Predicted arrival time prediction method and device, computer equipment and storage medium

Abstract

The invention discloses a predicted arrival time prediction method and device, computer equipment and a storage medium, and the method comprises the steps: dividing a target route into a plurality of sub-road segments, and obtaining the road basic attribute information of each sub-road segment; based on the road basic attribute information, predicting to obtain a basic prediction value corresponding to the target route; embedding the road marks corresponding to the sub-road segments to obtain hidden feature vectors of the sub-road segments; obtaining a compensation value corresponding to the target route based on the hidden feature vector of each sub-road segment; and obtaining predicted arrival time based on the basic predicted value and the compensation value. And obtaining a compensation value of the trajectory through a multi-layer neural network by mining hidden features of each sub-road in a deeper level. And the compensation value of the trajectory and the basic prediction value are accumulated as the final predicted arrival time, so that the fitting effect is better, and the prediction accuracy can be improved.

Description

Technical Field

[0001] The present application relates to the field of deep learning technology, and in particular to an estimated arrival time prediction method, apparatus, computer equipment and storage medium. Background Art

[0002] Estimated Time of Arrival (ETA) is a basic function in map software and is widely used in transportation, logistics distribution, and many other fields involving itinerary planning and time estimation.

[0003] At present, the commonly used ETA prediction methods can be mainly divided into methods based on historical data statistical analysis or traditional machine learning models. Methods based on historical data statistical analysis usually rely on a large amount of past travel data records. For example, in the field of transportation, the average driving time of a specific road section in different time periods is counted, and the ETA is estimated in combination with the current time period information. However, since it is based only on historical average data, once an abnormal situation occurs, the prediction results often have large deviations and cannot accurately reflect the actual travel time. Methods based on traditional machine learning models, such as linear regression and decision trees, can consider multiple factors, such as road type, distance, time period and other characteristics. However, traditional machine learning models have limited capabilities in dealing with complex nonlinear relationships. This leads to problems such as the predicted results not fitting the trajectory ATA well enough, being prone to overall underestimation or overestimation, and being not sensitive enough to specific road sections (prone to congestion). Summary of the invention

[0004] Based on this, it is necessary to provide an estimated arrival time prediction method, device, computer equipment and storage medium to address the above technical problems, so as to solve at least one problem existing in the prior art.

[0005] In a first aspect, a method for predicting an estimated arrival time is provided, comprising the following steps:

[0006] Divide the target route into multiple sub-road segments, and obtain basic road attribute information of each sub-road segment;

[0007] Based on the basic road attribute information, a basic prediction value corresponding to the target route is predicted;

[0008] Embed the road signs corresponding to each sub-road segment to obtain the hidden feature vector of each sub-road segment;

[0009] Based on the hidden feature vectors of each sub-road segment, obtaining a compensation value corresponding to the target route;

[0010] An estimated arrival time is obtained based on the basic prediction value and the compensation value.

[0011] In one embodiment, the predicting of a basic prediction value corresponding to the target route based on the basic road attribute information includes:

[0012] Dividing the basic road attribute information into sparse features and dense features;

[0013] Performing splicing processing on the sparse features and the dense features to obtain spliced ​​features;

[0014] Based on the splicing features, a basic prediction value corresponding to the target route is predicted.

[0015] In one embodiment, the step of performing splicing processing on the sparse features and the dense features to obtain spliced ​​features includes:

[0016] Performing embedding processing on each of the sparse features;

[0017] Normalizing each of the dense features;

[0018] The sparse features after embedding processing and the dense features after normalization processing are concatenated to obtain the concatenated features;

[0019] In one embodiment, predicting a basic prediction value corresponding to the target route based on the splicing feature includes:

[0020] Based on the splicing features, predicting the road features corresponding to each of the sub-road segments;

[0021] Predicting the road features corresponding to each of the sub-road segments to obtain a prediction value corresponding to each of the sub-road segments;

[0022] The prediction values ​​corresponding to the sub-road segments are accumulated to obtain a basic prediction value corresponding to the target route.

[0023] In one embodiment, predicting the road features corresponding to each of the sub-road segments based on the splicing features includes:

[0024] Performing convolution processing on the concatenated features to obtain input features;

[0025] According to the time step, each input feature is gated and updated in turn to obtain the road feature corresponding to each sub-road segment.

[0026] In one embodiment, obtaining the compensation value corresponding to the target route based on the hidden feature vectors of each sub-road segment includes:

[0027] Accumulating the hidden feature vectors of each of the sub-road segments to obtain a trajectory compensation value vector;

[0028] Based on the trajectory compensation value vector, a compensation value corresponding to the target route is predicted.

