Traffic prediction method and system based on regional signal optimization of traffic big model

Abstract

The application relates to the technical field of intelligent traffic, and discloses a traffic flow prediction method and system based on regional signal optimization of a traffic big model, the method comprising the following steps: acquiring road network structure data and vehicle flow density distribution data; performing correlation analysis and feature fusion according to the road network structure data and the vehicle flow density distribution data, obtaining a regional traffic feature vector, and predicting vehicle flow based on the regional traffic feature vector to obtain a vehicle flow prediction value; evaluating a crossroad congestion trend according to the vehicle flow prediction value, and performing spatial dependence and network topology characteristic fusion processing in combination with the regional traffic feature vector to obtain an initial parameter set subjected to spatial constraints; and performing optimization solution calculation on the initial parameter set subjected to spatial constraints to obtain an individualized signal control strategy. The method can reduce the optimization range by introducing physical topology constraints, prevent congestion overflow, and significantly improve the response speed of signal control.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, and in particular to a traffic flow prediction method and system based on regional signal optimization of a large traffic model. Background Technology

[0002] Currently, in the field of intelligent traffic control, in order to alleviate the increasingly serious urban congestion, traffic flow prediction data is usually used in conjunction with signal control algorithms to dynamically adjust the timing of traffic lights at intersections in order to improve regional traffic efficiency.

[0003] In existing technologies, heuristic search-based online optimization algorithms are commonly used to optimize signal parameters based on predicted traffic flow data. These methods typically define a broad, fixed parameter space (e.g., specifying green light durations between 20 and 90 seconds) and perform random searches or iterations within this space to find a timing scheme that minimizes total delay time or maximizes traffic volume. Traffic flow density distribution data is dynamic information collected in real-time by sensor arrays deployed in urban transportation networks. Sensor types can include inductive loop detectors, microwave radar, video analytics-based traffic cameras, or satellite remote sensing data. These sensors report data such as traffic flow, average speed, and lane occupancy at a fixed frequency (e.g., per minute) at the detection sections. As the number of road network nodes and control phases increases within a region, the combinations of signal parameters explode exponentially. Blindly searching within a broad, unconstrained space leads to excessively long algorithm convergence times, making it difficult to meet the millisecond-level real-time response requirements of traffic control. Furthermore, existing algorithms often only aim to maximize traffic flow, ignoring the hard limitations of the physical storage capacity of downstream road segments (i.e., network topology characteristics). This can easily lead to the generation of invalid strategies that, while theoretically having high throughput, actually cause vehicles to queue and overflow to upstream intersections due to insufficient downstream space, resulting in deadlocks.

[0004] Existing technologies suffer from the problem of insufficient real-time performance due to an excessively large search space. Summary of the Invention

[0005] This invention provides a traffic flow prediction method and system based on regional signal optimization of a large traffic model, in order to solve the technical problem of insufficient real-time performance caused by excessively large search space in the prior art.

[0006] Firstly, to address the aforementioned technical problems, this invention provides a traffic flow prediction method based on regional signal optimization using a large traffic model, comprising: Obtain road network structure data and traffic flow density distribution data, and perform correlation analysis and feature fusion on the road network structure data and the traffic flow density distribution data to obtain regional traffic feature vectors; Based on the regional traffic feature vector, vehicle flow prediction is performed to obtain the predicted vehicle flow value. Based on the predicted vehicle flow, the intersection congestion trend is assessed to determine the optimization adjustment range. The spatial dependence characteristics and network topology characteristics of the optimization adjustment range and the regional traffic feature vector are fused to obtain the prediction-based search boundary conditions. Based on the predicted search boundary conditions, dynamic traffic flow change characteristics are integrated to obtain a spatially constrained set of initial parameters. The initial parameter set subject to spatial constraints is optimized and calculated to obtain a personalized signal control strategy.

[0007] In one optional implementation, the step of acquiring road network structure data and traffic flow density distribution data, and performing correlation analysis and feature fusion on the road network structure data and the traffic flow density distribution data to obtain a regional traffic feature vector, includes: Acquire road network structure data and traffic flow density distribution data; The road network structure data and the traffic flow density distribution data are cleaned and preliminarily integrated to obtain a regional description dataset; The correlation analysis between intersection nodes is performed on the aforementioned regional description dataset to obtain preliminary features; The preliminary features are compared with data and the weight parameters are adjusted to obtain the regional traffic feature vector.

[0008] In one optional implementation, the step of predicting vehicle flow based on the regional traffic feature vector to obtain a predicted vehicle flow value includes: Get real-time traffic information; Scene recognition and pattern analysis are performed on the traffic feature vector of the area to obtain the vehicle flow change pattern; Based on the aforementioned vehicle traffic flow variation patterns, a time-series correlation analysis was performed to obtain a preliminary predicted trend; Based on the real-time traffic information, the preliminary prediction trend is calibrated to obtain the vehicle flow prediction value.

[0009] In one optional implementation, the step of assessing the intersection congestion trend based on the predicted vehicle flow rate to obtain an optimized adjustment range includes: Queue length is estimated from the predicted vehicle flow to obtain intersection congestion data; If the intersection congestion data exceeds the preset congestion threshold, a preset strategy mapping is performed to obtain a preliminary parameter adjustment plan. The initial parameter adjustment plan is then subjected to critical path analysis based on road network structure data to obtain the optimized adjustment range.

[0010] In one optional implementation, the step of fusing spatial dependency characteristics and network topology characteristics of the optimized adjustment range and the regional traffic feature vector to obtain prediction-based search boundary conditions includes: The spatial dependency characteristics of the regional traffic feature vector and the optimization adjustment range are integrated to obtain enhanced optimization input data; The enhanced optimized input data is subjected to hierarchical processing of change trend information to obtain the initial constraint conditions of the search boundary; The initial constraints of the search boundary are calibrated using network topology characteristics to obtain the predicted search boundary conditions.

[0011] In one optional implementation, the step of integrating dynamic traffic flow change characteristics based on the predicted search boundary conditions to obtain a spatially constrained initial parameter set includes: The predicted search boundary conditions are processed by traffic distribution layering to obtain a preliminary parameter set framework. The initial parameter set framework is dynamically calibrated based on real-time traffic flow changes to obtain the calibrated parameter set. The calibrated parameter set is then subjected to dynamic traffic flow matching degree verification to obtain an initial parameter set subject to spatial constraints.

[0012] In one optional implementation, the step of optimizing the spatially constrained initial parameter set to obtain a personalized signal control strategy includes: The initial parameter set subject to spatial constraints is iteratively processed to generate candidate signal control parameter combinations; Traffic efficiency simulation evaluation is performed on the candidate signal control parameter combinations to obtain the evaluation score for each candidate combination; Based on the evaluation score, the optimal combination of candidate signal control parameters is selected to obtain a personalized signal control strategy.

[0013] Secondly, the present invention provides a traffic flow prediction system based on regional signal optimization using a large traffic model, comprising: The feature generation module is used to acquire road network structure data and traffic flow density distribution data, and to perform correlation analysis and feature fusion on the road network structure data and the traffic flow density distribution data to obtain regional traffic feature vectors. The traffic flow prediction module is used to predict vehicle traffic flow based on the regional traffic feature vector to obtain the predicted vehicle traffic flow value. The congestion assessment module is used to assess the intersection congestion trend based on the predicted vehicle flow rate and obtain the optimization adjustment range. The boundary definition module is used to perform spatial dependency and network topology fusion processing on the optimization adjustment range and the regional traffic feature vector to obtain prediction-based search boundary conditions. The parameter generation module is used to integrate the dynamic traffic flow change characteristics based on the prediction-based search boundary conditions to obtain a spatially constrained initial parameter set. The strategy generation module is used to optimize and calculate the initial parameter set subject to spatial constraints to obtain a personalized signal control strategy.

