Method and system for early warning of large-flow congestion in traffic confluence area
By constructing highly realistic scenarios in traffic merging zones and using STL time-series decomposition and the Informer model to predict congestion levels, the problem of accurately judging and promptly responding to congestion in traditional methods is solved, enabling accurate forecasting and emergency response for the next hour.
Patent Information
- Application Number
- CN202511560535.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-29
- Publication Date
- 2026-01-20
AI Technical Summary
Traditional congestion warning methods are difficult to accurately determine the congestion situation in traffic merging areas, and cannot provide timely feedback and solutions. They are also susceptible to factors such as weather, leading to frequent traffic accidents.
Traffic scenarios are constructed using the SUMO simulation platform. Congestion factor sequences are decoupled using STL time-series decomposition technology. A frequency-domain adaptive normalization layer and an Informer model are introduced for independent prediction. Congestion levels are classified using the K-means clustering algorithm to achieve accurate prediction of congestion levels for the next hour.
It improves the accuracy and timeliness of congestion warnings, can dynamically adapt to the periodicity and non-stationarity of traffic flow, reduces computational complexity, reduces noise interference, and ensures the rationality and reliability of predictions.
Smart Images
Figure CN121366493A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of intelligent traffic control, in particular to a traffic flow merging area large flow congestion early warning method and system. BACKGROUND
[0002] With the acceleration of urbanization and the surge in the number of motor vehicles, the highway flow merging area has become a traffic bottleneck due to frequent traffic interweaving, which easily causes local congestion and leads to a chain of traffic paralysis. Traditional congestion early warning methods are mostly based on fixed thresholds or simple historical statistics, which are difficult to ensure accurate judgment and early processing of traffic congestion, resulting in an increased probability of frequent congestion or even accidents in the traffic flow merging area.
[0003] Therefore, the prior art uses traffic sensors based on geomagnetic or millimeter wave radar to detect basic data such as traffic flow and occupancy rate, and then calculates the traffic flow to determine the congestion condition. Or use target detection algorithms such as yolo11 to identify vehicle types, trajectories, and abnormal events, and model the congestion relationship of the road section. However, sensor data processing and congestion judgment require time, which is difficult to timely determine and feedback the congestion section and propose solutions. Target detection algorithms cannot predict and warn congestion mirrors, and are easily affected by weather and other factors, resulting in reduced robustness. Therefore, there is an urgent need for a new traffic flow merging area congestion early warning and prediction scheme that can accurately and in advance predict potential congestion conditions, facilitating subsequent timely emergency handling plans. SUMMARY
[0004] The present application provides a traffic flow merging area large flow congestion early warning method and system, which constructs a high-fidelity traffic scene through a SUMO simulation platform, accurately obtains multi-dimensional traffic flow parameters such as average flow, speed, spatial occupancy rate, and queuing time, uses STL time series decomposition technology to decouple the congestion factor sequence, introduces a frequency domain adaptive normalization (FAN) processing layer, uses the sequence prediction ability of the Informer model to independently predict and reconstruct each component, and accurately predicts the congestion level in the next hour, facilitating subsequent timely emergency handling plans.
[0005] TECHNICAL SCHEME
[0006] The present application discloses a traffic flow merging area large flow congestion early warning method, which comprises the following steps:
[0007] S1: Collecting multi-dimensional traffic flow parameters in the highway flow merging area;
[0008] S2: Based on the traffic flow parameters collected by the data collection unit, a comprehensive congestion factor is calculated based on a preset weight coefficient, a complete model training data set is constructed, and through a K-means clustering algorithm, two-level early warning thresholds of the congestion factor are determined, and the traffic state is divided into "smooth", "moderate congestion", and "severe congestion" three congestion levels;
[0009] S3: Using the STL algorithm, the comprehensive congestion factor sequence is decomposed to obtain three component sequences of trend, season, and residual, which are respectively introduced into the frequency domain adaptive normalization layer for processing;
[0010] S4: Each component is independently trained, the component sequence is input into the trained informer model, and the result is reconstructed to obtain the final congestion factor prediction result, and the output signal warning is based on the congestion level output signal in S2.
[0011] Further, the comprehensive congestion factor in S2 is calculated as follows:
[0012] wherein, is the average queuing time of the vehicle, is the road space occupancy rate, Aq is the average road flow, and Av is the average vehicle speed, is the comprehensive congestion factor.