[0029] In one embodiment, after obtaining the estimated arrival time based on the basic prediction value and the compensation value, the method further includes:

[0030] Based on the basic prediction value and the first loss function, a first loss value is calculated;

[0031] Based on the compensation value and the second loss function, a second loss value is calculated;

[0032] Based on the first loss value, the second loss value and the parameter penalty weight, a total loss value is calculated to optimize the model based on the total loss value.

[0033] In a second aspect, a device for predicting an estimated arrival time is provided, comprising:

[0034] A road basic attribute information acquisition unit, used to divide the target route into a plurality of sub-road segments and acquire the road basic attribute information of each sub-road segment;

[0035] A basic prediction value prediction unit, used to predict a basic prediction value corresponding to the target route based on the basic road attribute information;

[0036] A hidden feature vector acquisition unit is used to embed the road signs corresponding to each sub-road segment to obtain the hidden feature vector of each sub-road segment;

[0037] A compensation value obtaining unit, configured to obtain a compensation value corresponding to the target route based on the hidden feature vectors of each sub-road segment;

[0038] The estimated arrival time acquisition unit is used to obtain the estimated arrival time based on the basic prediction value and the compensation value.

[0039] In a third aspect, a computer device is provided, comprising a memory, a processor, and computer-readable instructions stored in the memory and running on the processor, wherein the processor implements the estimated arrival time prediction method as described above when executing the computer-readable instructions.

[0040] In a fourth aspect, a readable storage medium is provided, on which computer-readable instructions are stored. When the computer-readable instructions are executed by a processor, the estimated arrival time prediction method as described above is implemented.

[0041] The above-mentioned estimated arrival time prediction method, device, computer equipment and storage medium, the method implementation includes: dividing the target route into multiple sub-road segments, and obtaining the basic attribute information of each sub-road segment; based on the basic attribute information of the road, predicting the basic prediction value corresponding to the target route; embedding the road signs corresponding to each sub-road segment to obtain the hidden feature vector of each sub-road segment; based on the hidden feature vector of each sub-road segment, obtaining the compensation value corresponding to the target route; based on the basic prediction value and the compensation value, obtaining the estimated arrival time. In the embodiment of the present application, the prediction value of each sub-road segment is predicted by reading the basic attribute information of each sub-road segment, and the prediction value of each sub-road segment is accumulated as the vector of the entire trajectory to obtain the basic prediction value of the trajectory. Then, embedding is performed based on the index of each sub-road segment identifier to obtain the hidden feature vector of each sub-road segment, and the trajectory compensation value is obtained through a multi-layer neural network. Finally, the basic prediction value plus the compensation value is used as the final estimated arrival time. By digging deeper into the hidden features of each sub-road, the prediction result fits the target route better, which can improve the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the description of the embodiments of the present application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0043] Figure 1 This is a flow chart of an estimated arrival time prediction method in one embodiment of the present application;

[0044] Figure 2 This is a schematic diagram of a model architecture of an estimated arrival time prediction model in one embodiment of the present application;

[0045] Figure 3 It is a schematic diagram of an application environment for the training and prediction process of the estimated arrival time prediction model in one embodiment of the present application;

[0046] Figure 4 is a structural schematic diagram of an estimated arrival time prediction device in one embodiment of the present application;

[0047] Figure 5 It is a schematic diagram of a computer device in one embodiment of the present application. DETAILED DESCRIPTION

[0048] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0049] In one embodiment, if Figure 1 As shown, a method for predicting an estimated arrival time is provided, comprising the following steps:

[0050] In step S110, the target route is divided into a plurality of sub-road segments, and basic road attribute information of each sub-road segment is obtained;

[0051] In an embodiment of the present application, the basic road attribute information may include road ID, road length, road grade, congestion level, traffic lights and other information.

[0052] Optionally, the starting location information and the destination location information may be obtained, and based on the starting location information and the destination location information, a driving trajectory may be constructed as a target route, and the target route may be composed of multiple sub-road segments (links). The road attribute information corresponding to each link may be queried, and an attribute data set may be constructed for subsequent use.

[0053] In step S120, based on the basic road attribute information, a basic prediction value corresponding to the target route is predicted;

[0054] Optionally, the basic road attribute information of each sub-road segment in the attribute data set can be divided into sparse features and dense features. The sparse features (sparse) can be embedded (embedding), and the results can be spliced ​​with dense features (dense) features. Then, they can be input into the feature extraction model constructed by the convolutional neural network (CNN) and the long short-term memory network (LSTM) for processing to obtain the corresponding vector (vector) of each sub-road segment. After that, the obtained vector (vector) is predicted through the fully connected network to obtain the corresponding prediction value of each sub-road segment. The corresponding prediction values ​​of all sub-road segments are accumulated to obtain the basic prediction value pre_ETA of the target route.