[0014] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention achieves dynamic pruning of the optimization solution space by combining regional traffic characteristics and predicted trends to pre-construct prediction-based search boundary conditions. Unlike existing technologies that blindly search within a broad space, this invention dynamically generates narrowed search boundary conditions based on the gradient of changes in predicted vehicle flow values. This constraint-then-search processing mechanism locks the optimization range of subsequent optimization steps such as genetic algorithms within the most likely effective interval, greatly reducing the number of invalid iterations, thereby significantly improving the convergence speed and computational efficiency of the algorithm, and effectively solving the problem of slow real-time signal control response under complex road networks.

[0015] (2) This invention effectively avoids the risk of queuing overflow in regional signal control by introducing physical constraint calibration of network topology characteristics when generating the initial parameter set. This invention not only relies on traffic flow data, but also transforms network topology data such as the physical length of downstream road segments and lane capacity into mathematical constraints to calibrate the search boundary. This ensures that the generated signal control strategy is strictly limited within the physical carrying capacity of the road (for example, limiting the green light duration to prevent downstream queuing overflow), fundamentally preventing the phenomenon of traffic flow back to the upstream intersection due to blindly pursuing traffic efficiency, and ensuring the physical smoothness of the regional road network.

[0016] (3) This invention constructs a regional traffic feature vector containing spatiotemporal transmission logic by performing correlation analysis and feature fusion on the road network structure and traffic density distribution. This invention utilizes the extracted traffic transmission rules between intersections (such as time-delay correlations mined based on the Apriori algorithm) to enable traffic prediction and subsequent congestion assessment to no longer be limited to a single isolated intersection, but to deeply capture the dynamic propagation chain of congestion within the region. This makes the final personalized signal control strategy forward-looking, capable of adjusting downstream strategies in advance according to upstream traffic changes, and improving the system's adaptive matching capability for traffic flow characteristics in different functional areas such as commercial areas and residential areas. Attached Figure Description

[0017] Figure 1This is a schematic diagram of a traffic flow prediction method based on regional signal optimization of a large traffic model provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of a traffic flow prediction system based on regional signal optimization of a large traffic model, provided in the second embodiment of the present invention. Detailed Implementation

[0018] 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.

[0019] Reference Figure 1 The first embodiment of the present invention provides a traffic flow prediction method based on regional signal optimization of a large traffic model, comprising the following steps: S11, acquire road network structure data and traffic flow density distribution data, and perform correlation analysis and feature fusion on the road network structure data and the traffic flow density distribution data to obtain regional traffic feature vectors; S12, Based on the regional traffic feature vector, perform vehicle flow prediction to obtain the vehicle flow prediction value; S13, Based on the predicted vehicle flow rate, assess the intersection congestion trend and obtain the optimization adjustment range; S14, perform spatial dependency and network topology fusion processing on the optimized adjustment range and the regional traffic feature vector to obtain the prediction-based search boundary conditions. S15, Based on the predicted search boundary conditions, the dynamic traffic flow change characteristics are integrated to obtain a set of initial parameters subject to spatial constraints. S16, perform optimization calculations on the spatially constrained initial parameter set to obtain a personalized signal control strategy.

[0020] In step S11, road network structure data and traffic flow density distribution data are acquired, and correlation analysis and feature fusion are performed on the road network structure data and the traffic flow density distribution data to obtain a regional traffic feature vector, including: Acquire road network structure data and traffic flow density distribution data; The road network structure data and the traffic flow density distribution data are cleaned and preliminarily integrated to obtain a regional description dataset; The correlation analysis between intersection nodes is performed on the aforementioned regional description dataset to obtain preliminary features; The preliminary features are compared with data and the weight parameters are adjusted to obtain the regional traffic feature vector.

[0021] It's important to note that acquiring road network structure data and traffic flow density distribution data is fundamental to constructing a regional traffic profile. Road network structure data originates from a pre-established Geographic Information System (GIS) database, which digitally stores static physical information such as road topology, number of lanes, speed limits, intersection locations, and traffic light phases. Traffic flow density distribution data, on the other hand, is dynamic information collected in real-time by sensor arrays deployed throughout the urban transportation network. Sensor types can include inductive loop detectors, microwave radar, or video analytics-based traffic cameras, which report data such as traffic volume, average speed, and lane occupancy at monitored sections at a fixed frequency (e.g., every minute).

[0022] For example, in the traffic network of a city's core business district, the system retrieved structural data from the GIS database for the intersection of "Renmin Road" and "Jiefang Road," showing four lanes running east-west and three lanes running north-south. Meanwhile, the inductive loop detectors deployed beneath each approach lane of this intersection reported traffic density data at 8:30 AM on September 22, 2025: an average lane occupancy rate of 85% for the east approach lane and 78% for the west approach lane.

[0023] It should be noted that the road network structure data and traffic flow density distribution data undergo data cleaning and preliminary integration to eliminate noise and outliers in the original data and establish a mapping relationship between dynamic traffic data and the static road network. Data cleaning employs the 3-sigma principle, identifying and removing data points that deviate from their time series mean by more than three standard deviations to handle extreme outliers caused by sensor malfunctions or communication errors. Preliminary integration involves spatially matching the cleaned traffic flow data with specific road segments or intersections in the GIS database based on their collection location (e.g., latitude and longitude or sensor ID).

[0024] For example, suppose a ground loop detector reports a traffic flow of 500 vehicles in a certain minute, while the historical average traffic flow for the same period on that road segment is only 40 vehicles, with a standard deviation of 10 vehicles. Since 500 is far more than three times the standard deviation of the mean (40 + 3 * 10 = 70), this data point will be identified as an outlier by the system and removed. After integration, the cleaned data (such as "East entrance lane occupancy rate 85%)" is accurately appended to the "East entrance lane attribute of the intersection of Renmin Road and Jiefang Road" in the GIS database, forming a high-quality dataset containing accurate spatiotemporal labels, i.e., a regional description dataset.

[0025] It should be noted that the analysis of the relationships between intersection nodes in the aforementioned regional description dataset aims to uncover hidden and regular mutual influences between different nodes in the traffic network, such as the time lag in the transmission of congestion from upstream intersections to downstream intersections. This embodiment employs the Apriori algorithm in association rule mining. This algorithm iteratively searches for frequent itemsets in massive historical traffic state data by setting two thresholds: minimum support and minimum confidence. Ultimately, it generates a series of strong association rules in the form of "If {Event A} occurs, then {Event B} is very likely to occur."

[0026] It's worth noting that the Apriori algorithm specifically includes the following steps: First, the continuous regional description dataset is discretized into a series of "traffic state transactions." Each transaction represents a snapshot of time (e.g., a five-minute window on a specific day), and the "items" in the transaction are the specific traffic events that occur at that point in time (e.g., "Northbound congestion at intersection A," "Westbound free flow at intersection C"). The algorithm relies on two core thresholds: minimum support, which defines the lower bound of the frequency of an event combination (itemset) appearing in all transactions; the higher the support of an itemset, the more prevalent the pattern in the traffic network. Minimum confidence, which defines the credibility of association rules generated from an itemset; for example, the confidence of the rule "{A}→{B}" measures the probability that a transaction containing event A also contains event B. The algorithm mainly consists of two stages: the first stage iteratively discovers frequent itemsets. This stage utilizes the Apriori principle (i.e., any subset of an itemset must also be frequent) to prune layer by layer, reducing computational cost. The algorithm first scans all transactions, calculates the support of all individual events (1-itemsets), and filters out frequent 1-itemsets that meet the minimum support. Then, it generates candidate 2-itemsets based on these frequent 1-itemsets, scans transactions again, calculates their support, and filters again. This process is iterated until no longer frequent itemsets can be found. The second stage generates strong association rules from the frequent itemsets. For each frequent itemset found in the first stage, the algorithm generates all possible binary partitions, forming candidate rules. Then, it calculates the confidence of each candidate rule and filters out those rules with a confidence level not lower than the minimum confidence level as the final strong association rules.

