduled.

[0120] The dispatch session information contains dynamic data on traffic flow to be dispatched and is constructed according to the specific requirements of the traffic flow dispatch task. For example, for traffic flow dispatch tasks during peak hours, the dispatch session information will focus on the distribution of vehicle traffic during peak hours, the behavior patterns of different types of vehicles during peak hours, and how to meet the basic needs of vehicles in service areas while ensuring smooth traffic flow. If the traffic flow dispatch task is to ensure the safety and smooth flow of highways during large-scale events, the dispatch session information will consider more the dispatch of special vehicles (such as event-related transport vehicles, security vehicles, etc.) and how to avoid conflicts between ordinary vehicles and special vehicles. Taking the highway traffic flow dispatch during a large-scale sporting event as an example, for vehicles entering service areas, the dispatch session information will be constructed according to the event schedule (such as different vehicle traffic flow and demand patterns before, during, and after the event). Before the start of the event, a large number of spectators may drive to the venue. At this time, the dispatch session information will focus on the traffic flow and parking demand of small cars, and how to guide these vehicles to quickly pass through service areas or stop for rest. During the competition, traffic flow may decrease relatively, but some special vehicles (such as vehicles for event staff, emergency rescue vehicles, etc.) still need to be dispatched. The dispatch session information will then develop special dispatch strategies for these special vehicles. After the competition, there will be a return peak, and the dispatch session information needs to consider how to coordinate the return of different types of vehicles to avoid congestion.

[0121] Step S113: Load the scheduling session information into the previously generated traffic flow scheduling knowledge network, and use the obtained traffic flow scheduling knowledge network as an invalid annotation scheduling knowledge description of the dynamic data of the traffic flow to be scheduled.

[0122] The traffic flow scheduling knowledge network is built upon historical traffic data, basic traffic rules, and some preliminary scheduling strategies. This network contains numerous nodes and connections. Each node represents a traffic element (such as vehicle type, road conditions, service area facilities, etc.), and connections represent the relationships and interactions between these elements. When scheduling session information is loaded into this traffic flow scheduling knowledge network, the network generates a scheduling knowledge description based on its internal algorithms and logic. However, this description may be based on incomplete information or simplistic rules, and is therefore considered an invalid labeled scheduling knowledge description. For example, the traffic flow scheduling knowledge network might allocate parking spaces solely based on vehicle type and the order in which they enter the service area, without fully considering the vehicle's purpose of stay, the behavior patterns of passengers, or the current usage status of various facilities within the service area. In the case of large-scale sporting events, such an invalid labeled scheduling knowledge description might simply guide all cars to a single parking lot without distinguishing between vehicles going to or leaving the stadium, or considering the distribution of available parking spaces and the ease of vehicle entry and exit.

[0123] Step S114: Based on the adjustment instruction generated for the invalid labeled scheduling knowledge description of the dynamic data of the traffic flow to be scheduled, obtain the valid labeled scheduling knowledge description of the dynamic data of the traffic flow to be scheduled.

[0124] Adjustment instructions are generated by analyzing invalid labeled scheduling knowledge descriptions, combined with more comprehensive traffic data, practical operational experience, and more complex traffic demand assessments. For example, in cases where invalid labeled scheduling knowledge descriptions allocate parking positions according to simple rules, adjustment instructions may consider vehicle destination information, purpose of stay, and actual parking usage within the service area. During major sporting events, if invalid labeled scheduling knowledge descriptions guide all cars to a single parking lot, adjustment instructions might, based on whether the vehicles are heading to or leaving the stadium, guide vehicles heading to the parking lot near the highway exit for quick departure after the event, and guide vehicles leaving the stadium to the parking lot near service area facilities (such as food courts and restrooms) for passengers to rest and make purchases before leaving. Simultaneously, adjustment instructions also consider the distribution of available spaces in different areas of the parking lot, guiding vehicles to areas with more available spaces and convenient access. In this way, the valid labeled scheduling knowledge descriptions generated based on adjustment instructions can more comprehensively and rationally plan scheduling based on dynamic traffic flow data.

