A traffic signal light strategy dynamic regulation method based on space-time dimensions

Abstract

The application relates to the technical field of intelligent traffic control, and discloses a traffic signal lamp strategy dynamic regulation and control method based on space-time dimensions. The method captures multi-dimensional feature data of traffic flow, fuses weather and road topological information, constructs a dynamic traffic state tensor, carries out multi-scale space-time feature extraction on the tensor, generates a space-time mode atlas, analyzes the interaction strength between different traffic units based on the atlas through a graph neural network, divides a cooperative evolution area and a conflict frequent area, outputs a regional interaction label set, integrates signal lamp historical data and real-time delay, calculates a sensitivity coefficient of flow fluctuation on signal timing, forms a signal sensitivity analysis report, extracts conflict frequent area nodes with sensitivity exceeding an adaptive threshold, constructs a key interference node set, combines the space-time features of the set, uses a risk propagation model to evaluate a chain reaction probability, and generates a traffic signal optimization scheme containing a signal parameter optimization strategy and an emergency regulation and control instruction.

Description

Technical Field

[0001] This invention relates to the field of intelligent traffic control technology, specifically to a method for dynamic adjustment of traffic signal strategies based on spatiotemporal dimensions. Background Technology

[0002] Urban transportation systems are highly complex and dynamically changing mega-systems, and their operational efficiency and reliability largely depend on traffic light control strategies. Traditional signal control methods, such as fixed timing schemes, while simple and stable, cannot adapt to the significant non-uniformity and random fluctuations of traffic flow in time and space, often leading to empty intersections or exacerbated congestion. With the application of technologies such as inductive control and adaptive coordinated control, the level of intelligence in signal control has improved. These methods typically rely on local traffic parameters acquired by devices such as inductive loop detectors, radar, or video detectors, and adjust the signal timing of individual intersections or arterial roads based on preset rules or optimization algorithms. However, existing methods still have significant limitations. Their perception dimension is relatively singular, mostly focusing on traffic flow changes over time, lacking a deep understanding of the spatial distribution characteristics of traffic flow, and failing to effectively integrate external influencing factors such as road geometry and weather conditions.

[0003] In modern urban road networks, traffic congestion is rarely an isolated phenomenon, but rather spatially interconnected and temporally progressive. Oversaturation at one intersection can rapidly spread upstream, potentially triggering a chain reaction at adjacent intersections and ultimately leading to regional traffic paralysis. While current control strategies have partially achieved coordination on arterial roads or within small areas, they are typically based on relatively static "green wave" designs or simple neighbor communication mechanisms, making it difficult to characterize and understand the complex and dynamically evolving interactions between multiple traffic units in large-scale road networks. These interactions include both the synergistic effects expected from continuous traffic flow and competition and conflicts arising from path selection conflicts and phase mismatches. The lack of quantitative analysis of these spatiotemporal correlation mechanisms often limits control strategies to local optimization, making it difficult to anticipate and suppress the spread of traffic disturbances within the road network from a global perspective.

[0004] Traffic systems contain various sensitive nodes. Certain intersections or road segments, due to their unique topological locations or traffic compositions, are exceptionally sensitive to changes in signal timing; even minor adjustments can trigger significant changes in the road network's operational status. Existing control methods generally lack effective mechanisms for identifying these critical nodes and the ability to predict the potential impact of local adjustments on the global state. Control decisions are often based on current or recent historical conditions, representing a relatively passive response and lacking foresight regarding the risk of impending congestion propagation. When sudden events trigger traffic flow anomalies, the system struggles to adjust its strategies in a timely manner to curb the spread of the impact. Therefore, there is an urgent need for a method that can deeply mine traffic flow evolution patterns from spatiotemporal multidimensional data, quantify the dynamic interactions between intersections, accurately identify vulnerable links in the network, and generate forward-looking and interference-resistant signal optimization strategies to improve the overall resilience and efficiency of urban road networks. Summary of the Invention

[0005] The purpose of this invention is to provide a dynamic control method for traffic signal strategies based on spatiotemporal dimensions, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, this invention provides a method for dynamic control of traffic signal strategies based on spatiotemporal dimensions, the method comprising: By continuously capturing multidimensional feature data of traffic flow through an intelligent sensing node network, including vehicle trajectory coordinates, time series index, instantaneous flow density and speed distribution, and integrating weather conditions and road topology information, a dynamic traffic state tensor is constructed. Multi-scale spatiotemporal feature extraction is performed on the dynamic traffic state tensor, and a deep convolutional network is used to identify the periodic and sudden patterns of traffic flow to generate a spatiotemporal pattern map. Based on the spatiotemporal pattern map, the interaction intensity between different traffic units is analyzed by graph neural network, and the co-evolution zone and the conflict-prone zone are divided, and the regional interaction label set is output. For the collaborative evolution zone in the aforementioned regional interactive tag set, historical traffic light phase data and real-time response delay are integrated to calculate the sensitivity coefficient of traffic fluctuations to signal timing, and a signal sensitivity analysis report is generated. Nodes in frequently conflicting areas whose sensitivity exceeds the adaptive threshold are extracted from the signal sensitivity analysis report to construct a set of key interference nodes; Based on the spatial distribution and temporal evolution characteristics of the key interference node set, a risk propagation model is used to assess the probability of chain reactions, and a traffic signal optimization scheme including signal parameter optimization strategies and emergency control instructions is generated.

[0007] Preferably, the dynamic traffic state tensor includes a normalized flow matrix, velocity field distribution, environmental factor encoding, and road topology map; the spatiotemporal pattern map includes a periodic pattern feature map, a sudden pattern heatmap, and a spatiotemporal correlation matrix; the regional interaction label set specifically includes co-evolution zone identifiers, conflict-prone zone identifiers, and interaction intensity weights; the signal sensitivity analysis report includes a flow fluctuation sensitivity coefficient, a signal response delay gradient, and a timing adjustment priority score; the key interference node set includes node spatial topological location, node temporal activity index, and node sensitivity and flow correlation factors; and the traffic signal optimization scheme includes an optimized node list, a multi-dimensional risk diffusion map, and an adaptive parameter tuning sequence.

[0008] Preferably, the step of continuously capturing multi-dimensional feature data of traffic flow through an intelligent sensing node network, including vehicle trajectory coordinates, time series indexes, instantaneous flow density, and speed distribution, and fusing weather conditions and road topology information to construct a dynamic traffic state tensor includes: Deploy high-precision sensor arrays and video recognition units to collect the three-dimensional spatial coordinates and millisecond-level timestamps of each traffic monitoring point in real time, and simultaneously record the counting data and speed sampling values ​​of passing vehicles. At the same time, integrate visibility, precipitation intensity and temperature data from meteorological sensors, as well as the connection relationship and slope information of road geometry, to form a raw multi-source data stream. The original multi-source data stream is spatiotemporally aligned and missing value interpolated. The vehicle trajectory coordinates are converted into gridded traffic density, the speed distribution is fitted into a continuous field function, the weather conditions are quantized into environmental impact coefficients, the road topology is encoded into a graph structure, and the data is integrated into a dynamic traffic state tensor through tensor splicing operations.

[0009] Preferably, the step of extracting multi-scale spatiotemporal features from the dynamic traffic state tensor, identifying periodic and sudden patterns of traffic flow using a deep convolutional network, and generating a spatiotemporal pattern map includes: The dynamic traffic state tensor is input into a multi-branch convolutional neural network, where the spatial branch uses a three-dimensional convolutional kernel to extract local traffic flow features, the temporal branch uses a long short-term memory network to capture long-term dependencies, and the multi-scale branch fuses spatiotemporal information of different granularities through a pyramid pooling layer to output a multi-scale feature map. Pattern recognition analysis is performed on the multi-scale feature maps, and periodic patterns and burst patterns are separated by clustering algorithms. Visual heatmaps and feature vector matrices are generated and combined to form a spatiotemporal pattern map.

[0010] Preferably, the step of analyzing the interaction intensity between different traffic units based on the spatiotemporal pattern map using a graph neural network, dividing the area into co-evolution zones and conflict-prone zones, and outputting a set of regional interaction labels includes: The traffic units in the spatiotemporal pattern graph are modeled as graph nodes, and the traffic flow correlation between nodes is used as the edge weight to construct a spatiotemporal interaction graph. The graph attention network is used to calculate the influence weight of each node on its neighbors. Global information is aggregated through a message passing mechanism to output the interaction strength score of each node. Based on the distribution of interaction intensity scores, regions with scores above the collaboration threshold are marked as collaborative evolution zones, and regions with scores below the conflict threshold are marked as conflict-prone zones. The boundary coordinates and intensity weights of each region are recorded to generate a set of regional interaction labels.