[0055] In step S130, embedding processing is performed on the road signs corresponding to each sub-road segment to obtain a hidden feature vector of each sub-road segment;

[0056] In the embodiment of the present application, each sub-road segment has a unique identifier, namely link-id, which can be a digital number, an alphanumeric combination, etc. Since sub-road identifiers are usually expressed in discrete, symbolic forms (such as digital numbers, alphanumeric combinations, etc.), it is difficult to use them directly to participate in subsequent complex calculations and to perform effective fusion and analysis operations with other road-related features. Through embedding processing (embedding), these discrete indexes can be mapped to a low-dimensional continuous vector space, thereby mining their potential hidden features. For example, hidden features such as whether the road is often congested or has fewer vehicles.

[0057] Optionally, a suitable vector space dimension may be set first, for example, to map the sub-road identifier index to a vector space of dimension d (d can be set according to actual needs and data scale, such as 64-dimensional, 128-dimensional, etc.). Then, an initial mapping table of vector representations corresponding to all sub-road identifier indexes may be created. For example, if the sub-road identifier index is "A1", "A2", "B1", etc., a randomly generated vector of dimension d may be assigned to each sub-road identifier, such as "A1" corresponding to the vector [v11, v12, ..., v1d]. By querying the mapping table, the hidden feature vector corresponding to each sub-road segment may be determined.

[0058] It should be noted that the above mapping table can be learned based on road-related information, such as traffic flow, congestion, road grade, etc. For example, training can be performed based on a model structure such as a neural network, and the vector corresponding to the sub-road mark and other road-related features can be input as training data to train the model, thereby obtaining a mapping relationship between the sub-road mark index and the vector.

[0059] In step S140, based on the hidden feature vectors of each sub-road segment, a compensation value corresponding to the target route is obtained;

[0060] Optionally, the hidden feature vectors of the sub-road segments may be added together to obtain a compensation value vector, and then the compensation value vector may be predicted by a preset network, such as a fully connected network, to obtain a compensation value (offset). It is understandable that other networks may also be used to obtain the compensation value, such as a convolutional neural network (CNN), a recurrent neural network (RNN), etc.

[0061] In step S150, an estimated arrival time is obtained based on the basic prediction value and the compensation value.

[0062] Optionally, if the data format of the basic prediction value and the compensation value is the same, for example, the basic prediction value is a time length in minutes, the compensation value should also be a time correction in minutes. For example, the basic prediction value is that the estimated driving time from A to B is 30 minutes, and the compensation value is a delay of 5 minutes due to road construction, then the final estimated arrival time is 30+5=35 minutes. It should be noted that there are cases where the compensation value is negative, for example, if there are fewer vehicles on the road and the traffic is smoother, the driving time will be shortened, for example, 5 minutes, then the estimated arrival time is 30-5=25 minutes.

[0063] Optionally, if the data formats of the basic prediction value and the compensation value are different, for example, the basic prediction value is accurate to minutes and the compensation value is accurate to seconds, they need to be unified to the same accuracy before adding them. At the same time, the result after addition may need to be adjusted in accuracy according to the actual application scenario. For example, the final estimated arrival time may only need to be retained to integer minutes in some cases, but may need to be accurate to seconds in scenarios with higher requirements.

[0064] In an embodiment of the present application, a method for predicting an estimated arrival time is provided, comprising: dividing a target route into a plurality of sub-road segments, and obtaining basic road attribute information of each sub-road segment; predicting a basic prediction value corresponding to the target route based on the basic road attribute information; embedding the road signs corresponding to each sub-road segment to obtain a hidden feature vector of each sub-road segment; obtaining a compensation value corresponding to the target route based on the hidden feature vector of each sub-road segment; and obtaining an estimated arrival time based on the basic prediction value and the compensation value. In an embodiment of the present application, the prediction value of each sub-road segment is predicted by reading the basic attribute information of each sub-road segment, and the prediction value of each sub-road segment is accumulated as the vector of the entire trajectory to obtain the basic prediction value of the trajectory. Then, embedding is performed based on the index of each sub-road segment identifier to obtain the hidden feature vector of each sub-road segment, and the trajectory compensation value is obtained through a multi-layer neural network. Finally, the basic prediction value plus the compensation value is used as the final estimated arrival time. By digging deeper into the hidden features of each sub-road, the prediction result fits the target route better, which can improve the prediction accuracy.

