Eta prediction and model training method, device, medium and product

Abstract

Embodiments of the present disclosure disclose an ETA prediction and model training method, device, medium and product. The method comprises: obtaining a road section type of each road section in a target route; calculating a predicted value of a travel time of each road section at each future time slice after a departure time based on a road section level prediction model corresponding to the road section type of each road section; determining an estimated entering time of each road section based on the predicted value of the travel time of each road section at each future time slice; accumulating the predicted value of the travel time of each road section at the estimated entering time to obtain an estimated travel time of the target route at the road section level; and predicting an estimated arrival time corresponding to the target route based on the estimated travel time of the target route at the road section level and a related travel feature of the target route using a route level prediction model. The technical solution can more accurately predict the estimated arrival time corresponding to the target route.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of data processing, in particular to an ETA prediction and model training method, device, medium and product. BACKGROUND

[0002] ETA (Estimated Time of Arrival) is an estimate of the duration of a planned route. Generally, navigation applications also use ETA prediction algorithms to calculate the ETA of a user reaching a destination by a corresponding navigation route when planning a navigation route for the user. For navigation applications with high daily activity, the ETA prediction algorithm has a considerable call frequency, so the accuracy and calculation efficiency of the ETA prediction algorithm are highly required. The current ETA prediction algorithm usually directly uses a prediction model to predict the current travel time of each road segment on the navigation route, and accumulates to obtain the ETA of the navigation route. In this prediction algorithm, the current travel time of each different road segment is predicted by a unified prediction model, but in fact, many stable travel time road segments can quickly predict the results by a simple model, so the existing ETA prediction algorithm uses a unified complex prediction model for different types of road segments, which wastes computing resources and has low calculation efficiency, and the existing model predicts the current travel time of each road segment, without considering the specific time of entering the road segment, so the ETA predicted in this way is inaccurate. SUMMARY

[0003] To solve the problems in the related art, the embodiments of the present disclosure provide an ETA prediction and model training method, device, medium and product.

[0004] In a first aspect, an ETA prediction method is provided in the embodiments of the present disclosure.

[0005] Specifically, the ETA prediction method comprises:

[0006] obtaining road segment types of road segments in a target route;

[0007] calculating travel time prediction values of the road segments at each future time slice after a departure time based on road segment level prediction models corresponding to the road segment types of the road segments;

[0008] determining predicted entering time of each road segment based on the travel time prediction values of the road segments at each future time slice;

[0009] accumulating the travel time prediction values of the road segments at corresponding target time slices to obtain an estimated travel time of the target route at a road segment level, wherein the target time slices corresponding to the road segments are future time slices in which the predicted entering time of each road segment is located;

[0010] The route-level prediction model is used to predict an estimated arrival time corresponding to the target route based on the estimated travel time of the route segments of the target route and the relevant travel characteristics of the target route.

[0011] In a possible implementation, the route segment types include at least one of the following types: a low-frequency route segment, a stable low-flow route segment, a short-time burst congestion route segment, and a long-time burst congestion route segment.

[0012] When the route segment type of the route segment is a low-frequency route segment, the historical average travel time of the route segment is used to obtain the travel time prediction value of the route segment at each future time slice.

[0013] When the route segment type of the route segment is a stable low-flow route segment, a linear model corresponding to the scenario in which the route segment is located is used to obtain the travel time prediction value of the route segment at each future time slice based on the travel condition characteristics of the route segment.

[0014] When the route segment type of the route segment is a short-time burst congestion route segment, a short-term prediction deep model is used to obtain the travel time prediction value of the route segment at each future time slice based on the short-time condition characteristics of the route segment.

[0015] When the route segment type of the route segment is a long-time burst congestion route segment, a long-term prediction deep model is used to obtain the travel time prediction value of the route segment at each future time slice based on the long-time condition characteristics of the route segment.

[0016] In a possible implementation, the method further includes:

[0017] The scenario in which the route segment is located is determined based on at least one of the following characteristics of the route segment: a region characteristic in which the route segment is located, a time type of the departure time, and a road level of the route segment.

[0018] In a possible implementation, the route segment types include at least one of the following types under a first-level route segment type: a low-frequency route segment, a stable low-flow route segment, a short-time burst congestion route segment, and a long-time burst congestion route segment, and the first-level route segment type includes an ordinary route segment or a road intersection direction route segment.

[0019] When the road section type of the road section is a stable low-flow road section under a corresponding primary road section type, a linear model corresponding to the stable low-flow road section under the corresponding primary road section type is used to obtain predicted values of travel time of the road section at each future time slice based on travel road condition features of the road section under the corresponding primary road section type, wherein the travel road condition features of the road section under the ordinary road section type include at least one of real-time travel features and a historical average travel time sequence corresponding to each future time slice; and the travel road condition features of the road section under the intersection directional road section type include at least one of real-time travel features, a historical average travel time sequence corresponding to each future time slice, an intersection type, and a turning action type.

[0020] When the road section type of the road section is a short-time burst congestion road section under a corresponding primary road section type, a short-term prediction depth model corresponding to the corresponding primary road section type is used to obtain predicted values of travel time of the road section at each future time slice based on short-time road condition features of the road section under the corresponding primary road section type, wherein the short-time road condition features of the road section under the ordinary road section type include at least one of a real-time average travel time sequence in a short term, a historical average travel time sequence corresponding to each future time slice, and a real-time vehicle flow time sequence in a short term; and the short-time road condition features of the road section under the intersection directional road section type include at least one of a real-time average travel time sequence in a short term, a historical average travel time sequence corresponding to each future time slice, a real-time vehicle flow time sequence in a short term, an intersection type, and a turning action type.

[0021] When the road section type of the road section is a long-time burst congestion road section under a corresponding primary road section type, a long-term prediction depth model corresponding to the corresponding primary road section type is used to obtain predicted values of travel time of the road section at each future time slice based on long-time road condition features of the road section under the corresponding primary road section type, wherein the long-time road condition features of the road section under the ordinary road section type include at least one of a real-time average travel time sequence in a long term, a historical average travel time sequence corresponding to each future time slice, and a real-time vehicle flow time sequence in a long term; and the long-time road condition features of the road section under the intersection directional road section type include at least one of a real-time average travel time sequence in a long term, a historical average travel time sequence corresponding to each future time slice, a real-time vehicle flow time sequence in a long term, an intersection type, and a turning action type.

[0022] In a possible implementation, the short-term prediction depth model or the long-term prediction depth model includes a first time domain gated convolution layer, a graph attention layer, a second time domain gated convolution layer, a third time domain gated convolution layer, and a multi-layer perception (MLP) layer connected in sequence.

[0023] In a possible implementation, the method further includes:

[0024] training, based on first sample data, a linear model corresponding to the target scene, wherein the first sample data includes a traffic condition feature of a first sample road section at a first sample time and real values of travel times of each time slice after the first sample time, the first sample road section being a stable low-flow road section in the target scene;

[0025] training, based on second sample data, a corresponding short-term prediction deep model, wherein the second sample data includes a short-time traffic condition feature of a second sample road section at a second sample time and real values of travel times of each time slice after the second sample time, the second sample road section being a short-time burst congestion road section;

[0026] training, based on third sample data, a corresponding long-term prediction deep model, wherein the third sample data includes a long-time traffic condition feature of a third sample road section at a third sample time and real values of travel times of each time slice after the third sample time, the third sample road section being a long-time burst congestion road section.

[0027] In a possible implementation, the related traffic features of the target route include at least one of the following features: a start-end city of the target route, a departure time, lengths of road sections of various road grades, numbers of various types of road intersections, and numbers of various types of turning actions.

[0028] In a possible implementation, the method further includes:

[0029] training, based on fourth sample data, the route-level prediction model, the fourth sample data including estimated travel times of road sections of a sample route, related traffic features of the sample route, and a real arrival time of the sample route.

[0030] In a second aspect, an ETA prediction model training method is provided in the embodiments of the present disclosure, and the method includes:

[0031] Specifically, the ETA prediction model training method includes:

[0032] training, based on sample data of various road section types, a road section-level prediction model corresponding to each road section type;

[0033] calculating, based on the road section-level prediction model corresponding to each road section type, a travel time prediction value of each sample road section of a sample route at each time slice after a sample departure time;

[0034] determining, based on the travel time prediction value of each sample road section of the sample route at each time slice, an estimated entering time of each sample road section of the sample route;

[0035] accumulate the travel time prediction values of the sample road segments in the corresponding target time slice to obtain the estimated travel time of the sample route at the road segment level, wherein the target time slice corresponding to each sample road segment is the time slice in which the predicted entering time of each sample road segment is located;

[0036] based on the estimated travel time of the sample route at the road segment level, the related travel characteristics of the sample route, and the real arrival time of the sample route, a route-level prediction model is trained.

[0037] In a third aspect, an ETA prediction device is provided in the embodiments of the present disclosure, and the device comprises:

[0038] Specifically, the ETA prediction device comprises:

[0039] The obtaining module is configured to obtain the road segment types of road segments in a target route;

[0040] The calculation module is configured to calculate, based on the road segment-level prediction model corresponding to the road segment type of each road segment, the travel time prediction values of each road segment in future time slices after the departure time;

[0041] The determination module is configured to determine, based on the travel time prediction values of each road segment in the future time slices, the predicted entering time of each road segment;

[0042] The accumulation module is configured to accumulate the travel time prediction values of the road segments in the corresponding target time slice to obtain the estimated travel time of the target route at the road segment level, wherein the target time slice corresponding to each road segment is the future time slice in which the predicted entering time of each road segment is located;

[0043] The prediction module is configured to use a route-level prediction model to predict the predicted arrival time corresponding to the target route based on the estimated travel time of the target route at the road segment level and the related travel characteristics of the target route.

