Road speed prediction method fusing multi-source spatio-temporal features, medium and electronic device

Abstract

The application provides a road speed prediction method based on fusion of multi-source space-time characteristics, a medium and an electronic device. The method comprises: acquiring road speed characteristic data, time characteristic data, weather characteristic data and road network correlation characteristic data of a target road respectively, and constructing a multi-dimensional feature input sequence based on the road speed characteristic data, the time characteristic data, the weather characteristic data and the road network correlation characteristic data; inputting the multi-dimensional feature input vector into a pre-trained multi-input branch network model, processing each sequence in the multi-dimensional feature input vector by each branch of the multi-input branch network model respectively, and performing characteristic fusion and splicing on the sequences processed by each branch through a fusion layer to obtain a fusion feature; and inputting the fusion feature into a full connection layer to output speed prediction values of the target road at multiple future time steps. The application can effectively solve the technical problems of insufficient multi-source feature fusion, insufficient road network correlation modeling and low prediction accuracy in the prior art.

Description

Technical Field

[0001] This application belongs to the field of intelligent transportation and traffic big data analysis technology, and in particular relates to the field of road speed prediction technology. Background Technology

[0002] With the continuous expansion of logistics and transportation in the port area, road traffic conditions are becoming increasingly dynamic and uncertain. Accurately predicting road speeds is crucial for improving the efficiency of port operations and alleviating traffic congestion.

[0003] Existing road speed prediction methods mainly include statistical model-based methods and deep learning-based methods. Traditional statistical models struggle to characterize complex nonlinear spatiotemporal dependencies; while existing deep learning methods, although improving prediction accuracy to some extent, still have the following shortcomings: first, their fusion methods for multi-source heterogeneous data (such as time, weather, and road network structure) are relatively simple, lacking effective feature interaction mechanisms; second, they are insufficient in modeling the special road network structure of port areas (such as truck transport routes and key node coupling relationships); and third, they lack the ability to collaboratively model different time scales (short-term fluctuations and periodic changes).

[0004] Therefore, there is an urgent need for a road speed prediction method that can effectively integrate multi-source spatiotemporal information and characterize the complex road network relationships. Summary of the Invention

[0005] This application provides a road speed prediction method, medium, and electronic device that integrates multi-source spatiotemporal features to solve the technical problems of insufficient multi-source feature fusion, inadequate road network correlation modeling, and low prediction accuracy in the prior art.

[0006] In a first aspect, embodiments of this application provide a road speed prediction method that integrates multi-source spatiotemporal features, comprising: acquiring road speed feature data, time feature data, weather feature data, and road network association feature data of a target road respectively; constructing a multi-dimensional feature input sequence based on the road speed feature data, the time feature data, the weather feature data, and the road network association feature data; inputting the multi-dimensional feature input vector into a pre-trained multi-input branch network model; processing each sequence in the multi-dimensional feature input vector through each branch of the multi-input branch network model; and performing feature fusion and splicing on the sequences processed by each branch through a fusion layer to obtain a fused feature; and inputting the fused feature into a fully connected layer to output the speed prediction value of the target road at multiple future time steps.

[0007] In one implementation of the first aspect, obtaining the road network association feature data includes: obtaining the spatiotemporal coding vectors of adjacent roads and calculating the spatial attention weights of each adjacent road; and using the spatial attention weights to perform weighted aggregation on the spatiotemporal coding vectors of the adjacent roads to obtain the road network association features.

[0008] In one implementation of the first aspect, obtaining the spatiotemporal encoding vector of adjacent roads includes: constructing a directed road network graph based on the road topology and determining the set of adjacent roads of the target road; calculating the constraint time delay relative to the target road for each adjacent road in the set of adjacent roads; using the constraint time delay to time-align the historical speed sequence of the adjacent roads to obtain the time-delay-aligned adjacent sequence; and inputting the adjacent sequence into a long short-term memory network to obtain the spatiotemporal encoding vector of the adjacent roads.

[0009] In one implementation of the first aspect, the calculation of the spatial attention weight of each adjacent road includes: obtaining the prior connection strength between roads in the directed road network graph, the historical correlation between adjacent roads and the target road, and the traffic propagation delay parameter determined by the constraint delay, as well as their respective weight parameters; and calculating the spatial attention weight of each adjacent road based on the prior connection strength, historical correlation, traffic propagation delay parameter, and their respective weight parameters.

[0010] In one implementation of the first aspect, the directed road network graph uses roads as nodes, the actual travel directions between roads as edges, and includes the prior connection strength between roads; wherein, the prior connection strength between roads is a functional relationship between the distance between roads, the capacity parameter, and the topological connection type.

[0011] In one implementation of the first aspect, the constraint time delay is calculated by maximizing the cross-correlation coefficient between the historical speed sequences of the target road and adjacent roads, or by the ratio of the distance between roads to the average speed; and the constraint time delay does not exceed the length of the time window.

[0012] In one implementation of the first aspect, the method further includes: encoding the historical speed time window sequence of the target road using a long short-term memory network to obtain its own spatiotemporal encoding vector; and concatenating the difference between the weighted aggregation vector and the self-spatiotemporal encoding vector with the weighted aggregation vector to form a road network association feature with enhanced neighborhood influence and self-difference.

[0013] In one implementation of the first aspect, the multi-input branch network model is trained under supervision using historical data, and the loss function of the multi-input branch network model is the mean squared error function; the multi-input branch network model introduces a sliding time window to construct a training sample sequence during the training process.

[0014] In a second aspect, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the road speed prediction method fused with multi-source spatiotemporal features as described in any one of the first aspects of this application.

[0015] Thirdly, embodiments of this application provide an electronic device, the electronic device including a processor and a memory; the memory stores program instructions; the processor is used to run the program instructions to execute the road speed prediction method fusion of multi-source spatiotemporal features as described in any one of the first aspects of this application.

