Urban traffic big data congestion prediction method based on deep learning

Abstract

The invention discloses an urban traffic big data congestion prediction method based on deep learning, and the method comprises the following steps: collecting multi-source traffic data in an urban traffic operation process, and constructing a standardized space-time traffic data set; respectively inputting the prediction results into an improved Autoformer model and an improved GKAN model, and respectively generating a first prediction result and a second prediction result; calculating a numerical difference between the first prediction result and the second prediction result, and generating an inconsistency score; executing weighted fusion operation to generate a fusion prediction result; calculating an error value between the fusion prediction result and the observation data, and generating candidate prediction results according to the error value and a preset error threshold value; and comparing the disturbance score with a preset disturbance threshold value to generate a final prediction result. The method realizes joint modeling of model prediction consistency and traffic disturbance state, and has the technical advantages of high prediction precision, strong adaptability and excellent error control capability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of traffic information processing, and particularly relates to a city traffic big data congestion prediction method based on deep learning. BACKGROUND

[0002] With the acceleration of urbanization and the continuous growth of traffic demand, the congestion problem in the urban traffic system is increasingly prominent, which has a significant impact on commuting efficiency, environmental quality and urban operating costs. In recent years, with the development of traffic perception devices and data collection technologies, city congestion prediction based on multi-source traffic data has become a research hotspot. Existing technologies mostly use time series models such as LSTM, Transformer or spatial models such as GCN, GAT to model and predict urban traffic, and some methods attempt to integrate time modeling and spatial topological information to improve overall prediction accuracy. However, in practical applications, the following problems still exist universally:

[0003] Different models have different modeling capabilities for the same traffic scene. Time series models focus on capturing local fluctuations and periodic trends, but they are difficult to accurately identify sudden structural changes. Graph neural network models can perceive traffic network topological relationships, but they are insufficient in responding to dynamic disturbances under non-stationary states, resulting in inconsistent outputs of prediction results in high complexity scenarios, lack of stability and interpretability. In addition, existing fusion methods mostly use simple weighting or concatenation mechanisms, ignoring the impact of output differences between multiple models on prediction accuracy, and cannot flexibly adjust the fusion strategy according to model prediction bias. On this basis, there is a lack of a mechanism to quantitatively depict prediction inconsistency, making the fusion prediction process lack of pertinence and adjustment ability, and thus affecting the reliability and robustness of the final congestion state evaluation.

[0004] Therefore, how to provide a city traffic big data congestion prediction method based on deep learning is a problem that those skilled in the art need to solve. SUMMARY

[0005] One object of the present application is to provide a city traffic big data congestion prediction method based on deep learning. The present application combines an improved Autoformer model and an improved GKAN model, fully utilizes multi-source traffic data in the process of urban traffic operation, dynamically adjusts the fusion weight and outputs a robust prediction result by constructing a spatio-temporal traffic data set, performing time series modeling and spatial topological modeling, fusing multi-model prediction results and introducing an inconsistency scoring mechanism. The present application describes in detail the whole process from multi-source data collection, inconsistency scoring construction, disturbance perception feature extraction to final prediction result generation, and has the advantages of high prediction accuracy, sensitive response to sudden disturbances, flexible fusion mechanism and strong interpretability.

[0006] According to an embodiment of the present application, a deep learning-based urban traffic big data congestion prediction method comprises the following steps:

[0007] Step one: collect multi-source traffic data in the process of urban traffic operation, and construct a standardized spatio-temporal traffic data set;

[0008] Step two: input the spatio-temporal traffic data set into the improved Autoformer model and the improved GKAN model respectively, perform time series modeling process and spatial topology modeling process, and generate first prediction result and second prediction result respectively;

[0009] Step three: calculate the numerical difference between the first prediction result and the second prediction result, and extract the traffic state information corresponding to the target time step in the spatio-temporal traffic data set to generate inconsistency score;

[0010] Step four: calculate the fusion weight of the first prediction result and the second prediction result based on the inconsistency score, perform weighted fusion operation, and generate fusion prediction result;

[0011] Step five: determine whether the error value between the fusion prediction result and the observation data is greater than the preset error threshold; if the error value is greater than the preset error threshold, select the prediction result with the minimum error value between the first prediction result and the second prediction result and the observation data as the candidate prediction result;

[0012] Step six: extract the disturbance feature of the traffic state information to generate disturbance score; if the disturbance score is greater than the preset disturbance threshold, the first prediction result is taken as the final prediction result; if the disturbance score is less than or equal to the preset disturbance threshold, the candidate prediction result is taken as the final prediction result.

[0013] Optionally, the step one is specifically:

[0014] The multi-source traffic data includes traffic flow data, road average speed data, traffic event data, weather condition data and traffic signal control data;

[0015] The multi-source traffic data is synchronously processed according to time index and space position index to generate a unified data structure covering continuous time period and space grid;

[0016] The traffic flow data is subjected to unit conversion operation and numerical smoothing processing; the road average speed data is subjected to missing value filling and outlier correction operation; the traffic event data is subjected to event classification marking and time encoding conversion operation; the weather condition data is subjected to continuous variable standardization and discrete variable mapping operation; and the traffic signal control data is subjected to period normalization and phase offset encoding processing;

[0017] The processed whole traffic data is constructed into a standardized space-time traffic data set according to the time dimension and the space dimension.

[0018] Optionally, the step two is specifically:

[0019] The standardized space-time traffic data set is divided into a plurality of time window sequences according to the time dimension, and is input into the improved Autoformer model; the improved Autoformer model is provided with a disturbance perception gating module before a trend-season decomposition structure, and the activation strength of a trend path is controlled based on the traffic state change amplitude; a dynamic convolution module is arranged in the decoder structure to extract multi-scale time change features; a traffic state prediction value at a continuous time step is output to form a first prediction result; the trend-season decomposition structure refers to performing trend component extraction and season component separation operations on the input time sequence to form a trend feature sequence and a season feature sequence which can be independently modeled;

[0020] The standardized space-time traffic data set is constructed into a graph structure input containing a road topology structure according to the space dimension, and is input into the improved GKAN model; the improved GKAN model includes a traffic state driven edge weight scoring adjustment module, an edge weight adjustment coefficient is calculated based on the adjacent road segment speed gradient, the vehicle density difference and the signal timing change; a graph-level residual enhancement path is arranged in the graph convolution calculation structure to maintain the prediction stability under the dynamic change of the graph structure; a traffic state prediction value of each node at a target time step is output to form a second prediction result.