[0011] Preferably, the step of integrating historical traffic light phase data and real-time response delay into the co-evolution zone of the regional interaction tag set, calculating the sensitivity coefficient of traffic fluctuations to signal timing, and generating a signal sensitivity analysis report includes: The phase duration, cycle history, and real-time operation delay data of traffic lights in the co-evolution zone are exported from the traffic signal control system and aligned with the traffic flow time series of the corresponding area. Using a sequence-to-sequence prediction model, with the flow fluctuation sequence as input and the signal timing change as output, the model is trained to predict the sensitivity of flow changes to signal parameters, and the sensitivity coefficient is calculated as the gradient value. The sensitivity coefficients of all co-evolutionary regions are integrated, and the priority is adjusted by combining real-time response delay calculation to generate a signal sensitivity analysis report.

[0012] Preferably, the step of extracting nodes in frequently occurring conflict areas whose sensitivity exceeds an adaptive threshold from the signal sensitivity analysis report and constructing a set of key interference nodes includes: Analyze the distribution of sensitivity coefficients in the signal sensitivity analysis report, dynamically calculate the rolling average sensitivity threshold, and screen out nodes whose sensitivity coefficients exceed the threshold and are located in conflict-prone areas; Extract complete time-series data of nodes whose sensitivity coefficients exceed the threshold and are located in conflict-prone areas, including traffic peaks, velocity variation rates, and traffic light status change records, to form a node behavior profile; The stability of node behavior is verified by anomaly detection algorithm. Nodes that continuously exhibit abnormal fluctuations are marked as key interference nodes and compiled into a set of key interference nodes.

[0013] Preferably, the step of using a graph attention network to calculate the influence weight of each node on its neighbors includes: using a multi-head attention mechanism to calculate the similarity between node features, normalizing the attention weights using a softmax function, and weighted summing of the neighbor node features to update the current node representation.

[0014] Preferably, the method of using a sequence-to-sequence prediction model to calculate the sensitivity coefficient of traffic fluctuations to signal timing includes: adopting an encoder-decoder architecture, where the encoder encodes the traffic sequence into a hidden state, the decoder predicts the signal timing changes based on the hidden state, and calculates the gradient of the traffic input with respect to the predicted output as the sensitivity coefficient through backpropagation.

[0015] Preferably, the step of verifying the stability of node behavior through anomaly detection algorithm includes: applying the isolated forest algorithm to score the node time series for anomalies, marking moments with scores exceeding an adaptive anomaly threshold as anomalies, and counting the number of consecutive anomalies to confirm the effectiveness of key interference nodes.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention integrates multi-dimensional data such as vehicle trajectory, time index, flow density, and speed distribution, and fuses weather and road topology information to construct a dynamic traffic state tensor that provides a panoramic and structured description of the traffic system's operational status. This data foundation surpasses traditional single-section flow detection, simultaneously reflecting the distribution and variation characteristics of traffic flow in both time and space, providing a rich source of information for in-depth analysis.

[0017] This method employs deep convolutional networks to extract multi-scale spatiotemporal features from dynamic traffic state tensors. It can automatically learn and identify complex patterns inherent in traffic flow, such as the periodicity of morning and evening rush hours and sudden congestion patterns caused by events. The generated spatiotemporal pattern map transforms abstract traffic states into visual and computable feature representations, profoundly revealing the spatiotemporal dynamics behind traffic phenomena.

[0018] The introduction of graph neural networks to analyze the interaction strength between different traffic units is a significant advancement. It treats the road network as a dynamically evolving graph structure, enabling quantitative assessment of the strength of the influence relationship between any two units, thus objectively delineating areas of co-evolution and conflict-prone zones. This data-driven regional division provides a scientific basis for implementing differentiated and refined control strategies.

[0019] Signal sensitivity analysis is conducted in the co-evolution zone to correlate traffic flow fluctuations with changes in signal timing parameters. By calculating sensitivity coefficients, it is possible to identify which areas or intersections respond most significantly to signal adjustments. This allows for more targeted allocation of control resources, prioritizing control variables that have the greatest impact on system performance.

[0020] The steps to construct a set of key interference nodes focus on nodes that are highly sensitive and located in conflict-prone areas. These nodes are potential weak links in the road network, and even minor deterioration in their condition can trigger large-scale chain reactions through spatial correlation. Identifying these key nodes is equivalent to grasping the crucial hubs controlling the overall operation of the road network.

[0021] By employing a risk propagation model combined with the spatiotemporal characteristics of key nodes, the probability of a local adjustment or disturbance triggering a chain reaction throughout the road network is assessed. This allows the generated traffic signal optimization scheme to not only consider the optimization of local effects but also to fully evaluate the global impact and potential risks of the strategy. The scheme includes both conventional parameter optimization strategies and contingency control instructions for handling unexpected situations, demonstrating the robustness and forward-looking nature of the control system. The entire method forms a complete technical chain from multi-dimensional perception, pattern recognition, interaction analysis, sensitivity assessment, key node localization to risk prediction and strategy generation, realizing the transformation of traffic signal control from local, passive, and experience-based to global, proactive, and data-driven, enhancing the ability of urban traffic systems to cope with complexity and uncertainty. Attached Figure Description

[0022] Figure 1 This is a schematic diagram illustrating the working principle of the dynamic control method for traffic signal strategies based on spatiotemporal dimensions described in this invention. Figure 2 A flowchart showing the dynamic traffic state tensor and its related outputs; Figure 3 This is a chart showing the cyclical trend of traffic flow over 24 hours. Figure 4 A flowchart of the steps for generating a set of interactive tags for a region; Figure 5 A bar chart classifying the intensity of traffic interaction between regions. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Please see Figure 1This invention provides a dynamic traffic signal strategy control method based on spatiotemporal dimensions. The method includes: continuously capturing multi-dimensional feature data of traffic flow through a network of intelligent sensing nodes deployed in a road network, including vehicle trajectory coordinates, time-series indexes, instantaneous flow density, and speed distribution, while integrating external data sources such as weather conditions and road topology information; after spatiotemporal alignment and preprocessing, this data is integrated into a dynamic traffic state tensor, which serves as the basic data structure for subsequent analysis. Multi-scale spatiotemporal feature extraction is performed on the dynamic traffic state tensor, and a deep convolutional network is used to identify the implicit periodic and sudden patterns in the traffic flow, generating a visualized spatiotemporal pattern map. Based on the spatiotemporal pattern map, a graph neural network is used to model the interaction relationships between different traffic units, calculate the interaction intensity, thereby dividing the area into co-evolution zones and conflict-prone zones, and outputting a set of regional interaction labels. For the co-evolution zones in the regional interaction label set, historical traffic light phase data and real-time response delay information are integrated, and a predictive model is used to calculate the sensitivity coefficient of traffic flow fluctuations to signal timing parameters, generating a signal sensitivity analysis report. Nodes in frequently conflicting areas whose sensitivity exceeds an adaptive threshold are extracted from the signal sensitivity analysis report to construct a set of key interference nodes. This set includes attributes such as the spatial topological location and temporal activity index of the nodes. Combining the spatial distribution and temporal evolution characteristics of the key interference node set, a risk propagation model is used to assess the probability of traffic flow chain reactions, generating a traffic signal optimization scheme that includes signal parameter optimization strategies and emergency control instructions to achieve dynamic control.

[0025] Example 1: See Figure 2 In a specific implementation, the construction of the dynamic traffic state tensor involves data integration across multiple dimensions. The normalized traffic flow matrix is ​​generated by preprocessing the raw traffic flow data, which originates from a sensor network deployed along the road. These sensors collect data at fixed time intervals on the number of vehicles passing through designated cross-sections. The collected raw data is processed using a min-max normalization method to linearly transform the values ​​to the range of zero to one, eliminating the influence of dimensions to facilitate subsequent multi-source data fusion. The processing of the velocity field distribution employs spatial interpolation techniques to transform the instantaneous velocity sampling point data acquired by discretely arranged detectors into a normalized flow matrix. A continuous two-dimensional velocity field is generated, for example, by using the Kriging interpolation algorithm to estimate the velocity values ​​at unsampled locations, generating a velocity surface covering the entire road network. Environmental factor coding requires converting the structured weather data provided by the meteorological department into numerical influence coefficients. In the specific implementation, a mapping table will be established to map parameters such as visibility, precipitation intensity, and temperature to scale values ​​between zero and one. The construction of the road topology map depends on the basic road network data provided by the geographic information system. Nodes and edges in graph theory are used to represent intersections and road segments. Node attributes include latitude and longitude coordinates and type, while edge attributes include road grade, length, number of lanes, and direction.