[0013] Further, based on the comprehensive congestion factor , the threshold interval is distinguished by the K-means clustering algorithm, the high value interval is "severe congestion", the medium value interval is "moderate congestion", and the value interval is "smooth", and the objective function is as follows:
[0014] minΣ(i=1to3)Σ(x∈Cᵢ)||x-μᵢ||²
[0015] The threshold is determined as follows:
[0016] T1=(μ1+μ2) / 2
[0017] T2=(μ2+μ3) / 2
[0018] Wherein, Cᵢ is the i-th cluster, μᵢ is the cluster center, X is the congestion factor data set, μ1, μ2, μ3 are three cluster centers satisfying μ1<μ2<μ3, T1 is the demarcation threshold between smooth state and moderate congestion state, and T2 is the demarcation threshold between moderate congestion state and severe congestion state.
[0019] Further, the decomposition of the comprehensive congestion factor sequence in S3 is as follows:
[0020] The STL algorithm will use the sequence The sequence is divided into three components: trend component, seasonal component, and residual component. A frequency domain adaptive normalization layer is introduced for processing. In the spectrum, the algorithm identifies the trend mode and non-stationary mode. After filtering to retain the principal components, the algorithm reconstructs them to the time domain through inverse discrete Fourier transform to form new principal signal components. The remaining residual components are used as the input after normalization and become the modeling target of the subsequent model.
[0021] Furthermore, the structure of the Informer model described in S4 is as follows:
[0022] The encoder uses a cascaded multi-head probabilistic sparse self-attention module to process the input sequence. It extracts long-range temporal dependency features through a sparse attention mechanism with a computational complexity of LlogL. It uses a dual-path attention mechanism as the decoder. Masked multi-head probabilistic sparse self-attention ensures that the decoding process is based only on historical information. The standard multi-head attention module dynamically fuses the deep features passed by the encoder and outputs multi-step prediction results through a fully connected layer.
[0023] Furthermore, the result reconstruction and signal warning described in S4 are as follows:
[0024] Result reconstruction:
[0025] The Informer performs predictions on the three subsequences and then reconstructs the results to restore the true predicted values:
[0026]
[0027] like < If so, a green alert for "unimpeded traffic" will be output;
[0028] like ≤ < If so, a yellow alert for "moderate congestion" will be issued;
[0029] like ≥ If so, a red alert for "severe traffic congestion" will be issued;
[0030] in: It is the final prediction result of the reconstruction, derived from the predicted... , , The result is obtained by summing the three subsequences.
[0031] Furthermore, the present invention also discloses a system for early warning of high-volume congestion in traffic merging areas, the system including a data collection unit for collecting multi-dimensional traffic flow parameters of highway merging areas, including average flow, speed, space occupancy rate and queuing time;
[0032] Factor design and classification unit: used for acquiring traffic flow parameters collected by the data collection unit, calculating a comprehensive congestion factor, determining two-level early warning thresholds of the congestion factor through a K-means clustering algorithm, and dividing the traffic state into three levels of "smooth", "moderate congestion" and "heavy congestion";
[0033] Prediction and early warning unit: used for decomposing a comprehensive congestion factor sequence input by the factor design and classification unit through an STL algorithm to obtain three component sequences of trend, season and residual, respectively introducing a frequency domain adaptive normalization layer for processing, then performing prediction by an Informer model, and fusing the results to obtain a congestion factor prediction value for the next hour, according to which a congestion early warning signal is output;
[0034] Decision support unit: used for receiving a congestion level output by the early warning unit, providing a hierarchical control strategy recommendation for a traffic control platform, and realizing a control mode change from passive response to active intervention.
[0035] Further, the system downloads open source road network data of a target merging area based on an osmWebWizard tool of a SUMO simulation platform, and converts the data into a road network file format recognizable by SUMO to complete the construction of a basic simulation environment.
[0036] Beneficial effects:
[0037] 1. The STL time series decomposition technology is adopted to decouple the congestion factor sequence, so that the mixed original signal is decoupled into components of different time scales, thereby optimizing the prediction target of the subsequent FAN layer and Informer model, and further improving the accuracy of the subsequent prediction result.