[0021] Optionally, the improved Autoformer model is specifically:

[0022] The standardized space-time traffic data set is divided into a plurality of time window sequences with consistent lengths according to the time dimension to form an input sequence matrix;

[0023] A linear projection operation is performed on the input sequence matrix to generate an initial time step embedding representation;

[0024] A disturbance perception gating module is introduced on the basis of the initial time step embedding representation to build a gating structure containing an input channel disturbance score and a trend channel activation factor; the disturbance score is generated by normalizing the traffic state change amplitude and is input as a control gate function; the activation factor is output by the gate function to control the activation strength of the features in the trend path in each dimension;

[0025] The sequence adjusted by the disturbance perception gating module is input into a trend-season decomposition structure to perform a local mean calculation operation and a residual separation operation in a sliding window to obtain a trend component sequence and a season component sequence, respectively;

[0026] Residual attention modeling operations are respectively performed on the trend component sequence and the seasonal component sequence to capture local trend features and periodic change features within the same time window;

[0027] In the decoding stage, a dynamic convolution module is input with the trend component sequence, a group of convolution kernels based on different time window lengths is constructed, multi-scale convolution operations are performed to extract short-term, medium-term and long-term trend features;

[0028] Feature splicing operations are performed on the output features of the dynamic convolution module and the attention output features of the seasonal component to generate fusion feature representations;

[0029] On the basis of the fusion feature representations, multi-head self-attention layers, feedforward network layers and layer normalization structures are sequentially executed to complete feature transformation and normalization processing of each layer of stacked structures;

[0030] After all the decoding layer stacks are completed, the traffic state prediction values corresponding to future time steps are output to form the first prediction results.

[0031] Optionally, the improved GKAN model specifically comprises:

[0032] The spatial dimension information in the standardized spatiotemporal traffic data set is collected, the road network is constructed as a graph structure, the nodes of the graph structure represent road segments, the edges represent the connection relationship between adjacent road segments, and the graph structure is attached with a node feature matrix and an edge feature matrix;

[0033] The speed gradient between nodes, the density difference and the signal timing difference are calculated according to the vehicle speed sequence, the vehicle density sequence and the signal timing sequence contained in the node feature matrix;

[0034] The speed gradient, the density difference and the signal timing difference are input into an edge weight score adjustment module to calculate an edge weight disturbance index;

[0035] Linear mapping and normalization processing are performed on the edge weight disturbance index to generate an edge weight adjustment coefficient;

[0036] Numerical correction is performed on the initial edge weight matrix based on the edge weight adjustment coefficient to construct an adaptive edge weight graph structure;

[0037] A graph neural network backbone structure is set in the adaptive edge weight graph structure, and the graph neural network backbone structure comprises a plurality of graph convolution units;

[0038] Each graph convolution unit includes a self-attention mechanism and an edge weight modulation mechanism, and feature update operations are respectively performed on the message passing intensity between nodes and the edge connection importance;

[0039] A graph-level residual enhancement path is arranged between the graph convolution units, the graph convolution outputs of different layers are connected, the splicing and normalization processing of the shallow structure information and the deep abstract features are performed, and a cross-layer fusion feature is obtained;

[0040] Node features corresponding to the target time step are extracted from the cross-layer fusion features, the node features include the speed state, the traffic density state, the signal control state and the high-dimensional embedding representation of the road segment in the graph structure at the target time step, and constitute a state feature vector set;

[0041] A prediction process is performed based on the state feature vector set, and a traffic state prediction value corresponding to each node is output to form a second prediction result.

[0042] Optionally, the step three is specifically:

[0043] Difference calculation is performed on the traffic state values corresponding to the same road segment and the same time step in the first prediction result and the second prediction result to generate a node-level prediction deviation sequence;

[0044] Traffic state information corresponding to the target time step in the standardized spatio-temporal traffic data set is collected, and the traffic state information includes the actual speed value, the actual traffic density value and the signal timing state of each node;

[0045] A state correlation factor is constructed based on the prediction deviation sequence, the standard deviation, the coefficient of variation and the average change rate of the traffic state of each node in the target time step;

[0046] The prediction deviation sequence and the state correlation factor are combined to construct a three-dimensional input feature set, including the prediction deviation, the traffic state variation coefficient and the state change rate;

[0047] The three-dimensional input feature set is input into a weighted regression model, the model fitting is performed based on the historical traffic state data set, and an inconsistency scoring function is constructed;

[0048] The prediction inconsistency value of each node at the target time step is calculated through the inconsistency scoring function;

[0049] The inconsistency values of all nodes are normalized to output a normalized inconsistency score.

[0050] Optionally, the step four is specifically:

[0051] The normalized inconsistency value of each graph structure node at the target time step in the inconsistency score is extracted to form a node-level inconsistency score sequence;

[0052] The fusion weight of the first prediction result and the second prediction result is calculated for each node respectively, the fusion weight is based on the inconsistency score to perform an inverse mapping operation, the higher the score value, the lower the corresponding prediction result fusion weight, and the lower the score value, the higher the corresponding fusion weight;

[0053] The first prediction result weight matrix and the second prediction result weight matrix are respectively constructed, corresponding to the fusion weight values of each node;

[0054] Based on the node-level weighted fusion rule, the operation of each node is specifically: the prediction values of the corresponding nodes in the first prediction result and the second prediction result are multiplied by the corresponding fusion weight values respectively, and the results are weighted and summed to generate the initial fusion prediction value of the node level; the initial fusion prediction value of all nodes is normalized and calibrated to construct a fusion prediction result vector as the fusion prediction result.

[0055] Optionally, the step five is specifically:

[0056] The observation data of each node in the target time step is collected to construct an observation state vector, the observation state vector includes the actual speed value, the actual traffic density value and the signal timing state of each node;

[0057] The difference calculation operation is performed between the fusion prediction result vector and the observation state vector to generate a node-level prediction error sequence;

[0058] The absolute value average operation is performed on the node-level prediction error sequence to calculate the error value;

[0059] It is judged whether the error value is greater than a preset error threshold; if the error value is greater than the preset error threshold, the average error value between the first prediction result and the observation state vector is calculated, and the average error value between the second prediction result and the observation state vector is calculated, and the prediction result with smaller error value between the first prediction result and the second prediction result and the observation state vector is selected as the candidate prediction result;

[0060] If the error value is less than or equal to the preset error threshold, the fusion prediction result is taken as the candidate prediction result.

[0061] Optionally, the step six is specifically:

[0062] The traffic state information corresponding to the target time step in the standardized spatio-temporal traffic data set is collected, the traffic state information includes the speed value, the traffic density value and the signal timing state of each node;

[0063] The time series sample is constructed based on the traffic state information in the continuous multiple time steps before and after the target time step, and the disturbance characteristic parameters of each node are calculated, the disturbance characteristic parameters include the speed disturbance amplitude, the density disturbance amplitude and the signal change frequency;

[0064] characteristic combination of the speed disturbance amplitude, the density disturbance amplitude and the signal change frequency, to construct a node-level disturbance feature vector set;

[0065] inputting the disturbance feature vector set into a disturbance scoring function, performing disturbance intensity estimation, and outputting a node-level disturbance score result;

[0066] performing normalization processing on the disturbance score results of all nodes to generate a disturbance score;

[0067] determining whether the disturbance score is greater than a preset disturbance threshold value; if the disturbance score is greater than the preset disturbance threshold value, the first prediction result is taken as the final prediction result; if the disturbance score is less than or equal to the preset disturbance threshold value, the candidate prediction result is taken as the final prediction result.