[0027] For example, the system analyzes a month-long dataset of regional descriptions, setting a minimum support of 0.2 and a minimum confidence of 0.8. In the first stage, the algorithm first identifies the single events "Northbound congestion at intersection A" and "Northbound congestion at intersection B" as frequent 1-itemsets. It then combines these two events into a candidate 2-itemset "{Northbound congestion at intersection A, Northbound congestion at intersection B}", and scans the dataset again. It finds that the frequency of this combination also exceeds the minimum support of 0.2, thus confirming it as a frequent 2-itemset. In the second stage, the algorithm generates candidate rules based on this frequent itemset, such as "{Northbound congestion at intersection A} → {Northbound congestion at intersection B}". Calculations show that this rule has a confidence of 0.85, higher than the minimum confidence threshold of 0.8, therefore it is output as a strong association rule. This rule indicates that when northbound congestion occurs at intersection A, there is an 85% probability that it will cause northbound congestion at the downstream intersection B within 5 minutes. These discovered association rules, along with other statistical features (such as average vehicle speed and queue length), together constitute the preliminary features.

[0028] It should be noted that the data comparison and weight parameter adjustment of the preliminary features are intended to ensure that the final generated feature vector dynamically reflects the uniqueness of the current traffic conditions. This process compares the preliminary features obtained in the previous step with benchmark features stored in the historical database for the same time period (such as a weekday morning rush hour). If the current feature deviates significantly from the historical benchmark (for example, traffic flow is much lower than the historical average due to a traffic accident), the system will adopt an error feedback-based weight update strategy to dynamically adjust the weight of that feature in the final vector to highlight its anomaly.

[0029] It should be noted that, in this embodiment, the weight update strategy based on error feedback is implemented as follows: First, the currently observed preliminary feature values ​​are... Compared with the baseline feature values ​​in the historical database Perform max-min normalization, then calculate the absolute value of the normalized deviation (i.e., error) between the two. ,in These are the normalized values ​​of the initial eigenvalues; The normalized values ​​are the baseline feature values. Secondly, a non-linear activation function, such as the sigmoid function, is used to normalize the calculated error. Mapped to a weight value between 0 and 1 The calculation formula is as follows:

[0030] in, It is a preset deviation sensitivity threshold. It is a gain coefficient. The effect of this function is that when the normalized bias... Less than the threshold At that time, weight If the value is close to a baseline (e.g., 0.5), it indicates that the system considers this to be normal fluctuation; when the error... Exceeding the threshold At that time, weight It will quickly approach 1, indicating that the system considers this important new information. Ultimately, when constructing the regional traffic feature vector, the final value of this initial feature will be its current observation value plus the calculated weights. The product of.

[0031] It is worth noting that the aforementioned deviation sensitivity threshold The threshold was determined based on statistical analysis of error data from historical normal operating conditions. Specifically, the system selected the 95th percentile of the historical error distribution as the threshold. This means that any deviation exceeding this threshold is statistically considered a significant and noteworthy anomaly with a 95% confidence level. The gain coefficient... The value controls the drastic degree of weight changes. Its value is determined through offline optimization simulations on historical datasets, aiming to find an optimal value. The value is set so that the final weighted feature vector can enable the downstream prediction model to achieve the highest prediction accuracy.

[0032] For example, preliminary features show that the average vehicle speed on a main road during Friday evening rush hour is 40 km / h, while the baseline speed for that time period in the historical database is 15 km / h. The system identifies this significant positive discrepancy through data comparison. Assuming that after normalization, the current feature value... It is 0.8, the historical benchmark. It is 0.25. First, calculate the normalization bias, i.e. Assuming that, based on statistical analysis, a preset bias sensitivity threshold is established... The gain coefficient is 0.1, based on optimization simulation. The weight is 5. The system calculates the new weight according to the formula, that is... The calculated new weight of 0.90 is very close to 1, indicating that the system is highly confident that this is a significant traffic state change. In the final generated regional traffic feature vector, the contribution value of this feature will be... This clearly demonstrates the current smooth flow of traffic on the main road.

[0033] In step S12, based on the regional traffic feature vector, vehicle flow prediction is performed to obtain the predicted vehicle flow value, including: Get real-time traffic information; Scene recognition and pattern analysis are performed on the traffic feature vector of the area to obtain the vehicle flow change pattern; Based on the aforementioned vehicle traffic flow variation patterns, a time-series correlation analysis was performed to obtain a preliminary predicted trend; Based on the real-time traffic information, the preliminary prediction trend is calibrated to obtain the vehicle flow prediction value.

[0034] It should be noted that obtaining real-time traffic information is a prerequisite for achieving dynamic calibration of predictions. This step uses the same sensor network deployed in step S11 to continuously collect traffic flow data at a higher frequency (e.g., every 10 seconds). Unlike the minute-level aggregated data used to construct macroscopic feature vectors in S11, the real-time information obtained here aims to capture instantaneous, microscopic changes in traffic flow, providing the latest and most accurate on-site evidence for subsequent corrections to the preliminary predictions based on historical patterns.

[0035] For example, at a certain intersection, the feature vector generated in S11 might reflect a high average traffic density during the morning rush hour. In S12, real-time traffic information might show that at a precise moment of 8:45:10 AM, due to a minor accident ahead, the instantaneous speed of the upstream sensor dropped sharply to 5 km / h. This information will serve as a key input for subsequent calibration steps.

[0036] It should be noted that scene recognition and pattern analysis are performed on the aforementioned regional traffic feature vectors to match the current complex traffic conditions to the most similar historical patterns in order to extract pattern templates for prediction. This process consists of two steps: scene recognition and pattern analysis. Scene recognition uses a pre-trained decision tree classifier to analyze the regional traffic feature vectors generated in S11. This decision tree model is trained on a massive amount of historical feature vectors of labeled scenarios (such as "weekday morning rush hour" and "holiday evening shopping rush hour"), and can automatically classify them into the most suitable scene labels based on the various feature values ​​of the input vector.

[0037] Pattern analysis involves retrieving and extracting a standard time series template associated with a given scene label from a historical pattern database after the system has identified the scene label. This template is a curve that quantitatively describes the typical patterns of traffic flow evolution over time in that scene (e.g., how traffic accumulates, reaches its peak, and gradually dissipates).

[0038] It is worth noting that the construction and use of the decision tree classifier includes three parts: model structure, model training, and model usage. In the model structure part, the decision tree model consists of a root node, several internal nodes, and several leaf nodes. Each internal node represents a judgment on a certain dimension of the regional traffic feature vector (e.g., "average traffic density"); each branch represents a judgment result; and each leaf node represents a final scene classification label (e.g., "weekday morning rush hour"). The path from the root node to any leaf node constitutes a complete scene determination rule.

[0039] In the model training section, the model training is a supervised learning process aimed at automatically constructing the optimal tree structure using historical data. The training data is a historical dataset containing a massive number of samples, where each sample contains a regional traffic feature vector and a corresponding scene label verified by humans. This embodiment uses the CART (Classification and Regression Tree) algorithm to generate the tree. This algorithm recursively partitions the dataset based on minimizing Gini impurity. At each node, the algorithm traverses all possible split points for all features and selects the split that minimizes the sum of Gini impurities of the child nodes as the splitting rule for that node. This recursive process continues until a stopping condition is met (e.g., the number of samples in a node is below a preset threshold, or the tree depth reaches its upper limit), thus generating a complete decision tree.

[0040] In the model usage section, during real-time scene recognition, a traffic feature vector of a region to be classified is input from the root node. At each internal node, the system judges the corresponding feature value of the vector according to the node's partitioning rules and selects a branch to enter the next layer based on the judgment result. This process continues until a leaf node is reached, at which point the system uses the scene label represented by that leaf node as the final classification result for the current region's traffic feature vector.

[0041] For example, a regional traffic feature vector is input into the root node of a trained decision tree classifier. The root node performs its first judgment: "Is the average traffic density > 0.8?" (Yes), then proceeds to the left branch. The internal nodes of the left branch continue to perform judgments: "Is the association rule confidence > 0.85?" (Yes), then proceed to the next branch. After several judgments, the vector finally reaches a leaf node labeled "Weekday morning rush hour commuting," which is the final result of scene recognition. Subsequently, the system extracts the vehicle traffic flow change pattern corresponding to this scene from the database. This pattern shows that traffic flow typically begins to rise rapidly at 7:30, reaches its peak around 8:50, and then slowly declines.