[0125] Step S115: Generate template scheduling learning data containing the dynamic data of the traffic flow to be scheduled, the effective labeled scheduling knowledge description of the dynamic data of the traffic flow to be scheduled, and the invalid labeled scheduling knowledge description of the dynamic data of the traffic flow to be scheduled, to obtain a set of templates to be learned containing multiple template scheduling learning data.

[0126] For each traffic flow dynamic data to be scheduled, corresponding template scheduling learning data is generated following the steps described above. For example, based on traffic flow dynamic data from different time periods (such as weekdays, holidays, and special events) and different service areas, separate template scheduling learning data are generated. These template scheduling learning data contain rich information, including key traffic flow dynamic data to be scheduled, as well as valid and invalid labeled scheduling knowledge descriptions generated based on different logics. Combining these template scheduling learning data forms a set of templates to be learned, containing multiple template scheduling learning data. This set of templates to be learned will serve as the foundational data for subsequent traffic flow scheduling network training. By allowing the traffic flow scheduling network to learn the differences between valid and invalid labeled scheduling knowledge descriptions in these data, the network's scheduling capabilities are continuously optimized to achieve more efficient and rational highway traffic flow scheduling.

[0127] In one possible implementation, step S110 further includes:

[0128] Obtain the previously generated deep learning network.

[0129] Based on the set of templates to be learned, an initial set of templates to be learned is generated, which includes multiple initial template scheduling learning data. Each initial template scheduling learning data includes dynamic traffic flow data to be scheduled, and either a valid labeled scheduling knowledge description or an invalid labeled scheduling knowledge description of the dynamic traffic flow data to be scheduled.

[0130] The deep learning network is trained on the network parameters based on the initial set of templates to be learned, and the traffic flow scheduling network to be called is obtained.

[0131] In this embodiment, firstly, a pre-generated deep learning network is obtained. This deep learning network is constructed based on previous research or pre-training and has a specific network structure, such as a multi-layered neural network containing an input layer, several hidden layers, and an output layer. The number of neurons in the input layer is determined based on the number of features of the input data related to traffic flow scheduling that can be obtained. These features may include traffic volume, vehicle speed, and vehicle type ratios for different sections of the highway. The number of hidden layers and the number of neurons in each layer are designed and optimized to handle complex traffic flow scheduling relationships. The output layer is used to output decision results related to traffic flow scheduling, such as vehicle guidance directions and parking area allocations. The initial weights and biases of this deep learning network may be randomly initialized or determined based on some general initialization methods, the purpose of which is to provide an initial state for subsequent learning.

[0132] Based on the set of templates to be learned, an initial set of templates to be learned is generated, containing multiple initial template scheduling learning data. The set of templates to be learned contains rich traffic flow scheduling-related information, as mentioned earlier. Each template scheduling learning data includes dynamic traffic flow data to be scheduled, validly labeled scheduling knowledge descriptions of the dynamic traffic flow data to be scheduled, and invalidly labeled scheduling knowledge descriptions of the dynamic traffic flow data to be scheduled. When generating the initial set of templates to be learned, data is selected from the set to be learned to construct the initial template scheduling learning data. Each initial template scheduling learning data includes dynamic traffic flow data to be scheduled, and either a validly labeled scheduling knowledge description or an invalidly labeled scheduling knowledge description of the dynamic traffic flow data to be scheduled. For example, for traffic flow dynamic data of a highway service area, the initial template scheduling learning data might only contain traffic flow dynamic data and effectively labeled scheduling knowledge descriptions. The traffic flow dynamic data covers information such as the type, number, origin, and destination of vehicles entering and leaving the service area within a specific time period. The effectively labeled scheduling knowledge descriptions are reasonable scheduling strategies formulated based on various vehicle characteristics and the service area's facility usage, such as guiding vehicles to different parking areas according to vehicle type and purpose of stay, or planning their departure routes based on their destinations. Alternatively, another initial template scheduling learning data might contain traffic flow dynamic data and invalid labeled scheduling knowledge descriptions. The invalid labeled scheduling knowledge descriptions might be a simple scheduling strategy, such as allocating parking positions according to the order in which vehicles enter the service area, without considering other vehicle characteristics.