[0016] The road speed prediction method that integrates multi-source spatiotemporal features provided in this application has the following beneficial effects:

[0017] This application integrates time period, weather disturbance, road history, and road network correlation features to construct a unified multi-dimensional input, eliminating differences in heterogeneous data structures and accurately capturing the daily and weekly rhythms and sudden response patterns of traffic flow. Through traffic propagation time delay alignment and attention aggregation based on physical prior constraints, it characterizes the propagation delay and spatial diffusion effects of congestion waves in the road network, enhancing adaptability to complex topologies. Employing multi-branch LSTM parallel encoding and a time-dimensional attention mechanism, it dynamically adjusts the contribution weights of each feature in different scenarios, combined with direct multi-step output to avoid error accumulation. It maintains stable prediction accuracy in highly dynamic scenarios such as peak hours and sudden weather events, demonstrating strong generalization ability and high practical value. This effectively solves the technical problems of insufficient multi-source feature fusion, inadequate road network correlation modeling, and low prediction accuracy in existing technologies. Attached Figure Description

[0018] Figure 1 The diagram shown is an overall flowchart of a road speed prediction method that integrates multi-source spatiotemporal features according to an embodiment of this application.

[0019] Figure 2 The flowchart shown is a process for obtaining road network association features in a road speed prediction method that integrates multi-source spatiotemporal features according to an embodiment of this application.

[0020] Figure 3 The flowchart shown is a process for obtaining the spatiotemporal coding vector of adjacent roads in a road speed prediction method that integrates multi-source spatiotemporal features according to an embodiment of this application.

[0021] Figure 4The diagram shows the principle of obtaining spatial attention weights in a road speed prediction method that integrates multi-source spatiotemporal features according to an embodiment of this application.

[0022] Figure 5 The flowchart shown is a process for obtaining enhanced road network association features in a road speed prediction method that integrates multi-source spatiotemporal features according to an embodiment of this application.

[0023] Figure 6 The diagram shown is a schematic representation of the structure of an electronic device according to an embodiment of this application.

[0024] Component designation explanation

[0025] Electronic device - 100; memory - 101; processor - 102; display - 103; S100~S300 - steps; S110~S120 - steps; S111~S14 - steps; S510~S520 - steps. Detailed Implementation

[0026] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0027] Existing technologies for predicting road speeds in port areas typically rely solely on single historical speed time-series features for modeling, neglecting the significant impact of external disturbances such as weather and holidays on traffic conditions. Furthermore, traditional methods often simplify road network spatial relationships to static adjacency matrices or homogenization effects, failing to effectively characterize the physical time delays and dynamic diffusion differences in congestion waves propagating between different roads. In addition, fixed weights are often used in multi-feature fusion, lacking the ability to adaptively adjust the contribution of each factor under different traffic scenarios, resulting in insufficient prediction accuracy and robustness in highly dynamic and complex scenarios such as peak congestion and sudden weather events.

[0028] The purpose of this embodiment is to provide a road speed prediction method, medium, and electronic device that integrates multi-source spatiotemporal features. It proposes an LSTM road speed prediction method that integrates time features, weather features, and road network correlation features. The aim is to solve the technical problems of insufficient multi-source feature fusion, inadequate road network correlation modeling, and low prediction accuracy in the existing technology, so as to improve the accuracy and stability of road speed prediction in port areas.

[0029] The following will refer to the appendices in the embodiments of this application. Figure 1 To be continued Figure 6 This application provides a detailed description of the road speed prediction method, medium, and electronic device that integrates multi-source spatiotemporal features in the embodiments of this application. This allows those skilled in the art to understand and implement the road speed prediction method, medium, and electronic device that integrates multi-source spatiotemporal features in this embodiment without inventive effort.

[0030] This embodiment provides a road speed prediction method that integrates multi-source spatiotemporal features. It is applied to port area roads, but not limited to port area roads. It can be extended to urban expressways and main roads, highway sections and toll stations, road networks around large transportation hubs such as airports and railway stations, and dedicated roads in industrial and mining parks. It can achieve accurate prediction of road speed in different transportation systems and has good scene transferability and versatility.

[0031] Figure 1 The flowchart shown is a road speed prediction method that integrates multi-source spatiotemporal features, as described in an embodiment of this application. Figure 1 As shown in the embodiments of this application, the road speed prediction method that integrates multi-source spatiotemporal features includes the following steps S100 to S300.

[0032] Step S100: Obtain road speed feature data, time feature data, weather feature data, and road network association feature data of the target road respectively, and construct a multi-dimensional feature input sequence based on the road speed feature data, the time feature data, the weather feature data, and the road network association feature data;

[0033] Step S200: Input the multidimensional feature input vector into a pre-trained multi-input branch network model. Process each sequence in the multidimensional feature input vector through each branch of the multi-input branch network model. Then, perform feature fusion and splicing on the sequences processed by each branch through a fusion layer to obtain fused features.

[0034] Step S300: Input the fused features into the fully connected layer and output the speed prediction value of the target road in the future multiple time steps.

[0035] This embodiment of the road speed prediction method integrating multi-source spatiotemporal features constructs a multi-dimensional feature input sequence by unifying road speed features, time features, weather features, and road network correlation features. This achieves deep fusion of multi-source heterogeneous data at the input layer, enabling the model to simultaneously capture the temporal periodicity of traffic flow, the influence of external meteorological disturbances, and the spatial dependencies between roads. Based on this, a multi-input branch network model is used to independently encode and extract features from each feature sequence. A fusion layer organically combines cross-source features, effectively preserving the temporal evolution patterns and dynamic differences of different features, avoiding information loss caused by traditional single-sequence modeling. Finally, a fully connected layer directly outputs the predicted speed values ​​of the target road for multiple future time steps, completing multi-step prediction in an end-to-end manner and avoiding error accumulation from iterative prediction. Overall, this embodiment of the road speed prediction method integrating multi-source spatiotemporal features significantly improves the prediction accuracy and robustness of port area roads in highly dynamic and highly disturbed complex scenarios, and has good scene transferability and engineering practical value.

[0036] The following is in conjunction with the appendix Figure 2 To be continued Figure 5 The above steps S100 to S300 of the road speed prediction method that integrates multi-source spatiotemporal features in this embodiment will be described in detail.