[0042] It should be noted that the time-series correlation analysis performed based on the vehicle flow change patterns is to generate a preliminary prediction of the future based on historical patterns. This embodiment uses a pre-trained Long Short-Term Memory (LSTM) model to perform this step. As a recurrent neural network, LSTM's internal gating unit structure (input gate, forget gate, output gate) enables it to effectively learn and remember long-term dependencies in time-series data, making it particularly suitable for prediction tasks like traffic flow, which exhibit significant periodicity and trends. The model's input is the vehicle flow change pattern extracted in the previous step (a historical time series), and the output is a new time series of the same length as the input series. This output series represents the system's preliminary prediction of traffic flow changes over a future period.

[0043] It is worth noting that this embodiment uses a pre-trained Long Short-Term Memory (LSTM) model to perform this step. LSTM, as a special type of recurrent neural network, is particularly suitable for processing and predicting time series data. Its model structure, in this embodiment, consists of an input layer, two stacked LSTM layers, and a fully connected output layer. Each LSTM layer contains 64 LSTM units. The core of each unit is its "cell state," used to transmit long-term memory. The flow of information in the cell state is controlled by three sophisticated "gate" structures. The forget gate determines which information is discarded from the cell state at the previous time step; the input gate determines which part of the new input information at the current time step is updated in the cell state; and the output gate determines which information is output from the currently updated cell state as the prediction result for the current time step. A fully connected (dense) layer is output to transform the output of the last LSTM unit into a final output that conforms to the prediction target (i.e., the flow values ​​at multiple future time points).

[0044] The model training is a supervised learning process designed to teach the model the mapping relationship between historical traffic flow sequences and future traffic flow sequences. Training data comes from a time-series database containing massive amounts of historical traffic flow samples. This database is processed using a sliding window technique to generate a large number of training sample pairs. Each sample pair consists of an input sequence (e.g., 12 consecutive traffic data points sampled every 5 minutes within the past 60 minutes) and a corresponding target sequence (e.g., 6 consecutive traffic data points sampled every 5 minutes within the next 30 minutes). Training aims to minimize the Mean Squared Error (MSE) loss function, using the Adam optimizer for parameter updates. The system batches the training sample pairs into the model, calculates the MSE loss between the model's predicted output and the true target sequence, and backpropagates the error based on this loss to iteratively update the weight parameters within the network. Training terminates when the model's loss on the independent validation set no longer decreases significantly to prevent overfitting.

[0045] For example, during the model training phase, the system generates tens of thousands of training samples using historical data. One sample's input might be "12 traffic readings from 8:00 AM to 8:55 AM," with a target output of "6 traffic readings from 9:00 AM to 9:25 AM." During the model usage phase, the extracted "weekday morning rush hour" pattern (a sequence containing 12 time points and their corresponding traffic values ​​from the past hour) is input into the trained LSTM model. The model uses its internal memory units to calculate and output a prediction sequence for the next 30 minutes, for example, [predicted traffic of 120 vehicles / minute at 8:50 AM, 115 vehicles / minute at 8:55 AM, ...]. This sequence containing predictions for multiple future time points represents the initial trend prediction.

[0046] It should be noted that calibrating the preliminary predicted trend based on the real-time traffic information is to combine the macro-prediction based on historical patterns with the actual, instantaneous traffic dynamics to generate a more accurate final prediction result. This embodiment uses the Kalman filter algorithm to implement this calibration process. The algorithm takes the preliminary predicted trend generated by the LSTM model as input to its "prediction" stage and the real-time traffic information obtained in the first sub-step as input to its "measurement update" stage. The Kalman filter calculates the Kalman gain by calculating the deviation (i.e., innovation) between the predicted value and the actual measured value, and combines this with the preset system noise and measurement noise covariance. Finally, it outputs an optimal estimate of the actual state, i.e., the final vehicle flow prediction value, calculated as follows:

[0047] in, The Kalman gain at the current moment; The prior error covariance at the current moment is determined by the error at the previous moment and the process noise covariance. The observation matrix; To measure the noise covariance. The essence of this formula is to calculate an optimal weight to balance the uncertainty of the predicted values ​​(reflected in...). (in the middle) and the uncertainty of the measured values ​​(reflected in) (In Chinese). The noise covariance parameter is determined based on offline statistical analysis of historical prediction errors and the sensor's own noise characteristics.

[0048] For example, at 8:45:10 AM, the initial trend forecast indicated a traffic flow of 110 vehicles per minute. However, real-time traffic information showed an actual measurement of 125 vehicles per minute, indicating a sudden congestion not captured by historical patterns. After receiving these two inputs, the Kalman filter calculates based on preset noise parameters, optimally fuses the two data points, and finally outputs a calibrated traffic flow prediction of 121 vehicles per minute. This value, calibrated with real-time data, is smoother and more reliable than simple historical predictions or instantaneous measurements.

[0049] In step S13, based on the predicted vehicle flow rate, the intersection congestion trend is assessed to obtain the optimization adjustment range, including: Queue length is estimated from the predicted vehicle flow to obtain intersection congestion data; If the intersection congestion data exceeds the preset congestion threshold, a preset strategy mapping is performed to obtain a preliminary parameter adjustment plan. The initial parameter adjustment plan is then subjected to critical path analysis based on road network structure data to obtain the optimized adjustment range.

[0050] It should be noted that estimating queue length for the predicted vehicle flow is to transform the abstract flow data (unit: vehicles / minute) into a more physically meaningful and intuitively assessable indicator (unit: meters) of congestion. This embodiment uses a deterministic queuing model based on Traffic Shockwave Theory to achieve this estimation. This model treats vehicle arrivals and departures as a continuous flow, where the vehicle arrival rate ( The vehicle service rate (i.e., the maximum departure rate during the green light period, or saturation flow rate) is determined by the predicted vehicle flow rate. The lane physical attributes in the road network structure data are determined by the lane physical attributes. This is achieved by analyzing vehicle movement at red lights (…). The accumulation during the period and the dissipation during the green light period can be used to calculate the maximum number of vehicles queuing within the cycle. ):

[0051] This formula calculates the total number of vehicles arriving and accumulating within the valid red light duration. The calculated maximum number of vehicles in the queue... Multiplying this by a standard average vehicle length (e.g., 5 meters) yields the quantified predicted queue length. This predicted queue length, along with its corresponding predicted traffic flow, constitutes the intersection congestion data.

[0052] For example, the system receives a vehicle traffic forecast of 20 vehicles per minute (i.e., arrival rate). =20). Based on the current signal timing, the effective red light duration for this approach lane is... The expected queue length is 50 seconds (5 / 6 minutes). Based on a deterministic queuing model, the expected number of vehicles queuing at this approach lane is calculated to be 17. Assuming an average vehicle length of 5 meters, the final estimated predicted queue length is 17 vehicles × 5 meters / vehicle = 85 meters. This predicted queue length of 85 meters is the core indicator of intersection congestion.

[0053] It should be noted that after obtaining intersection congestion data, the system compares it with a preset congestion threshold to determine whether adjustments need to be initiated. The preset congestion threshold is determined based on the road's physical storage capacity. Specifically, the threshold is set to 80% of the physical length of the road segment upstream of the target intersection. This setting is based on common knowledge in the field and aims to ensure that queue length does not overflow to the previous intersection, thereby preventing regional traffic congestion. If the congestion data exceeds this threshold, the system executes a "preset strategy mapping," which is a lookup operation based on a rule base. The rule base pre-stores the mapping relationship between different congestion exceedance levels and preliminary adjustment strategies.

[0054] It is worth noting that the construction and structure of the rule base includes two parts: construction method and storage and structure.