[0133] The deep learning network learns its parameters based on an initial set of learning templates to obtain a traffic flow scheduling network. In this process, data from the initial learning template set is input into the deep learning network. For each initial template, the dynamic traffic flow data to be scheduled is input into the input layer of the deep learning network. The network performs forward propagation calculations based on its current weights and biases, obtaining an output result. This output result is compared with either valid or invalid labeled scheduling knowledge descriptions in the initial template learning data to calculate a loss value. For example, if the initial template learning data contains valid labeled scheduling knowledge descriptions, the output result of the deep learning network should be as close as possible to the optimal scheduling strategy represented by these descriptions. If the output result differs significantly from the descriptions, a large loss value will be generated. Based on this loss value, the weights and biases of the deep learning network are adjusted using the backpropagation algorithm. This process is repeated continuously. As more initial template learning data is input into the network, the weights and biases of the deep learning network are continuously optimized, thereby continuously improving the network's predictive ability for traffic flow scheduling. After learning and adjusting all the data in the initial set of templates to be learned multiple times, the final deep learning network is the traffic flow scheduling network to be called. This network can make effective predictions and decisions on traffic flow scheduling to a certain extent, providing support for subsequent traffic flow scheduling tasks.

[0134] In one possible implementation, step S120 includes:

[0135] Step S121: In the deep knowledge learning operation based on the template scheduling learning data during the target training phase, the first training confidence of the training traffic flow scheduling network corresponding to the template scheduling learning data and the first labeling confidence of the traffic flow scheduling network to be invoked to obtain the effective labeled scheduling knowledge description are determined.

[0136] In this embodiment, taking the traffic flow scheduling of a specific highway service area as an example, the template scheduling learning data involved includes important information such as the dynamic data of the traffic flow to be scheduled in the service area and the effective labeled scheduling knowledge description of the dynamic data of the traffic flow to be scheduled.

[0137] First, for the training traffic flow scheduling network, when processing the dynamic traffic flow data of the service area, it calculates the effective labeled scheduling knowledge description based on its internal network structure and parameters. The effective labeled scheduling knowledge description is an optimal scheduling strategy based on multiple factors, such as vehicle type, origin, destination, purpose of stay in the service area, stay time, and the real-time usage status of various facilities within the service area (such as parking lots, refueling areas, and catering areas). Assuming that the service area is in peak tourist season, the effective labeled scheduling knowledge description is to rationally allocate parking areas according to vehicle type and expected stay time to avoid traffic congestion and improve service efficiency. The training traffic flow scheduling network calculates the first training confidence level for this effective labeled scheduling knowledge description through its multi-layered neurons. For example, if the network calculates that its confidence level in this scheduling strategy is 0.75, this means that the first training confidence level is 0.75, implying that the training traffic flow scheduling network believes that the probability of correctly scheduling according to this effective labeled scheduling knowledge description is 75%.

[0138] Simultaneously, the traffic flow scheduling network to be invoked also calculates the first label confidence score for the same valid labeled scheduling knowledge description. This network, which has undergone prior learning and training, also evaluates the valid labeled scheduling knowledge description based on multiple factors according to its own network structure and parameters. Assuming the first label confidence score calculated by the network to be invoked is 0.65, this indicates that the network believes the probability of correctly scheduling according to this valid labeled scheduling knowledge description is 65%.

[0139] Step S122: Based on the error between the first training confidence and the first annotation confidence, determine the effective annotation training value of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be called.