[0026] In practice, the generation of spatiotemporal pattern maps is a multi-step computational process. Periodic pattern feature maps are generated by using signal processing techniques to extract recurring patterns from long-term traffic flow data. Fast Fourier Transform is applied to the speed or flow time series of each traffic unit to identify the frequencies corresponding to significant energy peaks. The periods corresponding to these frequencies are identified as the periodic patterns of traffic flow, such as daily or weekly cycles. Sudden pattern heatmaps are generated by monitoring the deviation of traffic flow parameters from historical baselines in real time. An anomaly detection algorithm based on a sliding window is used to calculate the Z-score value of flow or speed within the current time window. When the absolute value exceeds a preset threshold, it is determined to be a sudden event and visualized in the form of a heatmap in spatial location. The spatiotemporal correlation matrix is ​​used to quantify the degree of interdependence between different traffic units in the spatiotemporal dimension. The Pearson correlation coefficient between all pairs of traffic unit time series is calculated to form a symmetric matrix, with the rows and columns of the matrix corresponding to different spatial units and time lag values, respectively.

[0027] In practical implementation, the specific content of the regional interaction label set needs to be precisely defined. The identification of the co-evolution zone is achieved by thresholding the interaction intensity score calculated by the graph neural network. A high quantile threshold is set, and all continuous spatial regions with scores higher than the threshold are marked as co-evolution zones. A unique identifier is assigned to each region. The identification of the conflict-prone zone adopts the opposite logic. A low quantile threshold is set, and discrete nodes or regions with scores lower than the threshold are marked as conflict-prone zones. The interaction intensity weight is a continuous value, which is directly derived from the attention weight coefficient calculated by the graph attention network during message passing. This coefficient reflects the strength of the mutual influence between any two traffic units. The weight value is normalized to ensure that the sum of all weights is one. The generation of signal sensitivity analysis reports relies on the mining of historical data and the training of machine learning models. The calculation of the traffic fluctuation sensitivity coefficient requires the construction of a regression model, using historical traffic time series as input features and the change in signal light timing parameters for the corresponding time period as the prediction target. After training, the sensitivity coefficient is obtained by analyzing the model weights or calculating the partial derivatives of the input features with respect to the prediction results. The signal response delay gradient is extracted from the operation logs of the signal control system, recording the time difference between the issuance of instructions from the control center and the actual execution of actions by the intersection signal controller. These time difference data are grouped by region and their statistical distribution is calculated to obtain the delay gradient value for each region. The timing adjustment priority score is a comprehensive indicator, calculated by weighting and summing the traffic fluctuation sensitivity coefficient and the signal response delay gradient. The weight coefficients are set by traffic engineers according to actual management needs.

[0028] In practice, the construction of the key interference node set is a dynamic screening process. The spatial topological location information of the nodes is directly extracted from the road topology map, recorded as latitude and longitude coordinates in the geodetic coordinate system, and associated with specific road intersections or road segment markers. The node temporal activity index is calculated by analyzing the frequency of time points when nodes are marked as abnormal. The number of times each node is marked as abnormal by the anomaly detection algorithm within the rolling time window is counted, and this number is divided by the total number of detections within the time window to obtain a ratio value. The node sensitivity and traffic correlation factor is obtained by calculating the correlation coefficient between the node's sensitivity coefficient and its average traffic flow, which is used to determine whether there is a significant statistical correlation between changes in node sensitivity and changes in traffic flow. The final output is the generation of a traffic signal optimization scheme. The optimization node list is a text file that lists all the identifiers of the key interference nodes and their suggested adjustment types, such as timing scheme adjustments or phase sequence optimizations. The multi-dimensional risk diffusion map is a directed graph structure where nodes represent traffic units, edges represent risk propagation paths, and edge weights represent the probability of risk propagation. This graph is generated by analyzing historical event chains and simulating real-time traffic conditions. The adaptive parameter tuning sequence is a collection of timestamps and parameter key-value pairs that specifies in detail the specific values ​​of the signal control parameters that need to be adjusted to deal with specific key interference nodes at specific times, such as green light start time or phase interval duration.

[0029] In some embodiments, the normalized flow matrix of the dynamic traffic state tensor can be further subdivided into submatrices with smaller time granularities, such as dividing one hour of data into four 15-minute submatrices to capture traffic flow change characteristics at shorter time scales. The speed field distribution can include multiple statistics, including a variance field in addition to the mean, to characterize speed fluctuations. Environmental factor encoding can consider the combined effects of multiple weather elements, such as weighting visibility and precipitation intensity to form a comprehensive weather influence coefficient. The road topology map can add complex network indicators such as node degree centrality to better describe the importance of nodes in the network. The periodic pattern feature map of the spatiotemporal pattern atlas can contain the superimposed results of multiple periods, such as simultaneously displaying daily, weekly, and holiday periodic patterns. The sudden pattern heatmap can use multiple color coding schemes, such as using a gradient color system from blue to red to represent the change in anomaly intensity from low to high. The spatiotemporal correlation matrix can calculate the correlation under different time lags, thereby analyzing the propagation speed and direction of traffic flow impact. The co-evolution zone identifier of the regional interaction label set can include sub-region division, further dividing large co-regions into several co-regional subgroups. The conflict-frequent zone identifier can distinguish different types of conflicts, such as directional conflicts or flow saturation conflicts. The interaction intensity weight can be divided into outgoing weight and incoming weight, representing the influence of a node on other nodes and the degree of influence from other nodes, respectively.

[0030] It is understandable that the traffic flow fluctuation sensitivity coefficient in the signal sensitivity analysis report may need to be calculated separately for different traffic flow states. For example, it may be necessary to distinguish the sensitivity under saturated flow and free flow states. The signal response delay gradient may need to distinguish the delay under different control strategies, such as the delay difference under timed control and inductive control. The timing adjustment priority score may need to introduce a dynamic weighting mechanism to adjust the weight ratio of sensitivity and delay according to real-time traffic conditions. The spatial topology of key interference node sets may need to consider three-dimensional coordinates. For example, in cities with dense elevated roads, elevation information may need to be introduced. The node temporal activity index may need to consider the weight of different time periods. For example, anomalies during peak hours have a higher weight than anomalies during off-peak hours. The node sensitivity and traffic flow correlation factor may need to distinguish the impact of different vehicle types, such as the different impacts of truck traffic and passenger traffic on sensitivity.

[0031] It is understandable that the list of optimization nodes for traffic signal optimization schemes may need to be set with different processing priorities, such as three levels: immediate processing, planned processing, and observation and waiting. Multi-dimensional risk diffusion maps may need to be combined with real-time simulation technology to simulate the path and scope of risk diffusion. Adaptive parameter tuning sequences may need to be designed with a feedback mechanism to dynamically adjust subsequent parameter adjustment instructions based on the actual effect after the scheme is implemented.

[0032] Optionally, when constructing a dynamic traffic state tensor, if the data collection density in certain areas is low, data augmentation techniques, such as spatial correlation-based data imputation methods, can be used to improve the integrity and consistency of the tensor. When generating a spatiotemporal pattern map, if computational resources permit, more complex pattern recognition algorithms, such as variational autoencoders, can be used to learn the latent representation of traffic flow patterns. When dividing regional interaction label sets, if the distribution of interaction intensity scores exhibits multi-peak characteristics, clustering algorithms can be used to automatically determine the cooperation threshold and conflict threshold, without relying on fixed quantile rules. When calculating the signal sensitivity analysis report, if historical data is insufficient, transfer learning techniques can be used to fine-tune the model trained in other similar areas. When constructing a set of key interference nodes, if the false alarm rate of anomaly detection is high, a multi-algorithm voting mechanism can be introduced, combining the results of multiple algorithms such as isolated forest and local anomaly factor for comprehensive judgment. When generating traffic signal optimization schemes, if the computational complexity of the risk propagation model is too high, approximate algorithms or distributed computing techniques can be used to improve the generation speed of the schemes.

