Model-free iterative learning traffic signal control method considering disturbance compensation

Abstract

This invention discloses a model-free iterative learning traffic signal control method considering disturbance compensation, specifically comprising: Step 1: collecting traffic state data and constructing a network-level congestion imbalance index; Step 2: establishing a historical iterative dataset and training samples for a random forest model; Step 3: predicting the initial green light duration control quantity based on the random forest model; Step 4: estimating external disturbances based on an extended state observer and generating disturbance compensation quantities; Step 5: constructing a model-free adaptive iterative learning controller and a control optimization objective; Step 6: estimating pseudo-partial derivatives online and updating the model-free adaptive iterative learning control parameters; Step 7: updating the green light duration control quantity and generating a green light duration control strategy; Step 8: generating and executing signal timing instructions, forming a closed-loop iterative control process. This method solves the problems of difficulty in modeling traffic systems, difficulty in accurately estimating external disturbances, and reliance on empirically set initial control values ​​in existing technologies.

Description

Technical Field

[0001] This invention belongs to the field of intelligent traffic control technology, specifically relating to a model-free iterative learning traffic signal control method that considers disturbance compensation. Background Technology

[0002] With the continuous advancement of urbanization and the increasing number of motor vehicles, urban road traffic congestion has become an increasingly prominent problem. Traffic signal control, as an important means to improve the efficiency of urban intersections and alleviate traffic congestion, faces the need for further improvement in its intelligence level. Existing urban traffic signal control methods typically rely on accurate traffic flow models, constructing macroscopic or microscopic traffic flow models to characterize vehicle queuing and dissipation processes, and combining optimization algorithms to solve for signal timing strategies. However, real-world urban traffic systems exhibit significant nonlinear characteristics, uncertainties, and frequent external disturbances, which limits the application of model-based methods in engineering, mainly in the following aspects: Traditional traffic signal control models have developed relatively mature systems for describing periodic traffic flow patterns, providing a certain degree of timing optimization capability under stable operating conditions. However, these models may still exhibit response lag when dealing with sudden abnormal disturbances (such as unpredictable traffic flows caused by school dismissal in non-intersection areas, concentrated entry and exit from large parking lots, and temporary road closures due to road construction). Furthermore, model-based control methods typically rely on a defined system structure and stable model parameters, facing problems such as low update frequency and poor scalability, making it difficult to adapt efficiently to situations with frequently changing traffic demand or incomplete historical data. In addition, existing iterative learning control methods, when applied to traffic signal control, often rely on manual experience to set initial control variables, resulting in large initial tracking errors, excessive convergence iterations, and low utilization efficiency of historical iteration information. Therefore, constructing an intelligent traffic signal control method that integrates data-driven initialization and iterative domain disturbance estimation compensation mechanisms can help improve the robustness and adaptability of the system in dynamic and uncertain traffic environments. Summary of the Invention

[0003] The purpose of this invention is to provide a model-free iterative learning traffic signal control method that takes into account disturbance compensation, which solves the problems of difficulty in modeling traffic systems, difficulty in accurately estimating external disturbances, and dependence on empirical setting of control initial values ​​in the prior art.

[0004] The technical solution adopted in this invention is a model-free iterative learning traffic signal control method considering disturbance compensation, specifically: Step 1: Collect traffic status data and construct a network-level congestion imbalance index; Step 2: Establish the historical iteration dataset and training samples for the random forest model; Step 3: Predict the initial green light duration control value based on the random forest model; Step 4: Estimate external disturbances and generate disturbance compensation amounts based on the extended state observer; Step 5: Construct a model-free adaptive iterative learning controller and control optimization objective; Step 6: Estimate pseudo-partial derivatives online and update model-free adaptive iterative learning control parameters; Step 7: Update the green light duration control value and generate the green light duration control strategy; Step 8: Generate and execute signal timing instructions to form a closed-loop iterative control process.

[0005] The invention is further characterized in that: Step 1 is implemented as follows: Step 1.1: Determine the urban road traffic network to be controlled; wherein, the urban road traffic network includes multiple intersections, each intersection includes at least one approach link and a signal phase corresponding to the approach link; the signal phase includes east-west phase, north-south phase or other actually set release phases; Step 1.2: Collect real-time traffic data at each intersection; Specifically, real-time traffic data at each intersection is collected through a traffic detection system; the traffic detection system includes geomagnetic coil detectors, video detectors, microwave detectors, or other detection equipment capable of acquiring traffic operation status; The real-time traffic data includes the number of vehicles queuing at each intersection, the number of arriving vehicles, and the current green light duration. Step 1.3: Normalize the number of vehicles queuing at each intersection during each signal cycle to obtain the normalized congestion status of each intersection. Specifically, the normalized congestion state of the i-th intersection in the k-th signal cycle is defined as:

[0006] in, Intersection In the Normalized congestion status for each signal cycle; Intersection In the Total number of vehicles in queue per signal cycle Intersection The sum of effective storage capacities; Number the intersections; Number the signal period; Step 1.4: Based on the normalized congestion status of each intersection obtained in Step 1.3, calculate the normalized congestion difference between intersections and construct a network-level congestion imbalance index:

[0007] in, For the first Within the signal period, the first Normalized congestion differences corresponding to each intersection; and Intersections Intersection Normalized congestion status; Number the intersections; This represents the total number of crossover pairs in the network. Step 1.5: Use the network-level congestion imbalance index obtained in Step 1.4 as the output feedback signal; and input the output feedback signal, along with the current green light duration and the number of arriving vehicles, into the subsequent steps.