[0068] Optionally, the method further comprises: recording the first prediction result, the second prediction result, the fusion prediction result, the inconsistency score, the disturbance score and the final prediction result, constructing a training sample set, and performing a joint training process on the improved Autoformer model and the improved GKAN model to update the model parameters.

[0069] The beneficial effects of the present application are:

[0070] The present application solves the problems of split modeling capability, difficulty in inconsistent prediction result fusion and insensitive response to sudden disturbance in time modeling and space modeling of multi-source traffic data in the process of urban traffic operation by the synergistic fusion of the improved Autoformer model and the improved GKAN model, adopts standardized construction of spatiotemporal traffic data set and multi-model parallel prediction mechanism, combines the inconsistency scoring system based on numerical difference and the disturbance feature extraction method, introduces the dynamic weight adjustment strategy based on inconsistency scoring in the prediction fusion stage to realize adaptive weighted fusion of the first prediction result and the second prediction result; in the robustness adjustment stage, through the joint determination mechanism of the error value between the fusion prediction result and the observation data and the disturbance score, the candidate prediction result is dynamically selected and disturbance perception correction is performed to enhance the joint response capability of the model to periodic and sudden traffic fluctuations; finally, the prediction result with stability and sensitivity is output. The deep learning-based urban traffic big data congestion prediction method effectively improves the description capability of complex traffic state, has the beneficial effects of high prediction accuracy, sensitivity to inconsistency and disturbance change, and strong adaptability of fusion mechanism. BRIEF DESCRIPTION OF DRAWINGS

[0071] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, illustrate the present application together with the embodiments thereof, and explain the present application, and do not constitute a limitation of the present application. In the drawings:

[0072] Figure 1 A flow chart of a congestion prediction method based on deep learning for urban traffic big data is proposed in the present application;

[0073] Figure 2 A schematic diagram of dynamic selection of prediction results based on scoring driving is proposed in the present application.

[0074] Figure 3 A general structure diagram of a congestion prediction method based on deep learning for urban traffic big data is proposed in the present application. DETAILED DESCRIPTION

[0075] The present application will now be further described in detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams and only illustrate the basic structure of the present application in a schematic manner, and therefore only show the components related to the present application.

[0076] REFERENCE Figures 1-3 A congestion prediction method based on deep learning for urban traffic big data, comprising the following steps:

[0077] Step one: collect multi-source traffic data in the process of urban traffic operation, and construct a standardized spatio-temporal traffic data set;

[0078] Step two: input the spatio-temporal traffic data set into the improved Autoformer model and the improved GKAN model respectively, perform time series modeling process and spatial topology modeling process, and generate first prediction result and second prediction result respectively;

[0079] Step three: calculate the numerical difference between the first prediction result and the second prediction result, and extract the traffic state information corresponding to the target time step in the spatio-temporal traffic data set, and generate inconsistency score;

[0080] Step four: calculate the fusion weight of the first prediction result and the second prediction result based on the inconsistency score, perform weighted fusion operation, and generate fusion prediction result;

[0081] Step five: determine whether the error value between the fusion prediction result and the observation data is greater than the preset error threshold; if the error value is greater than the preset error threshold, select the prediction result with the smallest error value between the first prediction result and the second prediction result and the observation data as the candidate prediction result;

[0082] Step six: extract the disturbance feature of the traffic state information, generate disturbance score; if the disturbance score is greater than the preset disturbance threshold, the first prediction result is taken as the final prediction result; if the disturbance score is less than or equal to the preset disturbance threshold, the candidate prediction result is taken as the final prediction result.

[0083] The embodiment improves the consistency and processing efficiency of multi-source traffic data in the spatial and temporal dimensions by constructing a standardized spatiotemporal traffic dataset; by respectively inputting the spatiotemporal traffic dataset into an improved Autoformer model and an improved GKAN model, time series modeling and spatial topology modeling are performed, and the complementary advantages of the two types of models in periodic traffic trend modeling and spatial correlation capturing are fully utilized; by calculating the numerical difference between the two prediction results and combining the traffic state information of the target time step to generate an inconsistency score, the subsequent fusion weight can reflect the reliability difference between the model predictions; the fusion link adopts a dynamic weighting mechanism based on the inconsistency score, which increases the proportion of more reliable results when the prediction results are significantly different, and realizes adaptive regulation and control of model prediction fusion; based on the error judgment mechanism between the fusion result and the observation data, the optimal prediction result is dynamically selected as the candidate result, which ensures that the fusion process will not introduce error amplification; by further extracting traffic state disturbance features and constructing a disturbance scoring system, the response sensitivity to sudden traffic events is enhanced; finally, the first prediction result or the candidate result is adaptively selected as the final output according to the disturbance score, which improves the conservatism of the system in stable scenarios and the rapid response capability in disturbance scenarios, and has the technical advantages of high prediction accuracy, fast fluctuation response and strong model fusion intelligence.

[0084] In the embodiment, the step one is specifically:

[0085] The multi-source traffic data includes traffic flow data, road average speed data, traffic event data, weather condition data and traffic signal control data;

[0086] The multi-source traffic data is synchronously processed according to the time index and the spatial position index, and a unified data structure covering continuous time periods and spatial grids is generated;

[0087] The traffic flow data is subjected to unit conversion operation and numerical smoothing processing; the road average speed data is subjected to missing value filling and outlier correction operation; the traffic event data is subjected to event classification marking and time encoding conversion operation; the weather condition data is subjected to continuous variable standardization and discrete variable mapping operation; the traffic signal control data is subjected to period normalization and phase offset encoding processing;

[0088] The processed traffic data is constructed into a standardized spatiotemporal traffic dataset according to the time dimension and the spatial dimension.

[0089] In the embodiment, the step two is specifically:

[0090] The standardized spatiotemporal traffic data set is divided into a plurality of time window sequences in the time dimension, and is input into the improved Autoformer model; the improved Autoformer model sets a disturbance perception gating module in front of the original trend-season decomposition structure, constructs a disturbance vector based on the traffic state change amplitude corresponding to each time window, and obtains a gating weight function through disturbance vector fitting, which is used to adjust the activation strength of the trend path, so as to enhance the trend modeling capability of abnormal fluctuations or sudden changes; the trend-season decomposition structure performs trend component extraction and seasonal component separation operations on the input sequence of each time window, forming a trend feature sequence and a seasonal feature sequence that can be independently modeled; the trend feature sequence and the seasonal feature sequence are respectively input into a decoder structure, and a dynamic convolution module is introduced in the decoder, wherein the dynamic convolution module generates a dynamic convolution kernel weight function based on multi-scale feature responses within the time window, and the weight function is obtained through traffic state difference distribution fitting, which is used to extract multi-scale time variation features; the feature sequence output by the decoder is mapped into the traffic state prediction value of the corresponding time step to form a first prediction result.