[0038] 2. The frequency domain adaptive normalization (FAN) layer is introduced for processing, and then the Informer model is used for prediction, so that the model can dynamically adapt to periodicity and non-stationarity in the time series, overcome the dependence on fixed stationarity assumption of the traditional method, and make the subsequent Informer layer modeling more accurate.
[0039] 3. The informer model structure designed in the application adopts a cascaded multi-head self-attention module to process source sequence input, reduces the inherent computational complexity of the traditional self-attention mechanism, saves the prediction computing power cost, uses the mask multi-head self-attention mechanism to further filter invalid data and prevent future information leakage, further reduces noise interference, and ensures the rationality of prediction. BRIEF DESCRIPTION OF DRAWINGS
[0040] Figure 1 is a specific flowchart of the method of the application;
[0041] Figure 2 STL algorithm workflow diagram;
[0042] Figure 3 The frequency domain adaptive normalization layer (FAN) schematic diagram of the present application;
[0043] Figure 4 The Informer model schematic diagram of the present application. DETAILED DESCRIPTION
[0044] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.
[0045] As shown in Figure 1 , the embodiments of the present application disclose a traffic merging area large-flow congestion early warning method, and the method steps are as follows:
[0046] The road network file is initialized, the data is obtained by running the road network file according to the python script, the vehicle flow file is obtained, the vehicle flow is added in the routes.rou.xml file, the large flow is simulated, and the simulation running time is modified. The sumo simulation is run using the python script, and in the python script, the average flow, speed, space occupancy rate and queuing time and other parameters of the merging area are taken.
[0047] The data collection frequency is taken once every 900 steps (about 15 minutes) of simulation.
[0048] The comprehensive congestion factor is calculated according to the collected data, and the setting rule is expressed by the formula as follows:
[0049] Among them, is the average queuing time of vehicles; is the road space occupancy rate; Aq is the average flow of the road; Av is the average speed of the vehicle flow; is the comprehensive congestion factor.
[0050] The characteristic values of the data collected at this moment are as follows: , , Aq, Av, .
[0051] Further, the K-means clustering algorithm is used to automatically determine the three-level threshold system for dividing the traffic state according to the historical comprehensive congestion factor (Factor) data.
[0052] The objective function is:
[0053]
[0054] The formula for determining the threshold value is:
[0055]
[0056] where Cᵢ is the i-th cluster, μᵢ is the cluster center, X is the congestion factor dataset, μ1, μ2, μ3 are the three cluster centers satisfying μ1 < μ2 < μ3, T1 is the threshold between the free state and the moderate congestion state, and T2 is the threshold between the moderate congestion state and the severe congestion state.
[0057] Specifically, the algorithm aggregates data points into three cluster centers through unsupervised learning, and takes the middle value of adjacent cluster centers as the classification boundary, thereby defining the "free", "moderate congestion" and "severe congestion" three levels, and realizing the automatic division of congestion state.
[0058] The collected raw data needs to be preprocessed. The preprocessing steps include:
[0059] Data standardization: convert the data to a unified dimension for subsequent processing.
[0060] In the data standardization step, the data uniformization method adopted by the present application is the min-max normalization method, and the formula is:
[0061]
[0062] where, is the normalized data; is the original data. is the minimum and maximum value in the data.
[0063] The preprocessed data is stored in the database for historical data analysis and long-term trend prediction.
[0064] Further, as shown in Figure 2 , the original comprehensive congestion factor sequence is decoupled into three component sequences of trend, seasonality and residual by the STL algorithm. This decomposition enables the Informer model to learn different time series patterns for different components, improving the prediction accuracy:
[0065]
[0066] where, is the congestion factor observation value at time t; is the trend component, reflecting the long-term change law of the congestion factor; for seasonal component, embodying the periodicity of traffic flow; for residual component, representing random fluctuation and noise.
[0067] STL time series classification is realized by inner loop and outer loop. Inner loop handles seasonal component and trend component by alternating optimization strategy. The process first fixes the trend component, extracts the seasonal component based on the residual within a cycle using LOESS smoothing method; then refits the trend component after removing the extracted seasonal component. This alternating optimization mechanism is executed through multiple iterations until a stable component decomposition result is obtained; the outer loop adopts a robust weighting mechanism to dynamically adjust the weights of data points, and the weight distribution is based on the residual size after each iteration. This mechanism effectively reduces the influence of abnormal observations on the decomposition result by reducing the weight proportion of abnormal values, thereby significantly enhancing the overall robustness and stability of the algorithm.