[0008] Step 2 shall be implemented in accordance with the following: Step 2.1: Store the traffic data processed in Step 1 according to the signal cycle; specifically, the stored data includes normalized congestion status, network-level congestion imbalance index, number of arriving vehicles, and current green light duration. Step 2.2: Iterate the traffic data stored in Step 2.1 according to the iterative learning control method. After each iteration cycle, record the normalized congestion status, tracking error and green light duration control value corresponding to each time step in the current iteration cycle to form historical input data and historical output data. Step 2.3: Filter the historical iteration data; Specifically, records with tracking errors lower than the average tracking error are selected from each iteration time step as historical samples with better control effect; the historical tracking errors are sorted from smallest to largest, and samples with errors in the first 80th percentile are retained, while samples with errors in the last 20th percentile are removed or marked as abnormal samples. Step 2.4: Construct the training sample set for the random forest model; The input features of the training sample set include the number of arriving vehicles, the current green light duration, the normalized congestion status, the network-level congestion imbalance index, and the historical tracking error; the output label of the training sample set is the green light duration control value in the corresponding historical sample.

[0009] Step 3 shall be implemented in accordance with the following: Step 3.1: Construct a regression-based random forest model; The random forest model consists of multiple regression decision trees, each of which is trained based on the training sample set constructed in step 2.4. Each regression decision tree establishes a regression relationship between the intersection queue length and the green light duration control based on randomly selected samples and features. Step 3.2: Combine the current number of arriving vehicles, current green light duration, normalized congestion status, network-level congestion imbalance index, and tracking error corresponding to the most recent iteration obtained in Step 1 to form the input feature vector. , This represents the number of iterations. Step 3.3, input feature vector The results are input into each trained regression decision tree to obtain the prediction results of each regression decision tree for the initial green light duration of the current iteration cycle. , For the first Each regression decision tree is used for the current input feature vector The output is the predicted green light duration control value; Step 3.4: Prediction results of each regression decision tree. By performing average integration, the initial green light duration control value for the current iteration cycle is obtained:

[0010] in, For the first The initial green light duration control value at the start of the next iteration; The total number of decision trees in the random forest; For the first Each regression decision tree is used for the current input feature vector The output is the predicted green light duration control value; Number the regression decision tree; This represents the number of iterations.

[0011] Step 4 shall be implemented in accordance with the following: Step 4.1, identify external disturbances; where external disturbances include traffic flow changes caused by school dismissal, concentrated entry and exit from large parking lots, temporary road construction occupying roads, sudden traffic flow in non-intersection areas, or other non-modeling factors; Step 4.2: Employ the extended state observer method; Specifically, to measure the control effect of the network-level congestion imbalance index, a target reference sequence corresponding to the network-level congestion imbalance index is first set. ,in, The target network-level congestion imbalance index is represented as a reference value; then the extended state observer uses the network-level congestion imbalance index calculated in step 1 as the feedback output and the target network-level congestion imbalance index as the reference input to estimate external disturbances online. Step 4.3, the deviation between the actual observed network-level congestion imbalance index and the target network-level congestion imbalance index is shown below:

[0012] in, For the first In the nth iteration The perturbation estimate for each time step; For the first The perturbation estimate of the previous time step in the next iteration; For observer gain; For the first In the nth iteration The network-level congestion imbalance index is calculated at each time step based on the normalized congestion status. For the first The target network-level congestion imbalance index corresponding to the next iteration; Numbering the time steps in the iteration process; Step 4.4: Apply saturation constraints to the disturbance estimates obtained in Step 4.3 to obtain the disturbance estimates after saturation constraint processing. The details are as follows:

[0013] in, This is the saturation constant for the disturbance estimate; when the disturbance estimate is greater than... At that time, the disturbance estimate is restricted to When the disturbance estimate is less than At that time, the disturbance estimate is restricted to .