[0091] The standardized spatiotemporal traffic data set is divided into a plurality of time window sequences in the time dimension, and is input into the improved Autoformer model; the improved Autoformer model sets a disturbance perception gating module in front of the original trend-season decomposition structure, constructs a disturbance vector based on the traffic state change amplitude corresponding to each time window, and obtains a gating weight function through disturbance vector fitting, which is used to adjust the activation strength of the trend path, so as to enhance the trend modeling capability of abnormal fluctuations or sudden changes; the trend-season decomposition structure performs trend component extraction and seasonal component separation operations on the input sequence of each time window, forming a trend feature sequence and a seasonal feature sequence that can be independently modeled; the trend feature sequence and the seasonal feature sequence are respectively input into a decoder structure, and a dynamic convolution module is introduced in the decoder, wherein the dynamic convolution module generates a dynamic convolution kernel weight function based on multi-scale feature responses within the time window, and the weight function is obtained through traffic state difference distribution fitting, which is used to extract multi-scale time variation features; the feature sequence output by the decoder is mapped into the traffic state prediction value of the corresponding time step to form a first prediction result.

[0092] The first prediction result and the second prediction result respectively model the time regularity and the spatial topological dependence in the traffic evolution process in the embodiment, thereby maintaining the time series prediction accuracy while improving the response capability to local traffic structure changes.

[0093] In the embodiment, the improved Autoformer model is specifically:

[0094] The standardized spatiotemporal traffic data set is divided into a plurality of time window sequences with consistent length in the time dimension, to form a two-dimensional input sequence matrix, each row representing a time sequence of a traffic variable, and each column representing a plurality of variable values at the same time step; a linear projection operation is performed on the input sequence matrix, an affine transformation structure composed of a weight matrix and a bias term is used to map the original features to a fixed-dimensional embedding space, to generate an initial time step embedding representation, facilitating subsequent modeling of sequence dependency between time steps;

[0095] A perturbation-aware gating module is set based on the initial time step embedding representation, the module including a perturbation score generation structure and a trend path activation adjustment structure; the perturbation score generation structure calculates a perturbation amount based on the fluctuation amplitude of each variable within a sliding window, and after normalization processing, the perturbation score is used as a gating control signal; the specific perturbation score is obtained by normalizing the change rate of the same variable between adjacent time steps; the trend path activation adjustment structure receives the perturbation score and generates an activation factor of the trend channel by constructing a gating function, the activation factor is used as a weight coefficient acting on the feature channel in the trend path, to control the activation degree of the high-perturbation variable to the trend modeling path, thereby enhancing the trend feature modeling capability when the traffic state changes dramatically;

[0096] The adjusted sequence is input into a trend-seasonal decomposition structure, which extracts a trend component by a local mean calculation operation within a sliding window, and separates a seasonal component by a residual extraction method; the local mean calculation operation obtains a trend dominant part by performing a sliding average operation within a preset window length, and the residual separation operation obtains a periodic change part by subtracting the trend component from the original input, to complete the separation of the trend component sequence and the seasonal component sequence; this structure realizes decoupled modeling of long-term stable trends and periodic short-term fluctuations, effectively improving the perception ability of the model to different temporal features;

[0097] Residual attention modeling operations are performed on the trend component sequence and the seasonal component sequence respectively; the residual attention modeling operation includes a multi-head residual self-attention layer and a position encoding structure, which are used to model the dependency between the trend and the periodic component within the same time window, and to enhance the recognition ability of local time change patterns; the attention weight matrix is calculated by linear transformation of the input sequence, and is fitted by matrix multiplication and scaling inner product calculation method;

[0098] In the decoding stage, the trend component sequence is input into a dynamic convolution module; the dynamic convolution module constructs a plurality of convolution kernel groups with different lengths, respectively corresponding to different time scale trend modeling requirements; each group of convolution kernels is generated by learnable parameters, and performs multi-scale one-dimensional convolution operation to extract short-term, medium-term and long-term trend features; the convolution output is processed by an activation function and normalization to form a trend multi-scale feature vector;

[0099] The multi-scale trend feature vector output by the dynamic convolution module and the residual attention output feature vector of the seasonal component are spliced along the feature dimension to generate a fusion feature representation, which enhances the joint modeling capability of trend and periodic changes;

[0100] On the basis of the fusion feature representation, a stack module composed of a multi-head self-attention layer, a feedforward network layer and a layer normalization structure is sequentially executed; the multi-head self-attention layer is used to model the global dependency relationship between different time steps; the feedforward network layer adopts a two-layer fully connected structure with residual connection, which is used for feature nonlinear transformation and dimension remapping; the layer normalization structure performs standardization processing on the output of each layer to ensure the training stability and feature scale consistency;

[0101] After all the decoding layer stack structures are completed, the traffic state prediction values corresponding to the future multiple time steps are output to form the first prediction result.

[0102] The embodiment introduces a disturbance perception gating module and a dynamic convolution module to enhance the modeling capability of traffic state fluctuation scenarios on the basis of the original trend-seasonal decoupling modeling, realize high-precision, multi-scale time series prediction output, and have significant robustness and generalization capability improvement effect.

[0103] In the embodiment, the improved GKAN model is specifically:

[0104] The spatial dimension information in the standardized spatiotemporal traffic data set is collected, and the road network is constructed as a graph structure, the nodes of the graph structure representing road segments, and the edges representing the connection relationship between adjacent road segments, the graph structure being attached with a node feature matrix and an edge feature matrix;

[0105] The speed gradient, density difference and signal timing difference between nodes are calculated according to the vehicle speed sequence, vehicle density sequence and signal timing sequence contained in the node feature matrix, the speed gradient indicating the speed change amplitude of adjacent road segments at the same time step, the density difference being the difference in the number of vehicles per unit area in adjacent segments, and the signal timing difference indicating the difference degree of the red-green light period or phase control;

[0106] The speed gradient, density difference and signal timing difference are input into an edge weight score adjustment module to perform normalization processing and feature combination operation to generate an edge weight disturbance index, the disturbance index reflecting the sensitivity and dynamic change trend of local connection to traffic state propagation;