[0068] Three datasets are obtained through STL classification. One is the residual component subsequence dataset, whose feature value is , and the target value is ; one is the seasonal component subsequence dataset, whose feature value is , and the target value is , and one is the trend component subsequence dataset, whose feature value is , and the target value is .
[0069] Further, as shown in Figure 3 , each component subsequence dataset is introduced into the frequency domain adaptive normalization layer. The frequency domain adaptive normalization (FAN) mechanism is an innovative time series processing method, which is characterized by explicitly separating and dynamically adapting non-stationary components in time series through frequency domain transformation. This mechanism starts with frequency domain normalization processing: the original signal is mapped to the frequency domain through discrete Fourier transform, and the dominant frequency mode representing trend and non-stationarity is automatically identified and extracted in the frequency spectrum space. These dominant components are retained after filtering and reconstructed into time domain signals through inverse Fourier transform to form the main signal component; while the remaining residual component is input into the downstream model as a stationary representation. In the prediction phase, this mechanism adopts a divide-and-conquer strategy: the main frequency component is processed by a lightweight static predictor, while the non-dominant residual is captured by a complex model. The prediction results of the two are fused and reconstructed in the frequency domain, and finally the time domain prediction value is generated through inverse transform. This process dynamically predicts the evolution of non-stationary patterns through the pattern adaptation module, completely overcoming the dependence on fixed stationarity assumption of traditional methods, as shown in Figure 3 . The introduction processing formula is as follows:
[0070] , , are subsequences processed by FAN respectively.
[0071] Each component after FAN processing is independently trained and input into the trained informer model, as shown in Figure 4 The encoder of the informer adopts a cascaded multi-head self-attention module to process the source sequence input, extracts long-range temporal dependence features through a sparse attention mechanism with a complexity of only LlogL, L being the length of the sequence, and this multi-layer structure effectively extracts key temporal features while significantly reducing the inherent computational complexity of the traditional self-attention mechanism. The extracted deep features are transmitted to the decoder through splicing feature maps. The decoder adopts a double-path attention design. On the one hand, the masked multi-head self-attention mechanism is used to ensure that the decoding process only relies on historical information and prevents future information leakage; on the other hand, the standard multi-head attention module is used to dynamically integrate the feature information provided by the encoder. Finally, the model outputs multi-step prediction results through a fully connected layer:
[0072]
[0073] wherein, is the congestion factor observation value at time t; is the trend component, reflecting the long-term change rule of the congestion factor; is the seasonal component, embodying the periodicity of the traffic flow; is the residual component, representing random fluctuations and noise; Informer() is a kind of time series prediction model, which is used in combination with the STL decomposer for prediction in the present application; is the prediction result of the trend component at the next time; is the prediction result of the seasonal component at the next time; is the prediction result of the residual component at the next time.
[0074] The predicted congestion factor value is compared with the threshold value determined in advance through cluster analysis, so as to determine the congestion level to which it belongs.
[0075] If , the "smooth" green warning is output.
[0076] If ≤ , the "moderate congestion" yellow warning is output.
[0077] If ≥ , the "heavy congestion" red warning is output.
[0078] The following strategies are implemented according to the different levels inferred:
[0079] For "smooth", no policy is executed, for "moderate congestion", the execution of opening the ramp red light interval change is prompted to adjust traffic flow, and for "heavy congestion", the execution of strict restriction or even temporary closure of upstream entrance ramp is prompted to prevent further deterioration of congestion.
[0080] Further, the embodiment of the present application also discloses a system for large-flow congestion early warning of a traffic merging area, which comprises a data collection unit for collecting multi-dimensional traffic flow parameters of the merging area of the expressway, including average flow, speed, space occupancy and queuing time;
[0081] a factor design and classification unit for obtaining the traffic flow parameters collected by the data collection unit, calculating a comprehensive congestion factor, determining two-level early warning thresholds of the congestion factor through a K-means clustering algorithm, and dividing the traffic state into three levels of "smooth", "moderate congestion" and "heavy congestion";
[0082] a prediction and early warning unit for decomposing the sequence of the comprehensive congestion factor input by the factor design and classification unit through an STL algorithm to obtain three component sequences of trend, season and residual error, respectively introducing a frequency domain adaptive normalization layer for processing, then performing prediction by an Informer model, and fusing the results to obtain a congestion factor prediction value in the next one hour, and outputting a congestion early warning signal according to the congestion factor prediction value;
[0083] a decision support unit for receiving the congestion level output by the early warning unit, providing a hierarchical control strategy recommendation for a traffic control platform, and realizing a control mode change from passive response to active intervention.