[0014] Step 5 shall be implemented in accordance with the following: Step 5.1, define the iterative increment of the green light duration control quantity, the specific expression is as follows:

[0015] in, No. In the nth iteration The change in green light duration control amount at each time step relative to the previous time step; For the first In the nth iteration The green light duration control quantity for each time step; For the first The green light duration control value of the previous time step in the next iteration; Step 5.2, define the normalized congestion difference tracking error, the specific expression of which is as follows:

[0016] in, For the first In the nth iteration Normalized congestion difference tracking error at each time step; For the first The target network-level congestion imbalance index corresponding to the next iteration; For the first In the nth iteration The network-level congestion imbalance index is calculated at each time step based on the normalized congestion status. Step 5.3: Based on the iterative increment of the green light duration control quantity obtained in Step 5.1, the tracking error obtained in Step 5.2, and the disturbance estimation result obtained in Step 4, construct the objective function for the green light duration control quantity:

[0017] in, The objective function for controlling the green light duration; This is a pseudo-partial derivative used to characterize the impact of changes in green light duration control on changes in normalized congestion differences; The disturbance estimate obtained from step 4 is input into the controller; To control the weight of the smoothing term, it is used to limit the variation of the green light duration control amount between adjacent time steps; This is an energy consumption penalty factor used to limit the adjustment range of the green light duration control. Step 5.4, adjust the objective function from step 5.3 with respect to the control input. Taking the partial derivative and setting it to zero, we obtain the optimal control law for the green light duration control variable:

[0018] In formula (9), the first term controls the degree of inheritance of the green light duration of the current iteration relative to the green light duration of the previous time step; the second term is used to correct the green light duration based on the normalized congestion difference tracking error and the disturbance estimate; the meanings of the remaining symbols are the same as those in formula (8).

[0019] Step 6 shall be implemented in accordance with the following: Step 6.1, due to the pseudo-partial derivatives Since it does not rely on precise traffic flow models in advance, the pseudo-partial derivatives are estimated online based on the changes in green light duration control and network-level congestion imbalance indicators in step 5, thus constructing a cost function for online pseudo-partial derivative estimation. :

[0020] in, For the first In the nth iteration The increment of the network-level congestion imbalance index calculated based on the normalized congestion status at each time step; The increment for controlling the green light duration; This is the estimate of the disturbance from the previous time step; This is the pseudo-partial derivative estimate from the previous time step; To update the smoothing factor, which is used to limit abrupt changes in pseudo-partial derivative estimates; Step 6.2: By minimizing the pseudo-partial derivative estimation cost function in Step 6.1, the pseudo-partial derivative estimate is obtained according to the following update formula:

[0021] in, For the first In the nth iteration The pseudo-partial derivative estimates at each time step; To adjust the step size; This represents the observation increment used to characterize the changes in normalized congestion differences during the current iteration.

[0022] Step 7 shall be implemented in accordance with the following: Step 7.1, use the initial green light duration control value obtained in Step 3. The disturbance estimate obtained in step 4 The normalized congestion difference tracking error obtained in step 5.2 The pseudo-partial derivative estimation results obtained in step 6 are input together into the model-free adaptive iterative learning controller. Step 7.2, considering the disturbance estimation, update the green light duration control quantity in the following concise form:

[0023] in, For the first In the nth iteration The green light duration control quantity for each time step; For the first The green light duration control value of the previous time step in the next iteration; This is a historical retention factor used to adjust the degree to which the current green light duration control value inherits from the historical green light duration control value; This is the error amplification factor, used to adjust the sensitivity of the control strategy to the response of tracking error and disturbance compensation results; To normalize the tracking error for congestion differences; This is the disturbance estimate after saturation constraint processing; Step 7.3: Perform boundary constraint processing on the green light duration control quantity obtained in Step 7.2 to ensure that the green light duration meets the safe operation requirements of the actual traffic signal controller. Specifically, the green light duration will be limited to the minimum green light duration. With maximum green light duration Within the constraints; when the calculated green light duration control value is less than At that time, take When the calculated green light duration control value is greater than At that time, take ; in, The minimum green light duration is preset; This is the preset maximum green light duration; Step 7.4: Determine the green light duration control quantity after boundary constraint processing as the green light duration control strategy for the corresponding signal phase of each intersection; the control strategy includes the green light duration of the east-west phase and the corresponding green light duration of the north-south phase, or includes the green light duration of other phases configured by the actual traffic signal controller.

[0024] Step 8 shall be implemented in accordance with the following: Step 8.1: Based on the green light duration control strategy for the corresponding signal phases of each intersection obtained in Step 7, generate signal timing instructions; wherein, the signal timing instructions include the intersection number, signal phase number, green light duration control amount, and corresponding execution signal cycle; Step 8.2: Send the signal timing command generated in step 8.1 to the signal controller of the corresponding intersection, and the signal controller will adjust the green light duration of each phase. Step 8.3: After the signal controller adjusts the green light duration of each phase, real-time traffic status data of each intersection is collected again, and the data is returned to Step 1 for normalization processing and network-level congestion imbalance index calculation. Step 8.4: Based on the newly collected traffic state data, repeat steps 2 to 8 to form a closed-loop control process of "traffic state collection - historical data update - initial control quantity prediction - disturbance estimation compensation - model-free iterative learning control - timing instruction execution", so that the network-level congestion imbalance index gradually converges to the target network-level congestion imbalance index.