[0107] The edge weight disturbance index is subjected to linear mapping operation and combined with a standardized weight parameter to generate an edge weight adjustment coefficient, the edge weight adjustment coefficient being obtained by fitting a disturbance mapping function, and used for dynamically adjusting the connection strength between adjacent edges;

[0108] Perform numerical correction on the original adjacency matrix based on the edge weight adjustment coefficient, construct an adaptive edge weight graph structure after dynamic updating of the edge weight, and improve the modeling robustness of the graph structure under dynamic changing traffic conditions;

[0109] A graph neural network backbone structure is set in the adaptive edge weight graph structure, and the backbone structure includes a plurality of graph convolution units, each of which performs information aggregation and propagation based on joint input of node features and edge features;

[0110] A self-attention mechanism is set in each graph convolution unit to calculate the dependency weight between the target node and the neighbor node to form a node message passing strength matrix, and an edge weight modulation mechanism is set to perform a weighted adjustment operation on the message passing strength matrix according to the edge weight adjustment coefficient to generate a modulated aggregated feature;

[0111] A graph-level residual enhancement path is set between each graph convolution unit to perform feature splicing operation by connecting the graph convolution output of the previous layer and the current layer, and to maintain the information channel continuity and multi-scale structure stability between different layers by combining with the normalization layer processing;

[0112] The cross-layer fusion features output by the graph-level residual enhancement path are input into a feature decoding module to extract node features corresponding to the target time step, which include the speed state, traffic density state, signal control state of the road segment at the target time step and its high-dimensional embedding representation in the graph structure, to form a state feature vector set;

[0113] A traffic state prediction process is performed based on the state feature vector set, the node features of the target time step are input into a fully connected regression module, and the traffic state prediction value of each node is output to form a second prediction result.

[0114] The embodiment enhances the modeling capability of the dynamic changes of the traffic graph through the edge weight adjustment method, improves the stability and expression capability of the node features through the graph-level residual enhancement path, and thus improves the generalization performance and prediction accuracy in the spatial modeling process.

[0115] In the embodiment, the step three is specifically:

[0116] The first prediction result output by the improved Autoformer model and the second prediction result output by the improved GKAN model are aligned in time and space dimensions, the prediction values of the same road segment and the same time step are correspondingly matched, an element-by-element difference calculation operation is performed, and a node-level prediction deviation sequence including all road segments at all target time steps is generated;

[0117] Collect real traffic state information corresponding to the target time step in the standardized spatio-temporal traffic data set, which includes actual vehicle speed values, actual traffic density values and current signal timing states of each road segment, and perform node-level normalization processing to obtain a standardized target state sequence;

[0118] Based on the node-level prediction bias sequence and the standardized target state sequence, calculate the traffic state standard deviation, coefficient of variation and average change rate of each node at the target time step, the standard deviation reflects the dispersion degree of the traffic state, the coefficient of variation measures the relative fluctuation intensity of the traffic state based on the standard deviation, and the average change rate represents the change speed of the state per unit time;

[0119] Concatenate the node-level prediction bias value, traffic state coefficient of variation and state change rate according to the node and time dimensions to construct a three-dimensional input feature set, each item of the three-dimensional input feature set represents the prediction difference and state instability of a node at the target time step;

[0120] Input the three-dimensional input feature set into the weighted regression model, perform parameter fitting operation on the historical traffic state data set, and construct an inconsistency score function through a weighted feature combination function, the inconsistency score function is obtained through weighted linear fitting and is used to evaluate the cooperative relationship between prediction bias and traffic state fluctuation;

[0121] Score each node's three-dimensional feature item at the target time step using the inconsistency score function, and calculate the inconsistency value of each node, the higher the inconsistency value, the greater the difference between the two model prediction results and the more significant the state fluctuation;

[0122] Normalize all node inconsistency values by the maximum value to output an inconsistency score vector with a range of [0, 1], which forms a reference index for weight distribution in the model fusion stage.

[0123] In this embodiment, step four is specifically:

[0124] Extract the normalized score value of each graph structure node at the target time step from the inconsistency score vector output by the inconsistency score function, and construct a node-level inconsistency score sequence, each item in the sequence represents the cooperative intensity of prediction difference and state disturbance of the node at the current time step;

[0125] Based on the node-level inconsistency score sequence, a reverse weight mapping operation is performed to construct a fusion weight calculation function, which is obtained by fitting the reverse corresponding relationship between the node score value and the prediction reliability. The higher the score value, the stronger the prediction inconsistency of the current node, the lower the confidence in fusion, and the smaller the corresponding fusion weight value. The lower the score value, the better the consistency of the model prediction result, the higher the confidence in fusion, and the larger the corresponding fusion weight value;

[0126] According to the fusion weight calculation function, the fusion weight values corresponding to the first prediction result and the second prediction result are calculated for each node respectively, and two node-level fusion weight vectors are constructed to represent the confidence weight proportion of the first model output and the second model output in the weighted fusion process respectively.

[0127] The two fusion weight vectors are constructed into a first prediction result weight matrix and a second prediction result weight matrix according to the node order. The matrix structure is consistent with the prediction result vector, and is used to perform corresponding weight adjustment operations on the model output values of each node.

[0128] Based on the node-level weighted fusion rule, the prediction values of each node are multiplied by the corresponding weight values respectively. The first prediction result is multiplied by the first prediction result weight matrix, and the second prediction result is multiplied by the second prediction result weight matrix. The two weighted results are summed at the node level to generate an initial fusion prediction value sequence.

[0129] The initial fusion prediction value sequence is subjected to a normalization consistency calibration process, and the calibration process is obtained by fitting the historical fusion error and prediction error fluctuation relationship in the training set. The fusion prediction result is output.

[0130] In the present embodiment, the step five is specifically:

[0131] The actual observation data of all nodes in the graph structure at the target time step is collected to construct an observation state vector. The observation state vector contains the traffic running state information of each node, including the actual speed value, the actual traffic density value and the signal timing state. The signal timing state is represented as a combination feature of the discretized timing number and the duration after standardization processing.

[0132] Difference calculation operation is performed on the fusion prediction result vector and the observation state vector to extract the deviation between the prediction value and the actual value of each node at the corresponding time step, and a node-level prediction error sequence is generated.