[0033] Example 2: In this specific implementation, deploying a high-precision sensor array and a video recognition unit is the fundamental step in data acquisition. The high-precision sensor array includes millimeter-wave radar and lidar devices. These devices output point cloud data covering the detection area at a period of 100 milliseconds. The point cloud data is processed by a target tracking algorithm to obtain vehicle trajectory coordinates. The video recognition unit uses high-definition cameras deployed on the roadside. The cameras acquire video streams at a rate of 25 frames per second. The video streams are input into a pre-trained YOLOv5 deep learning model to detect and track vehicles in real time, generating a sequence of vehicle bounding boxes with time-series indexes. The three-dimensional spatial coordinates are obtained by fusing GPS differential positioning technology and inertial measurement unit data, achieving centimeter-level positioning accuracy. Millisecond-level timestamps are obtained from a high-precision network time protocol server. The system acquires and ensures that the time synchronization error of all sensing nodes is less than ten milliseconds. The vehicle counting data is generated by induction coil detectors buried in the road surface. The induction coil detectors generate electromagnetic induction signals and count when vehicles pass by. The speed sampling value is calculated using the dual radar speed measurement method. The instantaneous speed is calculated by comparing the arrival time difference between two radar detection points. Meteorological sensors are deployed on representative road sections. The visibility sensor adopts the forward scattering measurement principle. The precipitation intensity sensor adopts a tipping bucket rain gauge. The temperature sensor adopts a platinum resistance thermometer. The connection relationship of the road geometry is obtained from the digital urban road design drawings. The slope information is calculated by the elevation model generated by the laser scanner. The resulting raw multi-source data stream is transmitted to the data aggregation platform in real time in Apache Avro serialization format.

[0034] In practical implementation, spatiotemporal alignment and missing value imputation of the original multi-source data stream are key steps in data preprocessing. Spatiotemporal alignment uses a dynamic time warping algorithm to process data sources with different sampling frequencies, aligning high-frequency vehicle trajectory data with low-frequency meteorological data in the time dimension. Missing value imputation addresses data loss caused by transient sensor malfunctions, using time series prediction methods such as the ARIMA model for interpolation filling. Converting vehicle trajectory coordinates to gridded flow density requires defining spatial grid division rules, dividing the monitoring area into regular 10-meter by 10-meter grids, and counting the number of vehicle trajectory points appearing within a five-minute time window in each grid to form a gridded flow density matrix. Fitting the velocity distribution to a continuous field function uses an inverse distance weighted interpolation algorithm, using the velocity values ​​of known detection points as observation samples to calculate the velocity at unsampled locations. Speed ​​estimates are generated, and a speed contour map covering the entire road network is created. Weather conditions are quantified into environmental impact coefficients, requiring the establishment of a quantification model. Visibility values ​​are mapped to a visibility impact coefficient ranging from zero to one. Precipitation intensity is mapped to a road friction coefficient correction value, and temperature values ​​are mapped to an asphalt pavement performance impact factor. Road topology encoding is implemented as a graph structure using the NetworkX graph computing library. Road intersections are abstracted as graph nodes, and road segments are abstracted as graph edges. The edge weights are set as functions of segment length and number of lanes. The above processing results are integrated into a dynamic traffic state tensor through tensor splicing operations. The first dimension of the tensor is the time axis, the second dimension is the row index of the spatial grid, the third dimension is the column index of the spatial grid, and the fourth dimension is the feature channel, which successively includes traffic density, speed value, environmental coefficient, and graph structure encoding.

[0035] In the specific implementation, the dynamic traffic state tensor is input into a multi-branch convolutional neural network for feature extraction. The multi-branch convolutional neural network adopts a three-parallel sub-network structure. The spatial branch uses a three-dimensional convolutional kernel that slides along the spatiotemporal dimension. The kernel size is set to 3x3x3, corresponding to the three dimensions of time, spatial rows, and spatial columns, respectively. Local traffic flow features are extracted step by step through three layers of three-dimensional convolutional operations. The temporal branch first flattens the dynamic traffic state tensor into a time series along the spatial dimension and inputs it into a two-layer bidirectional long short-term memory network. The number of hidden layer units in the bidirectional long short-term memory network is set to 256 to capture the long-term dependencies and contextual information of traffic flow. The multi-scale branch processes the dynamic traffic state tensor through a pyramid pooling layer. The pyramid pooling layer contains three pooling windows of different scales, namely 2x2, 4x4, and 8x8, to perform multi-resolution downsampling on the input feature map. Then, the feature maps of different scales are upsampled to restore the original size and concatenated. The output of the multi-scale branch is concatenated with the outputs of the spatial branch and the temporal branch in the feature dimension to form a feature map containing multi-scale spatiotemporal information.

[0036] In practical implementation, pattern recognition analysis of multi-scale feature maps adopts a combination of unsupervised and supervised learning methods. When separating periodic and burst patterns through clustering algorithms, a Gaussian mixture model is used to model the probability density of the multi-scale feature maps. Cluster centers that conform to the Gaussian distribution are identified as periodic patterns, and outliers located at the tail of the distribution are identified as burst patterns. When generating a visual heatmap, the multi-scale feature maps are projected onto a two-dimensional plane, and the Jet color map is used to map the feature values ​​to color intensities. Red represents high feature value regions, and blue represents low feature value regions. The feature vector matrix is ​​obtained by dimensionality reduction of the multi-scale feature maps through principal component analysis. Principal components with a cumulative contribution rate of more than 95% are retained as feature vectors. The spatiotemporal pattern map is finally stored in HDF5 file format, containing three datasets: periodic pattern feature maps, burst pattern heatmaps, and spatiotemporal correlation matrices. The periodic pattern feature maps are stored as multi-frame image sequences, with each frame corresponding to the pattern distribution of a time segment. The burst pattern heatmaps are stored as anomaly intensity matrices with timestamps, and the spatiotemporal correlation matrix is ​​stored as a symmetric correlation coefficient matrix.

[0037] In some embodiments, the deployment density of the high-precision sensor array can be adjusted according to the road grade. Radar sensors are deployed at 50-meter intervals on urban expressways and at 100-meter intervals on main roads. The installation location of the video recognition unit needs to take into account the field of view coverage and lighting conditions. It is preferred to install it on street light poles or dedicated monitoring poles, with a height generally maintained at six to eight meters. The acquisition of three-dimensional spatial coordinates can be combined with RTK-GPS technology to improve positioning accuracy. In areas with severe satellite signal obstruction, UWB ultra-wideband positioning technology can be used as a supplement. The synchronization of millisecond-level timestamps can be achieved through the IEEE 1588 precision time protocol, and nanosecond-level time synchronization can be established between the master clock and the slave clock. Preprocessing of raw multi-source data streams can be implemented using streaming computing frameworks. Apache Flink is used for real-time data cleaning and format conversion. Spatiotemporal alignment operations can be performed by adding a time window mechanism during stream processing. The window size of the dynamic time warping algorithm is set to 60 seconds. Missing value imputation can be performed by using multiple imputation methods to generate multiple complete datasets to improve imputation accuracy. When converting vehicle trajectory coordinates into gridded traffic density, the grid size can be adjusted according to the actual application scenario. A fine grid of 5 meters by 5 meters is used in urban center areas, while a sparse grid of 20 meters by 20 meters is used in suburban areas. When fitting the velocity distribution to a continuous field function, in addition to inverse distance weighted interpolation, Kriging interpolation can also be used to consider spatial autocorrelation. When quantifying weather conditions into environmental impact coefficients, machine learning models can be introduced, and the influence relationship between meteorological elements and traffic flow can be learned through random forest regression algorithm.

[0038] It is understandable that training a multi-branch convolutional neural network requires a large amount of labeled data. Historical traffic datasets can be used for pre-training, and an end-to-end training approach can be adopted. The loss function is a weighted sum of cross-entropy loss and mean squared error loss. The spatial branch's 3D convolutional kernels are initialized using the He normal initialization method, and the activation function is the ReLU function. Batch normalization layers are placed after the convolutional layers. The temporal branch's bidirectional long short-term memory network uses dropout layers with a dropout rate of 10% to prevent overfitting. The hidden state of the last time step is used as the representation of the entire sequence. After the pyramid pooling layers of the multi-scale branch output feature maps of different resolutions, they are upsampled and restored to their original size through transposed convolution operations to ensure that the spatial size of the feature maps is consistent before the stitching operation can be performed. In pattern recognition analysis, Gaussian mixture models require specifying the number of clusters. The optimal number of clusters can be automatically determined using the Bayesian information criterion. Identification of periodic patterns requires combining time periodicity testing, and Fourier transform of the clustering results to verify whether they have significant periodic components. Generating heatmaps for bursty patterns requires setting an anomaly threshold, typically using the 95th percentile of the eigenvalue distribution as the anomaly judgment criterion. The generation of eigenvector matrices requires ensuring the orthogonality of eigenvectors, which is processed using the Schmitt orthogonalization method to process the eigenvectors obtained from principal component analysis. The storage format of spatiotemporal pattern maps needs to support fast retrieval. A hierarchical index structure can be established in the HDF5 file to query the spatial pattern distribution of a specific time period by time range.