[0025] The beneficial effects of this invention are: (1) The method of the present invention collects traffic data such as the queuing length of the approach link, the number of arriving vehicles, and the green light duration at each intersection in real time. It combines a model-free adaptive iterative learning control algorithm based on random forest and extended state observer to dynamically adjust the signal timing strategy of each intersection, drive the network normalized congestion imbalance to converge toward the target trajectory, and realize the coordinated optimization control of traffic lights at multiple intersections in the city.

[0026] (2) The method of the present invention significantly reduces the difficulty of modeling and parameter identification, does not require a precise traffic flow mathematical model, and is suitable for complex, nonlinear, and time-varying urban traffic system environments.

[0027] (3) The method of the present invention processes historical iteration data and real-time input data through a random forest model, and outputs an optimized initial green light duration control quantity for controller initialization, which can effectively reduce the initial timing error, improve control accuracy and accelerate the convergence speed of the control strategy.

[0028] (4) The method of the present invention can estimate external sudden disturbance factors such as school dismissal and parking lot entry and exit in real time by extending the state observer, and use them for iterative domain control compensation to improve the system's operational stability and timing robustness under non-ideal or extreme traffic conditions.

[0029] (5) The method of the present invention solves the problems of difficulty in modeling traffic systems, difficulty in accurately estimating external disturbances, and dependence on experience in setting initial control values ​​in the prior art. It can be transformed into a corresponding modular structure design, and the control algorithm can be embedded into the existing traffic signal control automation platform. It has good promotion value and practical application prospects. Attached Figure Description

[0030] Figure 1 The flowchart is for the model-free iterative learning traffic signal control method considering disturbance compensation in this invention; Figure 2 This is a road network diagram of Changfeng Street and Xuefu Street in Taiyuan City, as shown in Example 1. Figure 3 This is a comparison chart of the standard deviation of tracking error under slight disturbances in Example 1; Figure 4 This is a comparison chart of the standard deviation of tracking error under moderate disturbance in Example 1; Figure 5 This is a comparison chart of the standard deviation of tracking error under severe disturbance in Example 1; Figure 6 This is a comparison chart of average velocities under slight disturbances in Example 1; Figure 7 This is a comparison chart of average velocities under moderate disturbances in Example 1; Figure 8 This is a comparison chart of average velocities under severe disturbance in Example 1; Figure 9 This is a comparison chart of average delay times under slight disturbances in Example 1; Figure 10 This is a comparison chart of average delay times under moderate disturbances in Example 1; Figure 11 This is a comparison chart of average delay time under severe disturbances in Example 1. Detailed Implementation

[0031] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0032] This invention provides a model-free iterative learning traffic signal control method that considers disturbance compensation, such as... Figure 1 As shown, the method is applied to urban road traffic networks with multiple intersections. By collecting traffic state data of each intersection, the initial value of the green light duration control variable is obtained using a random forest model. External disturbances are estimated using an extended state observer. The green light duration is iteratively updated under a model-free adaptive iterative learning control framework, thereby converging the network congestion imbalance towards a preset target trajectory. The method specifically includes the following steps.

[0033] Step 1: Collect traffic status data and construct a network-level congestion imbalance index; Step 1 is implemented as follows: Step 1.1: Determine the urban road traffic network to be controlled; wherein, the urban road traffic network includes multiple intersections, each intersection includes at least one approach link and a signal phase corresponding to the approach link; the signal phase includes east-west phase, north-south phase or other actually set release phases; Step 1.2: Collect real-time traffic data at each intersection; Specifically, real-time traffic data at each intersection is collected through a traffic detection system; the traffic detection system includes geomagnetic coil detectors, video detectors, microwave detectors, or other detection equipment capable of acquiring traffic operation status; Real-time traffic data includes the number of vehicles queuing at each intersection, the number of arriving vehicles, and the current green light duration; Step 1.3: Normalize the number of vehicles queuing at each intersection during each signal cycle to obtain the normalized congestion status of each intersection. Specifically, the normalized congestion state of the i-th intersection in the k-th signal cycle is defined as:

[0034] in, Intersection In the Normalized congestion status for each signal cycle; Intersection In the Total number of vehicles in queue per signal cycle Intersection The sum of effective storage capacities; Number the intersections; Number the signal period; Step 1.4: Based on the normalized congestion status of each intersection obtained in Step 1.3, calculate the normalized congestion difference between intersections and construct a network-level congestion imbalance index:

[0035] in, For the first Within the signal period, the first Normalized congestion differences corresponding to each intersection; and Intersections Intersection Normalized congestion status; Number the intersections; This represents the total number of crossover pairs in the network. Step 1.5: Use the network-level congestion imbalance index obtained in Step 1.4 as the output feedback signal; and input the output feedback signal, along with the current green light duration and the number of arriving vehicles, into the subsequent steps.