[0044] In a third aspect, an ETA prediction device is provided in the embodiments of the present disclosure, and the device comprises:

[0045] When the road segment type of a road segment is a low-frequency road segment, the travel time prediction values of the road segment in future time slices are obtained based on the historical average travel time of the road segment;

[0046] When the road segment type of a road segment is a stable low-flow road segment, a linear model corresponding to the scenario in which the road segment is located is used to obtain the travel time prediction values of the road segment in future time slices based on the travel road condition characteristics of the road segment.

[0047] when the road section type of the road section is a short-time burst congestion road section, using a short-time prediction deep model to obtain a predicted value of the travel time of the road section at each future time slice based on short-time traffic features of the road section;

[0048] when the road section type of the road section is a long-time burst congestion road section, using a long-time prediction deep model to obtain a predicted value of the travel time of the road section at each future time slice based on long-time traffic features of the road section.

[0049] In a possible implementation, the device further includes:

[0050] a scene determination module configured to determine a scene in which the road section is located based on at least one of the following features of the road section: a region feature in which the road section is located, a time type of a departure time, and a road level of the road section.

[0051] In a possible implementation, the road section type includes at least one of the following types under a first-level road section type: a low-frequency road section, a stable low-flow road section, a short-time burst congestion road section, and a long-time burst congestion road section, the first-level road section type including an ordinary road section or a road intersection directional road section; and the calculation module is configured to:

[0052] when the road section type of the road section is a stable low-flow road section under a corresponding first-level road section type, using a linear model corresponding to the scene in which the road section is located to obtain a predicted value of the travel time of the road section at each future time slice based on travel traffic features of the road section under the corresponding first-level road section type, the linear model being a model corresponding to the stable low-flow road section under the corresponding first-level road section type, wherein the travel traffic features of the road section under the ordinary road section type include at least one of the following: real-time travel features and a historical average travel time sequence corresponding to the future time slices; and the travel traffic features of the road section under the road intersection directional road section type include at least one of the following: real-time travel features, a historical average travel time sequence corresponding to the future time slices, a road intersection type, and a turning action type.

[0053] when the road section type of the road section is a short-time burst congestion road section under a corresponding first-level road section type, using a short-time prediction deep model corresponding to the corresponding first-level road section type, and obtaining a predicted value of the travel time of the road section at each future time slice based on short-time traffic characteristics of the road section under the corresponding first-level road section type, wherein the short-time traffic characteristics of the road section under the ordinary road section type include at least one of the following: a real-time average travel time sequence in a short term, a historical average travel time sequence corresponding to each future time slice, and a real-time vehicle flow time sequence in a short term; and the short-time traffic characteristics of the road section under the intersection directional road section type include at least one of the following: a real-time average travel time sequence in a short term, a historical average travel time sequence corresponding to each future time slice, a real-time vehicle flow time sequence in a short term, an intersection type, and a turning action type.

[0054] when the road section type of the road section is a long-time burst congestion road section under a corresponding first-level road section type, using a long-time prediction deep model corresponding to the corresponding first-level road section type, and obtaining a predicted value of the travel time of the road section at each future time slice based on long-time traffic characteristics of the road section under the corresponding first-level road section type, wherein the long-time traffic characteristics of the road section under the ordinary road section type include at least one of the following: a real-time average travel time sequence in a long term, a historical average travel time sequence corresponding to each future time slice, and a real-time vehicle flow time sequence in a long term; and the long-time traffic characteristics of the road section under the intersection directional road section type include at least one of the following: a real-time average travel time sequence in a long term, a historical average travel time sequence corresponding to each future time slice, a real-time vehicle flow time sequence in a long term, an intersection type, and a turning action type.

[0055] In a possible implementation, the short-time prediction deep model or the long-time prediction deep model includes a first time domain gated convolution layer, a graph attention layer, a second time domain gated convolution layer, a third time domain gated convolution layer, and a multi-layer perception (MLP) layer connected in sequence.

[0056] In a possible implementation, the device further includes:

[0057] a first training module configured to train a linear model corresponding to a target scene based on first sample data, wherein the first sample data includes travel traffic characteristics of a first sample road section at a first sample time and real values of travel time at each time slice after the first sample time, and the first sample road section is a stable low-flow road section located in the target scene;

[0058] a second training module configured to train a corresponding short-term prediction depth model based on second sample data, wherein the second sample data comprises short-time traffic feature of a second sample road section at a second sample time, and true values of travel time of each time slice after the second sample time, and the second sample road section is a short-time burst congestion road section;

[0059] a third training module configured to train a corresponding long-term prediction depth model based on third sample data, wherein the third sample data comprises long-time traffic feature of a third sample road section at a third sample time, and true values of travel time of each time slice after the third sample time, and the third sample road section is a long-time burst congestion road section.

[0060] In a possible implementation, the related travel features of the target route comprise at least one of the following features: starting and ending city of the target route, departure time, length of each road grade, number of each type of intersection, and number of each type of turning action.

[0061] In a possible implementation, the device further comprises:

[0062] a fourth training module configured to train the route-level prediction model based on fourth sample data, wherein the fourth sample data comprises estimated travel time of a sample route at a road section level, related travel features of the sample route, and true arrival time of the sample route.

[0063] In a fourth aspect, an ETA prediction model training device is provided in the embodiments of the present disclosure.

[0064] Specifically, the ETA prediction model training device comprises:

[0065] a road section-level training module configured to train a road section-level prediction model corresponding to each road section type based on sample data of each road section type;

[0066] an intermediate calculation module configured to calculate, based on the road section-level prediction model corresponding to each road section type, travel time prediction values of each sample road section of a sample route at each time slice after a sample departure time; determine an estimated entering time of each sample road section of the sample route based on the travel time prediction values of each sample road section of the sample route at each time slice; and accumulate the travel time prediction values of the each sample road section at a corresponding target time slice of the sample route to obtain estimated travel time of the sample route at a road section level, wherein the target time slice corresponding to each sample road section is a time slice in which the estimated entering time of each sample road section is located.

[0067] The route-level training module is configured to train a route-level prediction model based on the estimated travel time of the route segment level of the sample route, the relevant travel characteristics of the sample route, and the actual arrival time of the sample route.

[0068] In a fifth aspect, an electronic device is provided, including a memory and a processor, wherein the memory is configured to store one or more computer instructions, wherein the one or more computer instructions are executed by the processor to implement the method of any one of the aspects.

[0069] In a sixth aspect, a computer-readable storage medium is provided, having stored thereon computer instructions, which, when executed by a processor, implement the method of any one of the aspects.

[0070] In a seventh aspect, a computer program product is provided, including computer instructions, which, when executed by a processor, implement the method steps of any one of the aspects.

[0071] In a seventh aspect, a navigation method is provided, wherein a navigation route is obtained based on at least a starting point, an ending point, and road condition information, an ETA of the navigation route is predicted and displayed, and navigation guidance is performed based on the navigation route, and the prediction of the ETA of the navigation route is implemented based on any one of the methods of the first aspect.

[0072] According to the technical solutions provided by the embodiments of the present disclosure, different route segment-level prediction models can be used to predict different route segment types, simple prediction scenarios are handed over to efficient simple models for prediction, the problem of high overhead and low computational efficiency of complex models is solved, the estimated travel time of the target route at the route segment level considering the predicted entering time of each route segment is taken as the input of the route-level prediction model, the prediction model can more accurately consider the future road conditions of the user entering the corresponding route segment at the future time, rather than the current road conditions, and thus the prediction accuracy is improved.

[0073] It should be understood that the foregoing general description and the following detailed description are only exemplary and explanatory, and are not limiting to the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0074] Other features, objects, and advantages of the present disclosure will become more apparent from the following detailed description of non-limiting embodiments, taken in conjunction with the accompanying drawings. In the drawings:

[0075] Figure 1 A flowchart of an ETA prediction method according to an embodiment of the present disclosure is shown.

[0076] Figure 2 A flowchart of an ETA prediction model training method according to an embodiment of the present disclosure is shown.

[0077] Figure 3 A structural block diagram of an ETA prediction device according to an embodiment of the present disclosure is shown.

[0078] Figure 4 A structural block diagram of an ETA prediction model training device according to an embodiment of the present disclosure is shown.

[0079] Figure 5 An application schematic diagram in a navigation application scenario according to an embodiment of the present disclosure is shown.

[0080] Figure 6 A structural block diagram of a server according to an embodiment of the present disclosure is shown.

[0081] Figure 7 A structural schematic diagram of a system of a server adapted to implement the method according to an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0082] Hereinafter, exemplary embodiments of the present disclosure will be described in detail with reference to the accompanying drawings so as to be easily implemented by those skilled in the art. In addition, parts irrelevant to the description of the exemplary embodiments are omitted in the accompanying drawings for the sake of clarity.

[0083] In the present disclosure, it should be understood that terms such as “include” or “have” are intended to indicate that there are features, numbers, steps, actions, components, parts or combinations thereof disclosed in the specification, and do not exclude the possibility that one or more other features, numbers, steps, actions, components, parts or combinations thereof exist or are added.

[0084] It should also be noted that the embodiments in the present disclosure and the features in the embodiments can be combined with each other without conflict. The present disclosure will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0085] In the present disclosure, the acquisition of user information or user data is an operation authorized, confirmed by the user, or actively selected by the user.

[0086] Figure 1 A flowchart of a predicted arrival time prediction method according to an embodiment of the present disclosure is shown. As shown in Figure 1 The predicted arrival time prediction method includes the following steps S101-S105:

[0087] In step S101, the road section types of each road section in the target route are acquired;

[0088] In step S102, based on the road section level prediction model corresponding to the road section type of each road section, the passing time prediction value of each time slice in the future after the departure time of each road section is calculated;

[0089] In step S103, a predicted entering time of each road segment is determined based on the predicted travel time of each road segment at each future time slice;

[0090] In step S104, the predicted travel time of each road segment at a corresponding target time slice is accumulated to obtain a road segment level estimated travel time of the target route, wherein the target time slice corresponding to each road segment is the future time slice in which the predicted entering time of each road segment is located.

[0091] In step S105, a route level prediction model is used to predict a predicted arrival time corresponding to the target route based on the road segment level estimated travel time of the target route and the related travel characteristics of the target route.