[0039] Optionally, in the sensor data acquisition stage, if fixed sensors cannot be deployed in certain areas, mobile sensing devices can be used as a supplement, such as installing GPS track recorders on buses to collect floating car data. In the data preprocessing stage, if the data quality of some sensors remains poor, the Kalman filter algorithm can be used to smooth the data, improving its reliability and stability. In the design of multi-branch convolutional neural networks, if computational resources are limited, depthwise separable convolution can be used instead of standard 3D convolution to reduce the number of model parameters and computational load. In the pattern recognition and analysis stage, if the traffic flow pattern is complex, hierarchical clustering algorithms can be used instead of Gaussian mixture models, performing coarse-grained clustering first and then fine-grained clustering. In the spatiotemporal pattern map generation process, for applications requiring high real-time performance, incremental principal component analysis can be used to update the feature vector matrix, avoiding recalculating all feature vectors each time. In multi-scale feature mapping processing, if the feature dimension is too high, leading to computational difficulties, an autoencoder can be used for feature dimensionality reduction, extracting the most representative low-dimensional feature representations.

[0040] See Figure 3This graph, a core achievement in the spatiotemporal pattern mapping phase, focuses on the periodic patterns of traffic flow over time. The horizontal axis covers the entire 24-hour period in hourly units, while the vertical axis represents traffic flow. The blue broken line and filled areas visually represent the fluctuations in traffic flow over time. This graph extracts multi-scale spatiotemporal features from the dynamic traffic state tensor, capturing long-term dependencies in traffic flow through the temporal branch of a deep convolutional network, and then combining this with signal processing techniques to identify periodic patterns. This mining of temporal patterns provides core temporal evidence for subsequent graph neural network analysis of regional interactions, signal sensitivity calculation, and key node identification. The identification of periodic patterns also lays a data foundation for predicting congestion propagation and developing proactive control strategies, ultimately helping traffic lights achieve precise temporal responses.

[0041] Example 3: See Figure 4 In a specific implementation, modeling traffic units in the spatiotemporal pattern graph as graph nodes is the fundamental step in constructing the spatiotemporal interaction graph. A traffic unit is defined as an intersection or the center point of a road segment in the road network. Each traffic unit extracts a feature vector from the spatiotemporal pattern graph as a node feature. The feature vector contains multiple dimensions such as the intensity of periodic patterns, the frequency of sudden patterns, and the spatiotemporal correlation coefficient. The traffic flow correlation between nodes is used as the edge weight, which is determined by calculating the mutual information value of the flow sequences of two traffic units in the same time period. The higher the mutual information value, the stronger the traffic flow correlation between the two traffic units. The spatiotemporal interaction graph is stored using an adjacency list data structure. Each node records its list of neighboring nodes and the corresponding edge weight value. The size of the graph is determined by the number of traffic units in the actual road network. For a large road network containing hundreds or thousands of traffic units, the spatiotemporal interaction graph may contain tens of thousands of edges.

[0042] In practical implementation, calculating the influence weights of each node on its neighbors using a graph attention network involves the specific implementation of a multi-head attention mechanism. The input to the graph attention network is the feature matrix of all nodes in the spatiotemporal interaction graph. Each row of the feature matrix corresponds to the feature vector of a node. The graph attention network first projects the feature vector of each node into a new feature space through a learnable linear transformation matrix. The projected feature vector is used to calculate the attention score. For each node pair, the attention score is calculated using an additive attention mechanism, with the specific formula as follows: in: Represents a node For nodes Importance score It is a shared linear transformation weight matrix. and These are nodes and nodes The original feature vector, It is the parameter vector of the attention mechanism. This represents a vector concatenation operation. It is a rectified linear unit activation function with a negative slope. After calculating the attention scores of all neighboring nodes, it is normalized using the softmax function to obtain the attention weight of each neighboring node. ,in Represents a node The set of neighboring nodes is used to update the current node representation by weighted summation of the neighboring node features. ,in It is a non-linear activation function such as the ELU function.

[0043] In practice, aggregating global information through a message passing mechanism requires multiple iterations. In the first iteration, each node receives information from its direct neighbor nodes and updates its own representation. In the second iteration, nodes can indirectly receive information from their neighbors' neighbors, and so on. After K iterations, each node can aggregate global information within a range of K hop neighbors. The specific process of message passing is as follows: In each iteration, each node sends its updated feature representation to all neighbor nodes and receives feature representations from all neighbor nodes. Then, it aggregates the received feature representations with its own features. Graph attention networks typically set two to three rounds of message passing iterations. Too many iterations may lead to oversmoothing, i.e., the feature representations of all nodes tend to be the same.

[0044] In practice, outputting the interaction strength score for each node requires mapping the updated node features to scalar values. The interaction strength score is calculated through a fully connected layer. The input of the fully connected layer is the final feature representation of the node after multiple rounds of message passing, and the output is a single scalar value. The weights of the fully connected layer are learned through the training process. There is no limit to the range of the interaction strength score. The higher the score, the greater the influence of the node in the entire spatiotemporal interaction graph. After the interaction strength score is calculated, the scores of all nodes are combined into a score vector, which is used for subsequent region partitioning.

[0045] In practical implementation, dividing the co-evolution zone and the conflict-prone zone based on the distribution of interaction strength scores requires determining appropriate thresholds. The co-evolution threshold is determined by calculating the upper quartile of the interaction strength scores of all nodes, i.e., the nodes with the highest scores in the top 25% are considered high-scoring nodes. The conflict threshold is determined by calculating the lower quartile of the interaction strength scores of all nodes, i.e., the nodes with the lowest scores in the bottom 25% are considered low-scoring nodes. Continuous regions with scores higher than the co-evolution threshold are marked as co-evolution zones. The definition of a continuous region is based on spatial connectivity analysis, which divides spatially adjacent nodes with scores higher than the co-evolution threshold into the same region. Nodes with scores lower than the conflict threshold are marked as conflict-prone zones. Conflict-prone zones can be discrete sets of nodes and do not necessarily require spatial continuity. Recording the boundary coordinates and intensity weights of each region is a crucial step in generating the region interaction label set. The region boundary coordinates are obtained by calculating the minimum convex hull containing all nodes within the region. The convex hull algorithm uses either the Graham scan method or the Jarvis step method to connect the outermost points of the node set to form a polygon boundary. The intensity weights are divided into intra-region weights and inter-region weights. Intra-region weights are calculated as the average of the interaction intensity scores of all nodes within the region, while inter-region weights are calculated as the average of the weights of all edges between the two regions. The region interaction label set is ultimately stored in JSON format, containing a unique identifier for each region, the region type (co-evolutionary region or conflict-prone region), a list of boundary coordinates, intensity weight values, and a list of included nodes.

[0046] In some embodiments, graph attention networks can be configured with multiple attention heads to improve model expressiveness. Each attention head independently calculates a set of attention weights, and then the outputs of multiple attention heads are concatenated or averaged. The number of attention heads is typically set to four to eight; too many attention heads will increase computational complexity but may bring performance improvements. Each attention head focuses on different aspects of the relationship between nodes; some attention heads may focus on the similarity of traffic patterns, while others may focus on the proximity of spatial distances. The message passing mechanism can consider the directionality of edges to construct a directed graph. If traffic flow has obvious directional characteristics, the spatiotemporal interaction graph can be defined as a directed graph. The edge weight from node i to node j represents the traffic flow intensity from traffic unit i to traffic unit j. In a directed graph attention network, message passing can only occur along the direction of the edge. Node i can only receive information from its neighboring nodes that point to it, and cannot receive information from the nodes it points to. This directed message passing is more consistent with the actual directional characteristics of traffic flow. More complex clustering algorithms can be used to replace simple thresholding methods for region partitioning. Spectral clustering can divide spatiotemporal interaction graphs into multiple communities, with tightly connected nodes within a community and sparsely connected nodes between communities. Spectral clustering first constructs the Laplacian matrix of the graph, then performs eigenvalue decomposition on the Laplacian matrix, and finally performs K-means clustering on the eigenvectors. Spectral clustering can discover community structures of arbitrary shapes and is not limited by the continuity of regions.

[0047] It's understandable that training graph attention networks requires a large amount of labeled data, but obtaining node-level interaction strength annotations is very difficult. Unsupervised learning methods such as graph autoencoders can be used for pre-training. Graph autoencoders map node features to a low-dimensional space through an encoder, and then reconstruct the graph's adjacency matrix through a decoder. Meaningful node representations are learned by minimizing the reconstruction error. The pre-trained graph attention network can then be fine-tuned on a small amount of labeled data. The calculation of interaction strength scores may be affected by the scale of node features. Before inputting the graph attention network, node features need to be standardized by subtracting the mean from the value of each feature dimension and dividing by the standard deviation, ensuring that all feature dimensions have zero mean and unit variance. Feature standardization can accelerate model convergence and improve training stability.