[0036] Step 2: Establish a historical iteration dataset and construct training samples for the random forest model; Step 2 shall be implemented in accordance with the following: Step 2.1: Store the traffic data processed in Step 1 according to the signal cycle; specifically, the stored data includes normalized congestion status, network-level congestion imbalance index, number of arriving vehicles, and current green light duration. Step 2.2: Iterate the traffic data stored in Step 2.1 according to the iterative learning control method. After each iteration cycle, record the normalized congestion status, tracking error and green light duration control value corresponding to each time step in the current iteration cycle to form historical input data and historical output data. Step 2.3: Filter the historical iteration data; Specifically, records with tracking errors lower than the average tracking error are selected from each iteration time step as historical samples with better control effects; the historical tracking errors are sorted from smallest to largest, and samples with errors in the first 80th percentile are retained, while samples with errors in the last 20th percentile are removed or marked as outliers.

[0037] Step 2.4: Construct the training sample set for the random forest model; The input features of the training sample set include the number of arriving vehicles, the current green light duration, the normalized congestion status, the network-level congestion imbalance index, and the historical tracking error; the output label of the training sample set is the green light duration control value in the corresponding historical sample. Step 3: Predict the initial green light duration control value based on the random forest model; Step 3 shall be implemented in accordance with the following: Step 3.1: Construct a regression-based random forest model; The random forest model consists of multiple regression decision trees, each of which is trained based on the training sample set constructed in step 2.4. Each regression decision tree establishes a regression relationship between the intersection queue length and the green light duration control based on randomly selected samples and features.

[0038] Step 3.2: Combine the current number of arriving vehicles, current green light duration, normalized congestion status, network-level congestion imbalance index, and tracking error corresponding to the most recent iteration obtained in Step 1 to form the input feature vector. , This represents the number of iterations. Step 3.3, input feature vector The results are input into each trained regression decision tree to obtain the prediction results of each regression decision tree for the initial green light duration of the current iteration cycle. , For the first Each regression decision tree is used for the current input feature vector The output is the predicted green light duration control value; Step 3.4: Prediction results of each regression decision tree. By performing average integration, the initial green light duration control value for the current iteration cycle is obtained:

[0039] in, For the first The initial green light duration control value at the start of the next iteration; The total number of decision trees in the random forest; For the first Each regression decision tree is used for the current input feature vector The output is the predicted green light duration control value; Number the regression decision tree; This represents the number of iterations.

[0040] Step 4: Estimate external disturbances and generate disturbance compensation amounts based on the extended state observer; Step 4 shall be implemented in accordance with the following: Step 4.1, determine the external disturbances; where external disturbances include traffic flow changes caused by school dismissal, concentrated entry and exit of large parking lots, temporary road construction occupying the road, sudden traffic flow in non-intersection areas, or other non-modeling factors; external disturbances will cause fluctuations in the network-level congestion imbalance index constructed in Step 1, thereby affecting the tracking effect of the traffic signal control method on the target operating state; Step 4.2: Employ the extended state observer method; Specifically, to measure the control effect of the network-level congestion imbalance index, a target reference sequence corresponding to the network-level congestion imbalance index is first set. ,in, The target network-level congestion imbalance index is represented as a reference value; then the extended state observer uses the network-level congestion imbalance index calculated in step 1 as the feedback output and the target network-level congestion imbalance index as the reference input to estimate external disturbances online. Step 4.3, the deviation between the actual observed network-level congestion imbalance index and the target network-level congestion imbalance index is shown below:

[0041] in, For the first In the nth iteration The perturbation estimate for each time step; For the first The perturbation estimate of the previous time step in the next iteration; For observer gain; For the first In the nth iteration The network-level congestion imbalance index is calculated at each time step based on the normalized congestion status. For the first The target network-level congestion imbalance index corresponding to the next iteration; Numbering the time steps in the iteration process; Step 4.4: Apply saturation constraints to the disturbance estimate obtained in Step 4.3 to prevent excessively large disturbance estimates from causing significant oscillations in the green light duration control value, thus obtaining the disturbance estimate after saturation constraint processing. The details are as follows:

[0042] in, This is the saturation constant for the disturbance estimate; when the disturbance estimate is greater than... At that time, the disturbance estimate is restricted to When the disturbance estimate is less than At that time, the disturbance estimate is restricted to .