[0092] As mentioned above, ETA (Estimated Time of Arrival) is an estimate of the travel time of a planned route. Generally, navigation applications also use ETA prediction algorithms to calculate the ETA of a user reaching a destination by a corresponding navigation route when planning a navigation route for the user. For navigation applications with high daily activity, the accuracy and computational efficiency of the ETA prediction algorithm are highly required. Current ETA prediction algorithms usually directly use a prediction model to predict the current travel time of each road segment on a navigation route, and then accumulate the ETA of the navigation route. In this prediction algorithm, the current travel time of each different road segment is predicted by a unified prediction model. However, in fact, the travel time of many road segments is stable, and the results can be quickly predicted by a simple model. Therefore, the use of a unified complex prediction model for different types of road segments in the existing ETA prediction algorithm wastes computational resources and has low computational efficiency. Moreover, the current model predicts the current travel time of each road segment, and does not consider the specific time of entering the road segment. Therefore, the ETA predicted in this way is not accurate.

[0093] In view of the above problems, the present disclosure provides an ETA prediction method, which can use different road segment level prediction models for different types of road segments. The simple prediction scenarios are predicted by efficient simple models, which solves the problem of high overhead and low computational efficiency of complex models. The road segment level estimated travel time of the target route considering the predicted entering time of each road segment is used as the input of the route level prediction model, which enables the prediction model to more accurately consider the future road conditions of the user entering the corresponding road segment at a future time, rather than the current road conditions, thereby improving the accuracy of the prediction.

[0094] In a possible implementation, the ETA prediction method can be applied to a computer, a computing device, an electronic device, a server, a server cluster, or the like for ETA prediction.

[0095] In a possible implementation, the target route can be a recommended navigation route from the starting location to the destination recommended by the navigation application for the user, or can also be a navigation route from the starting location to the destination selected by the user.

[0096] In a possible implementation, a road segment (also referred to as a link) refers to a basic unit of a road network, generally a small part of a road, and a road segment has only one entrance and one exit, and the entrance and the exit at each end of the road segment can have a topological connection relationship with at least one other road segment. A target route can include a series of continuous road segments.

[0097] In a possible implementation, the type of the road segment can be divided according to the stability of the travel time of the travel tool on the road segment, and after the travel road segment is divided, the road segment level prediction model corresponding to the type of the road segment type can be trained for the type of the road segment. For example, if the travel time of each travel tool on the road segment is regular or stable at each time, the future travel time of such a road segment is easy to predict, and the road segment can be divided into a type of road segment, and a simple road segment level prediction model can be trained for the type of road segment. If the travel time of each travel tool on the road segment is unstable at each time, and sudden congestion and the like are prone to occur, the future travel time of such a road segment is not easy to predict, and the road segment can be divided into another type of road segment, and a complex road segment level prediction model can be trained for the type of road segment to accurately predict the future travel time of the type of road segment.

[0098] In a possible implementation, the road segment level prediction model is used to predict the travel time prediction value of each road segment at each future time slice after the starting time, and the time slice refers to a time period. For example, the length of each time slice can be 5 minutes, and each future time slice after the starting time can be [0min-5min), [5min-10min), [(5n-1)min-5n min), and the like.

[0099] In a possible implementation, still taking the above example, each road segment in the target route can be sorted in order from the starting location to the destination, the estimated entering time of the first road segment is the starting time, the travel time prediction value of the first road segment in the time slice [0min-5min) in which the starting time is located is 10 minutes, and the estimated entering time of the second road segment is 10 minutes after the starting time. The travel time prediction value of the second road segment in the time slice [10min-15min) in which the starting time is located is 8 minutes, and the estimated entering time of the third road segment is 18 minutes after the starting time. Thus, the estimated entering time of each road segment in the target route can be calculated in turn.

[0100] In a possible implementation, the estimated travel time of the target route at the road segment level can be obtained by adding together the travel time prediction values of all road segments at the predicted entering time. In this case, the future time slice in which the predicted entering time of each road segment is located can be determined as the target time slice, and the travel time prediction value of each road segment at the corresponding target time slice can be the travel time prediction value of the road segment at the predicted entering time. For example, the target route includes three road segments, i.e., a first road segment, a second road segment, and a third road segment. The predicted entering time of the first road segment is the departure time, the target time slice corresponding to the first road segment is the first time slice [0min-5min) after the departure time, the travel time prediction value of the first road segment at the target time slice [0min-5min) is 10min, the target time slice corresponding to the second road segment is [10min-15min), the travel time prediction value of the second road segment at the target time slice [10min-15min) is 8min, the target time slice corresponding to the third road segment is [15min-20min), and the travel time prediction value of the third road segment at the target time slice [15min-20min) is 10min. In this case, the estimated travel time of the target route at the road segment level is 10min+8min+10min=28min.

[0101] In a possible implementation, the input of the route-level prediction model is the estimated travel time of the target route at the road segment level and the related travel features of the target route, and the output is the ETA corresponding to the target route, i.e., the predicted time required to travel from the departure location to the destination of the target route. After the estimated travel time of the target route at the road segment level and the related travel features of the target route are input into the route-level prediction model, the route-level prediction model is executed, and the ETA corresponding to the target route can be obtained.

[0102] In a possible implementation, the related travel features of the target route can include at least one of the following features: the start and end city of the target route, the departure time, the length of each road grade, the number of each type of intersection, and the number of each type of turning action. The start and end city can be the city code of the city where the departure location is located and the city code of the city where the destination is located. The road grade can include various grades such as expressway, national highway, provincial highway, county road, urban road, and rural road. The type of intersection includes a pre-defined complex intersection and a simple intersection. The type of intersection in the road network can be pre-defined, and the type of each intersection in the target route can be directly obtained, and then the number of each type of intersection can be obtained. The turning action can include left turning, right turning, and U-turn of the travel tool on the target route.

[0103] The embodiment configures different link-level prediction models for different link types, can hand over simple link prediction to an efficient simple model for prediction, solves the problem of large overhead and low calculation efficiency of a complex model, and thus, after obtaining the link types of the links in the target route, can calculate the link-level prediction model corresponding to the link type to obtain the link passing time prediction value of each time slice in the future after the departure time of each link, and then, based on the link passing time prediction value of each link in each time slice, determines the predicted entering time of each link and accumulates the link passing time prediction value of each link at the predicted entering time to obtain the link-level estimated passing time of the target route. At this time, the link-level estimated passing time obtained takes into account the predicted entering time of each link, so that when the route-level prediction model is used to predict the ETA corresponding to the target route based on the link-level estimated passing time of the target route and the related passing characteristics of the target route, the predicted ETA takes into account the predicted entering time of each link, so that the route-level prediction model can more accurately take into account the future road conditions of the user entering the corresponding link at the future time, rather than the current road conditions, thereby improving the prediction accuracy.

[0104] In a possible implementation, the link types can further include at least one of the following types: a low-frequency link, a stable low-flow link, a short-time sudden congestion link, and a long-time sudden congestion link. The low-frequency link refers to a link through which vehicles occasionally pass, and the low-frequency link is basically smooth without congestion in each time slice. The stable low-flow link refers to a link with less and more regular congestion and no sudden congestion in history, such as a link that is congested only during the peak period of holidays. The short-time sudden congestion link refers to a link that has experienced sudden congestion in history but the duration of the congestion is relatively short. The long-time sudden congestion link refers to a link that has experienced sudden congestion with a relatively long duration in history.

[0105] In a possible implementation, step S102 in the ETA prediction method, that is, calculating the link passing time prediction value of each link in each time slice in the future after the departure time based on the link-level prediction model corresponding to the link type of each link, can include the following steps:

[0106] When the link type of a link is a low-frequency link, the link passing time prediction value of the link in each time slice in the future is obtained based on the historical average passing time of the link;

[0107] When the link type of a link is a stable low-flow link, a linear model corresponding to the scene where the link is located is used to obtain the link passing time prediction value of the link in each time slice in the future based on the passing road condition characteristics of the link;

[0108] When the link type of the link is a short-time burst congestion link, a short-time prediction deep model is used to obtain the predicted travel time of the link at each future time slice based on short-time traffic features of the link.

[0109] When the link type of the link is a long-time burst congestion link, a long-time prediction deep model is used to obtain the predicted travel time of the link at each future time slice based on long-time traffic features of the link.

[0110] In this implementation, different link-level prediction models can be configured for different link types. For example, different link-level prediction models can be configured for low-frequency links, stable low-flow links, short-time burst congestion links, and long-time burst congestion links.

[0111] In this implementation, the simplest link-level prediction model can be configured for a low-frequency link. The simplest link-level prediction model can be a model for calculating the historical average travel time of the link. When the link type of the link is a low-frequency link, the link is always smooth at each time slice because there are only occasional vehicles on the link. Therefore, the historical average travel time of the link can be used as the predicted travel time of the link at each future time slice.

[0112] In this implementation, a relatively simple linear model can be configured for a stable low-flow link. The stable low-flow link can be divided into scenes according to scene features, and a linear model can be trained for each scene. The scene in which the link is located can be determined based on at least one of the following features of the link: the region in which the link is located (such as in which city), the time type of the departure time (such as morning peak, flat peak, or evening peak), and the road grade of the link. The linear model corresponding to different scenes can predict the predicted travel time of the link at each future time slice based on the travel traffic features of the link in different scenes.

[0113] In this implementation, the travel traffic features of the link can include at least one of the following: real-time travel features and a historical average travel time sequence corresponding to each future time slice. The real-time travel features refer to the travel time of the link at the current time. Assuming that the departure time is Monday at 9:00, the future time slices are [9:00, 9:10), [9:10, 9:20), and so on. The historical average travel time sequence corresponding to each future time slice refers to the average historical travel time of [9:00, 9:10) on Monday, the average historical travel time of [9:10, 9:20) on Monday, and so on. The travel traffic features of the link are input into the linear model corresponding to the scene in which the link is located, and the predicted travel time of the link at each future time slice output by the linear model can be obtained.