[0048] Optionally, if the spatiotemporal interaction graph is very large, containing tens of thousands of nodes, the computation of the complete graph attention network may be insufficient in memory. A sampling method can be used to reduce computational complexity. The GraphSAGE algorithm is trained by sampling a fixed number of neighbors for each node, which significantly reduces memory consumption while maintaining performance. Neighbor sampling can be uniform sampling or importance-based sampling. Importance sampling prioritizes sampling neighbors with larger weights based on the edge weights.

[0049] See Figure 5This graph is the core visualization output of the regional interaction label set generation stage based on graph neural network analysis, focusing on the quantification and classification of interaction intensity between different traffic areas. The horizontal axis represents the 12 traffic area numbers, and the vertical axis represents the interaction intensity score, with red, orange, and green corresponding to high, medium, and low intensity interaction levels, respectively. This graph models traffic units in the spatiotemporal pattern map as graph nodes, constructs a spatiotemporal interaction graph using traffic correlation as edge weights, calculates node interaction intensity scores through the message passing mechanism of a graph attention network, and finally completes regional classification based on the quantile threshold of the score distribution. This graph provides a direct basis for regional division for subsequent signal sensitivity analysis and key interference node identification, serving as a crucial bridge from pattern recognition to strategy formulation. It helps traffic engineers accurately identify the spatial distribution of cooperation and conflict in the road network, laying the decision-making foundation for the spatial interaction dimension of differentiated traffic light control strategies.

[0050] Example 4: In a specific implementation, deriving the phase duration, cycle history, and real-time operation delay data of traffic lights within the co-evolution zone from the traffic signal control system is fundamental to data analysis. Traffic signal control systems are typically equipped with standard data interfaces, such as the NTCIP standard interface based on the TCP / IP protocol. Structured data is read from the intersection signal controllers by calling functions such as GetPhaseTiming, GetCycleHistory, and GetOperationDelay. The phase duration data includes the green light time, yellow light time, and all-red time for each phase, stored as a time series in seconds. The cycle history data stores the... The change log for traffic light timing schemes includes scheme number, activation timestamp, and scheme parameters. Real-time operation delay data records the time difference between the control center issuing instructions and the actual execution of actions by the traffic signal controller at the intersection. After the data is exported, it needs to be aligned with the traffic flow time series of the corresponding area. The alignment operation is based on a unified time axis, using the second-level timestamp as the association key, and using an inner join method to match the traffic light data with the traffic flow time series extracted from the dynamic traffic state tensor. This ensures that the signal status at each time point corresponds one-to-one with the traffic flow status. The matched dataset is stored in a time series database, and each record contains fields such as timestamp, phase duration, cycle identifier, operation delay, and traffic value.

[0051] In practical implementation, calculating the sensitivity coefficient of traffic flow fluctuations to signal timing using a sequence-to-sequence prediction model requires constructing an encoder-decoder architecture. The encoder part is implemented using a multi-layer long short-term memory network. The input of the long short-term memory network is a standardized traffic flow fluctuation sequence within a past time window. The traffic flow fluctuation sequence is the difference sequence between the current traffic flow and the historical average traffic flow. The encoder's long short-term memory network encodes the variable-length input sequence into a fixed-dimensional context vector, capturing the temporal features of the traffic flow sequence. The decoder part also uses a multi-layer long short-term memory network. The initial state of the decoder's long short-term memory network is set to the context vector output by the encoder. The decoder progressively predicts the signal timing changes at multiple future time steps. The signal timing changes are relative to the current timing parameters. When training the sequence-to-sequence prediction model, a teacher-forced strategy is adopted. That is, during training, the actual signal timing changes are used as the input of the decoder, while during inference, the output of the decoder in the previous step is used as the input of the current step. After the model is trained, the gradient of the traffic flow fluctuation sequence input to the predicted output of the signal timing changes is calculated using automatic differentiation technology. The magnitude of the gradient value reflects the sensitivity of traffic flow fluctuations to signal timing, and this gradient value is the sensitivity coefficient.

[0052] In practical implementation, integrating the sensitivity coefficients of all co-evolutionary zones requires establishing a unified data structure. A data record is created for each co-evolutionary zone, containing a zone identifier, a sensitivity coefficient vector, a real-time response delay value, and a calculation timestamp. The sensitivity coefficient vector is a multi-dimensional vector, with each dimension corresponding to the sensitivity of a signal timing parameter (such as green light duration, cycle length, etc.). The real-time response delay value is extracted from the monitoring logs of the signal control system and calculated as the average of the operation delay over a recent period. The priority adjustment calculation adopts a weighted scoring method, scoring each co-evolutionary zone based on the sensitivity coefficient and the real-time response delay. The weight of the sensitivity coefficient is set to 0.7, and the weight of the real-time response delay is set to 0.3. The scoring formula is: Priority Score = 0.7 × Normalized Sensitivity Coefficient + 0.3 × Normalized Real-Time Response Delay. The normalization process uses the min-max normalization method, mapping the original value to the [0,1] interval. The signal sensitivity analysis report is generated in both PDF and JSON formats. The PDF format is easy for human reading and includes text descriptions, tables, and trend charts, while the JSON format is easy for the system to parse automatically and contains structured data fields.

[0053] In practical implementation, the specific configuration parameters of the sequence-to-sequence prediction model have a significant impact on model performance. The number of layers in both the encoder and decoder Long Short-Term Memory networks is set to three, the number of hidden layer units is set to 128, the input sequence length is set to 12 time steps (each time step is 5 minutes), and the output sequence length is set to 6 time steps. The model training uses the Adam optimizer, with an initial learning rate of 0.001, a training batch size of 32, and 100 training epochs. Training is terminated early when the validation set loss no longer decreases for 10 consecutive epochs. Gradient calculation adopts the backpropagation mode and is implemented through the automatic differentiation function of the deep learning framework. The sensitivity coefficient is taken as the Frobenius norm of the gradient matrix to obtain a comprehensive sensitivity scalar value.

[0054] In practice, referring to Table 1, the signal sensitivity analysis report includes multiple analytical dimensions. The beginning of the report is an executive summary, briefly describing the time range of the analysis, the number of co-evolutionary zones involved, and the main conclusions. Then comes the detailed data analysis section, including sensitivity coefficient distribution statistics, response delay trend analysis, and priority ranking list. The last part of the report provides adjustment recommendations, which are determined based on priority scores. For high-priority areas, it is recommended to adjust signal parameters immediately; for medium-priority areas, it is recommended to closely monitor them; and for low-priority areas, it is recommended to maintain the status quo. After the report is generated, it is automatically uploaded to the traffic signal management platform and sent to relevant management personnel via email.

[0055] Table 1: Data Structure of Signal Sensitivity Analysis Report In some embodiments, the attention mechanism of the sequence-to-sequence prediction model can enhance the processing capability of long sequences. The attention mechanism allows the decoder to focus on different parts of the input sequence when generating each output, improving the model's ability to capture long-distance dependencies. The additive attention mechanism calculates the association weights between the hidden states of the encoder at all time steps and the current state of the decoder, and the weighted summation of the encoder's hidden states is used as the context vector. The calculation of attention weights is implemented using a single-layer feedforward network, and the output dimension is the same as the encoder's hidden state dimension.

[0056] It is understandable that the calculation of the sensitivity coefficient may be affected by the magnitude of traffic fluctuations. Large traffic fluctuations may lead to gradient explosion. Before training the sequence to the sequence prediction model, the traffic fluctuation sequence needs to be standardized by dividing the traffic fluctuation value by the standard deviation of the historical fluctuation range to make the input data distribution more stable. When calculating the gradient, gradient clipping techniques can be used to set an upper limit threshold for the gradient norm to prevent the gradient value from being too large and affecting the rationality of the sensitivity coefficient.

[0057] Optionally, if the amount of training data is insufficient, a transfer learning strategy can be adopted. First, a sequence-to-sequence prediction model is pre-trained on a large-scale synthetic traffic flow dataset, and then fine-tuned on a smaller, actually collected dataset. Synthetic traffic flow data is generated using microscopic traffic simulation software to simulate traffic flow fluctuations and signal timing changes under various traffic scenarios. The pre-trained model only requires a small amount of real-world data to achieve good performance. The generation frequency of signal sensitivity analysis reports can be adjusted according to actual needs. During peak hours, reports can be generated every 15 minutes, and during off-peak hours, reports can be generated every hour. The report generation frequency setting needs to consider the balance between computational resource consumption and actual management needs. Too high a report generation frequency may increase the system burden, while too low a frequency may fail to reflect changes in traffic conditions in a timely manner.