[0043] Step 5: Construct a model-free adaptive iterative learning controller and establish a control optimization objective based on the initial control input and disturbance estimates; Step 5 shall be implemented in accordance with the following: Step 5.1, define the iterative increment of the green light duration control quantity, the specific expression is as follows:

[0044] in, No. In the nth iteration The change in green light duration control amount at each time step relative to the previous time step; For the first In the nth iteration The green light duration control quantity for each time step; For the first The green light duration control value of the previous time step in the next iteration; Step 5.2, define the normalized congestion difference tracking error, the specific expression of which is as follows:

[0045] in, For the first In the nth iteration Normalized congestion difference tracking error at each time step; For the first The target network-level congestion imbalance index corresponding to the next iteration; For the first In the nth iteration The network-level congestion imbalance index is calculated at each time step based on the normalized congestion status. Step 5.3: Based on the iterative increment of the green light duration control quantity obtained in Step 5.1, the tracking error obtained in Step 5.2, and the disturbance estimation result obtained in Step 4, construct the objective function for the green light duration control quantity:

[0046] in, The objective function for controlling the green light duration; This is a pseudo-partial derivative used to characterize the impact of changes in green light duration control on changes in normalized congestion differences; The disturbance estimate obtained from step 4 is input into the controller; To control the weight of the smoothing term, it is used to limit the variation of the green light duration control amount between adjacent time steps; This is an energy consumption penalty factor used to limit the adjustment range of the green light duration control.

[0047] Step 5.4, adjust the objective function from step 5.3 with respect to the control input. Taking the partial derivative and setting it to zero, we obtain the optimal control law for the green light duration control variable:

[0048] In formula (9), the first term controls the degree of inheritance of the green light duration of the current iteration relative to the green light duration of the previous time step; the second term is used to correct the green light duration based on the normalized congestion difference tracking error and the disturbance estimate; the meanings of the remaining symbols are the same as those in formula (8).

[0049] Step 6: Estimate pseudo-partial derivatives online and update model-free adaptive iterative learning control parameters.

[0050] Step 6 shall be implemented in accordance with the following: Step 6.1, due to the pseudo-partial derivatives Since it does not rely on precise traffic flow models in advance, the pseudo-partial derivatives are estimated online based on the changes in green light duration control and network-level congestion imbalance indicators in step 5, thus constructing a cost function for online pseudo-partial derivative estimation. :

[0051] in, For the first In the nth iteration The increment of the network-level congestion imbalance index calculated based on the normalized congestion status at each time step; The increment for controlling the green light duration; This is the estimate of the disturbance from the previous time step; This is the pseudo-partial derivative estimate from the previous time step; To update the smoothing factor, which is used to limit abrupt changes in pseudo-partial derivative estimates; Step 6.2: By minimizing the pseudo-partial derivative estimation cost function in Step 6.1, the pseudo-partial derivative estimate is obtained according to the following update formula:

[0052] in, For the first In the nth iteration The pseudo-partial derivative estimates at each time step; To adjust the step size; This represents the observation increment used to characterize the changes in normalized congestion differences during the current iteration.

[0053] Step 7: Update the green light duration control quantity based on the disturbance compensation results, and generate the green light duration control strategy.

[0054] Step 7 shall be implemented in accordance with the following: Step 7.1, use the initial green light duration control value obtained in Step 3. The disturbance estimate obtained in step 4 The normalized congestion difference tracking error obtained in step 5.2 The pseudo-partial derivative estimation results obtained in step 6 are input together into the model-free adaptive iterative learning controller. Step 7.2, considering the disturbance estimation, update the green light duration control quantity in the following concise form:

[0055] in, For the first In the nth iteration The green light duration control quantity for each time step; For the first The green light duration control value of the previous time step in the next iteration; This is a historical retention factor used to adjust the degree to which the current green light duration control value inherits from the historical green light duration control value; This is the error amplification factor, used to adjust the sensitivity of the control strategy to the response of tracking error and disturbance compensation results; To normalize the tracking error for congestion differences; This is the disturbance estimate after saturation constraint processing; Step 7.3: Perform boundary constraint processing on the green light duration control quantity obtained in Step 7.2 to ensure that the green light duration meets the safe operation requirements of the actual traffic signal controller. Specifically, the green light duration will be limited to the minimum green light duration. With maximum green light duration Within the constraints; when the calculated green light duration control value is less than At that time, take When the calculated green light duration control value is greater than At that time, take ; in, The minimum green light duration is preset; This is the preset maximum green light duration; Step 7.4: Determine the green light duration control quantity after boundary constraint processing as the green light duration control strategy for the corresponding signal phase of each intersection; the control strategy includes the green light duration of the east-west phase and the corresponding green light duration of the north-south phase, or includes the green light duration of other phases configured by the actual traffic signal controller.

[0056] Step 8: Generate and execute signal timing instructions to form a closed-loop iterative control process.