[0037] Step S100: Obtain road speed feature data, time feature data, weather feature data, and road network association feature data of the target road respectively, and construct a multi-dimensional feature input sequence based on the road speed feature data, the time feature data, the weather feature data, and the road network association feature data.

[0038] In this embodiment, in the prediction of road speed in the port area, the traffic state of the target road at any time is affected by the time evolution law, the spatial dependence of the road network and the external environmental disturbance. In this embodiment, multi-source heterogeneous data are uniformly constructed into a multi-dimensional feature input vector to achieve deep fusion of the time series features, spatial dependence features and external disturbance features.

[0039] For the target road At time t, the multidimensional feature input sequence The expression is:

[0040]

[0041] in, This is the time feature vector corresponding to time t, used to characterize the periodic temporal evolution of traffic flow; This is the weather feature vector corresponding to time t, used to characterize the real-time impact of external meteorological disturbances on road traffic conditions; For the target road The road speed feature vector at time t is used to characterize the historical speed temporal state of the target road itself; For the target road The road network association feature vector at time t is used to characterize the spatial dependency and traffic propagation impact between the target road and adjacent roads.

[0042] This embodiment unifies time series features, external disturbance features, road state features, and spatial dependency features into a multi-dimensional feature input vector, enabling the model to simultaneously learn the complementary relationships and interactions between multiple data sources within a single input framework, thereby improving the dynamic prediction capability of port area road speeds.

[0043] In this embodiment, historical road data of the target road is collected. The historical road data includes the speed value, traffic flow data, lane occupancy data and road attribute data of the target road at the current time and at historical time. The road attribute data includes at least one or more of lane number information and speed limit information.

[0044] The speed values ​​of the target road at each historical moment are extracted from the historical road data. Based on a preset time window length, the extracted speed values ​​at each historical moment are constructed into a time window sequence, the expression of which is:

[0045] in, to These represent the target road at time [time]. -1 to time -n represents the historical velocity values, where n is the length of the time window. This is the road speed feature sequence of the target road at time t.

[0046] In this embodiment, the time feature data includes, but is not limited to, one or more combinations of time-based encoding information, weekday information, holiday identification information, and periodic encoding information.

[0047] The time-based encoding information is used to characterize the hourly and minute positions of the current time within a day, in order to capture the distribution patterns of traffic flow at different times of the day; the weekday information is used to distinguish whether the current date is a weekday or a weekend, in order to reflect the systematic differences in traffic travel patterns under different date types; the holiday identification information is used to identify whether the current date is a statutory holiday or a special holiday, in order to capture the characteristics of abnormal changes in traffic flow during holidays; the periodic encoding information maps discrete time variables into continuous periodic vectors through sine and / or cosine functions, in order to capture the intraday periodic variation patterns of traffic flow.

[0048] In this embodiment, the periodic mapping processing of the time feature data transforms discrete time into a continuous vector representation. For example, the number of hours in the time-time encoding information is periodically mapped using a sine function or a cosine function to obtain the corresponding sine periodic encoding value or cosine periodic encoding value. For instance, the periodic encoding information is determined according to the following formula:

[0049]

[0050] In the formula, t represents the number of hours corresponding to the current moment. The sinusoidal period encoding value corresponds to the current time. Through the sinusoidal period encoding value, adjacent hourly times are kept continuous in the vector space, enabling the model to learn the smooth transition characteristics of traffic flow over time within a day.

[0051] In this embodiment, the weather feature data includes, but is not limited to, one or more combinations of precipitation information, temperature information, wind speed information, and visibility information.

[0052] The precipitation information is used to characterize the precipitation intensity in the target area at the current moment, reflecting the direct impact of precipitation on road surface slipperiness and vehicle speed; the temperature information is used to characterize the ambient temperature value in the target area at the current moment, capturing the potential impact of high or low temperatures on vehicle performance and traffic conditions; the wind speed information is used to characterize the wind intensity in the target area at the current moment, reflecting the impact of strong winds on vehicle stability and speed, especially in areas near the port where the wind effect is more significant; the visibility information is used to characterize the horizontal visibility distance in the target area at the current moment, reflecting the degree to which weather phenomena such as fog, haze, and precipitation restrict driver visibility and road capacity.

[0053] In this embodiment, the method further includes: normalizing the precipitation information, the temperature information, the wind speed information, and the visibility information respectively, mapping the values ​​of each weather feature data to a preset numerical range, so as to eliminate the problem of inconsistent numerical scales caused by the difference in units between different weather features.

[0054] The normalization process ensures that all weather features are of the same numerical magnitude before being input into the model, preventing features with large numerical ranges from dominating gradient updates during model training, thereby improving the stability and convergence speed of model training.

[0055] In this embodiment, the road network association features are used to characterize the impact of the traffic conditions of adjacent roads on the future speed of the target road, enabling the model to capture the spatial propagation and diffusion effects of traffic flow in the road network. The road network association features are feature vectors obtained by encoding and aggregating the historical speed sequences of each adjacent road in the set of adjacent roads of the target road.

[0056] Figure 2 The flowchart shown is a process for obtaining road network association features in a road speed prediction method that integrates multi-source spatiotemporal features, according to an embodiment of this application. Figure 2 As shown, in one implementation of this embodiment, obtaining the road network associated feature data includes the following steps S110 to S120.

[0057] Step S110: Obtain the spatiotemporal encoding vectors of adjacent roads and calculate the spatial attention weights of each adjacent road.

[0058] Step S120: The spatial attention weights are used to perform weighted aggregation of the spatiotemporal coding vectors of the adjacent roads to obtain road network association features.

[0059] In this embodiment, firstly, the spatiotemporal encoding vectors of each adjacent road in the set of adjacent roads of the target road are obtained. Then, based on the correlation strength between the target road and each adjacent road, the spatial attention weights corresponding to each adjacent road are calculated. The spatiotemporal encoding vectors are extracted from the historical speed sequences of each adjacent road after time-delay alignment using a time-series coding network. The spatial attention weights are determined based on at least one of the following factors: prior connection strength between roads, historical correlation, and traffic propagation characteristics. Next, the spatiotemporal encoding vectors of each adjacent road are weighted and aggregated using the spatial attention weights to obtain the road network correlation feature vector corresponding to the target road at that time. This weighted aggregation allows adjacent roads with a closer correlation to the target road to have a greater impact on its future speed prediction, while the impact of adjacent roads with weaker correlations is correspondingly suppressed.