[0140] In the example above, the first training confidence score is 0.75, and the first annotation confidence score is 0.65. The error between the two is 0.75 - 0.65 = 0.1. This 0.1 represents the effective annotation training value of the trained traffic flow scheduling network corresponding to the template scheduling learning data compared to the traffic flow scheduling network to be called. This effective annotation training value reflects the improvement of the trained traffic flow scheduling network in learning effective annotation scheduling knowledge descriptions compared to the traffic flow scheduling network to be called, under the template scheduling learning data. If this value is large, it indicates that the trained traffic flow scheduling network has significant room for improvement or has already made significant progress in learning this effective annotation scheduling knowledge description. This can serve as an important basis for adjusting network parameters during subsequent network training, helping to further optimize the trained traffic flow scheduling network's ability to learn effective annotation scheduling knowledge descriptions, thereby improving the accuracy and effectiveness of the entire traffic flow scheduling network in traffic flow scheduling tasks.

[0141] In one possible implementation, the method further includes:

[0142] For each template scheduling learning data, during the deep knowledge learning operation based on the template scheduling learning data in the target training phase, the invalid labeled training value of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be invoked is determined.

[0143] Step S140 includes:

[0144] When the training error in the target training stage no longer decreases, based on the effective labeled training value, invalid labeled training value, and invalid labeled global training value corresponding to each template scheduling learning data, deep knowledge learning is terminated, and the target traffic flow scheduling network corresponding to the traffic flow scheduling network to be called is obtained.

[0145] In this embodiment, taking the traffic flow scheduling of a highway service area as an example, during the target training phase, each template scheduling learning data includes dynamic data of the traffic flow to be scheduled in the service area, invalid labeled scheduling knowledge descriptions of the dynamic data of the traffic flow to be scheduled, and other contents.

[0146] When performing deep knowledge learning on template scheduling training data, both the training traffic flow scheduling network and the traffic flow scheduling network to be called will process invalid labeled scheduling knowledge descriptions. Invalid labeled scheduling knowledge descriptions are relatively simple or not optimized scheduling strategies. For example, in service area traffic flow scheduling, invalid labeled scheduling knowledge descriptions may allocate parking positions sequentially according to the order in which vehicles enter the service area, without considering complex factors such as vehicle type and purpose of stay.

[0147] The training network for traffic flow scheduling calculates the results based on its network structure and parameters using this invalid labeled scheduling knowledge description. For example, suppose that when processing traffic flow dynamic data for a certain service area, the training network calculates a confidence level of 0.4 for this scheduling method using this simple sequential parking invalid labeled scheduling knowledge description through the calculation of neurons within the network. This is a result of the training network for traffic flow scheduling based on template scheduling learning data obtaining the invalid labeled scheduling knowledge description, which can be understood as a manifestation of confidence.

[0148] Simultaneously, the traffic flow scheduling network to be invoked will also calculate the same invalid labeled scheduling knowledge descriptions and obtain the corresponding results. Assume the confidence level obtained by the traffic flow scheduling network to be invoked is 0.35.

[0149] Then, based on the difference between these two results, the invalid label training value of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be called is determined. In this example, the difference is 0.4 - 0.35 = 0.05, and this 0.05 is the invalid label training value of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be called. This invalid label training value reflects the difference between the training traffic flow scheduling network and the traffic flow scheduling network to be called in terms of invalid label scheduling knowledge description, which is of great significance for evaluating the learning progress of the training traffic flow scheduling network and adjusting the network training strategy.

[0150] Next, based on the effective labeled training value, invalid labeled training value, and invalid labeled global training value corresponding to each template scheduling learning data, when it is determined that the training error in the target training stage no longer continues to decrease, deep knowledge learning is terminated, and the target traffic flow scheduling network corresponding to the traffic flow scheduling network to be called is obtained.

[0151] Throughout the target training phase, numerous template scheduling learning datasets are used for training. For each template scheduling learning dataset, its corresponding effective labeled training value, invalid labeled training value, and invalid labeled global training value are calculated. These values ​​reflect the performance of the trained traffic flow scheduling network relative to the traffic flow scheduling network to be invoked in different aspects.