[0065] In an embodiment of the present application, the step of predicting a basic prediction value corresponding to the target route based on the basic road attribute information includes:

[0066] Dividing the basic road attribute information into sparse features and dense features;

[0067] Performing splicing processing on the sparse features and the dense features to obtain spliced ​​features;

[0068] Based on the splicing features, a basic prediction value corresponding to the target route is predicted.

[0069] It should be noted that sparse features refer to features in which most of the values ​​in the road attribute dataset are zero (or empty values, etc., indicating that the road does not exist or does not appear), and only a few of them have non-zero values. For example, in traffic data, for each road (link), there may be many different types of traffic event markers (such as traffic accidents, road construction, etc.), but most of the time these events do not occur on most links. In this case, the features related to these traffic events are sparse, and only have specific values ​​when the corresponding events occur at a specific time and on a specific road section. Overall, they are sparsely distributed in the entire dataset. Dense features refer to features that are relatively evenly distributed in the road attribute dataset. In most cases, they have specific and valid values, and there are no large numbers of empty or zero values. For example, certain attributes such as the length of the road, the width of the road, and the number of lanes.

[0070] Optionally, after obtaining sparse features and dense features, the sparse features and dense features can be spliced ​​together. For example, the vector dimension corresponding to the sparse features is [m] (m-dimensional vector), and the vector dimension composed of the dense features is [n] (n-dimensional vector). The two can be directly combined into a new vector with a dimension of [m+n] as a spliced ​​feature in the manner of vector splicing.

[0071] In an embodiment of the present application, the step of performing splicing processing on the sparse features and the dense features to obtain spliced ​​features includes:

[0072] Performing embedding processing on each of the sparse features;

[0073] Normalizing each of the dense features;

[0074] The sparse features after embedding processing and the dense features after normalization processing are concatenated to obtain the concatenated features;

[0075] Optionally, sparse features can be embedded. For example, a large amount of sample data containing sparse features can be collected, such as data sets with various sparse feature values ​​for many road segments in different time periods. Then, a statistical model can be constructed. For example, if two sparse features often appear at the same time in the same section or a similar section, it means that they should be close in the mapped vector space. Based on this principle, the vector representation corresponding to each sparse feature is continuously adjusted through an optimization algorithm, and finally a low-dimensional vector after each sparse feature mapping is obtained, and a model for embedding processing can be obtained. The currently acquired sparse features can be input into the model to obtain the embedding results corresponding to the sparse features. Each dense feature can be normalized separately, and then the normalized dense features and the sparse features after embedding are spliced ​​to obtain spliced ​​features.

[0076] In an embodiment of the present application, the step of predicting a basic prediction value corresponding to the target route based on the splicing feature includes:

[0077] Based on the splicing features, predicting the road features corresponding to each of the sub-road segments;

[0078] Predicting the road features corresponding to each of the sub-road segments to obtain a prediction value corresponding to each of the sub-road segments;

[0079] The prediction values ​​corresponding to the sub-road segments are accumulated to obtain a basic prediction value corresponding to the target route.

[0080] Optionally, the concatenated features can be processed sequentially through a convolutional neural network (CNN) and a long short-term memory network (LSTM) to obtain the road features (vector) corresponding to each sub-road segment. The road features can then be input into a fully connected network with several layers (1 to 3 layers) for prediction processing to obtain the predicted value of each sub-road segment (link). The predicted values ​​of all sub-road segments (links) are accumulated to obtain the basic prediction value pre_ETA.

[0081] The CNN may be a one-layer one-dimensional CNN, and the LSTM may be a one-layer LSTM. It should be noted that the number of layers and dimensions of the CNN can be adjusted according to actual conditions, and similarly, the number of layers of the LSTM can also be adjusted according to actual conditions.

[0082] It should be noted that one-dimensional CNN is mainly used to process sequence data. After receiving the vector composed of the spliced ​​sparse features and dense features, the convolution kernel will slide along this one-dimensional sequence to perform convolution operations. The convolution kernel has a certain width and specific weight parameters. During the sliding process, it performs dot product operations with the vector elements at the corresponding position, so that local features can be extracted. Then, after the feature sequence output from the one-dimensional CNN is passed to the LSTM, as the time step (that is, step by step along the feature sequence) progresses, the LSTM will process the features corresponding to each element transmitted from the one-dimensional CNN in turn, and finally generate a vector representation of a fixed dimension, namely the road feature. The cell state and hidden state in the LSTM will be continuously updated to memorize and transmit the sequence information. The key structures of the input gate, forget gate and output gate are used to control the inflow, forgetting and output of information.