[0055] In this construction methodology, the rule base is built by combining traffic engineering knowledge with offline microscopic traffic simulation. First, based on congestion mechanisms, a series of basic congestion scenario levels (e.g., mild, moderate, and severe congestion) and corresponding preliminary intervention strategies (e.g., slightly increasing green light duration, significantly extending green light duration, and adjusting phase sequence) are defined. Then, using microscopic traffic simulation software such as VISSIM or SUMO, a simulation environment matching the real road network is constructed to simulate congestion scenarios under thousands of different traffic flow combinations. In the simulation, the actual effects of various intervention strategies on congestion relief are systematically tested and quantified. Finally, the optimal mapping relationship between "congestion severity level - preliminary adjustment strategy" is determined and solidified into rules.

[0056] In terms of storage and structure, this rule base is stored in the form of a database, such as a table named "Preliminary Strategy Mapping". Each row of this table represents a mapping rule, and its core column attributes include the following five attributes: RuleID, which is a unique identifier for each rule; MetricType, which defines the congestion metric that the rule targets, such as "queue length exceeding the limit (meters)" or "queue length exceeding the limit by a percentage (%)"; ConditionLowerBound, which is the lower limit of the range of congestion metric values ​​that trigger the rule; ConditionUpperBound, which is the upper limit of the range of congestion metric values ​​that trigger the rule; ActionType, which corresponds to the category of the preliminary adjustment action, such as "extend green light duration" or "shorten green light duration"; and ActionParameter, which is the specific value of the adjustment action, such as "10%" or "5 seconds".

[0057] For example, suppose the physical length of the upstream road segment at an intersection is 100 meters, and its preset congestion threshold is set to 80 meters. The predicted queue length estimated in the previous step is 95 meters. Because 95 meters exceeds the 80-meter threshold, the system triggers an adjustment. The system uses the "queue length exceeding value" of 95 meters as the adjustment threshold. Using 80 = 15 meters as the query condition, the system searches the "Preliminary Strategy Mapping Table." The system matches a rule whose congestion metric type is "queue length exceeding the limit (meters)," with a lower limit of 10 and an upper limit of 20. The system then reads the corresponding adjustment action type as "extend green light duration," with an adjustment parameter of "10%." This instruction to "increase green light duration by 10%" is the preliminary parameter adjustment plan obtained in this step.

[0058] It should be noted that the critical path analysis based on road network structure data is performed on the preliminary parameter adjustment scheme to ensure that the optimization adjustment of local intersections will not negatively affect the coordination of the entire traffic corridor. This analysis process abstracts the intersection network as a graph, where intersections are nodes and roads are edges. When a preliminary scheme (such as "increase the green light at intersection A by 5 seconds") is generated, the system identifies the critical traffic path (e.g., an urban arterial road) where intersection A is located. Subsequently, a micro-traffic flow simulation model is used to evaluate the potential impact of the scheme on adjacent upstream and downstream intersections along the path, such as whether it will cause downstream intersections to be overwhelmed by traffic surges, or whether it will disrupt the existing green wave coordination with upstream intersections.

[0059] It is worth noting that the construction and calibration process of the microscopic traffic flow simulation model includes two parts: model construction and model calibration and verification. In the model construction process, the simulation model in this embodiment can be built based on the open-source platform SUMO (Simulation of Urban MObility). The core of the model lies in simulating each vehicle as an independent, autonomous intelligent agent. Internally, it mainly consists of the following sub-models: a car-following model, used to define how a vehicle follows the vehicle in front. The Krauss car-following model is adopted, which comprehensively considers parameters such as the vehicle's expected speed, maximum acceleration, safe distance, and driver reaction time; a lane-changing model, used to define under what conditions a vehicle decides to change lanes, for example, to overtake or prepare for an exit; and a driver behavior model, which assigns random but statistically consistent driving behavior parameters to each simulated vehicle (e.g., small differences in expected speed, reaction time speed) to reflect the diversity of real-world drivers.

[0060] In the model calibration and validation process, rigorous calibration and validation are essential to ensure that the simulation model accurately reproduces real-world traffic phenomena. First, real-world traffic data, such as cross-sectional flow rate, average travel time, and queue length, is collected from the critical path to be analyzed, serving as the "benchmark truth" for calibration. This process aims to adjust the parameters within the simulation model (such as expected speed and reaction time in a car-following model) to minimize the error between the simulation output and real-world data. This embodiment employs a genetic algorithm for automatic parameter optimization. The system runs thousands of simulations, each using a different set of parameters, and calculates the difference between the simulation results (such as simulated travel time) and the actual travel time. The genetic algorithm iterates continuously, simulating the process of natural selection, until it finds the optimal parameter combination that best matches the simulation results to reality. Using the calibrated optimal parameters, the simulation is run on a new, uncalibrated real-world traffic dataset. If the error between the key indicators (such as flow rate and speed) of the simulation output and the new dataset is within an acceptable range (e.g., the GEH statistic, generally accepted in traffic engineering, is less than 5), then the simulation model has been validated and can be used for subsequent strategy evaluation.

[0061] For example, the initial parameter adjustment plan is "increase the green light duration by 5 seconds for north-south traffic at intersection A". Critical path analysis identifies intersection A as being located on a north-south urban arterial road, which implements green wave coordination control. Traffic flow simulation results show that simply adding 5 seconds of green light to intersection A will disrupt its green wave band with the downstream intersection B, causing traffic congestion at intersection B before the red light. Therefore, the system modifies the initial plan, and the final optimized adjustment range is "increase the green light duration by 3 seconds for north-south traffic at intersection A, while simultaneously shifting the signal phase of the downstream intersection B forward by 2 seconds." This set of specific parameter adjustment instructions after co-optimization is the final output of this step.

[0062] In step S14, the spatial dependency characteristics and network topology characteristics of the optimized adjustment range and the regional traffic feature vector are fused to obtain the prediction-based search boundary conditions, including: The spatial dependency characteristics of the regional traffic feature vector and the optimization adjustment range are integrated to obtain enhanced optimization input data; The enhanced optimized input data is subjected to hierarchical processing of change trend information to obtain the initial constraint conditions of the search boundary; The initial constraints of the search boundary are calibrated using network topology characteristics to obtain the predicted search boundary conditions.

[0063] It should be noted that integrating the regional traffic feature vector and the optimization adjustment range with spatial dependency characteristics is to combine the optimization objective (adjustment range) with its surrounding traffic environment (spatial dependency characteristics), providing richer contextual information for subsequent boundary decisions. Here, "spatial dependency characteristics" originate from the flow transmission relationships between intersections discovered through association rule mining in step S11. The integration process is achieved by constructing an augmented feature matrix, which not only includes the adjustment parameters of the intersection to be optimized but also incorporates the real-time state characteristics (such as current queue length, flow rate, etc.) of adjacent intersections with strong spatial dependencies.

[0064] For example, suppose the "optimization adjustment range" determined in S13 is "{green light duration increase: 3-5 seconds}" for intersection A. Meanwhile, the "regional traffic feature vector" generated in S11 contains a strong correlation rule: "{intersection A congestion} → {intersection C congestion}". The "spatial dependency feature integration" operation will then generate an enhanced data record with the content: {target node: A, adjustment range: [+3s, +5s], associated node: C, current queue length of node C: 45 meters}. This more comprehensive data record, including the status of associated nodes, is the enhanced optimization input data.

[0065] It should be noted that the enhanced and optimized input data undergoes hierarchical processing based on the trend information of vehicle flow prediction values. This aims to dynamically set preliminary constraint boundaries according to the predicted traffic urgency. "Trend information" refers to the flow change gradient obtained by differentiating the vehicle flow prediction curve generated in S12 over a future period (e.g., 15 minutes). This gradient quantifies whether congestion is worsening, easing, or stabilizing. The "hierarchical processing" is a rule-based classification process that maps different change gradients to different urgency levels (e.g., "urgent," "moderate," "stable"). Each level corresponds to a set of preset initial constraints with varying degrees of leniency. The gradient thresholds that distinguish different levels are determined based on Receiver Operating Characteristic (ROC) curve analysis of historical traffic event data, aiming to find the optimal critical point that can distinguish different congestion evolution patterns with the highest accuracy.