[0152] Training error is a crucial metric for evaluating the learning effectiveness of a traffic flow scheduling network. During training, the training error gradually changes as the network learns from template scheduling data. For example, in the initial stages of training, the training error may be relatively large due to the network's initial weights and parameter settings. As the network learns from the template scheduling data, it continuously adjusts its weights and parameters, and the training error gradually decreases.

[0153] When considering the effective labeled training value, invalid labeled training value, and invalid labeled global training value of each template scheduling learning data, these values ​​will affect the trend of training error. For example, if the effective labeled training value is high, it indicates that the training traffic scheduling network has performed well in learning the effective labeled scheduling knowledge description, which may reduce the training error; if the invalid labeled training value or the invalid labeled global training value shows abnormal changes, it will also affect the training error.

[0154] When, after multiple rounds of training, the training error no longer decreases, it indicates that the traffic flow scheduling network has reached a relatively stable state under the current training data and strategy. At this point, the deep knowledge learning process is terminated. The network obtained through this training is the target traffic flow scheduling network corresponding to the traffic flow scheduling network to be called. This target traffic flow scheduling network can better handle highway traffic flow scheduling tasks. For example, when processing new service area traffic flow dynamic data, it can provide more reasonable scheduling decisions based on various vehicle characteristics and the actual situation of the service area, including vehicle parking guidance and route planning, thereby improving the efficiency and rationality of highway traffic flow scheduling.

[0155] Figure 2 The diagram illustrates the hardware structure of a highway service area traffic flow scheduling system 100 based on an intelligent algorithm, provided by an embodiment of the present invention, for implementing the aforementioned intelligent algorithm-based highway service area traffic flow scheduling method. Figure 2 As shown, the highway service area traffic flow scheduling system 100 based on intelligent algorithms may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

[0156] Machine-readable storage medium 120 may store data and / or instructions. In some embodiments, machine-readable storage medium 120 may store data acquired from an external terminal. In some embodiments, machine-readable storage medium 120 may store data and / or instructions used by the intelligent algorithm-based highway service area traffic flow scheduling system 100 to perform or use in order to accomplish the exemplary methods described in this invention.

[0157] In the specific implementation process, one or more processors 110 execute computer-executable instructions stored in machine-readable storage medium 120, so that processor 110 can execute the highway service area traffic flow scheduling method based on intelligent algorithm as described in the above method embodiment. Processor 110, machine-readable storage medium 120 and communication unit 140 are connected through bus 130. Processor 110 can be used to control the sending and receiving actions of communication unit 140.

[0158] The specific implementation process of processor 110 can be found in the various method embodiments executed by the above-mentioned highway service area traffic flow scheduling system 100 based on intelligent algorithms. The implementation principle and technical effect are similar, and will not be repeated here.

[0159] Furthermore, this embodiment of the invention also provides a readable storage medium, wherein computer-executable instructions are preset in the readable storage medium, and when the processor executes the computer-executable instructions, the above-mentioned highway service area traffic flow scheduling method based on intelligent algorithms is implemented.