[0114] In this implementation, a more complex short-term prediction deep model can be configured for a short-time burst congestion road segment, which can predict the travel time prediction value of the road segment at each future time slice based on the short-time traffic characteristics of the road segment, wherein the short-time traffic characteristics of the road segment include at least one of the following: a real-time average travel time sequence in a short term, a historical average travel time sequence corresponding to the future time slices, a real-time vehicle flow time sequence in a short term. The real-time average travel time sequence in a short term can be the average travel time corresponding to a column of time sequences in a short time period. For example, the real-time average travel time sequence in a short term can be a sequence composed of average travel times of each minute in the past 1 hour at the current time. The real-time vehicle flow time sequence in a short term can be the vehicle flow information (such as inflow, outflow, and residence) corresponding to a column of time sequences on the road segment in a short time period. For example, the real-time vehicle flow time sequence in a short term can be a sequence composed of inflow of travel tools, a sequence composed of outflow, and a sequence composed of residence in each minute in the past 1 hour at the current time.

[0115] In this implementation, a more complex long-term prediction deep model can be configured for a long-time burst congestion road segment, which can predict the travel time prediction value of the road segment at each future time slice based on the long-time traffic characteristics of the road segment, wherein the long-time traffic characteristics of the road segment include at least one of the following: a real-time average travel time sequence in a long term, a historical average travel time sequence corresponding to the future time slices, a real-time vehicle flow time sequence in a long term; the real-time average travel time sequence in a long term can be the average travel time corresponding to a column of time sequences in a long time period. For example, the real-time average travel time sequence in a long term can be a sequence composed of average travel times of each minute in the past 3 hours at the current time. The real-time vehicle flow time sequence in a long term can be the vehicle flow information (such as inflow, outflow, and residence) corresponding to a column of time sequences on the road segment in a long time period. For example, the real-time vehicle flow time sequence in a long term can be a sequence composed of inflow of travel tools, a sequence composed of outflow, and a sequence composed of residence in each minute in the past 3 hours at the current time.

[0116] In this implementation, different road segment level prediction models are configured for different road segment types, which realizes decoupling of prediction effects of different road segment types, solves the problem of high overhead of complex models in the prior art by using simple and efficient models to solve simple prediction problems, improves the calculation efficiency, and finally the prediction results of each type of road segment can be mixed for use, which improves the prediction accuracy.

[0117] In a possible implementation, the road section types include at least one of the following types under a first-level road section type: a low-frequency road section, a stable low-flow road section, a short-time burst congestion road section, and a long-time burst congestion road section, and the first-level road section type includes a common road section or a road intersection directional road section.

[0118] In this implementation, the road intersection directional road section refers to a road section whose exit connects at least two other road sections in different directions, and the speeds of different lanes in the road intersection directional road section are obviously different when a traffic tool travels on the road intersection directional road section. For example, a left-turn lane is more congested, the travel speed is slower, and the travel time is longer, while a straight lane is less congested, the travel speed is faster, and the travel time is shorter. The common road section refers to other road sections except the road intersection directional road section.

[0119] In this implementation, the road section types in the road network can include low-frequency road sections, stable low-flow road sections, short-time burst congestion road sections, and long-time burst congestion road sections under a common road section type, and low-frequency road sections, stable low-flow road sections, short-time burst congestion road sections, and long-time burst congestion road sections under a road intersection directional road section type.

[0120] In this implementation, compared with the common road section, the road intersection directional road section needs to consider more factors when predicting the future travel time, such as the type of the road intersection, the turning action of the road intersection, and the like. Therefore, different road section-level prediction models need to be set for the road intersection directional road section and the common road section. For example, different road section-level prediction models can be configured for low-frequency road sections, stable low-flow road sections, short-time burst congestion road sections, and long-time burst congestion road sections under the common road section type, and low-frequency road sections, stable low-flow road sections, short-time burst congestion road sections, and long-time burst congestion road sections under the road intersection directional road section type.

[0121] In this implementation, when the road section type of a road section is a low-frequency road section under a corresponding first-level road section type, the simplest road section-level prediction model can be configured for low-frequency road sections under the two first-level road section types.

[0122] In the implementation, when the road section type of the road section is a stable low-flow road section under a corresponding first-level road section type, different first-level road section types need to consider different influencing factors when predicting, and therefore different linear models need to be trained for different first-level road section types. In the implementation, a linear model needs to be trained for different scenes under different first-level road section types. The passing road condition features of the road section under the corresponding first-level road section type are used to obtain the passing time prediction value of the road section at each future time slice by using the linear model corresponding to the scene in which the road section is located, wherein the passing road condition features of the road section under the ordinary road section type include at least one of the following: real-time passing features and a historical average passing time sequence corresponding to each future time slice; the passing road condition features of the road section under the intersection directional road section type include at least one of the following: real-time passing features, a historical average passing time sequence corresponding to each future time slice, an intersection type, and a turning action type. The intersection type includes a complex intersection and a simple intersection, and the turning action type refers to a turning action type of a passing vehicle when the vehicle exits the exit of the road section along the target route, including actions such as left turning, right turning, and U-turning.

[0123] In the implementation, when the road section type of the road section is a short-time burst congestion road section under a corresponding first-level road section type, a more complex short-term prediction deep model can be configured for the short-time burst congestion road section under the first-level road section type. Different first-level road section types need to consider different influencing factors when predicting, and therefore different short-term prediction deep models need to be trained for different first-level road section types. Different short-term prediction deep models corresponding to different first-level road section types can predict the passing time prediction value of the road section at each future time slice based on short-time road condition features of the road section under different road section types. In this way, the short-term prediction deep model corresponding to the corresponding first-level road section type is used to obtain the passing time prediction value of the road section at each future time slice based on the short-time road condition features of the road section under the corresponding first-level road section type, wherein the short-time road condition features of the road section under the ordinary road section type include at least one of the following: a real-time average passing time sequence in a short term, a historical average passing time sequence corresponding to each future time slice, and a real-time vehicle flow time sequence in a short term; the short-time road condition features of the road section under the intersection directional road section type include at least one of the following: a real-time average passing time sequence in a short term, a historical average passing time sequence corresponding to each future time slice, a real-time vehicle flow time sequence in a short term, an intersection type, and a turning action type.

[0124] In this implementation, when the road section type of a road section is a long-time burst congestion road section under a corresponding primary road section type, a more complex long-term prediction deep model can be configured for the long-time burst congestion road section under the primary road section type, different primary road section types need to consider different influencing factors when predicting, and therefore different long-term prediction deep models need to be trained for different primary road section types. Different long-term prediction deep models corresponding to different primary road section types can predict the travel time prediction values of the road section at future time slices based on long-time traffic characteristics of the road section under different road section types. Using the long-term prediction deep model corresponding to the corresponding primary road section type, the travel time prediction values of the road section at future time slices can be obtained based on the long-time traffic characteristics of the road section under the corresponding primary road section type, wherein the long-time traffic characteristics of the road section under the ordinary road section type include at least one of the following: a real-time average travel time sequence in a long term, a historical average travel time sequence corresponding to the future time slices, and a real-time vehicle flow time sequence in a long term; the long-time traffic characteristics of the road section under the intersection directional road section type include at least one of the following: a real-time average travel time sequence in a long term, a historical average travel time sequence corresponding to the future time slices, a real-time vehicle flow time sequence in a long term, an intersection type, and a turning action type.

[0125] In a possible implementation, the short-term prediction deep model or the long-term prediction deep model described above can include a first temporal gate convolution layer (Temporal gate Convolution), a graph attention layer, a second temporal gate convolution layer, a third temporal gate convolution layer, and an MLP (Multilayer Perceptron) layer connected in sequence.

[0126] In this implementation, the model structure of the short-term prediction deep model or the long-term prediction deep model can be that the output end of the first temporal gate convolution layer is connected to the input end of the graph attention layer, the output end of the graph attention layer is connected to the input end of the second temporal gate convolution layer, the output end of the second temporal gate convolution layer is connected to the input end of the third temporal gate convolution layer, and the output end of the third temporal gate convolution layer is connected to the input end of the MLP layer. In this way, the short-term prediction deep model or the long-term prediction deep model can better extract features in the various time sequences described above, and then make accurate predictions.

[0127] In a possible implementation, the ETA prediction method described above can also configure different linear models for stable low-flow road sections in different scenarios under different primary road section types, that is, the following steps can also be included:

[0128] The linear model corresponding to the target scene is trained based on first sample data, wherein the first sample data includes the traffic condition feature of the first sample road segment at the first sample time, and the true value of the travel time of each time slice after the first sample time, and the first sample road segment is a stable low-flow road segment in the target scene.

[0129] In this implementation, the first sample data of the first sample road segment in different scenes at the first sample time can be used to train the linear model corresponding to the stable low-flow road segment in different scenes. The traffic condition feature in the first sample data can be input into the initial linear model to obtain the predicted value of the travel time of each time slice after the first sample time output by the linear model. The parameters in the linear model are adjusted according to the error between the predicted value of the travel time of each time slice after the first sample time and the true value of the travel time of each time slice after the first sample time, until the error is reduced to a certain degree, and then the linear model corresponding to the stable low-flow road segment in the target scene is trained.

[0130] In this implementation, the road segment type can include at least one first-level road segment type: ordinary road segment and intersection direction road segment. Different first-level road segment types need to consider different influencing factors when predicting, so different linear models need to be trained for different first-level road segment types. In this implementation, a linear model needs to be trained for different scenes under different first-level road segment types. The corresponding linear model can be trained based on first sample data, wherein the first sample data includes the traffic condition feature of the first sample road segment at the first sample time under the corresponding first-level road segment type, and the true value of the travel time of each time slice after the first sample time, and the first sample road segment is a stable low-flow road segment under the corresponding first-level road segment type and is located in the target scene. The traffic condition feature of the first sample road segment at the first sample time under the ordinary road segment type includes at least one of the following: the travel feature at the first sample time and the historical average travel time sequence of each time slice after the first sample time; and the traffic condition feature of the first sample road segment at the first sample time under the intersection direction road segment type includes at least one of the following: the travel feature at the first sample time, the historical average travel time sequence of each time slice after the first sample time, the intersection type, and the turning action type.