[0083] The vector obtained after convolution processing and long short-term memory model (LSTM) processing is used as the input of the fully connected network to the fully connected network. A fully connected network is used as an example. Each layer of neurons in the fully connected network is connected to all neurons in the next layer. For the input vector, it will first be multiplied with the weight matrix of the layer to achieve linear transformation. After the linear transformation, the prediction value corresponding to each sub-road segment can be obtained. By accumulating the prediction value corresponding to each sub-road segment, the basic prediction value corresponding to the target route is obtained. For example, a target route contains link1, link2, link3, and the corresponding prediction values ​​are P1, P2, and P3, respectively. Then the basic prediction value pre_ETA=P1+P2+P3.

[0084] It should be noted that if the fully connected network includes multiple layers, the output of the first layer of the fully connected network will be used as the input of the second layer, and so on. Each layer is constantly further abstracting, integrating and transforming the features passed from the previous layer, and the basic prediction value can be output through the last layer.

[0085] In an embodiment of the present application, predicting the road features corresponding to each of the sub-road segments based on the splicing features includes:

[0086] Performing convolution processing on the concatenated features to obtain input features;

[0087] According to the time step, each input feature is gated and updated in turn to obtain the road feature corresponding to each sub-road segment.

[0088] Optionally, the convolution layer can use a set of convolution kernels (filters) to scan the input sequence. For a one-dimensional CNN, the width of the convolution kernel can correspond to the feature dimension of the input sequence. For example, if there are n features at each time step of the input data, the width of the convolution kernel is also n. The convolution kernel can slide along the length of the sequence, and each slide calculates the dot product of the convolution kernel and the corresponding input data part, adds the results of these dot products, and introduces nonlinearity through an activation function (such as ReLU) to obtain a new feature value, that is, the input feature. When the convolution kernel completes processing of all concatenated features, a new set of feature sequences will be generated. LSTM processes each time step in the feature sequence generated by CNN one by one, and continuously updates its internal hidden state and cell state. After processing the entire feature sequence, the hidden state of the last time step can be selected as the output. The road features corresponding to each of the sub-road segments can be obtained.

[0089] In an embodiment of the present application, obtaining the compensation value corresponding to the target route based on the hidden feature vectors of each sub-road segment includes:

[0090] Accumulating the hidden feature vectors of each of the sub-road segments to obtain a trajectory compensation value vector;

[0091] Based on the trajectory compensation value vector, a compensation value corresponding to the target route is predicted.

[0092] Optionally, the hidden feature vectors of all sub-road segments can be added in sequence according to the corresponding element positions. For example, assuming that there are m sub-road segments, the hidden feature vector of each sub-road segment is represented as v_i (where i takes values ​​from 1 to m), and the dimension of each vector is n, then for the first element of the vector, the first element of v_1, the first element of v_2, ..., and so on to the first element of v_m are added to obtain the first element of the accumulated vector; in the same way, the other elements of the vector (from the second to the nth) are also respectively subjected to such accumulation operations to obtain the trajectory supplement value vector. Then, several layers (1 to 3 layers) of fully connected networks are applied to the obtained compensation value vector to obtain the compensation value offset. Take a layer of fully connected network for illustration, where each layer of neurons in the fully connected network is connected to all neurons in the next layer. For the input trajectory compensation value vector, it will first be multiplied with the weight matrix of the layer to achieve linear transformation. After the linear transformation, the compensation value corresponding to each of the sub-road segments can be obtained.

[0093] It should be noted that if the fully connected network includes multiple layers, the output of the first layer of the fully connected network will be used as the input of the second layer, and so on. Each layer is constantly further abstracting, integrating, and linearly transforming the features passed from the previous layer, and the compensation value can be output through the last layer.

[0094] In the embodiment of the present application, after obtaining the estimated arrival time based on the basic prediction value and the compensation value, the method further includes:

[0095] Based on the basic prediction value and the first loss function, a first loss value is calculated;

[0096] Based on the compensation value and the second loss function, a second loss value is calculated;

[0097] Based on the first loss value, the second loss value and the parameter penalty weight, a total loss value is calculated to optimize the model based on the total loss value.