[0133] An absolute value extraction operation is performed on the node-level prediction error sequence, and an average operation is performed on the absolute error values of all nodes to calculate a global prediction error value, which is fitted by a weighted average of node importance weights to reduce the influence of edge node errors on overall evaluation;

[0134] It is judged whether the global prediction error value is greater than a preset error threshold value, which is a steady-state error limit calculated by a threshold fitting function based on a historical error distribution curve, for dynamically adjusting the error tolerance range under different traffic scenarios;

[0135] If the error value is greater than the preset error threshold value, it indicates that there is a significant deviation between the current fusion prediction result and the actual traffic state, and difference calculation and average error value extraction operations are performed on the first prediction result and the observed state vector and the second prediction result and the observed state vector to construct first and second error values;

[0136] The first and second error values are compared, and the prediction result with a smaller error value with the observed state vector is selected as the candidate prediction result, which retains the path with the smallest error in the current time step prediction structure;

[0137] If the error value is less than or equal to the preset error threshold value, it indicates that the fusion prediction result has met the accuracy requirement, and the fusion prediction result is directly taken as the candidate prediction result;

[0138] The embodiment effectively improves the reliability and robustness of the prediction result through accuracy evaluation and error correction between the fusion prediction result and the real observed state, and has the technical effects of suppressing abnormal fusion deviation and enhancing model generalization ability.

[0139] In the embodiment, step six is specifically:

[0140] The traffic state information corresponding to the target time step in the standardized spatio-temporal traffic data set is collected, including the speed value, traffic density value and signal timing state of each node in the graph structure, the speed value and density value are smoothed by sliding window average processing, and the signal timing state is converted into a standardized timing vector by discrete event encoding;

[0141] Based on a plurality of time steps before and after the target time step, node traffic state information in the continuous time sequence is extracted to construct time sequence samples, each sample containing a fixed-length node-level state change sequence, including a speed sequence, a density sequence and a timing state sequence;

[0142] performing a perturbation feature extraction operation on the time series samples, calculating a perturbation feature parameter of each node within the time series, the perturbation feature parameter including a speed perturbation amplitude, a density perturbation amplitude, and a signal change frequency, wherein the perturbation amplitude is fitted by a maximum-minimum difference and a standard deviation, and the signal change frequency is fitted by the number of timing changes in a unit time step;

[0143] standardizing the perturbation feature parameters and performing a feature splicing operation to construct a node-level perturbation feature vector set, each node corresponding to a group of perturbation feature vectors, for reflecting the traffic state fluctuation degree of the node within the time series;

[0144] inputting the perturbation feature vector set into a perturbation scoring function, the perturbation scoring function constructing a perturbation intensity estimation model based on a weighted combination of multi-dimensional perturbation features, outputting a node-level perturbation score result, and the scoring function weight parameters being fitted by historical high-error nodes;

[0145] performing maximum-minimum normalization processing on the perturbation score results of all nodes to generate a full-map perturbation score in a unified scale, the perturbation score being used to evaluate the overall fluctuation intensity of the current traffic state;

[0146] judging whether the perturbation score is greater than a preset perturbation threshold, the perturbation threshold being calculated by a percentile fitting function based on a historical perturbation score distribution;

[0147] if the perturbation score is greater than the preset perturbation threshold, indicating that the current traffic state fluctuates greatly and the fusion strategy is unstable, the first prediction result is taken as the final prediction result to avoid introducing prediction error amplification in the fusion process;

[0148] if the perturbation score is less than or equal to the preset perturbation threshold, indicating that the traffic state is in a stable interval, the candidate prediction result is taken as the final prediction result to improve the dynamic adaptability of the result.

[0149] In the embodiment, further comprising:

[0150] recording the first prediction result, the second prediction result, the fusion prediction result, the inconsistency score, the perturbation score, and the final prediction result, and associating and marking the standardized traffic state information of the corresponding time step, to construct a training sample set enhanced by a supervision signal, the training sample set including four types of information: prediction input features, model output results, fusion judgment basis, and final prediction labels;

[0151] Perform screening and enhancement processing on the training sample set, and take inconsistent samples with inconsistency scores greater than a set threshold and high disturbance samples with disturbance scores greater than a set threshold as a key sample set, the threshold is determined by a fitting function through a historical error distribution curve, and an oversampling processing is performed on the key sample set to improve the proportion of key scenes in training;

[0152] A joint training process is constructed, the improved Autoformer model and the improved GKAN model are respectively used as a time series modeling subnetwork and a spatial topology modeling subnetwork, and the input features are shared and the respective prediction results are output in the joint training process, and the fusion prediction result is output through a fusion module;

[0153] The error values of the first prediction result, the second prediction result and the fusion prediction result with respect to the final prediction result are calculated, and the error values are respectively input into a time subnetwork loss function, a spatial subnetwork loss function and a fusion judgment loss function, the three types of loss functions are combined by a joint loss function, and the weight coefficients of the joint loss function are obtained by fitting the performance of the verification set;

[0154] The back propagation process of the joint loss function is performed, the network parameters in the improved Autoformer model and the improved GKAN model are updated respectively, and the prediction error change trend of each module in each iteration is recorded, which is used to judge the training convergence and the module performance stability;

[0155] After the joint training is completed, the updated time series modeling submodel and the spatial topology modeling submodel are saved for subsequent actual prediction tasks.

[0156] The present embodiment can realize continuous optimization of model parameters based on dynamic feedback signals through training sample construction and joint training process, significantly enhance the adaptability of the model to inconsistent samples and high disturbance scenes, and has the technical effects of strong training sample coverage, high model generalization ability and stability improvement of fusion strategy.

[0157] Example 1:

[0158] In order to verify the feasibility of the present application in implementation, the present application is applied to the intelligent traffic management center of a certain A city, which faces serious congestion during peak hours, unstable prediction and delayed response all the year round, especially in the sudden congestion scenes before and after morning and evening peak hours and holidays, the traditional traffic prediction system lacks sensitivity to sudden events and spatio-temporal inconsistent data, resulting in delayed early warning and invalid signal optimization strategy, which seriously affects the efficiency of urban traffic scheduling and the travel experience of the public.

[0159] In this embodiment, the traffic management center of City A selects three main urban areas and five main roads in the core area of the city, covering 47 main intersections and 117 traffic monitoring nodes, as a pilot deployment area. The deployment period is three months. The data collection platform is based on the existing signal control system, traffic detection radar, video detector and GPS bus positioning system to form a structured spatio-temporal traffic data set, sampled every 5 minutes, with a cumulative effective time step of more than 260,000, containing speed, density, and timing traffic state information, with a total data volume of 4.7TB.

[0160] In this scenario, the method of the application first standardizes all historical and real-time data to build a standardized spatio-temporal traffic data set; then performs trend-seasonal decomposition and time series modeling through an improved Autoformer model, and completes graph structure modeling and spatial topology reasoning through an improved GKAN model; and finally outputs the fusion prediction result based on the inconsistency scoring mechanism, dynamically identifies the spatio-temporal prediction conflict area, automatically activates the disturbance scoring function, and outputs the fusion prediction result. It is worth emphasizing that the method will automatically switch the prediction trust model according to whether there is a sudden disturbance, and dynamically select a single model result as the final output when the model prediction error is large, significantly improving the response accuracy and stability of the system.