[0058] Example 5: In a specific implementation, analyzing the sensitivity coefficient distribution in the signal sensitivity analysis report requires processing data from multiple co-evolutionary regions. The signal sensitivity analysis report is stored in JSON format, containing fields such as the identifier, sensitivity coefficient value, and calculation timestamp for each co-evolutionary region. The sensitivity coefficient distribution analysis first arranges the sensitivity coefficients of all co-evolutionary regions in chronological order to form a time series dataset. The rolling average sensitivity threshold is dynamically calculated using a time window sliding method, with the time window size set to 60 minutes, containing 12 sampling points at 5-minute intervals. The rolling average sensitivity threshold is calculated as the arithmetic mean of all sensitivity coefficients within the time window plus a standard deviation. The standard deviation reflects the degree of fluctuation of the sensitivity coefficients. The filtering condition requires that the sensitivity coefficient of a node exceeds the rolling average sensitivity threshold and that the node is marked as a conflict-prone region in the regional interaction label set. The list of conflict-prone region nodes is read from the JSON file of the regional interaction label set, containing the spatial identifier and type attribute of the nodes. The filtering operation is implemented through database query statements, such as the WHERE clause in SQL, which simultaneously satisfies both the sensitivity coefficient being greater than the threshold and the region type being a conflict-prone region.

[0059] In practice, extracting complete time-series data from nodes with sensitivity coefficients exceeding the threshold and located in conflict-prone areas requires accessing multiple data sources. Peak traffic data is extracted from the dynamic traffic state tensor. By querying the historical traffic records of a specified node within a time window, the maximum value is taken as the peak traffic flow. The speed variation rate is calculated as the ratio of the standard deviation to the mean. First, the standard deviation of all speed samples of a node within the time window is calculated, and then divided by the speed mean to obtain the dimensionless coefficient of variation. Traffic light status change records are exported from the operation log of the signal control system, including the signal phase switching timestamp, phase type, and duration. Node behavior archives are stored in the form of structured data tables. Each record contains fields such as node identifier, timestamp, peak traffic flow, speed variation rate, and traffic light status code. The archive data is arranged in chronological order to form a complete time-series record. Verifying the stability of node behavior through anomaly detection algorithms involves the specific implementation of the Isolation Forest algorithm. The Isolation Forest algorithm constructs a forest of multiple random trees. Each random tree isolates data points by randomly selecting features and split values. The node time series data first needs preprocessing, converting multivariate time series such as traffic peaks, speed variation rates, and traffic light status changes into feature vectors. These feature vectors contain statistical features such as mean, variance, maximum, and minimum values, as well as temporal features such as autocorrelation coefficients and trend slopes. The Isolation Forest algorithm assigns anomaly scores to these feature vectors. The anomaly score is based on the path length required for data point isolation; a shorter path length indicates easier isolation and a higher degree of anomaly. The anomaly score is normalized to between 0 and 1; a score closer to 1 indicates a greater likelihood of an anomaly. The adaptive anomaly threshold is determined by calculating the 95th quantile of the anomaly score distribution. Moments where the score exceeds the adaptive anomaly threshold are marked as anomaly moments. Counting consecutive anomalies requires defining temporal continuity. A consecutive anomaly is defined as an anomaly score exceeding the threshold for three or more consecutive time steps within a sliding time window. Nodes meeting the consecutive anomaly condition are marked as critical interference nodes.

[0060] In practice, constructing a set of key interference nodes requires integrating multiple verification results. The set of key interference nodes is stored in list form. Each node record includes a node spatial identifier, the first anomaly timestamp, the anomaly duration, the anomaly type classification, and the impact assessment. The node spatial identifier uses a unique code in the road network, corresponding to the spatial coordinates in the geographic information system. The anomaly type classification is based on the anomaly performance characteristics, such as traffic anomaly, speed anomaly, or mixed anomaly. The impact assessment is based on a weighted score calculated based on the anomaly duration and anomaly score value. The set of key interference nodes is updated regularly, with the update frequency consistent with the signal sensitivity analysis report generation frequency, typically set to update once every 15 minutes. The update operations include adding newly identified key interference nodes, removing nodes that have recovered to normal, and updating the anomaly status of existing nodes.

[0061] In practical implementation, consider a specific example: Suppose a node Node_2057 located at an intersection of a main urban road has a sensitivity coefficient of 0.88 in the signal sensitivity analysis report, while the rolling average sensitivity threshold for the current time window is 0.72. This node is also marked as a conflict-prone area in the regional interaction label set, meeting the initial screening criteria. Extracting the complete time-series data of Node_2057 over the past 60 minutes shows a peak traffic flow of 1200 vehicles per hour, a speed variation rate of 0.35, and 8 phase switches in the traffic light status during the monitoring period. Inputting this time-series data into the Isolation Forest algorithm for anomaly detection, the anomaly score calculation results show that the anomaly score exceeds 0.92 for five consecutive time steps between 43 and 47 minutes. Since the adaptive anomaly threshold is set to 0.90, this period is marked as an anomaly period. Because the number of consecutive anomalies reaches 5, exceeding the set threshold of 3, Node_2057 is officially marked as a critical interference node and added to the critical interference node set.

[0062] In some embodiments, the rolling average sensitivity threshold can be calculated using a weighted moving average method, giving higher weight to recent data. The exponentially weighted moving average coefficient is set to 0.3, allowing the sensitivity coefficient at the most recent time point to have a greater impact on the threshold calculation. The weighted moving average can reflect the recent trend of the sensitivity coefficient more quickly, improving the threshold's response speed to changes in traffic conditions. Node behavior profiles can include more dimensional feature indicators. Features such as queue length, lane occupancy rate, and vehicle type ratio can enrich the description of node behavior. These additional features are extracted from video detector or radar sensor data, standardized, and added to the feature vector. Multidimensional features can improve the accuracy of anomaly detection and reduce false alarm rates.

[0063] It is understandable that the depth setting of the random trees in the Isolation Forest algorithm affects the sensitivity of anomaly detection. The maximum depth of the random trees is limited to 8 layers to avoid increasing the computational complexity due to excessively deep trees. The number of subtrees is set to 100 to ensure that the forest has sufficient diversity. The Isolation Forest algorithm supports incremental learning. When new time series data arrives, the existing random trees can be partially updated without retraining the entire forest. The incremental learning mechanism is suitable for streaming data processing scenarios.

[0064] Optionally, if computational resources are limited, a distance-based anomaly detection algorithm can be used instead of the isolated forest algorithm. The K-nearest neighbor algorithm determines the degree of anomaly by calculating the average distance from each data point to its K nearest neighbors; a larger distance indicates a higher probability of an anomaly. The value of K is usually set between 5 and 10. Distance-based algorithms are simple to implement, computationally efficient, and suitable for large-scale datasets. The counting of consecutive anomalies can be achieved using a sliding time window method. The time window size is set to 30 minutes, containing six 5-minute intervals. Within the time window, the number of consecutive anomaly scores exceeding the threshold is counted. The sliding time window method can prevent accidental anomalies at a single time point from being misjudged as key interference nodes. Persistent storage of the key interference node set can be implemented using a time-series database. InfluxDB time-series database is specifically designed for time-series data, supporting efficient data writing and query operations. Each node record in the key interference node set is accompanied by a timestamp label, facilitating historical queries and trend analysis by time range.

[0065] In practical implementation, when using a risk propagation model to assess the probability of a chain reaction, the set of key interference nodes is taken as the risk source, and quantitative calculations are performed based on the structural attributes of the spatiotemporal interaction graph. First, the initial risk value of the risk source is defined. : in, This represents the node sensitivity coefficient (derived from the set of key interfering nodes). This is the node time activity index (from the set of key interfering nodes). This represents the deviation between the peak node traffic and the historical average (derived from the dynamic traffic state tensor).

[0066] The probability of risk propagation in the spatial dimension Determined by the edge weights of the spatiotemporal interaction graph: in For nodes right Interaction strength weight, The spatial distance between the two nodes.

[0067] Chain reaction probability Defined as a time window Within 30 minutes, the percentage of nodes with a risk value exceeding the threshold of 0.7. It is the time variable within the window: in Through iterative updates: This represents the total set of nodes in the road network. The calculation results are directly used as input for generating traffic signal optimization schemes.

[0068] In specific implementations, the generation of traffic signal optimization schemes is directly linked to risk levels, and the schemes include two parts: signal parameter optimization strategies and emergency control instructions.

[0069] For low-risk scenarios ( Perform routine parameter optimization: for key interference nodes The green light duration adjustment amount is in For current traffic, To improve the traffic capacity of intersections; for the co-evolution zone, the phase difference is adjusted based on the spatiotemporal correlation matrix to expand the green wave bandwidth.