[0060] This embodiment uses a spatial attention mechanism to adaptively weight and aggregate information about adjacent roads, enabling the model to dynamically distinguish the differentiated impacts of different adjacent roads on the target road, rather than treating all adjacent roads as equal. At the same time, the road network association features effectively characterize the spatial propagation pattern of traffic flow and the congestion diffusion effect in the port area road network, allowing the model to perceive the transmission of traffic pressure from adjacent roads in advance, thereby improving the prediction accuracy and response speed of future speed changes of the target road.

[0061] Figure 3The flowchart shown is a process for obtaining the spatiotemporal coding vectors of adjacent roads in a road speed prediction method that integrates multi-source spatiotemporal features according to an embodiment of this application. In one implementation of this embodiment, obtaining the spatiotemporal coding vectors of adjacent roads includes the following steps S111 to S114.

[0062] Step S111: Construct a directed road network graph based on the road topology and determine the set of adjacent roads of the target road.

[0063] In one implementation of this embodiment, the directed road network graph uses roads as nodes, the actual travel directions between roads as edges, and includes the prior connection strength between roads; wherein, the prior connection strength between roads is a functional relationship between the distance between roads, the capacity parameter, and the topological connection type.

[0064] The port area road network is abstracted as a directed graph.

[0065] in: The set of road nodes is defined as follows: each node in the set of road nodes corresponds to a road in the port area road network, and each road is represented as an independent node. The set of directed edges represents an actual traffic connection between two roads, and the direction of the edge is consistent with the allowed vehicle traffic direction in the port area road network. It is used to depict the legal propagation path of traffic flow between roads. Let be the set of prior connection strengths between roads, where Represents road nodes With road nodes The prior connection strength between roads is used to characterize the degree of spatial association between roads based on the static topology of the road network, and provides spatial prior constraints for the weighted aggregation of subsequent adjacent road information.

[0066] In this embodiment, the prior connection strength between roads is determined by static modeling, and its expression is:

[0067] in, Represents road nodes With road nodes Spatial distance parameters between them Represents road nodes With road nodes Traffic capacity parameters between Represents road nodes With road nodes Topology connection type parameters between them This is the preset mapping function.

[0068] The spatial distance parameter This parameter is used to characterize the spatial proximity between roads, reflecting the distance attenuation effect of traffic flow in space; that is, the closer the spatial distance between roads, the more significant the mutual influence of their traffic conditions. According to road nodes With road nodes The number of lanes and / or road grade of each road are determined to characterize the traffic flow carrying capacity between roads; roads with greater capacity have a stronger impact on the traffic conditions of other roads; the topology connection type parameter Used to distinguish the road nodes With road nodes The connection methods between roads include at least direct connection and turning connection, where the traffic flow propagation resistance between roads is less under direct connection than under turning connection.

[0069] Through the static modeling, the road network structure information, basic information on the adjacency relationships between roads, and prior weights of spatial influence of the port area road network are output, providing prior spatial constraints based on physical topology for the subsequent dynamic aggregation of adjacent road information.

[0070] In this embodiment, obtaining the set of adjacent roads of the target road includes:

[0071] For the target road According to the directed road network diagram The set of directed edges in Extract the target road The target road is formed by all the road nodes pointed to by the directed edges starting from the origin. The set of adjacent roads is defined as follows:

[0072] in, Indicates the target road The set of adjacent roads, Represents the first in the port area road network a road, This indicates that there exists a path from the road node in the directed road network graph. Point to road node The directed edge.

[0073] The set of adjacent roads The target road was defined. The range of adjacent roads that have a direct spatial impact is defined as only including adjacent roads that have a direct passage connection with the target road and whose traffic flow can be directly reached by the target road in the spatial feature aggregation range. This avoids introducing road nodes unrelated to the target road into the calculation, reduces noise interference, and lowers computational complexity.

[0074] Step S112: For each adjacent road in the set of adjacent roads, calculate its constraint time delay relative to the target road.

[0075] The constraint time delay is used to ensure the consistency between the traffic propagation time delay and the time window length. Through this constraint time delay, it is ensured that during time delay alignment, the historical speed sequences of adjacent roads maintain a correspondence with the input time window of the target road in the time dimension. That is, the historical data of adjacent roads after time delay offset still falls within the historical time range covered by the time window, thereby ensuring that the extracted features of adjacent roads match the current input features of the target road in the time dimension, avoiding the loss of effective historical information due to time delays exceeding the window range.

[0076] In one implementation of this embodiment, the constraint time delay is calculated by maximizing the cross-correlation coefficient between the historical speed sequences of the target road and adjacent roads, and its expression is:

[0077]

[0078] in, To constrain time delay, Indicates the target road The velocity value at time t, Indicates the adjacent road At time t- speed value, Represents the cross-correlation function, The time-delay variable is to be optimized; by selecting the variable that maximizes the cross-correlation coefficient... The value is used as a traffic propagation time lag to reflect traffic flow from the adjacent road. propagate to the target road Required feature delay time.

[0079] In another embodiment of this invention, the constraint time delay can also be calculated using the ratio of the distance between roads to the average speed.

[0080] That is, through the target road Adjacent roads Road distance between With the preset average traffic speed The ratio is used to calculate the traffic propagation time delay, and its expression is:

[0081]

[0082] Furthermore, in this embodiment, the constraint delay does not exceed the length of the time window. A constraint condition is applied to the traffic propagation delay to ensure that the traffic propagation delay does not exceed the preset time window length. Its constraint expression is:

[0083]

[0084] The constraints ensure that the historical speed sequences of adjacent roads after time-delay alignment match the input time window of the target road in the time dimension, thus preventing the effective historical information from exceeding the window range and becoming unusable due to excessive time delay.

[0085] Step S113: Use the constraint time delay to perform time alignment on the historical speed sequence of adjacent roads to obtain the time-delay aligned adjacent sequence.