[0066] It is worth noting that the mapping rule is as follows: if the traffic flow change gradient is greater than +5.0 vehicles / minute², it is determined to be at the "emergency" level, and the system adopts a set of tightened initial constraints for rapid intervention, such as {minimum green light duration: 25 seconds, maximum green light duration: 40 seconds, maximum cycle time: 120 seconds}; if the traffic flow change gradient is between +1.0 and +5.0 vehicles / minute², it is determined to be at the "medium" level, and the system adopts a set of standard initial constraints for proactive adjustment, such as {minimum green light duration: 20 seconds, maximum green light duration: 60 seconds, maximum cycle time: 15 seconds}. If the traffic flow change gradient is between -1.0 and +1.0 vehicles / minute², it is determined to be at the "stable" level. The system adopts a set of relaxed initial constraints for conventional optimization to explore a better solution, such as {minimum green light duration: 15 seconds, maximum green light duration: 90 seconds, maximum cycle time: 180 seconds}. If the traffic flow change gradient is less than -1.0 vehicles / minute², it is determined to be at the "mitigation" level. The system adopts a set of conservative initial constraints to maintain the status quo, such as {minimum green light duration: 15 seconds, maximum green light duration: 50 seconds, maximum cycle time: 150 seconds}.

[0067] For example, the system analyzes the trend information of the predicted vehicle traffic flow and obtains a traffic flow change gradient of "+10 vehicles / minute²" for the next 15 minutes. According to the above mapping rule, this value is greater than +5.0 vehicles / minute², so the system classifies the current situation as "emergency" and matches it with a set of relatively strict initial constraints corresponding to this level, namely {minimum green light duration: 25 seconds, maximum green light duration: 40 seconds, maximum cycle duration: 120 seconds}. This set of strict parameter boundaries is the initial constraint condition of the search boundary obtained in this step.

[0068] It should be noted that the initial constraints of the search boundary are calibrated using network topology characteristics to ensure that the constraints generated by traffic data are feasible and safe at the physical road network level. "Network topology characteristics" originate from road physical attributes in the GIS database, such as the number of lanes, road segment length, and turning restrictions. The calibration process employs a constraint satisfaction algorithm. This algorithm treats each parameter in the initial constraints (such as maximum green light duration) as a variable and the road network topology characteristics as physical constraints (such as maximum queuing capacity), then checks whether the variables are within the constraints. If a parameter exceeds the physical constraint, the algorithm corrects it to ensure that the final boundary conditions do not lead to physical traffic overflow or deadlock. The core correction mechanism of the algorithm is as follows: The algorithm iterates through each parameter variable in the initial constraints. For each variable, the algorithm extracts all related physical constraints from the network topology characteristics for evaluation. For example, for the variable "maximum green light duration," its directly related physical constraint is the maximum vehicle storage capacity of the downstream road segment. If the check finds that the value of a parameter variable exceeds the boundary of its physical constraint, the algorithm performs a boundary reset operation. Specifically, if a parameter value is higher than a physical upper limit constraint (such as maximum capacity), the algorithm will directly correct its value to that upper limit; if a parameter value is lower than a physical lower limit constraint (such as minimum safe duration), its value will be corrected to that lower limit. This process ensures that all parameters ultimately fall within their physical feasible region.

[0069] It is worth noting that, through the aforementioned network topology calibration, the system reduces the originally broad theoretical space of signal parameters (e.g., 0-180 seconds) to an effective feasible region (e.g., 45-60 seconds) that conforms to the current physical road conditions. This pre-pruning of the solution space eliminates a large number of invalid solutions that, although theoretically existing, would lead to physical overflow, thus providing a clear search boundary for subsequent optimization algorithms.

[0070] For example, the "initial constraints of the search boundary" include a condition: "Maximum green light duration for north-south traffic at intersection A: 90 seconds." During verification, the network topology calibration module, based on GIS data, discovers that the north-south road segment downstream of intersection A is only 200 meters long, and its maximum physical vehicle storage capacity can only accommodate traffic released within a 60-second green light duration. To prevent vehicles from overflowing to the next intersection, the constraint satisfaction algorithm adjusts (downgrades) this constraint from 90 seconds to 60 seconds. This final set of parameter boundaries, after physical feasibility verification and correction, constitutes the predicted search boundary conditions.

[0071] In step S15, based on the predicted search boundary conditions, dynamic traffic flow change characteristics are integrated to obtain a spatially constrained initial parameter set, including: The predicted search boundary conditions are processed by traffic distribution layering to obtain a preliminary parameter set framework. The initial parameter set framework is dynamically calibrated based on real-time traffic flow changes to obtain the calibrated parameter set. The calibrated parameter set is then subjected to dynamic traffic flow matching degree verification to obtain an initial parameter set subject to spatial constraints.

[0072] It should be noted that the hierarchical processing of traffic distribution in the prediction-based search boundary conditions is to transform continuous, dynamic traffic flow prediction data into a discrete, structured parameter framework, serving as the basis for subsequent calibration. The prediction-based search boundary conditions include traffic distribution information derived from vehicle traffic flow predictions; the "hierarchical processing" here is implemented using the K-Means Clustering algorithm. This algorithm takes all predicted traffic flow data points within the boundary conditions as input and automatically divides them into K categories or "layers," each representing a typical traffic level (e.g., "low traffic," "medium traffic," "high traffic"). The number of clusters, K, is determined based on offline analysis of historical traffic data using the Elbow Method, aiming to find the optimal number of clusters to balance the model's interpretability and complexity. After clustering, the system matches a set of benchmark signal timing parameters (such as green light duration and cycle length) from a pre-set expert knowledge base for each traffic level. The set of all levels and their corresponding benchmark parameters constitutes the preliminary parameter set framework.

[0073] For example, the system determines that K=3 is the optimal number of clusters after analysis using the elbow method. The K-means clustering algorithm is then run on the predicted traffic distribution data, dividing it into three traffic levels: "low traffic" (0-200 vehicles / hour), "medium traffic" (201-500 vehicles / hour), and "high traffic" (>500 vehicles / hour). The system then matches baseline parameters for each level from the expert knowledge base, forming a preliminary parameter set framework. For example, one record is: {Traffic level: high traffic, baseline green light duration: 50 seconds, baseline cycle: 120 seconds}.

[0074] It should be noted that the real-time dynamic calibration of the initial parameter set framework is performed to correct the baseline parameters based on historical and predicted data using the latest on-site traffic data. This "dynamic calibration" is accomplished through a feedback control loop implemented by a proportional-integral (PI) controller. In this loop, the "baseline green light duration" corresponding to the current traffic flow level is the control target (Setpoint), while the "intersection queue length" collected by real-time sensors is the process variable. The PI controller calculates the error between the target queue length (expected from the baseline parameters) and the actual queue length, and calculates a correction amount based on preset proportional gain (Kp) and integral gain (Ki) to dynamically adjust the baseline parameters. The Kp and Ki gain parameters are determined based on the Ziegler-Nichols tuning method after extensive testing with historical data in a simulation environment, aiming to ensure that the control system has both fast response and operational stability.

[0075] For example, the current real-time traffic belongs to the "high traffic" level, with a baseline green light duration of 50 seconds. The system expects the queue length to be 60 meters under this duration. However, the actual queue length reported by the real-time sensor is 75 meters, resulting in an error of -15 meters. Based on this error and preset Kp and Ki values, the PI controller calculates a positive correction, instructing the green light duration to increase by 5 seconds. Therefore, in the calibrated parameter set, the green light duration for this level is updated to 55 seconds.