[0098] Alternatively, see Figure 2 , an estimated time of arrival prediction model can be constructed, and the estimated time of arrival prediction model may include a first embedding module for embedding sparse features, a normalization module for normalizing dense features, and a concatenation module (Comcat) for concatenating the embedded sparse features with the normalized dense features. A convolutional neural network (CNN) connected to the concatenation module, and a long short-term memory network (LSTM) connected to the convolutional neural network (CNN) to obtain the road features corresponding to each sub-road segment, and then the prediction value of each link can be obtained through the first fully connected network connected to the long short-term memory network (LSTM), and the prediction values ​​of all links are accumulated to obtain the basic prediction value pre_ETA. The estimated time of arrival prediction model can also include a second embedding module for embedding the index of each sub-road segment identifier to obtain a hidden feature vector, and an accumulation layer connected to the second embedding module for accumulating all hidden feature vectors to obtain a trajectory compensation value vector, and then the compensation value is obtained through the second fully connected network connected to the accumulation layer. The final estimated arrival time can be obtained by adding the basic prediction value to the compensation value pre_ETA+offset.

[0099] Alternatively, see Figure 3, the training process of the estimated arrival time prediction model is as follows: the road network and the current road conditions can be matched through the user trajectory on a daily basis and in real time to generate offline training data. After preprocessing the training data, for example, after filtering out abnormal situations such as abnormal stops, missing trajectories, and detours, the basic map data, historical road conditions, and trajectory mining features are used to generate training features corresponding to all links on the trajectory and save them as csv files. Then, the training features can be input into the model in sequence for iterative training, and the basic prediction value obtained in each iteration process is calculated through the first loss function to obtain the first loss value, and the compensation value is calculated through the second loss function to obtain the second loss value, and the total loss value is calculated based on the first loss value, the second loss value, and the parameter penalty weight. If the total loss value is greater than the preset loss threshold, the model parameters are adjusted, and the remaining training features are selected again to enter the model for prediction. Until the number of training times is greater than the preset number of times, or the total loss value is less than or equal to the preset loss threshold, the training is stopped to obtain a trained model.

[0100] Among them, the total loss value (Loss) is calculated by the following formula:

[0101]

[0102] Among them, mape is the loss value of the basic prediction value calculated by the mean absolute percentage error, Indicates the loss value of the compensation value; diff i represents the difference value, 0ffset represents the compensation value, α represents the weight parameter; L1 is the regularization term; eta i represents the predicted arrival time of the training sample, ata i represents the actual arrival time of the training sample, and λ represents the penalty weight;

[0103]

[0104] When the driver or user inputs the starting point and the end point through the application pre-installed on the terminal, the route calculation service can be called through the background. The route calculation service obtains the route through path planning, that is, the target route. The target route can be divided into multiple sub-road segments, and the basic road attribute information of each sub-road segment is obtained. Based on the basic road attribute information, the basic prediction value corresponding to the target route is predicted; the road signs corresponding to each sub-road segment are embedded to obtain the hidden feature vectors of each sub-road segment; based on the hidden feature vectors of each sub-road segment, the compensation value corresponding to the target route is obtained; based on the basic prediction value and the compensation value, the estimated arrival time, that is, the driving time of the target route, is obtained. The eta service calls the libtorch library through C++ code to load this model to provide the eta return value. Finally, it is returned to the user or driver's client through layer-by-layer calls.

[0105] It should be noted that there is a problem of too much data during the training process, so an innovative data loading solution is designed. First, the CSV is split by day, and then the training is loaded in a loop. This method saves memory and improves the model effect. If the data volume is still large, a certain proportion of the training data will be randomly sampled and loaded into the dataloader. In order to facilitate data alignment, the dataloader will arrange the trajectories from long to short, add 0 at the end to align the data, and use batch for training.

[0106] It is understandable that the indicators generally used to evaluate the accuracy of ETA are MAPE (mean absolute percentage error) and good rate. The traditional estimated arrival time prediction method generally has a MAPE of about 12.5%. By applying the technical solution of this application, MAPE can reach 11.5%. The good rate is increased from 76.4% to 84.6%. It is aligned with the leading level in the industry, and the model is simple and the response time is short.

[0107] In the embodiment of the present application, the prediction value of each sub-road segment is obtained by reading the basic attribute information of each sub-road segment, and the prediction value of each sub-road segment is accumulated as the vector of the entire trajectory to obtain the basic prediction value of the trajectory. Then, embedding is performed based on the index of each sub-road segment identifier to obtain the hidden feature vector of each sub-road segment, and the trajectory compensation value is obtained through a multi-layer neural network. Finally, the basic prediction value plus the compensation value is used as the final estimated arrival time. By digging deeper into the hidden features of each sub-road, the prediction result can better fit the target route, which can improve the prediction accuracy.