[0086] For the set of adjacent roads Each adjacent road in According to its corresponding traffic propagation time lag The adjacent roads Historical velocity sequence shifted backward along the time axis After several time steps, the time-delay aligned adjacency sequence is obtained, and its expression is:

[0087] in, Indicates adjacent roads The velocity sequence after time-delay alignment, where τ is the adjacent road Relative to the target road The traffic propagation time delay, i.e., τ = τij; Indicates the adjacent road At any moment Speed ​​value; A preset time-delay alignment sequence length is used to ensure that the time-delay aligned adjacent sequence is consistent with the input time window of the target road in terms of time range. Through this time alignment process, the adjacent roads... In the historical velocity sequence at time The traffic condition changes that occurred and the target road At any moment The traffic conditions correspond in time, thus explicitly encoding the delay effect of traffic flow propagation from adjacent roads to the target road into the input features, enabling the model to accurately learn the temporal causal relationship of congestion waves propagating in the road network.

[0088] After time-delay alignment, the time-delay aligned adjacency sequence corresponding to each adjacent road is output. The time-delay aligned adjacency sequence is used to realign the historical speed information of adjacent roads according to the time delay of traffic flow propagation, so that the traffic state changes of the adjacent roads at historical moments can accurately correspond to the traffic state of the target road at the current moment in terms of causality, providing time-delay corrected input data for the subsequent sequence encoding of adjacent road features.

[0089] Step S114: Input the adjacency sequence into the Long Short-Term Memory network to obtain the spatiotemporal encoding vector of the adjacency road.

[0090] The adjacency sequence after time delay alignment The adjacent roads are obtained by inputting the sequence into a Long Short-Term Memory (LSTM) network. spatiotemporal encoding vector Its expression is:

[0091]

[0092] , indicating the adjacent road The corresponding fixed-dimensional spatiotemporal encoding vector, Let be the hidden layer dimension of the Long Short-Term Memory network.

[0093] The Long Short-Term Memory (LSTM) network, through its internal forgetting gate, input gate, and output gate gating mechanism, adaptively extracts and encodes long-term dependencies and short-term fluctuation features in the time-delay aligned adjacency sequence, compressing the variable-length sequence data into a fixed-dimensional spatiotemporal encoding vector. This spatiotemporal encoding vector not only contains the adjacency paths... The speed change trend within the time window also implicitly encodes the temporal causal relationship of traffic flow propagation from the adjacent road to the target road through the time delay alignment process.

[0094] The traffic state representation of adjacent roads no longer uses the speed value at a single moment as the feature input. Instead, the complete historical speed sequence of each adjacent road after time-delay alignment is sequence-encoded through a Long Short-Term Memory network to obtain a fixed-dimensional spatiotemporal encoding vector as the feature representation of the adjacent road. Compared with the method of directly using a single speed value, the sequence encoding vector can completely preserve the speed change trend, fluctuation characteristics and dynamic evolution law of adjacent roads within a past time window, rather than just reflecting the instantaneous state at a certain isolated moment. This provides a richer and more robust representation of the dynamic behavior of adjacent roads for subsequent spatial feature aggregation, effectively avoiding spatial feature distortion caused by random fluctuations or noise interference of speed values ​​at a single moment, and significantly improving the information carrying capacity and temporal stability of road network associated features.

[0095] Figure 4 This diagram illustrates the principle of obtaining spatial attention weights in a road speed prediction method that integrates multi-source spatiotemporal features, as described in an embodiment of this application. Figure 4 As shown, in one implementation of this embodiment, the calculation of the spatial attention weight of each adjacent road includes: obtaining the prior connection strength between roads in the directed road network graph, the historical correlation between adjacent roads and the target road, and the traffic propagation delay parameters determined by the constraint delay, as well as their respective weight parameters; and calculating the spatial attention weight of each adjacent road based on the prior connection strength, historical correlation, traffic propagation delay parameters, and their respective weight parameters.

[0096] Obtain the target roads respectively With adjacent roads Prior connection strength between Historical relevance and the traffic propagation time delay parameters determined by the constraint time delay. And obtain the first weighting coefficient corresponding to the prior connection strength. The second weighting coefficient corresponding to the historical correlation. and the third weighting coefficient corresponding to the traffic propagation time delay parameter. Based on the prior connection strength, the historical correlation, the traffic propagation time delay parameters, and their respective weighting coefficients, the adjacent roads are calculated. The corresponding impact factor, i.e. Its expression is:

[0097] Wherein, the prior connection strength Used to characterize the degree of spatial correlation between roads based on the static topology of the road network; the historical correlation Used to characterize the target road Adjacent roads The degree of statistical correlation over historical velocity sequences; the traffic propagation time delay parameters According to the constraint time delay The time delay characteristic used to characterize the propagation of traffic flow from the adjacent road to the target road is defined as follows:

[0098]

[0099] Regarding the aforementioned influencing factors The adjacent roads are obtained by performing Softmax normalization. Corresponding spatial attention weights Its expression is:

[0100]

[0101] in, Indicates the target road The set of adjacent roads, This represents the normalized spatial attention weights.

[0102] Based on the aforementioned adjacent roads Corresponding spatial attention weights The spatiotemporal encoding vectors of each adjacent road The target road is obtained by performing a weighted summation. At any moment Weighted aggregation vector Its expression is:

[0103] in, Indicates the target road The set of adjacent roads, Indicates adjacent roads Relative to the target road Spatial attention weights, Indicates adjacent roads The spatiotemporal encoding vector, The weighted aggregation vector; the weighted aggregation vector The spatiotemporal encoding vectors of each adjacent road in the adjacent road set are dynamically fused according to their respective attention weights. This makes the contribution of adjacent roads with closer spatial association with the target road, stronger historical relevance, and smaller traffic propagation time lag more significant to the aggregation result, thereby achieving a precise quantitative representation of the spatial influence of the road network on the target road. The weighted aggregation vector is then used to... As the target road At any moment Road network association characteristics .