[0057] Step 8 shall be implemented in accordance with the following: Step 8.1: Based on the green light duration control strategy for the corresponding signal phases of each intersection obtained in Step 7, generate signal timing instructions; wherein, the signal timing instructions include the intersection number, signal phase number, green light duration control amount, and corresponding execution signal cycle; Step 8.2: Send the signal timing command generated in step 8.1 to the signal controller of the corresponding intersection, and the signal controller will adjust the green light duration of each phase. Step 8.3: After the signal controller adjusts the green light duration of each phase, real-time traffic status data of each intersection is collected again, and the data is returned to Step 1 for normalization processing and network-level congestion imbalance index calculation. Step 8.4: Based on the newly collected traffic state data, repeat steps 2 to 8 to form a closed-loop control process of "traffic state collection - historical data update - initial control quantity prediction - disturbance estimation compensation - model-free iterative learning control - timing instruction execution", so that the network-level congestion imbalance index gradually converges to the target network-level congestion imbalance index.

[0058] Example 1 To verify the actual control effect of the model-free iterative learning traffic signal control method considering disturbance compensation provided in this invention, such as... Figure 2 As shown, this embodiment selects the road area between Changfeng Street and Xuefu Street in Taiyuan City as the simulation object, constructs a city multi-intersection road traffic simulation network, and uses VISSIM traffic simulation software and Python control program to carry out simulation experiments.

[0059] This embodiment is implemented according to the following: Step 1: Construct a simulation network of urban multi-intersection road traffic.

[0060] Specifically, an urban road traffic simulation network was constructed in the actual road area between Changfeng Street and Xuefu Street in Taiyuan City, comprising 28 intersections, 17 traffic flow entrances, and 81 roads. Each intersection was configured with corresponding signal phase schemes and initial signal timing schemes based on actual road conditions.

[0061] Step 2: Set the basic parameters for traffic simulation and the time-segmented traffic demand.

[0062] Specifically, the signal cycle of each intersection is set to 120s, the total simulation duration is set to 7200s, the traffic state data sampling cycle is set to 120s, the simulation resolution is set to 1, and the vehicle turning ratio is set to a ratio of 1:3:1 for straight, left, and right turns.

[0063] The simulation time was divided into five periods: 0s–1440s, 1440s–2880s, 2880s–4320s, 4320s–5760s, and 5760s–7200s. Traffic flow inputs were set according to the traffic demand corresponding to each period. The period from 2880s to 4320s corresponds to peak traffic hours; the other periods correspond to periods of low or transitional traffic demand, simulating the time-varying traffic flow characteristics of the urban traffic network.

[0064] Step 3: Collect traffic status data and form a historical iterative dataset.

[0065] Specifically, after each signal cycle, the number of queuing vehicles, the number of arriving vehicles, and the current green light duration are collected at each intersection. Based on the number of queuing vehicles and the effective storage capacity of the approach links at each intersection, the normalized congestion status of each intersection is calculated, and a network-level congestion imbalance index is obtained based on the difference in normalized congestion status between intersections.

[0066] During the model-free iterative learning control process, the normalized congestion state, network-level congestion imbalance index, tracking error, and green light duration control value are recorded for each iteration time step. Records with smaller tracking errors are selected from each iteration time step to construct a historical iteration dataset.

[0067] Step 4: Train a random forest model based on the historical iteration dataset and predict the initial green light duration control amount.

[0068] Specifically, the historical iteration dataset obtained in step 3 is used as training samples to establish a regression-type random forest model. Normalized congestion status, network-level congestion imbalance index, number of arriving vehicles, current green light duration, and tracking error are used as input features of the random forest model, and the corresponding green light duration control value is used as the output label.

[0069] At the start of a new iteration, the current traffic state data is input into the trained random forest model to obtain the initial green light duration control values ​​for each intersection. and control the initial green light duration. The input model-free iterative learning controller is used as the starting control variable for subsequent iterative updates.

[0070] Step 5: Set the external disturbance conditions and estimate the disturbance using the extended state observer.

[0071] Specifically, additional input traffic flows are applied to roads x_3, x_33, x_41, x_54, x_65, x_38, and x_10 to simulate sudden traffic flows in non-intersection areas caused by school dismissal, concentrated vehicle entry and exit in commercial areas, concentrated exit from parking lots, or sudden traffic events.

[0072] Three operating conditions are set: mild disturbance, moderate disturbance, and severe disturbance. The mild disturbance is set in the time period from 0s to 1440s, the moderate disturbance is set in the peak traffic period from 2880s to 4320s, and the severe disturbance is set in the time period from 4320s to 5760s.

[0073] An additional 150 vehicles of traffic flow will be added under mild disturbance conditions; an additional 300 vehicles of traffic flow will be added under moderate disturbance conditions; and an additional 450 vehicles of traffic flow will be added under severe disturbance conditions.

[0074] Based on the network-level congestion imbalance index obtained in step 3 and the preset target network-level congestion imbalance index, the external disturbance is estimated using an extended state observer to obtain the disturbance estimate. The disturbance estimate is then input into the model-free iterative learning controller to compensate for control errors caused by sudden external traffic flows.

[0075] Step 6: Set up the comparison control method and conduct iterative control experiments.