[0108] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0109] In one embodiment, an estimated arrival time prediction device is provided, and the estimated arrival time prediction device corresponds to the estimated arrival time prediction method in the above embodiment. Figure 4 As shown, the estimated arrival time prediction device includes a road basic attribute information acquisition unit 10, a basic prediction value prediction unit 20, a hidden feature vector acquisition unit 30, a compensation value acquisition unit 40 and an estimated arrival time acquisition unit 50. Each functional module is described in detail as follows:

[0110] The basic road attribute information acquisition unit 10 is used to divide the target route into a plurality of sub-road segments and acquire the basic road attribute information of each sub-road segment;

[0111] A basic prediction value prediction unit 20, configured to predict a basic prediction value corresponding to the target route based on the basic road attribute information;

[0112] A hidden feature vector acquisition unit 30 is used to embed the road signs corresponding to each sub-road segment to obtain the hidden feature vector of each sub-road segment;

[0113] A compensation value acquisition unit 40, configured to obtain a compensation value corresponding to the target route based on the hidden feature vectors of each sub-road segment;

[0114] The estimated arrival time acquisition unit 50 is used to obtain an estimated arrival time based on the basic prediction value and the compensation value.

[0115] In one embodiment of the present application, the basic prediction value prediction unit 20 is further used to:

[0116] Dividing the basic road attribute information into sparse features and dense features;

[0117] Performing splicing processing on the sparse features and the dense features to obtain spliced ​​features;

[0118] Based on the splicing features, a basic prediction value corresponding to the target route is predicted.

[0119] In one embodiment of the present application, the basic prediction value prediction unit 20 is further used to:

[0120] Performing embedding processing on each of the sparse features;

[0121] Normalizing each of the dense features;

[0122] The sparse features after embedding processing and the dense features after normalization processing are concatenated to obtain the concatenated features;

[0123] In one embodiment of the present application, the basic prediction value prediction unit 20 is further used to:

[0124] Based on the splicing features, predicting the road features corresponding to each of the sub-road segments;

[0125] Predicting the road features corresponding to each of the sub-road segments to obtain a prediction value corresponding to each of the sub-road segments;

[0126] The prediction values ​​corresponding to the sub-road segments are accumulated to obtain a basic prediction value corresponding to the target route.

[0127] In one embodiment of the present application, the basic prediction value prediction unit 20 is further used to:

[0128] Performing convolution processing on the concatenated features to obtain input features;

[0129] According to the time step, each input feature is gated and updated in turn to obtain the road feature corresponding to each sub-road segment.

[0130] In one embodiment of the present application, the compensation value acquisition unit 40 is further configured to:

[0131] Accumulating the hidden feature vectors of each of the sub-road segments to obtain a trajectory compensation value vector;

[0132] Based on the trajectory compensation value vector, a compensation value corresponding to the target route is predicted.

[0133] In one embodiment of the present application, the device further includes a loss value calculation unit, which is used to:

[0134] Based on the basic prediction value and the first loss function, a first loss value is calculated;

[0135] Based on the compensation value and the second loss function, a second loss value is calculated;

[0136] Based on the first loss value, the second loss value and the parameter penalty weight, a total loss value is calculated to optimize the model based on the total loss value.

[0137] In the embodiment of the present application, the prediction value of each sub-road segment is obtained by reading the basic attribute information of each sub-road segment, and the prediction value of each sub-road segment is accumulated as the vector of the entire trajectory to obtain the basic prediction value of the trajectory. Then, embedding is performed based on the index of each sub-road segment identifier to obtain the hidden feature vector of each sub-road segment, and the trajectory compensation value is obtained through a multi-layer neural network. Finally, the basic prediction value plus the compensation value is used as the final estimated arrival time. By digging deeper into the hidden features of each sub-road, the prediction result can better fit the target route, which can improve the prediction accuracy.

[0138] For the specific definition of the estimated arrival time prediction device, please refer to the definition of the estimated arrival time prediction method above, which will not be repeated here. Each module in the above-mentioned estimated arrival time prediction device can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0139] In one embodiment, a computer device is provided. The computer device may be a terminal device, and its internal structure diagram may be as follows: Figure 5 As shown. The computer device includes a processor, a memory, and a network interface connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a readable storage medium. The readable storage medium stores computer-readable instructions. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer-readable instructions are executed by the processor, an estimated arrival time prediction method is implemented. The readable storage medium provided in this embodiment includes a non-volatile readable storage medium and a volatile readable storage medium.