[0104] Furthermore, the weighted aggregation vector can be concatenated or differentially processed with the spatiotemporal encoding vector of the target road itself to obtain enhanced road network association features, thereby further strengthening the ability to express the differences between the neighborhood influence and its own state.

[0105] Figure 5 The flowchart shown is a process for obtaining enhanced road network association features in a road speed prediction method that integrates multi-source spatiotemporal features, according to an embodiment of this application. Figure 5 As shown, in one implementation of this embodiment, it further includes:

[0106] Step S510: Encode the historical speed time window sequence of the target road using a long short-term memory network to obtain its own spatiotemporal encoding vector.

[0107] Step S520: The difference between the weighted aggregation vector and its own spatiotemporal encoding vector is concatenated with the weighted aggregation vector to form a road network association feature with neighborhood influence and enhanced self-difference.

[0108] Using a Long Short-Term Memory network for the target road Historical speed time window sequence Sequence encoding is performed to obtain the target road. Self-spatial encoding vector Its expression is:

[0109] =LSTM( )

[0110] The weighted aggregation vector With its own spatiotemporal encoding vector Perform a difference operation to obtain a difference vector representing the difference between the influence of the neighborhood and its own state. ; the weighted aggregation vector With the difference vector The parts are then pieced together to form the target road. At any moment The corresponding road network association features include enhanced neighborhood influence and self-differentiation. Its expression is:

[0111] By introducing the difference vector, the road network association feature not only includes the weighted aggregation information of the traffic states of adjacent roads, but also explicitly encodes the difference information between the target road's own state and the aggregated state of its neighborhood. When the traffic state change of the target road is inconsistent with the neighborhood trend, the difference vector can enhance the model's ability to perceive this abnormal change, making the model more sensitive to the asymmetry and local mutations in the propagation of traffic flow in the road network, thereby improving the recognition accuracy of key dynamic processes such as congestion initiation and dissipation.

[0112] After acquiring the road speed characteristic data, time characteristic data, weather characteristic data, and road network association characteristic data of the target road, a system for identifying the target road is constructed based on these data. The multidimensional feature input sequence at time t, the multidimensional feature input sequence The expression for the single-time feature is:

[0113]

[0114] The multidimensional feature input sequence is constructed using a sliding window approach for the target road. The expression for the input time series at time t is:

[0115] in, Indicates the target road The multidimensional feature input time series corresponding to time t; Indicates the target road The single-time multidimensional feature input vector at time t includes time feature vector, weather feature vector, road speed feature vector, and road network association feature vector; The preset time window length; Indicates the target road At any moment The single-time multidimensional feature input vector corresponding to each time point up to time t.

[0116] The target road is displayed using the sliding window method. In continuous The single-time multidimensional feature input vectors at time +1 are arranged in chronological order to construct a multidimensional feature input sequence for input into the multi-input branch network model. This enables the multi-input branch network model to learn the dynamic evolution of each feature in the time dimension from the multidimensional feature input sequence. At the same time, by continuously sliding the window over historical data, a large number of training samples are constructed, improving the model's ability to capture short-term fluctuation patterns.

[0117] Step S200: Input the multidimensional feature input vector into a pre-trained multi-input branch network model. Process each sequence in the multidimensional feature input vector through each branch of the multi-input branch network model. Then, perform feature fusion and splicing on the sequences processed by each branch through a fusion layer to obtain fused features.

[0118] In this embodiment, the multi-input branch network model includes a time branch, a road branch, a spatial branch, and a weather branch, each branch being used to process the feature sequences of the corresponding category in the multi-dimensional feature input sequence.

[0119] The time branch employs a first long short-term memory network to receive the time feature sequence, extract the temporal periodic evolution pattern of traffic flow, and output the time branch hidden state vector. The road branch employs a second long short-term memory network to receive the historical speed time window sequence of the target road, extract the dynamic change trend of the target road's own speed, and output the hidden state vector of the road branch. The spatial branch employs a third long short-term memory network to receive the road network association feature sequence, extract the temporal evolution pattern of the influence of adjacent roads on the target road, and output the spatial branch hidden state vector. The weather branch employs a fully connected network or a fourth long short-term memory network to receive the weather feature sequence, extract the disturbance features of external meteorological conditions on traffic conditions, and output the weather branch hidden state vector. The hidden state vectors output from each branch are fused and concatenated in the fusion layer to obtain a fused feature vector. Its expression is:

[0120] The fusion layer unifies the temporal features from different sources in the hidden state space, enabling the model to make collaborative predictions by integrating multi-source information such as time patterns, the historical state of the road itself, the spatial influence of the road network, and external meteorological disturbances.

[0121] Step S300: Input the fused features into the fully connected layer and output the speed prediction value of the target road in the future multiple time steps.

[0122] The fused feature vector The input is a fully connected layer, which performs a non-linear mapping on the fused feature vector to output the target road. The expression for the velocity prediction over multiple consecutive time steps in the future is:

[0123]

[0124] in, This represents the fused feature vector formed by concatenating the hidden state vectors of the time branch, road branch, spatial branch, and weather branch. Represents the mapping function of the fully connected layer; The vector of predicted speed values ​​for the target road over multiple future time steps is expressed as follows:

[0125] in, to The target roads are respectively represented In the future At that time Predicted speed values This is the preset number of prediction steps.

[0126] By outputting the predicted velocity values ​​for multiple consecutive time steps in the future through the fully connected layer, the prediction strategy of direct multi-step output avoids the problem of error accumulation caused by iterative rolling based on single-step prediction, thereby ensuring the stability of prediction accuracy within different prediction time ranges.

[0127] In one implementation of this embodiment, historical data is used to supervise the training of the multi-input branch network model, and the loss function of the multi-input branch network model is the mean squared error function; a sliding time window is introduced to construct a training sample sequence during the training process of the multi-input branch network model.

[0128] The training method for the multi-input branch network model is as follows: historical speed data of the target road is used as a supervision signal, and mean squared error is used as the loss function to supervise the training of the multi-input branch network model; the expression of the loss function is:

[0129]

[0130] By minimizing the mean squared error loss function, the Euclidean distance between the multi-step velocity prediction vector output by the multi-input branch network model and the corresponding true velocity vector is minimized, thereby achieving joint optimization of velocity values ​​for multiple future time steps and ensuring the overall accuracy of multi-step prediction.