[0131] In a possible implementation, the ETA prediction method described above can also configure different short-term prediction depth models for short-time burst congestion road segments under different first-level road segment types, that is, the following steps can also be included:

[0132] The second sample data includes short-time traffic characteristics of the second sample road segment at the second sample time, and real values of the travel time of each time slice after the second sample time. The second sample road segment is a short-time sudden congestion road segment.

[0133] In this implementation, the short-term prediction deep model can be trained by using the second sample data of the second sample road segment at the second sample time. The short-time traffic characteristics in the second sample data can be input into the initial short-term prediction deep model to obtain predicted values of the travel time of each time slice after the second sample time output by the short-term prediction deep model. The parameters in the short-term prediction deep model can be adjusted according to the error between the predicted values of the travel time of each time slice after the second sample time and the real values of the travel time of each time slice after the second sample time, until the error is reduced to a certain degree, and then the short-term prediction deep model is trained.

[0134] In this implementation, the road segment type can include at least one first-level road segment type, such as a common road segment and a road intersection directional road segment. Different first-level road segment types need to consider different influencing factors during prediction, so different short-term prediction deep models need to be trained for different first-level road segment types. In this implementation, a short-term prediction deep model needs to be trained for each first-level road segment type. The corresponding short-term prediction deep model can be trained based on second sample data. The second sample data includes short-time traffic characteristics of a second sample road segment of a corresponding first-level road segment type at a second sample time, and real values of the travel time of each time slice after the second sample time. The second sample road segment is a short-time sudden congestion road segment under the corresponding first-level road segment type. The short-time traffic characteristics of the second sample road segment of the common road segment type at the second sample time include at least one of the following: a sequence of average travel time in a short term before the second sample time, a sequence of historical average travel time corresponding to each time slice after the second sample time, and a sequence of vehicle flow in a short term before the second sample time. The short-time traffic characteristics of the second sample road segment of the road intersection directional road segment type at the second sample time include at least one of the following: a sequence of average travel time in a short term before the second sample time, a sequence of historical average travel time corresponding to each time slice after the second sample time, a sequence of vehicle flow in a short term before the second sample time, a road intersection type, and a turning action type.

[0135] In a possible implementation, the ETA prediction method can further configure different long-term prediction deep models for long-time sudden congestion road segments under different first-level road segment types, that is, the method can further include the following steps:

[0136] The third sample data of the third sample road section at the third sample time is used to train the corresponding long-term prediction deep model. The long-time traffic feature in the third sample data is input into the initial long-term prediction deep model to obtain the predicted travel time of each time slice after the third sample time output by the long-term prediction deep model. The parameters in the long-term prediction deep model are adjusted according to the error between the predicted travel time of each time slice after the third sample time and the real travel time of each time slice after the third sample time, until the error is reduced to a certain degree, and then the long-term prediction deep model is trained.

[0137] In this implementation, the third sample data of the third sample road section at the third sample time is used to train the corresponding long-term prediction deep model. The long-time traffic feature in the third sample data is input into the initial long-term prediction deep model to obtain the predicted travel time of each time slice after the third sample time output by the long-term prediction deep model. The parameters in the long-term prediction deep model are adjusted according to the error between the predicted travel time of each time slice after the third sample time and the real travel time of each time slice after the third sample time, until the error is reduced to a certain degree, and then the long-term prediction deep model is trained.

[0138] In this implementation, the road section type can include at least one of the following first-level road section types: ordinary road section and intersection direction road section. Different first-level road section types need to consider different influencing factors during prediction, so different long-term prediction deep models need to be trained for different first-level road section types. In this implementation, a long-term prediction deep model needs to be trained for each first-level road section type. The third sample data is used to train the corresponding long-term prediction deep model, wherein the third sample data includes the long-time traffic feature of the third sample road section of the corresponding first-level road section type at the third sample time, and the real travel time of each time slice after the third sample time. The third sample road section is a long-time sudden congestion road section under the corresponding first-level road section type. The long-time traffic feature of the third sample road section of the ordinary road section type at the third sample time includes at least one of the following: the average travel time sequence in a long period before the third sample time, the historical average travel time sequence corresponding to each time slice after the third sample time, and the vehicle flow time sequence in a long period before the third sample time. The long-time traffic feature of the third sample road section of the intersection direction road section type at the third sample time includes at least one of the following: the average travel time sequence in a long period before the third sample time, the historical average travel time sequence corresponding to each time slice after the third sample time, the vehicle flow time sequence in a long period before the third sample time, the intersection type, and the turning action type.

[0139] In a possible implementation, the ETA prediction method can further include the following steps:

[0140] The fourth sample data is used to train the route-level prediction model, wherein the fourth sample data includes the estimated travel time of the road section level of the sample route, the related travel feature of the sample route, and the real arrival time of the sample route.

[0141] In this implementation, the related traffic features of the sample route include at least one of the following features: the start and end cities of the sample route, the departure time, the length of each road grade, the number of each type of intersection, and the number of each type of turning action.

[0142] In this implementation, the estimated travel time of the sample route at the road segment level in the fourth sample data and the related traffic features of the sample route are input into an initial route-level prediction model to obtain the estimated arrival time of the sample route output by the route-level prediction model. The parameters in the route-level prediction model are adjusted according to the error between the estimated arrival time of the sample route and the actual arrival time of the sample route until the error is reduced to a certain degree, and then the corresponding route-level prediction model is trained.

[0143] In this implementation, the route-level prediction model can be a DNN (Deep Neural Networks) model.

[0144] In this implementation, a double-layer prediction model architecture of "road segment-level prediction-route-level prediction" is proposed. The deduction result predicted by the road segment-level prediction model is further optimized by the route-level prediction model to obtain the final predicted ETA, so that the route-level model can focus on the prediction modeling of the travel time on each type of road segment, and the route-level prediction model can achieve higher accuracy by further adjusting the result.

[0145] The present disclosure also provides an ETA prediction model training method, Figure 2 A flowchart of an ETA prediction model training method according to an embodiment of the present disclosure is shown. As shown in Figure 2 The ETA prediction model training method includes the following steps S201-S205:

[0146] In step S201, based on the sample data of each road segment type, a road segment-level prediction model corresponding to each road segment type is trained;

[0147] In step S202, based on the road segment-level prediction model corresponding to each road segment type, the travel time prediction value of each type of sample road segment in the sample route at each time slice after the sample departure time is calculated;

[0148] In step S203, based on the travel time prediction value of each sample road segment in the sample route at each time slice, the estimated entering time of each sample road segment in the sample route is determined;

[0149] In step S204, the travel time prediction values of the sample road segments in the corresponding target time slice are accumulated to obtain the estimated travel time of the sample road at the road segment level, wherein the target time slice corresponding to each sample road segment is the time slice in which the predicted entering time of each sample road segment is located.

[0150] In step S205, the route-level prediction model is trained based on the estimated travel time of the sample road at the road segment level, the related travel characteristics of the sample road, and the real arrival time of the sample road.

[0151] In a possible implementation, the ETA prediction model training method can be applied to a computer, a computing device, an electronic device, a server, a server cluster, or the like, which performs ETA prediction model training.

[0152] In a possible implementation, a road segment (also referred to as a link) refers to a basic unit of a road network, generally a small part of a road, and a road segment has only one entrance and one exit. The entrance and the exit at each end of the road segment can have a topological connection relationship with at least one other road segment. A sample road can include a series of continuous road segments.

[0153] In a possible implementation, the types of road segments can be classified according to the stability of the travel time of the travel tools on the road segments. After the travel road segments are classified, the road segment-level prediction model corresponding to the type of road segment can be trained for the type of road segment. For example, if the travel time of each travel tool on the road segment is regular or stable at each time, the future travel time of such a road segment is easy to predict, and such a road segment can be classified into a type of road segment, and a simple road segment-level prediction model can be trained for the type of road segment. If the travel time of each travel tool on the road segment is unstable at each time, and sudden congestion or the like is prone to occur, the future travel time of such a road segment is not easy to predict, and such a road segment can be classified into another type of road segment, and a complex road segment-level prediction model can be trained for the type of road segment to accurately predict the future travel time of the type of road segment.

[0154] In a possible implementation, the road segment-level prediction model is used to predict the travel time prediction values of each road segment at each time slice after the sample departure time of the sample road. The time slice refers to a time period. For example, the length of each time slice can be 5 min, and each time slice after the sample departure time can be [0 min-5 min), [5 min-10 min), [(5n-1) min-5n min), and the like.

[0155] In a possible implementation, still taking the above example, the road segments in the sample route can be sorted in the order from the starting point to the destination, the predicted entering time of the first sample road segment is the sample starting time, the predicted travel time of the first sample road segment in the time slice [0min-5min) in which the sample starting time is located is 10min, the predicted entering time of the second sample road segment is 10min after the sample starting time, the predicted travel time of the second sample road segment in the time slice [10min-15min) in which the sample starting time is located is 8min, and the predicted entering time of the third road segment is 18min after the sample starting time. In this way, the predicted entering time of each sample road segment in the sample route can be sequentially calculated.

[0156] In a possible implementation, the predicted travel times of the sample road segments at the predicted entering times are all added together to obtain the estimated travel time of the sample route at the road segment level. Here, the future time slice in which the predicted entering time of each sample road segment is located can be determined as a target time slice, and the predicted travel time of each sample road segment in the corresponding target time slice is the predicted travel time of the road segment at the predicted entering time. Still taking the above example, assume that the sample route includes three sample road segments, i.e., the first road segment, the second road segment, and the third road segment, the predicted entering time of the first road segment is the starting time, the target time slice corresponding to the first road segment is the first time slice [0min-5min) after the starting time, the predicted travel time of the first road segment in the target time slice [0min-5min) is 10min, the target time slice corresponding to the second road segment is [10min-15min), the predicted travel time of the second road segment in the target time slice [10min-15min) is 8min, the target time slice corresponding to the third road segment is [15min-20min), and the predicted travel time of the third road segment in the target time slice [15min-20min) is 10min. Therefore, the estimated travel time of the sample route at the road segment level is 10min+8min+10min=28min.