[0084] The hierarchical control strategy is executed by an execution unit built in the decision support unit, and the execution unit implements level determination and strategy interactive operation, and provides a strategy for different types of priorities according to the classification result of the decision support unit.
[0085] The operation steps of the execution unit are as follows:
[0086] Early warning instruction analysis:
[0087] The congestion level signal generated by the prediction and early warning unit is received and analyzed to obtain corresponding early warning levels, color codes and standardized prompt information.
[0088] Multi-channel early warning information release:
[0089] The early warning information is synchronously released through a variable information sign, a visual interface of the traffic control platform and a public information service channel.
[0090] Full-factor data synchronous monitoring:
[0091] Deploy virtual detector arrays at key locations in the merging area, synchronously collect traffic flow, speed, space occupancy, and queue time data.
[0092] Multi-dimensional evaluation of early warning effect:
[0093] Calculate early warning accuracy, information reach rate, and traffic behavior impact, etc. indicators, and generate early warning effect evaluation report.
[0094] Model parameter adaptive optimization:
[0095] Feedback evaluation data to STL+FAN+Informer prediction model, optimize model parameters and K-means clustering threshold, and improve prediction accuracy.
[0096] Intelligent monitoring of abnormal events:
[0097] Real-time detection of traffic accidents, equipment communication interruption and other abnormal events, trigger corresponding emergency warning plan.
[0098] Early warning whole process data archiving:
[0099] Store early warning instructions, traffic flow data and effect evaluation results, support historical data backtracking and early warning strategy optimization verification.
[0100] Prediction effect verification of the present application:
[0101] In the present application, the data set division strategy according to time sequence is adopted, and the ratio of training set and test set is 8:2. Among them, the training set accounts for 80% of the total sample amount, which is used for model training and parameter optimization; the test set accounts for 20%, which is used for evaluating the generalization performance and prediction ability of the model. The model performance is comprehensively evaluated by the following indexes:
[0102]
[0103] Among them:
[0104] n: total number of samples in the test set. : index of sample.
[0105] : real observation value at the th time point.
[0106] : predicted value of the model at the th time point.
[0107] : average value of all real observation values, often used as a simple benchmark prediction model.
[0108] In this study, the proposed STL+FAN+Informer model is compared with the benchmark models such as PatchTST and Informer, and the performance comparison results are shown in Table 1.
[0109] Table 1
[0110]
[0111] According to the model performance comparison results shown in the table, the STL+FAN+Informer fusion model proposed in this study exhibits performance on various evaluation indicators, surpassing all comparison models including STL+Informer. In terms of goodness of fit, the R² value of the STL+FAN+Informer model reaches the highest level of 0.948, higher than STL+Informer (0.938), better than single Informer model (0.900) and PatchTST model (0.871), proving that it can capture more complex dynamic characteristics in the data and has better explanation ability. In terms of prediction error, the RMSE (110.34) and MAE (77.533) of the model are the lowest values. Compared with the STL+Informer model with sub-optimal performance, the RMSE is reduced by about 19%, and the MAE is improved by about 11.3%, indicating that the introduction of the FAN module for frequency domain adaptive processing effectively improves the accuracy and stability of model prediction.
[0112] The above description of the embodiments enables those skilled in the art to implement or use the present application. Various modifications to the embodiments will be apparent to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application should not be limited to the embodiments shown herein, but should cover the widest range compatible with the principles disclosed herein and the novel features.
Claims
1. A method for traffic flow congestion warning in a merging area, characterized in that, The method comprises the following steps: S1: collecting multi-dimensional traffic flow parameters of a highway merging area; S2: calculating a comprehensive congestion factor based on a preset weight coefficient according to the traffic flow parameters collected by the data collection unit, constructing a complete model training data set, determining two-level early warning thresholds of the congestion factor through a K-means clustering algorithm, and dividing the traffic state into three congestion levels of "smooth", "moderate congestion" and "heavy congestion"; S3: decomposing the comprehensive congestion factor sequence by using an STL algorithm to obtain three component sequences of trend, season and residual, and respectively introducing a frequency domain adaptive normalization layer for processing; S4: independently training each component, inputting the component sequence into the trained informer model, and reconstructing the result to obtain the final congestion factor prediction result, and outputting a signal warning based on the congestion level in S2.