[0131] During the training of the multi-input branch network model, a sliding time window is used to slide over the historical velocity sequence with a fixed step size, and the window length at each step is [value missing]. The historical sequence is used as the input sample, and the future sequence after the window is used as the input sample. The true velocity values ​​at each time step are used as prediction labels to construct multiple training samples with overlapping relationships. By using the sliding time window construction method, the same short-term fluctuation event appears repeatedly in multiple training samples in different context positions, thereby enhancing the ability of the multi-input branch network model to capture short-term changes in traffic conditions.

[0132] In this embodiment, the online prediction process of the road speed prediction method that integrates multi-source spatiotemporal features includes the following steps:

[0133] 1) Acquire the latest traffic data and external feature data in real time, and update the multidimensional feature input sequence within the input time window;

[0134] 2) Calculate the set of adjacent roads corresponding to the target road at the current time and the time-delay alignment sequence of each adjacent road to obtain real-time road network association characteristics;

[0135] 3) Input the updated multidimensional feature input sequence into the trained multi-input branch network model for forward inference, and output the speed prediction value of the target road in the future multiple time steps to achieve rolling prediction.

[0136] 4) Among them, multi-step prediction supports the configuration of different prediction time ranges, which include at least one or more of 5 minutes, 15 minutes and 30 minutes, to adapt to the differentiated needs of prediction timeliness in different traffic management application scenarios.

[0137] The scope of protection for the road speed prediction method that integrates multi-source spatiotemporal features described in this application is not limited to the order of steps listed in this embodiment. Any solution implemented by adding, subtracting, or replacing steps in the prior art based on the principles of this application is included within the scope of protection of this application.

[0138] According to the method provided in the embodiments of this application, this application also provides a computer program product, which includes: computer program code, which, when run on a computer, causes the computer to execute the road speed prediction method that integrates multi-source spatiotemporal features provided in the embodiments of this application.

[0139] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the road speed prediction method that integrates multi-source spatiotemporal features provided in this application.

[0140] In this application embodiment, any combination of one or more storage media can be used. The storage medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, RAM, ROM, erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0141] This application also provides an electronic device. The road speed prediction method integrating multi-source spatiotemporal features described in this application is applied to the electronic device. In some embodiments, the electronic device may include, but is not limited to, any of the following types: server, industrial control computer, embedded terminal, edge computing device, cloud computing platform, or vehicle-mounted computing unit.

[0142] Figure 6 The diagram shown is a structural schematic of the electronic device 100 provided in an embodiment of this application. Figure 6 As shown, in some embodiments, the electronic device 100 includes a memory 101, a processor 102, and a computer program stored in the memory 101 and executable on the processor 102. When the processor 102 executes the computer program, it implements the steps of the road speed prediction method fusing multi-source spatiotemporal features described in the embodiments of this application.

[0143] Specifically, the processor 102 is connected to the memory 101 and is used to execute the computer program stored in the memory 101 so that the electronic device 100 executes the road speed prediction method that integrates multi-source spatiotemporal features provided in any embodiment of this application.

[0144] The processor 102 may be one or more of a central processing unit, a graphics processing unit, a tensor processor, a neural network processor, or a field-programmable gate array, and is used to execute instructions in the computer program to complete computational tasks such as model inference and online prediction.

[0145] Optionally, the processor 102 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0146] The memory 101 is used to store the computer program, network parameters of the multi-input branch network model, historical traffic data, and real-time acquired multi-source feature data. Specifically, the memory 101 may include a computer system readable medium in the form of volatile memory, such as random access memory (RAM) and / or cache memory. The electronic device 100 may further include other removable / non-removable, volatile / non-volatile computer system storage media. The memory 101 may include at least one program product having a set (e.g., at least one) of program modules configured to perform the functions of the embodiments of this application.

[0147] Optionally, the electronic device may also include a communication interface for interacting with external data sources, including but not limited to road traffic detectors, weather sensors, video detection equipment, or third-party traffic data platforms, to acquire road speed data, weather data, and road network topology information in real time.

[0148] Optionally, in this embodiment, the electronic device 100 may further include a display 103, which is communicatively connected to the processor 102 and the memory 101, and is used to display the relevant graphical user interface of the road speed prediction method that integrates multi-source spatiotemporal features.

[0149] The graphical user interface may include, but is not limited to, one or more of the following interface elements: prediction result display interface, road network topology visualization interface, real-time traffic status monitoring interface, and historical prediction accuracy analysis interface.

[0150] The prediction result display interface is used to intuitively present the predicted speed values ​​and confidence intervals of the target road at multiple time steps in the future in the form of line charts, bar charts, or dashboards, so that traffic management personnel can quickly grasp the road traffic trend.

[0151] The road network topology visualization interface is used to display the directed road network structure of the port area road network in the form of nodes and lines. It can dynamically present the spatial attention weight distribution between different road nodes through color rendering or changes in line thickness, making it possible to visualize, analyze and diagnose the spatial influence relationship of the road network.

[0152] The real-time traffic status monitoring interface is used to simultaneously display the current speed value, historical speed change curve, and adjacent sequence comparison information after time delay alignment of the target road and its adjacent roads, to assist operators in judging the rationality of the prediction results.

[0153] The historical prediction accuracy analysis interface is used to compare and display the deviation between the predicted speed value and the actual speed value within a historical period, so as to evaluate the prediction performance of the multi-input branch network model in different time periods and different scenarios.

[0154] By configuring the display 103, the prediction process and results of the road speed prediction method that integrates multi-source spatiotemporal features are made visually interactive, which improves the operability and user experience of the system in the actual application of port area traffic management.

[0155] In practical deployment, the electronic device can be deployed in the port area traffic management center, roadside edge computing nodes, or cloud servers. By connecting with the port area's existing traffic monitoring system, meteorological monitoring system, and vehicle detection system, it can receive real-time traffic flow data and external environmental data from the port area roads. It can also use the multi-input branch network model deployed on the electronic device to perform online rolling predictions and output the prediction results to traffic guidance screens, signal control systems, or the port area vehicle dispatching platform to provide decision support for port area traffic management and vehicle route planning.