[0157] In a possible implementation, the related travel features of the sample route include at least one of the following features: the starting and ending point cities of the sample route, the starting time, the lengths of road segments of different grades, the numbers of road intersections of different types, and the numbers of different turning actions.

[0158] In a possible implementation, the link-level estimated travel time of a sample route and the relevant travel features of the sample route can be input into an initial route-level prediction model to obtain an estimated arrival time of the sample route output by the route-level prediction model, and the parameters in the route-level prediction model are adjusted according to the error between the estimated arrival time of the sample route and the actual arrival time of the sample route until the error is reduced to a certain extent, and then the corresponding route-level prediction model is trained.

[0159] In a possible implementation, the link types can further include at least one of the following types: a low-frequency link, a stable low-flow link, a short-time burst congestion link, and a long-time burst congestion link. The low-frequency link refers to a link that is occasionally passed through by vehicles, and the low-frequency link is basically smooth without congestion at each time slice. The stable low-flow link refers to a link that has less and more regular congestion and has never had burst congestion in history, such as a link that is congested only during the peak period of holidays. The short-time burst congestion link refers to a link that has had burst congestion in history but the duration of the burst congestion is relatively short. The long-time burst congestion link refers to a link that has had burst congestion in history with a relatively long duration.

[0160] In a possible implementation, different link-level prediction models can be trained for different link types. For example, different link-level prediction models can be trained for the low-frequency link, the stable low-flow link, the short-time burst congestion link, and the long-time burst congestion link.

[0161] In this implementation, the simplest link-level prediction model can be configured for the low-frequency link, and the simplest link-level prediction model can be a model for calculating the historical average travel time of the link. When the link type of the link is a low-frequency link, the historical average travel time of the link can be used as the travel time prediction value of the link at each future time slice because there are occasional vehicles passing through the low-frequency link and the link is smooth at each time slice.

[0162] In a possible implementation, the training method described above can further train different linear models for stable low-flow links in different scenarios, that is, the method can further include the following steps:

[0163] Based on the first sample data, a linear model corresponding to a target scenario is trained, wherein the first sample data includes travel condition features of a first sample link at a first sample time and actual values of travel times at each time slice after the first sample time, and the first sample link is a stable low-flow link located in the target scenario.

[0164] In the implementation mode, the first sample data of the first sample section in different scenes at the first sample moment can be used to train the linear model corresponding to the stable low-flow section in different scenes. The traffic condition features in the first sample data can be input into the initial linear model to obtain the predicted traffic time values of each time slice after the first sample moment output by the linear model. The parameters in the linear model are adjusted according to the error between the predicted traffic time values of each time slice after the first sample moment and the real traffic time values of each time slice after the first sample moment, until the error is reduced to a certain degree, and then the linear model corresponding to the stable low-flow section in the target scene is trained.

[0165] In the implementation mode, the road section type can include at least one first-level road section type, i.e., ordinary road section and intersection directional road section. Different first-level road section types need to consider different influencing factors when predicting, so different linear models need to be trained for different first-level road section types. In the implementation mode, a linear model needs to be trained for different scenes under different first-level road section types. The corresponding linear model can be trained based on the first sample data, wherein the first sample data includes the traffic condition features of the first sample section in the corresponding first-level road section type at the first sample moment, and the real traffic time values of each time slice after the first sample moment. The first sample section is a stable low-flow section in the corresponding first-level road section type and is located in the target scene. The traffic condition features of the first sample section in the ordinary road section type at the first sample moment include at least one of the following: traffic features at the first sample moment and historical average traffic time sequence of each time slice after the first sample moment. The traffic condition features of the first sample section in the intersection directional road section type at the first sample moment include at least one of the following: traffic features at the first sample moment, historical average traffic time sequence of each time slice after the first sample moment, intersection type, and turning action type.

[0166] In a possible implementation mode, a more complex short-term prediction deep model can be trained for a short-time burst congestion section. Different first-level road section types need to consider different influencing factors when predicting, so different short-term prediction deep models need to be trained for different first-level road section types, i.e., the following steps can also be included:

[0167] The corresponding short-term prediction deep model can be trained based on the second sample data, wherein the second sample data includes short-time traffic condition features of the second sample section at the second sample moment and real traffic time values of each time slice after the second sample moment. The second sample section is a short-time burst congestion section.

[0168] In the implementation mode, the short-term prediction deep model can be trained by using second sample data of a second sample section at a second sample time. The short-time traffic feature in the second sample data can be input into the initial short-term prediction deep model to obtain predicted values of the travel time of each time slice after the second sample time, and the parameters in the short-term prediction deep model can be adjusted according to the error between the predicted values of the travel time of each time slice after the second sample time and the real values of the travel time of each time slice after the second sample time, until the error is reduced to a certain degree, and then the short-term prediction deep model is obtained.

[0169] In the implementation mode, the road section type can include at least one first-level road section type, and different first-level road section types need to consider different influencing factors when predicting. Therefore, different short-term prediction deep models need to be trained for different first-level road section types. In the implementation mode, one short-term prediction deep model needs to be trained for each first-level road section type. The corresponding short-term prediction deep model can be trained based on second sample data, wherein the second sample data includes short-time traffic features of a second sample section of a corresponding first-level road section type at a second sample time, and real values of the travel time of each time slice after the second sample time. The second sample section is a short-time sudden congestion road section of the corresponding first-level road section type. The short-time traffic features of the second sample section of the ordinary road section type at the second sample time include at least one of the following: an average travel time sequence in a short term before the second sample time, a historical average travel time sequence corresponding to each time slice after the second sample time, and a vehicle flow time sequence in a short term before the second sample time. The short-time traffic features of the second sample section of the intersection directional road section type at the second sample time include at least one of the following: an average travel time sequence in a short term before the second sample time, a historical average travel time sequence corresponding to each time slice after the second sample time, a vehicle flow time sequence in a short term before the second sample time, an intersection type, and a turning action type.

[0170] In a possible implementation mode, a more complex long-term prediction deep model can be trained for a long-time sudden congestion road section of a first-level road section type. Different first-level road section types need to consider different influencing factors when predicting. Therefore, the ETA prediction method can further train different long-term prediction deep models for long-time sudden congestion road sections of different first-level road section types, that is, the method can further include the following steps:

[0171] The corresponding long-term prediction deep model can be trained based on third sample data, wherein the third sample data includes long-time traffic features of a third sample section at a third sample time, and real values of the travel time of each time slice after the third sample time. The third sample section is a long-time sudden congestion road section.

[0172] In this implementation, the third sample data of the third sample section at the third sample time can be used to train the corresponding long-term prediction depth model. The long-time traffic feature in the third sample data can be input into the initial long-term prediction depth model to obtain the predicted travel time values of each time slice after the third sample time, and the parameters in the long-term prediction depth model can be adjusted according to the error between the predicted travel time values of each time slice after the third sample time and the real travel time values of each time slice after the third sample time until the error is reduced to a certain degree, and then the long-term prediction depth model is trained.

[0173] In this implementation, the section type can include at least one first section type, such as a normal section and a direction-divided intersection section. Different first section types need to consider different influencing factors when predicting, so different long-term prediction depth models need to be trained for different first section types. In this implementation, a long-term prediction depth model needs to be trained for each first section type. The corresponding long-term prediction depth model can be trained based on third sample data, wherein the third sample data includes the long-time traffic feature of a third sample section of a corresponding first section type at a third sample time, and the real travel time values of each time slice after the third sample time, and the third sample section is a long-time sudden congestion section of the corresponding first section type. The long-time traffic feature of the third sample section of the normal section type at the third sample time includes at least one of the following: an average travel time sequence in a long term before the third sample time, a historical average travel time sequence corresponding to each time slice after the third sample time, and a vehicle flow time sequence in a long term before the third sample time. The long-time traffic feature of the third sample section of the direction-divided intersection section type at the third sample time includes at least one of the following: an average travel time sequence in a long term before the third sample time, a historical average travel time sequence corresponding to each time slice after the third sample time, a vehicle flow time sequence in a long term before the third sample time, an intersection type, and a turning action type.

[0174] In a possible implementation, the short-term prediction depth model or the long-term prediction depth model described above can include a first temporal gate convolution layer (Temporal gate Convolution), a graph attention layer, a second temporal gate convolution layer, a third temporal gate convolution layer, and an MLP (Multilayer Perceptron) layer connected in sequence.

[0175] In the implementation, a model structure of the short-term prediction depth model or the long-term prediction depth model can be that an output end of a first time domain gated convolution layer is connected to an input end of a graph attention layer, an output end of the graph attention layer is connected to an input end of a second time domain gated convolution layer, an output end of the second time domain gated convolution layer is connected to an input end of a third time domain gated convolution layer, and an output end of the third time domain gated convolution layer is connected to an input end of the MLP layer. In this way, the short-term prediction depth model or the long-term prediction depth model can better extract features in the various time series mentioned above, and then make accurate prediction.

[0176] The ETA prediction model training method in the embodiment is the same as the explanation of the related content in the ETA prediction method, and specific details can be referred to the description in the ETA prediction method, which will not be repeated here.