2. The method for traffic flow congestion warning of traffic merging area according to claim 1, wherein, The comprehensive congestion factor in S2 is calculated as follows: ； wherein, is the average queue time for vehicles, is the road space occupancy, Aq is the average road flow, and Av is the average vehicle speed, is the comprehensive congestion factor.
3. The method for traffic flow congestion warning of traffic merging area according to claim 2, characterized in that, Based on comprehensive congestion factor , by K-means clustering algorithm to distinguish threshold interval, high value interval is "heavy congestion", medium value interval is "moderate congestion", value interval is "smooth", the objective function is as follows: min∑(i=1to3)∑(x∈C i ||x-μ i ||² The threshold is determined as follows: T1=(μ1+μ2) / 2 T2=(μ2+μ3) / 2 wherein C i is the ith cluster, μ i is the cluster center, X is the congestion factor dataset, μ1, μ2, μ3 are three cluster centers satisfying μ1 < μ2 < μ3, T1 is the dividing threshold between the free state and the moderate congestion state, and T2 is the dividing threshold between the moderate congestion state and the severe congestion state.
4. The method for traffic flow congestion warning of traffic merging area according to claim 1, wherein, The decomposition of the comprehensive congestion factor sequence in S3 is as follows: STL algorithm divides the sequence into three component sequences, namely trend component, seasonal component and residual component. A frequency domain adaptive normalization layer is introduced for processing. In the frequency spectrum, the algorithm identifies the trend mode and the non-stationary mode. After filtering to retain the principal component, it is reconstructed to the time domain through inverse discrete Fourier transform to form a new principal signal component. The remaining residual component is used as the input after normalization and becomes the modeling target of the subsequent model.
5. The method for traffic flow congestion warning of traffic merging area according to claim 1, wherein, The informer model structure in S4 is as follows: The encoder adopts a cascaded multi-head probability sparse self-attention module to process the input sequence, extracts long-range time sequence dependency features through a sparse attention mechanism with a complexity of LlogL, adopts a double-path attention mechanism as the decoder, and a mask multi-head probability sparse self-attention ensures that the decoding process is only based on historical information. The standard multi-head attention module dynamically fuses the deep features transmitted by the encoder, and outputs the multi-step prediction result through the full connection layer.
6. The method for traffic flow merging area large-flow congestion early warning according to claim 5, characterized in that, The result reconstruction and signal warning in S4 are as follows: Result reconstruction: After the three sub-sequences are predicted by the informer, the result is reconstructed to restore the true prediction value: ； If then output "smooth" green pre-warning; If ≤ < then output a yellow warning of "moderate congestion"; If ≥ then output "heavy congestion" red warning; wherein: is the final prediction result of the reconstruction, derived from the predicted , , three sub-sequences are accumulated.
7. The system for traffic merging area large-flow congestion warning according to any one of claims 1-6, characterized in that, The system comprises The data collection unit is used to collect multi-dimensional traffic flow parameters of the highway merging area, including average flow, speed, space occupancy and queuing time; The factor design and classification unit is used to obtain the traffic flow parameters output by the data collection unit, calculate the comprehensive congestion factor, determine the two-level early warning thresholds of the congestion factor through the K-means clustering algorithm, and divide the traffic state into three levels of "smooth", "moderate congestion" and "heavy congestion"; The prediction and warning unit is used to decompose the comprehensive congestion factor sequence input by the factor design and classification unit through the STL algorithm to obtain three component sequences of trend, season and residual, respectively introduce a frequency domain adaptive normalization layer for processing, then predict by the informer model, and fuse the results to obtain the congestion factor prediction value in the next hour, and output the congestion warning signal accordingly; The decision support unit is used to receive the congestion level output by the warning unit, provide hierarchical control strategy recommendation for the traffic control platform, and realize the transition of the control mode from passive response to active intervention.
8. The system for traffic merging area large-flow congestion warning according to claim 7, characterized in that, The system downloads the open source road network data of the target merging area based on the osmWebWizard tool of the SUMO simulation platform, and converts it into a road network file format recognizable by SUMO to complete the construction of the basic simulation environment.