[0160] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A method for traffic flow scheduling in highway service areas based on intelligent algorithms, characterized in that, The method includes: The system acquires the traffic flow scheduling network to be invoked and a set of templates to be learned, which includes multiple template scheduling learning data. The template scheduling learning data includes dynamic traffic flow data to be scheduled in highway service areas, invalid labeled scheduling knowledge descriptions of the dynamic traffic flow data to be scheduled, and valid labeled scheduling knowledge descriptions of the dynamic traffic flow data to be scheduled. The dynamic traffic flow data to be scheduled includes dynamic change data presented by different vehicle sources, different vehicle types, and different working time periods. The scheduling knowledge descriptions include the guidance route planning for vehicles entering the service area, the precise allocation of parking areas, the driving route planning for different vehicles, and the scheduling decision information for setting the signage for leaving the service area. For each template scheduling learning data in the set of templates to be learned, during the deep knowledge learning operation based on the template scheduling learning data in the target training phase, the effective labeled training value of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be called is determined; the training traffic flow scheduling network corresponding to the template scheduling learning data is obtained by deep knowledge learning based on the forward training allocation group of the training allocation group to which the template scheduling learning data belongs; in the first deep knowledge learning phase, the network learning target of the initial training allocation group is the traffic flow scheduling network to be called. Based on the template scheduling learning data and the deep knowledge learning stage preceding the template scheduling learning data, determine the global training value of the invalid annotations of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be invoked; When the training error in the target training phase no longer decreases, based on the effective labeled training value and the invalid labeled global training value corresponding to each template scheduling learning data, deep knowledge learning is terminated, and the target traffic flow scheduling network corresponding to the traffic flow scheduling network to be invoked is obtained; the target traffic flow scheduling network is used to determine the scheduling knowledge description of the target traffic flow dynamic data.

2. The method for traffic flow scheduling in highway service areas based on intelligent algorithms according to claim 1, characterized in that, The step of determining the invalid annotation global training value of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be invoked, based on the template scheduling learning data and the deep knowledge learning stage preceding the template scheduling learning data, includes: Determine the first sample performance index of the template scheduling learning data, and the second sample performance index of the previous template scheduling learning data; In the deep knowledge learning stage based on the previous template scheduling learning data, the global training value of the training traffic flow scheduling network corresponding to the previous template scheduling learning data relative to the traffic flow scheduling network to be called is obtained from the previous invalid annotations. In the deep knowledge learning stage based on the template scheduling learning data, the invalid labeled training value of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be invoked is determined. Based on the first sample performance index and the second sample performance index, the invalid label training value and the past invalid label global training value are fused and calculated to determine the invalid label global training value of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be called.

3. The method for traffic flow scheduling in highway service areas based on intelligent algorithms according to claim 2, characterized in that, In the deep knowledge learning stage based on the template scheduling learning data, determining the invalid labeled training value of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be invoked includes: In the deep knowledge learning operation based on the template scheduling learning data during the target training phase, the second training confidence of the training traffic flow scheduling network corresponding to the template scheduling learning data that obtains invalid labeled scheduling knowledge description, and the second label confidence of the traffic flow scheduling network to be invoked that obtains invalid labeled scheduling knowledge description; Based on the error between the second training confidence and the second label confidence, the invalid label training value of the template scheduling learning data corresponding to the training traffic flow scheduling network relative to the traffic flow scheduling network to be called is determined.

4. The highway service area traffic flow scheduling method based on intelligent algorithms according to claim 1, characterized in that, The step of determining the invalid annotation global training value of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be invoked, based on the template scheduling learning data and the deep knowledge learning stage preceding the template scheduling learning data, includes: Determine the target sample performance index of the training allocation group where the template scheduling learning data is located, and the past sample performance index of the previous training allocation groups before the training allocation group. In the deep knowledge learning phase based on the past training allocation group, the invalid labeled past training value of the training traffic flow scheduling network corresponding to the past training allocation group relative to the traffic flow scheduling network to be called is obtained during the target training phase. In the target training phase, based on the deep knowledge learning phase of the training allocation group, the invalid labeled target training value of the training traffic flow scheduling network corresponding to the training allocation group relative to the traffic flow scheduling network to be called is determined. Based on the target sample performance index and the past sample performance index, the invalid labeled target training value and invalid labeled past training value are fused and calculated to determine the invalid labeled global training value of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be called.

5. The highway service area traffic flow scheduling method based on intelligent algorithms according to claim 4, characterized in that, Determine the past sample performance metrics of the previous training assignment groups before the training assignment group, including: Determine the group spacing between previous training assignment groups and the current training assignment group; Based on the group spacing, the past sample performance indicators of the past training allocation groups are determined; except for the initial training allocation group, the past sample performance indicators of the remaining past training allocation groups are negatively correlated with the group spacing.