[0140] In an embodiment of the present application, a computer device is provided, including a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor, wherein the processor implements the steps of the estimated arrival time prediction method as described above when executing the computer-readable instructions.

[0141] In an embodiment of the application, a readable storage medium is provided, wherein the readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by a processor, the steps of the estimated arrival time prediction method as described above are implemented.

[0142] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing related hardware through computer-readable instructions, and the computer-readable instructions can be stored in a non-volatile readable storage medium or a volatile readable storage medium. When the computer-readable instructions are executed, they may include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0143] Those skilled in the art can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0144] The embodiments described above are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A method for predicting an estimated arrival time, characterized in that: The method comprises: Divide the target route into multiple sub-road segments, and obtain basic road attribute information of each sub-road segment; Based on the basic road attribute information, a basic prediction value corresponding to the target route is predicted; Embed the road signs corresponding to each sub-road segment to obtain the hidden feature vector of each sub-road segment; Based on the hidden feature vectors of each sub-road segment, obtaining a compensation value corresponding to the target route; An estimated arrival time is obtained based on the basic prediction value and the compensation value.

2. The estimated arrival time prediction method according to claim 1, characterized in that: The step of predicting a basic prediction value corresponding to the target route based on the basic road attribute information includes: Dividing the basic road attribute information into sparse features and dense features; Performing splicing processing on the sparse features and the dense features to obtain spliced ​​features; Based on the splicing features, a basic prediction value corresponding to the target route is predicted.

3. The estimated arrival time prediction method according to claim 2, characterized in that: The step of performing splicing processing on the sparse features and the dense features to obtain spliced ​​features includes: Performing embedding processing on each of the sparse features; Normalizing each of the dense features; The sparse features after the embedding process and the dense features after the normalization process are concatenated to obtain the concatenated features.

4. The estimated arrival time prediction method according to claim 2 or 3, characterized in that: The step of predicting a basic prediction value corresponding to the target route based on the splicing feature includes: Based on the splicing features, predicting the road features corresponding to each of the sub-road segments; Predicting the road features corresponding to each of the sub-road segments to obtain a prediction value corresponding to each of the sub-road segments; The prediction values ​​corresponding to the sub-road segments are accumulated to obtain a basic prediction value corresponding to the target route.

5. The estimated arrival time prediction method according to claim 4, characterized in that: The predicting of the road features corresponding to each of the sub-road segments based on the splicing features includes: Performing convolution processing on the concatenated features to obtain input features; According to the time step, gate operation and state update are performed on each of the input features in turn to obtain the road features corresponding to each of the sub-road segments.

6. The estimated arrival time prediction method according to claim 1, characterized in that: The obtaining of the compensation value corresponding to the target route based on the hidden feature vectors of each sub-road segment includes: Accumulating the hidden feature vectors of each of the sub-road segments to obtain a trajectory compensation value vector; Based on the trajectory compensation value vector, a compensation value corresponding to the target route is predicted.

7. The estimated arrival time prediction method according to claim 1, characterized in that: After obtaining the estimated arrival time based on the basic prediction value and the compensation value, the method further includes: Based on the basic prediction value and the first loss function, a first loss value is calculated; Based on the compensation value and the second loss function, a second loss value is calculated; Based on the first loss value, the second loss value and the parameter penalty weight, a total loss value is calculated to optimize the model based on the total loss value.

8. An estimated arrival time prediction device, characterized in that: The device comprises: A road basic attribute information acquisition unit, used to divide the target route into a plurality of sub-road segments and acquire the road basic attribute information of each sub-road segment; A basic prediction value prediction unit, used to predict a basic prediction value corresponding to the target route based on the basic road attribute information; A hidden feature vector acquisition unit is used to embed the road signs corresponding to each sub-road segment to obtain the hidden feature vector of each sub-road segment; A compensation value obtaining unit, configured to obtain a compensation value corresponding to the target route based on the hidden feature vectors of each sub-road segment; The estimated arrival time acquisition unit is used to obtain the estimated arrival time based on the basic prediction value and the compensation value.

9. A computer device comprising a memory, a processor, and computer-readable instructions stored in the memory and executed on the processor, characterized in that: When the processor executes the computer-readable instructions, the estimated arrival time prediction method according to any one of claims 1 to 7 is implemented.

10. A readable storage medium having computer readable instructions stored thereon, characterized in that: When the computer-readable instructions are executed by a processor, the estimated arrival time prediction method according to any one of claims 1 to 7 is implemented.

Images