[0161] In the actual deployment of three months, the traffic management center of City A has completed 11872 prediction tasks through the method, covering peak prediction, holiday prediction, and special weather prediction in the morning and evening peak periods. To verify the effect of the application, we select relevant indicators in typical prediction tasks during peak hours for statistics, and compare them with the LSTM model and GCN model used by the existing deployment system, and construct the following comparison table:

[0162] Table 1 Comparison of average performance of three types of models in peak period prediction tasks

[0163] Indicator name LSTM model GCN model The method of the invention Predicted average MAE (km / h) 6.82 6.14 4.21 Peak error rate (%) 24.7％ 20.3％ 12.1％ Prediction lead time (minutes) 4.1 5.0 7.2 Inconsistency recognition accuracy (%) — — 91.3％

[0164] From Table 1, it can be seen that the application is significantly better than traditional models in terms of average prediction error and peak error control ability in sudden congestion scenarios, especially after introducing the spatio-temporal inconsistency identification mechanism, the fusion prediction lead time is improved by more than 2 minutes, greatly enhancing the foresight of peak period prediction. The traditional model often fails to respond in sudden congestion or event intervention scenarios due to its inability to consider dynamic disturbances and prediction conflicts.

[0165] To further demonstrate the role of the disturbance identification module in improving prediction accuracy, we statistic the sudden disturbance event identification and response results, focusing on the disturbance score change and prediction adjustment effect of the application method, and obtain the following table:

[0166] Table 2 Disturbance score and predicted adjustment record table during sudden disturbance event (excerpts)

[0167]

[0168] As can be seen from Table 2, in the scenario where the disturbance score is higher than 0.8, the application automatically switches to the Autoformer model as the final prediction output, reducing the prediction error by more than 2 km / h on average, proving that the disturbance perception mechanism has significant discriminability and adjustment effect on sudden scenarios. In the scenario where the disturbance score is in the middle interval, the system maintains the fusion prediction strategy, showing good stability and adaptability.

[0169] The embodiment solves the core technical problems of prediction space-time conflict not easy to identify, fixed model fusion mechanism, and lag response to sudden disturbance in actual deployment. By introducing the dynamic inconsistency score and disturbance score mechanism, the application not only enhances the robustness of the model structure, but also significantly improves the prediction ability in high dynamic traffic scenarios, with wide generalizability and engineering practical value. Especially in complex scenarios such as holiday travel and sudden accidents on highways, the multi-stage fusion mechanism and scoring judgment strategy proposed by the application have irreplaceable technical advantages.

[0170] The above describes only the preferred specific embodiments of the application, but the protection scope of the application is not limited thereto, and any person skilled in the art can make equivalent substitutions or changes to the technical solutions and inventive concepts of the application within the technical scope disclosed by the application, which should be covered within the protection scope of the application.

Claims

1. A deep learning-based urban traffic big data congestion prediction method, characterized in that, The method comprises the following steps: Step one: collecting multi-source traffic data in the process of urban traffic operation, and constructing a standardized spatio-temporal traffic data set; Step two: inputting the spatio-temporal traffic data set into an improved Autoformer model and an improved GKAN model respectively, performing time series modeling and spatial topology modeling processes, and generating first and second prediction results respectively; Step three: calculating the numerical difference between the first and second prediction results, and extracting traffic state information corresponding to the target time step in the spatio-temporal traffic data set to generate an inconsistency score; Step four: calculating the fusion weight of the first and second prediction results based on the inconsistency score, performing a weighted fusion operation, and generating a fusion prediction result; Step five: determining whether the error value between the fusion prediction result and the observed data is greater than the preset error threshold; If the error value is greater than the preset error threshold, select the prediction result with the smallest error value between the first and second prediction results and the observed data as the candidate prediction result; Step six: extracting the perturbation feature of the traffic state information to generate a perturbation score; If the perturbation score is greater than the preset perturbation threshold, the first prediction result is taken as the final prediction result; If the perturbation score is less than or equal to the preset perturbation threshold, the candidate prediction result is taken as the final prediction result. 2.The congestion prediction method based on deep learning of urban traffic big data according to claim 1, characterized in that, The step one is specifically: The multi-source traffic data includes traffic flow data, road average speed data, traffic event data, weather condition data and traffic signal control data; The multi-source traffic data is synchronously processed according to time index and spatial position index to generate a unified data structure covering continuous time period and spatial grid; The traffic flow data is subjected to unit conversion and numerical smoothing processing; the road average speed data is subjected to missing value filling and outlier correction operation; the traffic event data is subjected to event classification marking and time encoding conversion operation; the weather condition data is subjected to continuous variable standardization and discrete variable mapping operation; the traffic signal control data is subjected to period normalization and phase offset encoding processing; All processed traffic data is constructed into a standardized spatio-temporal traffic data set according to time dimension and space dimension. 3.The congestion prediction method based on deep learning of urban traffic big data according to claim 2, characterized in that, The step two is specifically: The standardized spatio-temporal traffic data set is divided into multiple time window sequences according to time dimension and input into the improved Autoformer model; the improved Autoformer model sets a disturbance perception gating module before the trend-season decomposition structure, controls the trend path activation strength based on the traffic state change amplitude; sets a dynamic convolution module in the decoder structure to extract multi-scale time change features; Output the traffic state prediction value at the continuous time step to form the first prediction result; the trend and season decomposition structure refers to the trend component extraction and season component separation operation on the input time series to form the trend feature sequence and season feature sequence which can be independently modeled; The standardized spatiotemporal traffic data set is constructed into a graph structure input containing road topology structure according to the spatial dimension, and is input into the improved GKAN model; the improved GKAN model comprises a traffic state driven edge weight score adjustment module, an edge weight adjustment coefficient is calculated based on adjacent road section speed gradient, vehicle density difference and signal timing change; a graph level residual enhancement path is arranged in the graph convolution calculation structure, and the prediction stability under the dynamic change of the graph structure is maintained; a traffic state prediction value of each node at a target time step is output, and a second prediction result is formed. 4.The congestion prediction method based on deep learning of urban traffic big data according to claim 3, characterized in that, The improved Autoformer model is specifically: The standardized spatiotemporal traffic data set is divided into a plurality of time window sequences with consistent lengths according to the time dimension, and an input sequence matrix is formed; A linear projection operation is performed on the input sequence matrix to generate an initial time step embedding representation; A disturbance perception gating module is introduced on the basis of the initial time step embedding representation, a gating structure containing an input channel disturbance score and a trend channel activation factor is constructed; the disturbance score is generated by normalizing the traffic state change amplitude and is input as a control gate function; the activation factor is output by the gate function and controls the activation strength of each dimension feature in the trend path; The sequence adjusted by the disturbance perception gating module is input into a trend-seasonal decomposition structure, a local mean calculation operation and a residual separation operation in a sliding window are performed, and a trend component sequence and a seasonal component sequence are respectively obtained; Residual attention modeling operations are respectively performed on the trend component sequence and the seasonal component sequence to capture local trend features and periodic change features within the same time window; In the decoding stage, the trend component sequence is input into a dynamic convolution module, a convolution kernel group based on different time window lengths is constructed, and a multi-scale convolution operation is performed to extract short-term, medium-term and long-term trend features; Feature splicing operations are performed on the output features of the dynamic convolution module and the attention output features of the seasonal component to generate fusion feature representations; On the basis of the fusion feature representations, multi-head self-attention layers, feedforward network layers and layer normalization structures are sequentially executed to complete feature transformation and normalization processing of each layer stacking structure. After all decoding layer stacking is completed, traffic state prediction values corresponding to future time steps are output to form a first prediction result. 5.The congestion prediction method based on deep learning of urban traffic big data according to claim 4, characterized in that, The improved GKAN model is specifically: The spatial dimension information in the standardized spatiotemporal traffic data set is collected, and a road network is constructed into a graph structure, the nodes of the graph structure represent road segments, the edges represent the connection relationship between adjacent road segments, and the graph structure is attached with a node feature matrix and an edge feature matrix; The speed gradient, density difference and signal timing difference between nodes are calculated according to the vehicle speed sequence, vehicle density sequence and signal timing sequence contained in the node feature matrix; The speed gradient, density difference and signal timing difference are input into the edge weight score adjustment module to calculate the edge weight disturbance index; Linear mapping and normalization processing are performed on the edge weight disturbance index to generate an edge weight adjustment coefficient; Numerical correction is performed on the initial edge weight matrix based on the edge weight adjustment coefficient to construct an adaptive edge weight graph structure; A graph neural network backbone structure is arranged in the adaptive edge weight graph structure, and the graph neural network backbone structure includes a plurality of graph convolution units; Each graph convolution unit includes a self-attention mechanism and an edge weight modulation mechanism, which respectively perform feature update operations on the message passing strength between nodes and the importance of edge connection; A graph-level residual enhancement path is arranged between the graph convolution units, connecting the graph convolution outputs of different layers to perform splicing and normalization processing of shallow structure information and deep abstract features, to obtain cross-layer fusion features; Node features corresponding to the target time step are extracted from the cross-layer fusion features, including the speed state, traffic density state, signal control state of the road segment at the target time step, and the high-dimensional embedding representation in the graph structure, forming a state feature vector set; Based on the state feature vector set, a prediction process is performed to output the traffic state prediction value of each node, forming a second prediction result. 6.The congestion prediction method based on deep learning of urban traffic big data according to claim 5, characterized in that, The step three is specifically: Difference calculation is performed on the traffic state values in the first prediction result and the second prediction result corresponding to the same road segment and the same time step to generate a node-level prediction deviation sequence; Traffic state information corresponding to the target time step in the standardized spatiotemporal traffic data set is collected, including the actual speed value, actual traffic density value, and signal timing state of each node; A state correlation factor is constructed based on the prediction deviation sequence, the standard deviation, coefficient of variation, and average change rate of the traffic state of each node within the target time step; The prediction deviation sequence and the state correlation factor are combined to construct a three-dimensional input feature set, including the prediction deviation, traffic state variation coefficient, and state change rate; The three-dimensional input feature set is input into a weighted regression model, the model fitting is performed based on the historical traffic state data set, and an inconsistency scoring function is constructed; The prediction inconsistency value of each node at the target time step is calculated through the inconsistency scoring function; The inconsistency values of all nodes are normalized to output a normalized inconsistency score.