[0177] Figure 3 A structural block diagram of an ETA prediction apparatus according to an embodiment of the present disclosure is shown. The apparatus can be realized as part or all of an electronic device by software, hardware, or a combination of both. As shown in the structural block diagram, the ETA prediction apparatus includes: Figure 3

[0178] The acquisition module 301 is configured to acquire road segment types of road segments in a target route;

[0179] The calculation module 302 is configured to calculate, based on road segment level prediction models corresponding to the road segment types of the road segments, predicted travel time values of the road segments at each future time slice after a departure time;

[0180] The determination module 303 is configured to determine, based on the predicted travel time values of the road segments at each future time slice, predicted entering time points of the road segments;

[0181] The accumulation module 304 is configured to accumulate the predicted travel time values of the road segments at corresponding target time slices to obtain a road segment level estimated travel time of the target route, where the target time slices corresponding to the road segments are future time slices in which the predicted entering time points of the road segments are located;

[0182] The prediction module 305 is configured to predict, based on the road segment level estimated travel time of the target route and related travel characteristics of the target route, a predicted arrival time corresponding to the target route using a route level prediction model.

[0183] In a possible implementation, the road segment types include at least one of the following types: a low-frequency road segment, a stable low-flow road segment, a short-time sudden congestion road segment, and a long-time sudden congestion road segment; and the calculation module 302 is configured to:

[0184] ​when the road section type of the road section is a low-frequency road section, obtaining, based on a historical average travel time of the road section, a travel time prediction value of the road section at each time slice in the future;

[0185] when the road section type of the road section is a stable low-flow road section, using a linear model corresponding to a scene where the road section is located, and obtaining, based on a travel road condition feature of the road section, a travel time prediction value of the road section at each time slice in the future;

[0186] when the road section type of the road section is a short-time sudden congestion road section, using a short-term prediction deep model, and obtaining, based on a short-time road condition feature of the road section, a travel time prediction value of the road section at each time slice in the future;

[0187] when the road section type of the road section is a long-time sudden congestion road section, using a long-term prediction deep model, and obtaining, based on a long-time road condition feature of the road section, a travel time prediction value of the road section at each time slice in the future.

[0188] In a possible implementation, the device further includes:

[0189] a scene determination module configured to determine a scene where the road section is located based on at least one of the following features of the road section: a region feature where the road section is located, a time type of a departure time, and a road level of the road section.

[0190] In a possible implementation, the road section type includes at least one of the following types under a first-level road section type: a low-frequency road section, a stable low-flow road section, a short-time sudden congestion road section, and a long-time sudden congestion road section, the first-level road section type including an ordinary road section or a road intersection direction road section; and the calculation module 302 is configured to:

[0191] when the road section type of the road section is a stable low-flow road section under a corresponding first-level road section type, using a linear model corresponding to a scene where the road section is located, and obtaining, based on a travel road condition feature of the road section under the corresponding first-level road section type, a travel time prediction value of the road section at each time slice in the future, the linear model being a model corresponding to the stable low-flow road section under the corresponding first-level road section type, wherein the travel road condition feature of the road section under the ordinary road section type includes at least one of the following: a real-time travel feature and a historical average travel time sequence corresponding to the time slices in the future; and the travel road condition feature of the road section under the road intersection direction road section type includes at least one of the following: a real-time travel feature, a historical average travel time sequence corresponding to the time slices in the future, a road intersection type, and a turning action type;

[0192] when the road section type of the road section is a short-time burst congestion road section under a corresponding first-level road section type, using a short-time prediction deep model corresponding to the corresponding first-level road section type, and obtaining a predicted value of the travel time of the road section at each future time slice based on short-time traffic characteristics of the road section under the corresponding first-level road section type, wherein the short-time traffic characteristics of the road section under the ordinary road section type include at least one of the following: a real-time average travel time sequence in a short term, a historical average travel time sequence corresponding to each future time slice, and a real-time vehicle flow time sequence in a short term; and the short-time traffic characteristics of the road section under the intersection directional road section type include at least one of the following: a real-time average travel time sequence in a short term, a historical average travel time sequence corresponding to each future time slice, a real-time vehicle flow time sequence in a short term, an intersection type, and a turning action type.

[0193] when the road section type of the road section is a long-time burst congestion road section under a corresponding first-level road section type, using a long-time prediction deep model corresponding to the corresponding first-level road section type, and obtaining a predicted value of the travel time of the road section at each future time slice based on long-time traffic characteristics of the road section under the corresponding first-level road section type, wherein the long-time traffic characteristics of the road section under the ordinary road section type include at least one of the following: a real-time average travel time sequence in a long term, a historical average travel time sequence corresponding to each future time slice, and a real-time vehicle flow time sequence in a long term; and the long-time traffic characteristics of the road section under the intersection directional road section type include at least one of the following: a real-time average travel time sequence in a long term, a historical average travel time sequence corresponding to each future time slice, a real-time vehicle flow time sequence in a long term, an intersection type, and a turning action type.

[0194] In a possible implementation, the short-time prediction deep model or the long-time prediction deep model includes a first time domain gated convolution layer, a graph attention layer, a second time domain gated convolution layer, a third time domain gated convolution layer, and a multi-layer perception (MLP) layer connected in sequence.

[0195] In a possible implementation, the device further includes:

[0196] a first training module configured to train a linear model corresponding to a target scene based on first sample data, wherein the first sample data includes travel traffic characteristics of a first sample road section at a first sample time and real values of travel time at each time slice after the first sample time, and the first sample road section is a stable low-flow road section located in the target scene;

[0197] a second training module configured to train a corresponding short-term prediction depth model based on second sample data, wherein the second sample data comprises short-time traffic feature of a second sample road segment at a second sample time, and true values of travel time of each time slice after the second sample time, and the second sample road segment is a short-time burst congestion road segment;

[0198] a third training module configured to train a corresponding long-term prediction depth model based on third sample data, wherein the third sample data comprises long-time traffic feature of a third sample road segment at a third sample time, and true values of travel time of each time slice after the third sample time, and the third sample road segment is a long-time burst congestion road segment.

[0199] In a possible implementation, the related travel features of the target route comprise at least one of the following features: start and end cities of the target route, departure time, length of each road grade, number of each type of intersection, and number of each type of turning action.

[0200] In a possible implementation, the device further comprises:

[0201] a fourth training module configured to train the route-level prediction model based on fourth sample data, wherein the fourth sample data comprises estimated travel time of a sample route at a road segment level, related travel features of the sample route, and true arrival time of the sample route.

[0202] The ETA prediction device in this embodiment is consistent with the above-described ETA prediction method, and specific details can be referred to the description of the ETA prediction method, which will not be described here.

[0203] Figure 4 A structural block diagram of an ETA prediction model training device according to an embodiment of the present disclosure is shown. The device can be realized as part or all of an electronic device through software, hardware, or a combination of both. As shown in the figure, the ETA prediction model training device 400 comprises: Figure 4

[0204] a road segment-level training module 401 configured to train a road segment-level prediction model corresponding to each road segment type based on sample data of each road segment type;

[0205] ​The intermediate calculation module 402 is configured to calculate, based on the link-level prediction model corresponding to each link type, the predicted travel time of each time slice after the sample link of each type in the sample route at the sample departure time; determine the estimated entry time of each sample link in the sample route based on the predicted travel time of each sample link in the sample route at each time slice; and accumulate the predicted travel time of each sample link at the corresponding target time slice to obtain the estimated link-level travel time of the sample route, wherein the target time slice corresponding to each sample link is the time slice in which the estimated entry time of each sample link is located.

[0206] The route-level training module 403 is configured to train a route-level prediction model based on the estimated link-level travel time of the sample route, the relevant travel characteristics of the sample route, and the real arrival time of the sample route.

[0207] The ETA prediction model training device in this embodiment corresponds to the above-described ETA prediction model training method, and specific details can be referred to the description of the ETA prediction model training method above, which will not be described here again.

[0208] The disclosure also discloses a navigation service, wherein the ETA of the navigation route used by the navigated object is obtained based on the above-described ETA prediction method, and the navigation route selection service of the corresponding scenario is provided for the navigated object based on the ETA of the navigation route. The corresponding scenario is one or a combination of a combination of long-distance AR navigation, elevated navigation, or main and auxiliary road navigation.

[0209] The disclosure also discloses a navigation method, wherein a navigation route is obtained based on at least the starting point, the ending point, and the road condition, the ETA of the navigation route is predicted and displayed, and the navigation route is guided based on the navigation route. The prediction of the ETA of the navigation route is based on any one of the above-described ETA prediction methods.

[0210] In the disclosure, the ETA prediction method can be executed on a server such as a cloud server. The server can perform the ETA prediction according to the above-described ETA prediction method, and provide more accurate location services such as navigation services and path planning services for a mobile terminal. The mobile terminal can be a mobile phone, a pad, an IoT (Internet of Things) device, a vehicle-mounted terminal, and other devices that can display an electronic map. When the server performs navigation or path planning for the mobile terminal, the ETA corresponding to the navigation route or each planning route can be obtained based on the above-described method, and sent to the mobile terminal, so that the mobile terminal displays the ETA corresponding to the route for the user. The user can select a travel route or a travel time based on the ETA.

[0211] Figure 5 This diagram illustrates an application scenario in navigation based on an embodiment of the present disclosure. For example... Figure 5 As shown, the mobile terminal and the server communicate via a wireless network based on communication standards, such as WiFi, 2G, 3G, 4G / LTE, 5G, or a combination thereof. When a user enters their selected origin and destination on the mobile terminal for navigation, the mobile terminal sends these to the server. The server can then recommend navigation routes based on these origin and destination, obtaining multiple candidate routes. It then predicts the estimated arrival time for each candidate route using the aforementioned method and selects a predetermined number of recommended routes from the candidate routes based on these estimated arrival times. The server sends each recommended navigation route and its corresponding estimated arrival time to the mobile terminal, which can receive and display these routes. The user can then select their desired navigation route based on the estimated arrival time of each recommended route, such as choosing the route with the shortest estimated arrival time.

[0212] This disclosure also provides a server, Figure 6 A structural block diagram of a server according to an embodiment of the present disclosure is shown.

[0213] like Figure 6 As shown, the server 600 includes a memory 601 and a processor 602, wherein the memory 601 is used to store one or more computer instructions, wherein the one or more computer instructions are executed by the processor 602 to implement the method according to embodiments of the present disclosure.