[0156] In the above embodiments, the road speed prediction method integrating multi-source spatiotemporal features can be implemented entirely or partially through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented entirely or partially in the form of a computer program product. A computer program product includes one or more computer instructions (programs). When the computer program instructions (programs) are loaded and executed on a computer, they generate, in whole or in part, the flow or function according to the embodiments of this application. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., high-density digital video discs, DVDs), or semiconductor media (e.g., solid-state disks, SSDs, etc.).

[0157] The road speed prediction method integrating multi-source spatiotemporal features in this embodiment, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0158] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0159] In summary, the method in this embodiment unifies time-period features, weather-related external disturbances, historical road speed features, and road network spatial features into a multi-dimensional feature input sequence. This achieves end-to-end deep fusion of multi-source heterogeneous data at the input layer, enabling the model to simultaneously capture the periodic evolution of traffic flow, the impact of external environmental shocks, and the congestion propagation effect of the road network space. It eliminates structural differences between heterogeneous features from the data source, significantly improving speed prediction accuracy in complex scenarios. By introducing a traffic propagation time-delay alignment mechanism during adjacent road information aggregation and adaptively weighting aggregation based on prior connection strength, historical correlation, and propagation characteristics of roads, it effectively characterizes the physical delay and dynamic diffusion patterns of traffic flow propagation among port area roads, enhancing the method's adaptability to highly complex road network topologies such as port gates and intersections. A multi-branch parallel coding architecture with time, road, spatial, and weather branches is constructed. A temporal attention mechanism is introduced during the hidden state fusion stage of each branch, enabling dynamic adjustment of the contribution of features from different sources to the prediction results under different traffic scenarios. This allows the method to adaptively enhance its dependence on road network spatial diffusion features during peak congestion periods and autonomously emphasize the response to meteorological disturbance features under sudden severe weather conditions, thus significantly improving the robustness of the method under various complex traffic scenarios. Furthermore, a direct multi-step output strategy is adopted to output the speed prediction values ​​for multiple future time steps at once, avoiding the error accumulation problem caused by traditional recursive iterative prediction methods. This ensures the stability of prediction accuracy within different prediction time ranges such as 5 minutes, 15 minutes, and 30 minutes, giving the method in this embodiment good generalization ability and practical application value in highly dynamic and strongly disturbed scenarios such as port areas. Therefore, this application effectively overcomes the various shortcomings of the prior art and has high industrial application value.

[0160] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A road speed prediction method integrating multi-source spatiotemporal features, characterized in that, include: The road speed feature data, time feature data, weather feature data, and road network association feature data of the target road are obtained respectively, and a multi-dimensional feature input sequence is constructed based on the road speed feature data, the time feature data, the weather feature data, and the road network association feature data; The multidimensional feature input sequence is input into a pre-trained multi-input branch network model. Each branch of the multi-input branch network model processes each sequence in the multidimensional feature input sequence. The fusion layer then performs feature fusion and splicing on the sequences processed by each branch to obtain the fused features. The fused features are input into a fully connected layer, which outputs the predicted speed of the target road at multiple future time steps. Obtaining the road network association feature data includes: Obtain the spatiotemporal encoding vectors of adjacent roads and calculate the spatial attention weights of each adjacent road; The spatial attention weights are used to weight and aggregate the spatiotemporal encoding vectors of adjacent roads to obtain road network association features; The process of obtaining the spatiotemporal encoding vector of adjacent roads includes: A directed road network graph is constructed based on the road topology to determine the set of adjacent roads of the target road; For each adjacent road in the set of adjacent roads, calculate its constraint time delay relative to the target road; The historical speed sequences of adjacent roads are time-aligned using the aforementioned constraint time delay to obtain a time-delay-aligned adjacent sequence. The adjacency sequence is input into a long short-term memory network to obtain the spatiotemporal encoding vector of the adjacent roads; The calculation of the spatial attention weights for each adjacent road includes: The prior connection strength between roads in the directed road network graph, the historical correlation between adjacent roads and the target road, the traffic propagation time delay parameters determined by the constraint time delay, and their corresponding weight parameters are obtained respectively. The spatial attention weights of each adjacent road are calculated based on the prior connection strength, historical correlation, traffic propagation time delay parameters, and their respective weight parameters.

2. The road speed prediction method fusion of multi-source spatiotemporal features according to claim 1, characterized in that, The directed road network graph uses roads as nodes and the actual travel directions between roads as edges, and includes the prior connection strength between roads; wherein, the prior connection strength between roads is a functional relationship between the distance between roads, the capacity parameter, and the topological connection type.

3. The road speed prediction method fusion of multi-source spatiotemporal features according to claim 1, characterized in that, The constraint time delay is calculated as follows: The constraint is determined by maximizing the cross-correlation coefficient between the historical speed sequences of the target road and adjacent roads, or by calculating the ratio of the distance between roads to the average speed; and the constraint time delay does not exceed the length of the time window.

4. The road speed prediction method fusion of multi-source spatiotemporal features according to claim 1, characterized in that, Also includes: The historical speed time window sequence of the target road is encoded using a long short-term memory network to obtain its own spatiotemporal encoding vector; The difference between the weighted aggregation vector and its own spatiotemporal encoding vector is concatenated with the weighted aggregation vector to form a road network association feature with enhanced neighborhood influence and self-difference.

5. The road speed prediction method fusion of multi-source spatiotemporal features according to claim 1, characterized in that, The multi-input branch network model is trained under supervision using historical data, and the loss function of the multi-input branch network model is the mean squared error function; a sliding time window is introduced to construct a training sample sequence during the training process of the multi-input branch network model.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the road speed prediction method that integrates multi-source spatiotemporal features as described in any one of claims 1 to 5.

7. An electronic device, characterized in that, The electronic device includes: Processor and memory; The memory stores program instructions; The processor is configured to run the program instructions to execute the road speed prediction method that integrates multi-source spatiotemporal features as described in any one of claims 1 to 5.

Images