[0076] It should be noted that performing dynamic traffic flow matching degree verification on the calibrated parameter set is the final quality control step to ensure that the final output initial parameter set can highly match the real traffic dynamics. This "matching degree verification" is accomplished by calculating the similarity between two time series: the first series is the "simulated traffic pattern" obtained by running the "calibrated parameter set" through a microscopic traffic flow simulation model (as described in S13); the second series is the "real-time traffic pattern" actually collected over a past period (e.g., 5 minutes). Before performing similarity calculation, to eliminate the influence of differences in numerical ranges under different traffic flows, the system first applies Z-score normalization to both the "simulated traffic pattern" and "real-time traffic pattern" series. This process converts each series into a standard normal distribution with a mean of 0 and a standard deviation of 1, ensuring that the subsequent DTW distance calculation only focuses on the morphological similarity of the series, without being affected by the absolute numerical magnitude. This embodiment uses the Dynamic Time Warping (DTW) algorithm to calculate the similarity between these two normalized series. The DTW algorithm can effectively measure the similarity between two sequences that have time axis scaling or offset. If the calculated DTW distance is lower than a preset similarity threshold, it proves that the calibrated parameters can well reproduce real traffic dynamics, and the verification is successful. The similarity threshold is determined by statistical analysis of a large number of historical simulation and real data pairs to find the critical DTW distance that can distinguish between "match" and "mismatch" states with the highest accuracy.

[0077] It is worth noting that the method for determining the similarity threshold is based on receiver operating characteristic curve analysis. Specifically, the system first requires a historical validation set, which contains a large number of "simulated-real" traffic pattern data pairs, each labeled as "match" or "mismatch". Subsequently, the system tests at a series of candidate DTW distance thresholds and calculates the "true positive rate" (the proportion of correctly identified "match" pairs) and "false positive rate" (the proportion of "mismatch" pairs incorrectly identified as "matches") for each threshold. By plotting these (false positive rate, true positive rate) data points into an ROC curve, the system ultimately selects the DTW distance corresponding to the point on the curve closest to the upper left corner (i.e., achieving the highest true positive rate at the lowest false positive rate) as the optimal balance point for distinguishing between "match" and "mismatch" states and the final preset similarity threshold.

[0078] For example, the system uses a simulation model to obtain a "simulated traffic pattern" sequence under a 55-second green light duration. Subsequently, the DTW algorithm calculates the distance between this sequence and the "real-time traffic pattern" sequence from the past 5 minutes to be 0.18. Assuming that ROC analysis of historical data determines that setting the DTW distance threshold to 0.25 best distinguishes between matching and mismatched states, and since 0.18 is lower than 0.25, the matching degree verification is successful. Therefore, this calibrated parameter set is ultimately confirmed as a qualified, spatially constrained initial parameter set that can be used for the next optimization solution.

[0079] In step S16, the initial parameter set subject to spatial constraints is optimized and calculated to obtain a personalized signal control strategy, including: The initial parameter set subject to spatial constraints is iteratively processed to generate candidate signal control parameter combinations; Traffic efficiency simulation evaluation is performed on the candidate signal control parameter combinations to obtain the evaluation score for each candidate combination; Based on the evaluation score, the optimal combination of candidate signal control parameters is selected to obtain a personalized signal control strategy.

[0080] It should be noted that the iterative processing of the spatially constrained initial parameter set is to systematically search for the optimal combination of signal parameters within a preset boundary. This embodiment uses a Genetic Algorithm (GA) to execute this iterative search process. Since the spatially constrained initial parameter set has been determined in step S15, the GA in this step does not need to blindly search the entire parameter space, but only needs to fine-tune and optimize within the pruned convergence domain. This improves the iterative convergence speed of the algorithm by several times compared to traditional full-domain search, thus meeting the millisecond-level real-time control requirements. The process first uses the initial parameter set obtained in S15 as the "initial population," where each parameter set is considered an "individual" or "chromosome." Subsequently, the algorithm enters a cyclical iterative "evolution" process. In each generation, a new "offspring population," i.e., candidate signal control parameter combinations, is generated through the following operations: Selection is based on the evaluation score (fitness) calculated from the previous generation individuals, using a roulette wheel selection method. Individuals with higher scores have a greater probability of being selected as "parents" for reproduction. Crossover involves randomly exchanging some parameters (genes) of selected "parent" pairs using a single-point crossover method, thereby generating "offspring" individuals with new combinations of characteristics. Mutation involves randomly changing a parameter value of some newly generated "offspring" individuals with a small, pre-defined mutation probability, in order to increase population diversity and prevent the algorithm from getting stuck in local optima.

[0081] The hyperparameters, such as the crossover probability and mutation probability, were determined through offline grid search optimization on a standard traffic scenario test set, aiming to find the parameter combination that achieves the best balance between the algorithm's convergence speed and solution quality.

[0082] For example, the genetic algorithm selects two parent individuals from the initial population: individual 1 = {green light duration: 55 seconds, period: 120 seconds} and individual 2 = {green light duration: 60 seconds, period: 110 seconds}. After performing a crossover operation, a offspring individual, offspring 1 = {green light duration: 55 seconds, period: 110 seconds}, may be generated. Subsequently, this offspring mutates with a 2% probability, and its green light duration parameter is randomly increased by 1 second, ultimately resulting in a new candidate signal control parameter combination: {green light duration: 56 seconds, period: 110 seconds}.

[0083] It should be noted that the traffic efficiency simulation evaluation of the candidate signal control parameter combinations is to quantify the merits of each potential solution. This process involves running an independent simulation of each candidate parameter combination using the calibrated micro-traffic flow simulation model from step S13. The simulation will model the traffic conditions over a future period (e.g., 30 minutes) after applying the candidate parameters and output a series of key performance indicators, primarily including total throughput (…). Average delay () ) and average queue length ( To enable a comprehensive comparison of these performance indicators with different physical units, the system first normalizes them, converting them into dimensionless performance indices. This process involves comparing the simulation output values ​​with a set of benchmark values ​​(…). , , This is accomplished through comparison. The baseline value is obtained by running a simulation under the same predicted traffic conditions using unoptimized initial signal parameters (i.e., the output of S15), and represents the basic performance level without optimization. The specific performance index calculation formula is as follows:

[0084]

[0085]

[0086] The system uses a predefined fitness function to weight and summarize these performance metrics, calculating a comprehensive evaluation score. This fitness function can be defined as follows:

[0087] in, This is the final evaluation score; , , These are the performance indices for throughput, delay, and queue length calculated above, respectively. , , These are the weighting coefficients corresponding to each performance indicator. These weights are determined based on the current traffic management objectives for the region. For example, in a scenario where the primary goal is to alleviate congestion, (Delay weight) and (Queue length weight) will be assigned a higher value.

[0088] For example, a candidate parameter combination {green light duration: 56 seconds, cycle: 110 seconds} is input into the simulation model. After the simulation runs, the output performance metrics are: =800 vehicles, =45 seconds, =60 meters. Assume the system was simulated beforehand using initial parameters, and the baseline value obtained is: total traffic volume. =750 vehicles, average delay =55 seconds, average queue length =70 meters. First, calculate the performance index of each component:

[0089]

[0090]

[0091] Assuming the current management objective is to balance traffic efficiency and delays, the weights are set as follows: =0.5, =0.3, =0.2. Therefore, the final evaluation score for this candidate combination is... for: .

[0092] It should be noted that selecting the optimal candidate signal control parameter combination based on the evaluation score is the decision-making and termination step in the entire optimization process. In each iteration of the genetic algorithm, the calculated evaluation score is used to guide the selection operation of the next generation of the population, thereby achieving survival of the fittest and continuous improvement of population performance. The entire iteration process is not indefinite, but ends when any of the following termination conditions are met: the preset maximum number of generations has been reached (e.g., 100 generations); the evaluation score of the best individual in the population has not significantly improved for several consecutive generations (e.g., 10 generations), indicating that the algorithm has converged. When the algorithm terminates, the system selects the individual that obtained the highest evaluation score in the entire evolutionary history, and takes its corresponding signal control parameter combination as the final optimal solution.

[0093] For example, when the genetic algorithm reaches generation 58, the evaluation score of the optimal individual is 1.152. In the subsequent 10 generations (generations 59 to 68), the evaluation scores of all new individuals appearing in the population do not exceed 1.152. At this point, the convergence termination condition is met, and the algorithm stops. The system then outputs the optimal parameter combination found in generation 58, such as {green light duration: 58 seconds, period: 115 seconds, phase difference: 10 seconds}. This final set of parameters obtained through optimization is the personalized signal control strategy.