[0214] Figure 7 A schematic diagram of a system architecture suitable for implementing a method according to an embodiment of the present disclosure is shown.

[0215] like Figure 7 As shown, the server system 700 includes a processing unit 701, which can execute various processes described in the above embodiments according to a program stored in a read-only memory (ROM) 702 or a program loaded from a storage section 708 into a random access memory (RAM) 703. The RAM 703 also stores various programs and data required for the operation of the system 700. The processing unit 701, ROM 702, and RAM 703 are interconnected via a bus 704. An input / output (I / O) interface 705 is also connected to the bus 704.

[0216] The following components are connected to the I / O interface 705: an input section 706 including a keyboard, a mouse, etc.; an output section 707 including a display such as a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 708 including a hard disk, etc.; and a communication section 709 including a network interface card such as a LAN card, a modem, etc. The communication section 709 performs communication processing via a network such as the Internet. The drive 710 is also connected to the I / O interface 705 as necessary. A removable medium 711 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc. is attached to the drive 710 as necessary, so that a computer program read out therefrom is installed in the storage section 708 as necessary. The processing unit 701 can be implemented as a CPU, a GPU, a TPU, a FPGA, a NPU, etc.

[0217] In particular, the methods described above can be implemented as a computer software program according to embodiments of the present disclosure. For example, embodiments of the present disclosure include a computer program product comprising computer instructions which, when executed by a processor, implement the method steps described above. In such embodiments, the computer program product can be downloaded and installed from a network via the communication section 709, and / or installed from the removable medium 711.

[0218] The flow diagrams and the block diagrams in the drawings are illustrations of possible architectures, functions, and operations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flow diagrams and the block diagrams can represent a module, a segment, or a portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that in some alternative implementations, the functions noted in the blocks can occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently or the blocks may

[0219] The units or modules described in the embodiments of the present disclosure can be implemented by means of software, or by means of programmable hardware. The described units or modules can also be provided in a processor, and the names of these units or modules do not constitute a limitation on the units or modules themselves in some cases.

[0220] As another aspect, the disclosure also provides a computer readable storage medium, which can be the computer readable storage medium contained in the electronic device or the computer system in the above embodiments; or can be a computer readable storage medium existing separately and not assembled into a device. The computer readable storage medium stores one or more programs used by one or more processors to execute the method described in the disclosure.

[0221] The above description is merely the preferred embodiments of the disclosure and the explanation of the principles of the applied technology. It should be understood by those skilled in the art that the inventive scope of the disclosure is not limited to the technical solutions formed by the specific combinations of the above technical features, and should also cover other technical solutions formed by the combinations of the above technical features or equivalent features without departing from the inventive concept. For example, the technical solutions formed by the mutual replacement of the above features and the technical features disclosed in the disclosure (but not limited to) having similar functions.

Claims

1. A method for predicting an estimated time of arrival, comprising: obtaining a link type of each link in a target route, wherein the link type comprises at least one of: a low-frequency link, a stable low-flow link, a short-time burst congestion link, and a long-time burst congestion link; calculating a predicted travel time value of each link at each future time slice after a departure time based on a link-level prediction model corresponding to the link type of the link; determining an estimated entry time of each link based on the predicted travel time value of the link at each future time slice; accumulating the predicted travel time value of each link at a target time slice corresponding to the estimated entry time of the link to obtain a link-level estimated travel time of the target route, wherein the target time slice corresponding to each link is a future time slice in which the estimated entry time of the link is located; predicting an estimated time of arrival corresponding to the target route based on the link-level estimated travel time of the target route and a related travel feature of the target route using a route-level prediction model.

2. The method of claim 1, wherein, The calculating of the predicted travel time value of each link at each future time slice after the departure time based on the link-level prediction model corresponding to the link type of the link comprises: when the link type of a link is a low-frequency link, obtaining the predicted travel time value of the link at each future time slice based on a historical average travel time of the link; when the link type of a link is a stable low-flow link, obtaining the predicted travel time value of the link at each future time slice based on a travel condition feature of the link using a linear model corresponding to a scenario in which the link is located; when the link type of a link is a short-time burst congestion link, obtaining the predicted travel time value of the link at each future time slice based on a short-time travel condition feature of the link using a short-term prediction deep model; when the link type of a link is a long-time burst congestion link, obtaining the predicted travel time value of the link at each future time slice based on a long-time travel condition feature of the link using a long-term prediction deep model.

3. The method of claim 2, wherein, The method further comprises: determining a scenario in which the link is located based on at least one of the following features of the link: a region feature in which the link is located, a time type of the departure time, and a road level of the link.

4. The method of claim 2, wherein, The link type comprises at least one of the following types under a first-level link type: a low-frequency link, a stable low-flow link, a short-time burst congestion link, and a long-time burst congestion link, wherein the first-level link type comprises a common link or a link-directional link at an intersection. In a case where the road section type of the road section is a stable low-flow road section under a corresponding primary road section type, a linear model corresponding to the stable low-flow road section under the corresponding primary road section type is used to obtain predicted values of travel time of the road section at future time slices based on travel condition features of the road section under the corresponding primary road section type, wherein the travel condition features of the road section under the ordinary road section type include at least one of real-time travel features and a historical average travel time sequence corresponding to the future time slices, and the travel condition features of the road section under the intersection directional road section type include at least one of real-time travel features, a historical average travel time sequence corresponding to the future time slices, an intersection type and a turning action type. In a case where the road section type of the road section is a short-time burst congestion road section under a corresponding primary road section type, a short-term prediction deep model corresponding to the corresponding primary road section type is used to obtain predicted values of travel time of the road section at future time slices based on short-time road condition features of the road section under the corresponding primary road section type, wherein the short-time road condition features of the road section under the ordinary road section type include at least one of a real-time average travel time sequence in a short term, a historical average travel time sequence corresponding to the future time slices and a real-time vehicle flow time sequence in the short term, and the short-time road condition features of the road section under the intersection directional road section type include at least one of a real-time average travel time sequence in a short term, a historical average travel time sequence corresponding to the future time slices, a real-time vehicle flow time sequence in the short term, an intersection type and a turning action type. In a case where the road section type of the road section is a long-time burst congestion road section under a corresponding primary road section type, a long-term prediction deep model corresponding to the corresponding primary road section type is used to obtain predicted values of travel time of the road section at future time slices based on long-time road condition features of the road section under the corresponding primary road section type, wherein the long-time road condition features of the road section under the ordinary road section type include at least one of a real-time average travel time sequence in a long term, a historical average travel time sequence corresponding to the future time slices and a real-time vehicle flow time sequence in the long term, and the long-time road condition features of the road section under the intersection directional road section type include at least one of a real-time average travel time sequence in a long term, a historical average travel time sequence corresponding to the future time slices, a real-time vehicle flow time sequence in the long term, an intersection type and a turning action type.

5. The method of claim 2, wherein, The short-term prediction deep model or the long-term prediction deep model includes a first time domain gated convolution layer, a graph attention layer, a second time domain gated convolution layer, a third time domain gated convolution layer and a multi-layer perception (MLP) layer connected in sequence.

6. The method of claim 2, wherein, The method further includes: training a linear model corresponding to a target scene based on first sample data, wherein the first sample data includes travel condition features of a first sample road section at a first sample time and real values of travel time at each time slice after the first sample time, and the first sample road section is a stable low-flow road section located in the target scene. training a corresponding short-term prediction deep model based on second sample data, wherein the second sample data comprises short-time traffic feature of a second sample road segment at a second sample time, and real values of travel time of each time slice after the second sample time, and the second sample road segment is a short-time sudden congestion road segment; training a corresponding long-term prediction deep model based on third sample data, wherein the third sample data comprises long-time traffic feature of a third sample road segment at a third sample time, and real values of travel time of each time slice after the third sample time, and the third sample road segment is a long-time sudden congestion road segment.

7. The method of claim 1, wherein, The related travel features of the target route include at least one of the following features: start and end point cities of the target route, departure time, length of each road grade, number of each type of intersection, and number of each type of turning action.

8. The method of claim 1, wherein, The method further comprises: training the route-level prediction model based on fourth sample data, wherein the fourth sample data comprises estimated travel time of a sample route at a road segment level, related travel features of the sample route, and real arrival time of the sample route. 9.A method for training a predicted arrival time prediction model, comprising: training a road segment-level prediction model corresponding to each road segment type based on sample data of each road segment type, wherein the road segment type includes at least one of the following types: low-frequency road segment, stable low-flow road segment, short-time sudden congestion road segment, and long-time sudden congestion road segment; calculating predicted values of travel time of each sample road segment of a sample route at each time slice after a sample departure time based on the road segment-level prediction model corresponding to each road segment type; determining predicted entering time of each sample road segment of the sample route based on the predicted values of travel time of each sample road segment at each time slice; accumulating the predicted values of travel time of each sample road segment at a corresponding target time slice to obtain estimated travel time of the sample route at a road segment level, wherein the target time slice corresponding to each sample road segment is a time slice in which the predicted entering time of each sample road segment is located; training a route-level prediction model based on the estimated travel time of the sample route at the road segment level, related travel features of the sample route, and real arrival time of the sample route.

10. An electronic device comprising a memory and a processor; wherein, The memory is configured to store one or more computer instructions, wherein the one or more computer instructions are executed by the processor to implement the method steps of any one of claims 1 to 9. 11.A readable storage medium having computer instructions stored thereon, wherein the computer instructions are executed by a processor to implement the method steps of any one of claims 1 to 9. 12.A computer program product comprising computer instructions, wherein the computer instructions are executed by a processor to implement the method steps of any one of claims 1 to 9.

13. A navigation method, wherein, A navigation route is obtained based on at least a starting point, an end point, and a road condition, a predicted arrival time of the navigation route is predicted and displayed, and navigation guidance is performed based on the navigation route, and the prediction of the predicted arrival time of the navigation route is implemented based on any one of the methods of claims 1 to 9.

Images