7. The deep learning-based urban traffic big data congestion prediction method according to claim 6, characterized in that, The step four is specifically: The normalized inconsistency value of each graph structure node at the target time step in the inconsistency score is extracted to form a node-level inconsistency score sequence; The fusion weight of the first prediction result and the second prediction result is calculated for each node, respectively, and the fusion weight is based on the reverse mapping operation of the inconsistency score, that is, the higher the score value, the lower the fusion weight of the corresponding prediction result, and the lower the score value, the higher the fusion weight; First prediction result weight matrix and second prediction result weight matrix are respectively constructed, corresponding to the fusion weight values of each node; Based on the node-level weighted fusion rule, the operation of each node is specifically: the prediction values of the corresponding nodes in the first prediction result and the second prediction result are multiplied by the corresponding fusion weight values, respectively, and the results are weighted and summed to generate a node-level initial fusion prediction value; The initial fusion prediction values of all nodes are normalized and calibrated for consistency to construct a fusion prediction result vector as the fusion prediction result. 8.The congestion prediction method based on deep learning of urban traffic big data according to claim 7, characterized in that, The step five is specifically: Collect observation data of each node in the target time step, construct an observation state vector, which includes the actual speed value, actual traffic density value and signal timing state of each node; Perform difference calculation operation on the fusion prediction result vector and the observation state vector to generate a node-level prediction error sequence; Perform absolute value average operation on the node-level prediction error sequence to calculate an error value; Determine whether the error value is greater than a preset error threshold; If the error value is greater than the preset error threshold, calculate the average error value between the first prediction result and the observation state vector, and the average error value between the second prediction result and the observation state vector, respectively, and select the prediction result with smaller error value between the first prediction result and the second prediction result as the candidate prediction result; If the error value is less than or equal to the preset error threshold, the fusion prediction result is taken as the candidate prediction result. 9.The congestion prediction method based on deep learning of urban traffic big data according to claim 8, characterized in that, The step six is specifically: Collect traffic state information corresponding to the target time step in the standardized spatio-temporal traffic data set, which includes the speed value, traffic density value and signal timing state of each node; Based on the traffic state information in the continuous multiple time steps before and after the target time step, construct a time series sample, and calculate the disturbance characteristic parameter of each node, which includes the speed disturbance amplitude, density disturbance amplitude and signal change frequency; Feature combination is performed on the speed disturbance amplitude, density disturbance amplitude and signal change frequency to construct a node-level disturbance feature vector set; Input the disturbance feature vector set into the disturbance scoring function to perform disturbance intensity estimation and output the node-level disturbance score result; Perform normalization processing on the disturbance score results of all nodes to generate a disturbance score; Determine whether the disturbance score is greater than a preset disturbance threshold; If the disturbance score is greater than the preset disturbance threshold, the first prediction result is taken as the final prediction result; If the disturbance score is less than or equal to the preset disturbance threshold, the candidate prediction result is taken as the final prediction result. 10.The congestion prediction method based on deep learning of urban traffic big data according to claim 9, characterized in that, Also includes: Record the first prediction result, the second prediction result, the fusion prediction result, the inconsistency score, the disturbance score and the final prediction result, construct a training sample set, and perform joint training process on the improved Autoformer model and the improved GKAN model to update the model parameters.

Images