[0094] In summary, this invention constructs a closed-loop control method covering the entire process from deep fusion of regional traffic features, dynamic traffic flow prediction, adaptive search boundary generation to personalized control strategy optimization. This method deeply couples the traffic prediction capability driven by a large model with the topological characteristics of the regional road network. Furthermore, it innovatively introduces a mechanism for dynamically constraining the optimization range based on prediction results. This solves the technical problems in existing technologies, such as inaccurate traffic flow prediction and lack of specificity and adaptability in signal control strategies, caused by the use of global models and the neglect of regional traffic feature differences. This significantly improves the accuracy, dynamism, and intelligence of urban traffic signal control.

[0095] Reference Figure 2 The second embodiment of the present invention provides a traffic flow prediction system based on regional signal optimization using a large traffic model, comprising: The feature generation module is used to acquire road network structure data and traffic flow density distribution data, and to perform correlation analysis and feature fusion on the road network structure data and the traffic flow density distribution data to obtain regional traffic feature vectors. The traffic flow prediction module is used to predict vehicle traffic flow based on the regional traffic feature vector to obtain the predicted vehicle traffic flow value. The congestion assessment module is used to assess the intersection congestion trend based on the predicted vehicle flow rate and obtain the optimization adjustment range. The boundary definition module is used to perform spatial dependency and network topology fusion processing on the optimization adjustment range and the regional traffic feature vector to obtain prediction-based search boundary conditions. The parameter generation module is used to integrate the dynamic traffic flow change characteristics based on the prediction-based search boundary conditions to obtain a spatially constrained initial parameter set. The strategy generation module is used to optimize and calculate the initial parameter set subject to spatial constraints to obtain a personalized signal control strategy.

[0096] It should be noted that the traffic flow prediction system based on regional signal optimization of a large traffic model provided in this embodiment of the invention is used to execute all the process steps of the traffic flow prediction method based on regional signal optimization of a large traffic model in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0097] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a traffic flow prediction program based on a large-scale traffic model for regional signal optimization. When the processor executes the computer program, it implements the steps described in the embodiments of the traffic flow prediction method based on a large-scale traffic model for regional signal optimization, for example... Figure 1 The step S11 shown. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the above system embodiments, such as the feature generation module.

[0098] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0099] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0100] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting all parts of the electronic device via various interfaces and lines.

[0101] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0102] Wherein, if the modules / units integrated in the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0103] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0104] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A traffic flow prediction method based on regional signal optimization using a large traffic model, characterized in that, include: Obtain road network structure data and traffic flow density distribution data, and perform correlation analysis and feature fusion on the road network structure data and the traffic flow density distribution data to obtain regional traffic feature vectors; Based on the regional traffic feature vector, vehicle flow prediction is performed to obtain the predicted vehicle flow value. Based on the predicted vehicle flow, the intersection congestion trend is assessed to determine the optimization adjustment range. The optimization adjustment range and the regional traffic feature vector are fused with spatial dependency characteristics and network topology characteristics to obtain prediction-based search boundary conditions. This includes: integrating the spatial dependency characteristics of the regional traffic feature vector and the optimization adjustment range to obtain enhanced optimization input data; performing hierarchical processing of the change trend information on the enhanced optimization input data to obtain initial constraints of the search boundary; and calibrating the initial constraints of the search boundary with network topology characteristics to obtain prediction-based search boundary conditions. Based on the predicted search boundary conditions, dynamic traffic flow change characteristics are integrated to obtain a spatially constrained set of initial parameters. The initial parameter set subject to spatial constraints is optimized and calculated to obtain a personalized signal control strategy.

2. The traffic flow prediction method based on regional signal optimization using a large traffic model according to claim 1, characterized in that, The process of acquiring road network structure data and traffic flow density distribution data, and performing correlation analysis and feature fusion on the road network structure data and the traffic flow density distribution data to obtain a regional traffic feature vector includes: Acquire road network structure data and traffic flow density distribution data; The road network structure data and the traffic flow density distribution data are cleaned and preliminarily integrated to obtain a regional description dataset; The correlation analysis between intersection nodes is performed on the aforementioned regional description dataset to obtain preliminary features; The preliminary features are compared with data and the weight parameters are adjusted to obtain the regional traffic feature vector.

3. The traffic flow prediction method based on regional signal optimization using a large traffic model according to claim 1, characterized in that, The step of predicting vehicle flow based on the regional traffic feature vector to obtain the predicted vehicle flow value includes: Get real-time traffic information; Scene recognition and pattern analysis are performed on the traffic feature vector of the area to obtain the vehicle flow change pattern; Based on the aforementioned vehicle traffic flow variation patterns, a time-series correlation analysis was performed to obtain a preliminary predicted trend; Based on the real-time traffic information, the preliminary prediction trend is calibrated to obtain the vehicle flow prediction value.

4. The traffic flow prediction method based on regional signal optimization using a large traffic model according to claim 1, characterized in that, The step of assessing the intersection congestion trend based on the predicted vehicle flow rate and obtaining the optimization adjustment range includes: Queue length is estimated from the predicted vehicle flow to obtain intersection congestion data; If the intersection congestion data exceeds the preset congestion threshold, a preset strategy mapping is performed to obtain a preliminary parameter adjustment plan. The initial parameter adjustment plan is then subjected to critical path analysis based on road network structure data to obtain the optimized adjustment range.

5. The traffic flow prediction method based on regional signal optimization using a large traffic model according to claim 1, characterized in that, The step of integrating dynamic traffic flow change characteristics based on the predicted search boundary conditions to obtain a spatially constrained initial parameter set includes: The predicted search boundary conditions are processed by traffic distribution layering to obtain a preliminary parameter set framework. The initial parameter set framework is dynamically calibrated based on real-time traffic flow changes to obtain the calibrated parameter set. The calibrated parameter set is then subjected to dynamic traffic flow matching degree verification to obtain an initial parameter set subject to spatial constraints.

6. The traffic flow prediction method based on regional signal optimization using a large traffic model according to claim 1, characterized in that, The optimization calculation of the spatially constrained initial parameter set to obtain a personalized signal control strategy includes: The initial parameter set subject to spatial constraints is iteratively processed to generate candidate signal control parameter combinations; Traffic efficiency simulation evaluation is performed on the candidate signal control parameter combinations to obtain the evaluation score for each candidate combination; Based on the evaluation score, the optimal combination of candidate signal control parameters is selected to obtain a personalized signal control strategy.

7. A traffic flow prediction system based on regional signal optimization using a large traffic model, characterized in that, include: The feature generation module is used to acquire road network structure data and traffic flow density distribution data, and to perform correlation analysis and feature fusion on the road network structure data and the traffic flow density distribution data to obtain regional traffic feature vectors. The traffic flow prediction module is used to predict vehicle traffic flow based on the regional traffic feature vector to obtain the predicted vehicle traffic flow value. The congestion assessment module is used to assess the intersection congestion trend based on the predicted vehicle flow rate and obtain the optimization adjustment range. The boundary definition module is used to fuse the spatial dependency characteristics and network topology characteristics of the optimization adjustment range and the regional traffic feature vector to obtain prediction-based search boundary conditions. This includes: integrating the spatial dependency characteristics of the regional traffic feature vector and the optimization adjustment range to obtain enhanced optimization input data; performing hierarchical processing of the enhanced optimization input data based on change trend information to obtain initial constraints of the search boundary; and calibrating the initial constraints of the search boundary based on network topology characteristics to obtain prediction-based search boundary conditions. The parameter generation module is used to integrate the dynamic traffic flow change characteristics based on the prediction-based search boundary conditions to obtain a spatially constrained initial parameter set. The strategy generation module is used to optimize and calculate the initial parameter set subject to spatial constraints to obtain a personalized signal control strategy.

Images