6. The method for traffic flow scheduling in highway service areas based on intelligent algorithms according to claim 1, characterized in that, The steps to obtain a set of templates to be learned that contain multiple template scheduling learning data include: Acquire traffic flow scheduling tasks and dynamic data of multiple traffic flows to be scheduled; For each piece of dynamic traffic flow data to be scheduled, determine scheduling session information that matches the traffic flow scheduling task; the scheduling session information includes the dynamic traffic flow data to be scheduled. The scheduling session information is loaded into the previously generated traffic flow scheduling knowledge network, and the output of the traffic flow scheduling knowledge network is used as the invalid annotation scheduling knowledge description of the dynamic data of the traffic flow to be scheduled. Based on the adjustment instruction generated from the invalid labeled scheduling knowledge description of the dynamic data of the traffic flow to be scheduled, the valid labeled scheduling knowledge description of the dynamic data of the traffic flow to be scheduled is obtained. Generate template scheduling learning data that includes the dynamic data of the traffic flow to be scheduled, the effective labeled scheduling knowledge description of the dynamic data of the traffic flow to be scheduled, and the invalid labeled scheduling knowledge description of the dynamic data of the traffic flow to be scheduled, to obtain a set of templates to be learned that contains multiple template scheduling learning data.

7. The method for traffic flow scheduling in highway service areas based on intelligent algorithms according to claim 1, characterized in that, The steps to obtain the traffic flow scheduling network to be invoked include: Obtain the previously generated deep learning network; Based on the set of templates to be learned, an initial set of templates to be learned is generated, which contains multiple initial template scheduling learning data; each initial template scheduling learning data includes dynamic traffic flow data to be scheduled, and also includes either a valid labeled scheduling knowledge description or an invalid labeled scheduling knowledge description of the dynamic traffic flow data to be scheduled. The deep learning network is trained on the network parameters based on the initial set of templates to be learned, and the traffic flow scheduling network to be called is obtained.

8. The method for traffic flow scheduling in highway service areas based on intelligent algorithms according to claim 1, characterized in that, In the deep knowledge learning operation based on the template scheduling learning data during the target training phase, determining the effective labeled training value of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be invoked includes: In the deep knowledge learning operation based on the template scheduling learning data during the target training phase, the first training confidence of the training traffic flow scheduling network corresponding to the template scheduling learning data to obtain an effective labeled scheduling knowledge description, and the first labeling confidence of the traffic flow scheduling network to be invoked to obtain an effective labeled scheduling knowledge description; Based on the error between the first training confidence and the first annotation confidence, the effective annotation training value of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be called is determined.

9. The highway service area traffic flow scheduling method based on intelligent algorithms according to any one of claims 1-8, characterized in that, The method further includes: For each template scheduling learning data, during the deep knowledge learning operation based on the template scheduling learning data in the target training phase, the invalid labeled training value of the training traffic flow scheduling network corresponding to the template scheduling learning data relative to the traffic flow scheduling network to be called is determined; When the training error in the target training stage no longer decreases after determining, based on the effective labeled training value and the invalid labeled global training value corresponding to each template scheduling learning data, the deep knowledge learning is terminated, and the target traffic flow scheduling network corresponding to the traffic flow scheduling network to be invoked is obtained, including: When the training error in the target training stage no longer decreases, based on the effective labeled training value, invalid labeled training value, and invalid labeled global training value corresponding to each template scheduling learning data, deep knowledge learning is terminated, and the target traffic flow scheduling network corresponding to the traffic flow scheduling network to be called is obtained.

10. A highway service area traffic flow scheduling system based on intelligent algorithms, characterized in that, The intelligent algorithm-based highway service area traffic flow scheduling system includes a processor and a memory, the memory and the processor being connected. The memory is used to store programs, instructions or code, and the processor is used to execute the programs, instructions or code in the memory to implement the intelligent algorithm-based highway service area traffic flow scheduling method according to any one of claims 1-9.

Images