[0070] For high-risk scenarios ( The following emergency control instructions were generated: First, regional-level control was activated, temporarily prohibiting left turns in areas prone to conflict and opening emergency lanes; second, dynamic access control was implemented, setting variable speed limits 500m upstream of key interference nodes, with speed limits set according to... Dynamic adjustment ( The optimization scheme is implemented by: 1) determining the free-flow velocity; and 2) inserting a special priority phase to forcibly extend the green light by no more than 15 seconds for buses and emergency vehicles passing through intersections. The final optimized scheme is output as an adaptive parameter tuning sequence, including execution timestamps accurate to the second, parameter key-value pairs (e.g., {"node_id":"V_2057","phase":2,"green_time":45}), and the expected risk reduction (≥20%). After the scheme is uploaded to the traffic signal management platform, a closed-loop verification and fine-tuning are performed every 5 minutes based on the real-time transmitted dynamic traffic state tensor.

[0071] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0072] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A dynamic control method for traffic signal strategies based on spatiotemporal dimensions, characterized in that, The method includes: By continuously capturing multidimensional feature data of traffic flow through an intelligent sensing node network, including vehicle trajectory coordinates, time series index, instantaneous flow density and speed distribution, and integrating weather conditions and road topology information, a dynamic traffic state tensor is constructed. Multi-scale spatiotemporal feature extraction is performed on the dynamic traffic state tensor, and a deep convolutional network is used to identify the periodic and sudden patterns of traffic flow to generate a spatiotemporal pattern map. Based on the spatiotemporal pattern map, the interaction intensity between different traffic units is analyzed through graph neural networks, and co-evolution zones and conflict-prone zones are delineated, outputting a set of regional interaction labels. For the collaborative evolution zone in the aforementioned regional interactive tag set, historical traffic light phase data and real-time response delay are integrated to calculate the sensitivity coefficient of traffic fluctuations to signal timing, and a signal sensitivity analysis report is generated. Nodes in frequently conflicting areas whose sensitivity exceeds the adaptive threshold are extracted from the signal sensitivity analysis report to construct a set of key interference nodes; Based on the spatial distribution and temporal evolution characteristics of the key interference node set, a risk propagation model is used to assess the probability of chain reactions, and a traffic signal optimization scheme including signal parameter optimization strategies and emergency control instructions is generated.

2. The method for dynamic control of traffic signal strategies based on spatiotemporal dimensions according to claim 1, characterized in that, The dynamic traffic state tensor includes a normalized flow matrix, velocity field distribution, environmental factor encoding, and road topology map; the spatiotemporal pattern map includes a periodic pattern feature map, a sudden pattern heatmap, and a spatiotemporal correlation matrix; the regional interaction label set specifically includes co-evolution zone identifiers, conflict-prone zone identifiers, and interaction intensity weights; the signal sensitivity analysis report includes a flow fluctuation sensitivity coefficient, a signal response delay gradient, and a timing adjustment priority score; the key interference node set includes node spatial topological location, node temporal activity index, and node sensitivity and flow correlation factors; the traffic signal optimization scheme includes an optimized node list, a multi-dimensional risk diffusion map, and an adaptive parameter tuning sequence.

3. The method for dynamic control of traffic signal strategies based on spatiotemporal dimensions according to claim 1, characterized in that, The steps of continuously capturing multi-dimensional feature data of traffic flow through an intelligent sensing node network, including vehicle trajectory coordinates, time series indexes, instantaneous flow density, and speed distribution, and fusing weather conditions and road topology information to construct a dynamic traffic state tensor include: Deploy high-precision sensor arrays and video recognition units to collect the three-dimensional spatial coordinates and millisecond-level timestamps of each traffic monitoring point in real time, and simultaneously record the counting data and speed sampling values ​​of passing vehicles. At the same time, integrate visibility, precipitation intensity and temperature data from meteorological sensors, as well as the connection relationship and slope information of road geometry, to form a raw multi-source data stream. The original multi-source data stream is spatiotemporally aligned and missing value interpolated. The vehicle trajectory coordinates are converted into gridded traffic density, the speed distribution is fitted into a continuous field function, the weather conditions are quantized into environmental impact coefficients, the road topology is encoded into a graph structure, and the data is integrated into a dynamic traffic state tensor through tensor splicing operations.

4. The method for dynamic control of traffic signal strategies based on spatiotemporal dimensions according to claim 3, characterized in that, The steps of extracting multi-scale spatiotemporal features from the dynamic traffic state tensor, identifying periodic and sudden patterns of traffic flow using a deep convolutional network, and generating a spatiotemporal pattern map include: The dynamic traffic state tensor is input into a multi-branch convolutional neural network, where the spatial branch uses a three-dimensional convolutional kernel to extract local traffic flow features, the temporal branch uses a long short-term memory network to capture long-term dependencies, and the multi-scale branch fuses spatiotemporal information of different granularities through a pyramid pooling layer to output a multi-scale feature map. Pattern recognition analysis is performed on the multi-scale feature maps, and periodic patterns and burst patterns are separated by clustering algorithms. Visual heatmaps and feature vector matrices are generated and combined to form a spatiotemporal pattern map.

5. The method for dynamic control of traffic signal strategies based on spatiotemporal dimensions according to claim 4, characterized in that, The steps of analyzing the interaction intensity between different traffic units based on the spatiotemporal pattern map, dividing the area into co-evolution zones and conflict-prone zones, and outputting a set of regional interaction labels include: The traffic units in the spatiotemporal pattern graph are modeled as graph nodes, and the traffic flow correlation between nodes is used as the edge weight to construct a spatiotemporal interaction graph. The graph attention network is used to calculate the influence weight of each node on its neighbors. Global information is aggregated through a message passing mechanism to output the interaction strength score of each node. Based on the distribution of interaction intensity scores, regions with scores above the collaboration threshold are marked as collaborative evolution zones, and regions with scores below the conflict threshold are marked as conflict-prone zones. The boundary coordinates and intensity weights of each region are recorded to generate a set of regional interaction labels.

6. The method for dynamic control of traffic signal strategies based on spatiotemporal dimensions according to claim 5, characterized in that, The steps for integrating historical traffic light phase data and real-time response delay into the co-evolution zone of the regional interactive tag set, calculating the sensitivity coefficient of traffic fluctuations to signal timing, and generating a signal sensitivity analysis report include: The phase duration, cycle history, and real-time operation delay data of traffic lights in the co-evolution zone are exported from the traffic signal control system and aligned with the traffic flow time series of the corresponding area. Using a sequence-to-sequence prediction model, with the flow fluctuation sequence as input and the signal timing change as output, the model is trained to predict the sensitivity of flow changes to signal parameters, and the sensitivity coefficient is calculated as the gradient value. The sensitivity coefficients of all co-evolutionary regions are integrated, and the priority is adjusted by combining real-time response delay calculation to generate a signal sensitivity analysis report.

7. The method for dynamic control of traffic signal strategies based on spatiotemporal dimension according to claim 6, characterized in that, The step of extracting nodes in frequently conflicting areas whose sensitivity exceeds an adaptive threshold from the signal sensitivity analysis report and constructing a set of key interference nodes includes: Analyze the distribution of sensitivity coefficients in the signal sensitivity analysis report, dynamically calculate the rolling average sensitivity threshold, and screen out nodes whose sensitivity coefficients exceed the threshold and are located in conflict-prone areas; Extract complete time-series data of nodes with sensitivity coefficients exceeding the threshold and located in conflict-prone areas, including traffic peaks, speed variation rates, and traffic light status change records, to form node behavior profiles; The stability of node behavior is verified by anomaly detection algorithm. Nodes that continuously exhibit abnormal fluctuations are marked as key interference nodes and compiled into a set of key interference nodes.

8. The method for dynamic control of traffic signal strategies based on spatiotemporal dimension according to claim 5, characterized in that, The method of using graph attention network to calculate the influence weight of each node on its neighbors includes: using a multi-head attention mechanism to calculate the similarity between node features, normalizing the attention weights through the softmax function, and weighted summing of the neighbor node features to update the current node representation.

9. The method for dynamic control of traffic signal strategies based on spatiotemporal dimension according to claim 6, characterized in that, The sensitivity coefficient of traffic fluctuations to signal timing is calculated using a sequence-to-sequence prediction model. This involves an encoder-decoder architecture where the encoder encodes the traffic sequence into a hidden state, the decoder predicts changes in signal timing based on the hidden state, and the gradient of the traffic input with respect to the predicted output is calculated as the sensitivity coefficient through backpropagation.

10. The method for dynamic control of traffic signal strategies based on spatiotemporal dimension according to claim 7, characterized in that, The verification of node behavior stability through anomaly detection algorithms includes: applying the isolated forest algorithm to score the node time series for anomalies, marking moments with scores exceeding an adaptive anomaly threshold as anomalies, and counting the number of consecutive anomalies to confirm the effectiveness of key interference nodes.

Images