[0076] Specifically, the following four control methods are compared: The first type is the MFAILC method, which is a model-free adaptive iterative learning control method that does not use random forest initial control quantity prediction and extended state observer disturbance compensation. The second method is the KNN-MFAILC method, which is a model-free adaptive iterative learning control method that uses the K-nearest neighbor algorithm to predict the initial control quantity. The third type is the RF-MFAILC method, which is a model-free adaptive iterative learning control method that uses a random forest model to predict the initial green light duration control quantity but does not use extended state observer disturbance compensation. The fourth method is the ESO-RFMFAILC method, which is a model-free iterative learning traffic signal control method that uses a random forest model to predict the initial green light duration control quantity and uses an extended state observer to estimate and compensate for external disturbances.

[0077] Under mild, moderate, and severe disturbance conditions, the iterative control processes of the above four control methods are executed respectively, and the tracking error standard deviation, network average vehicle speed, and network average delay time of the network-level congestion imbalance index are recorded.

[0078] Step 7: Verify the improvement effect of the method of the present invention on the control convergence speed and control accuracy.

[0079] Specifically, the standard deviation of the tracking error of the network-level congestion imbalance index is used as the evaluation index for control convergence performance, where the tracking error is the deviation between the actual network-level congestion imbalance index and the target network-level congestion imbalance index. The changes in the standard deviation of the tracking error of the four control methods with the number of iterations are compared under mild, moderate, and severe disturbance conditions.

[0080] Depend on Figures 3-5 It can be seen that the standard deviation of the tracking error of the method of the present invention is lower than that of the MFAILC method, KNN-MFAILC method and RF-MFAILC method under mild, moderate and severe disturbance conditions.

[0081] Under moderate disturbance conditions, the tracking error standard deviation of the method of the present invention is 21.930, which is lower than 24.580 of the MFAILC method, representing a reduction of approximately 10.4% compared to the MFAILC method.

[0082] Furthermore, the method of the present invention typically reaches a stable state after 20 to 25 iterations under moderate and severe disturbance conditions; while the MFAILC method typically requires about 30 to 35 iterations to reach a similar stable state.

[0083] The experimental results above demonstrate that the random forest model can output a better initial green light duration control value based on historical iteration data. This reduces tracking error in the initial stage of control; the extended state observer can estimate and compensate for external disturbances, enabling the controller to maintain good convergence speed and control accuracy when external disturbances are present.

[0084] Step 8: Verify the effect of the method of the present invention on improving the average vehicle speed of the road network.

[0085] Specifically, using the network average vehicle speed as an evaluation index for the overall traffic efficiency of the road traffic network, the changes in the network average vehicle speed with the number of iterations were compared among the four control methods under mild, moderate, and severe disturbance conditions.

[0086] like Figures 6-8 As shown, the average network vehicle speed of the method of the present invention is higher than that of the MFAILC method, KNN-MFAILC method and RF-MFAILC method under various disturbance conditions.

[0087] Under severe disturbance conditions, the average vehicle speed of the network using the method of this invention is 48.620 km / h, which is higher than the 48.360 km / h of the MFAILC method.

[0088] The experimental results above demonstrate that the method of the present invention can adaptively adjust the green light duration of each phase based on the normalized congestion status of different intersections, network-level congestion imbalance index, and disturbance estimation results, thereby alleviating excessive congestion accumulation at local intersections and improving the overall traffic efficiency of the urban road network.

[0089] Step 9: Verify the improvement effect of the method of the present invention on vehicle delay and queuing risk.

[0090] Specifically, using the average network delay time as an evaluation index for vehicle waiting time, the changes in the average network delay time of four control methods with the number of iterations under mild, moderate, and severe disturbance conditions were compared.

[0091] Experimental results show that the average network delay time of the method of the present invention is lower than that of the MFAILC method, KNN-MFAILC method and RF-MFAILC method under various disturbance conditions.

[0092] from Figures 9-11 It can be seen that under moderate disturbance conditions, the average network delay time of the method of the present invention is 62.180s, which is lower than 65.150s of the MFAILC method; under severe disturbance conditions, the average network delay time of the method of the present invention is 72.050s, which is lower than 76.100s of the MFAILC method.

[0093] The experimental results above demonstrate that the method of the present invention can not only reduce the average vehicle delay time, but also reduce the delay level of high-delay vehicles and the risk of excessive queuing at local intersections, thereby reducing the possibility of queue spillover and congestion spread.

[0094] Step 10: The verification conclusion of this embodiment is obtained.

[0095] Specifically, based on the experimental results of steps 7 to 9, it can be seen that the model-free iterative learning traffic signal control method considering disturbance compensation provided by this invention can predict a better initial green light duration control value through a random forest model. It can reduce initial control error and accelerate iterative convergence; it can estimate and compensate for external disturbances by extending the state observer, thereby improving control stability and robustness under moderate and severe disturbance conditions; and it can increase the average vehicle speed of the road network, reduce the average vehicle delay time, and reduce the maximum queue length at local intersections, thereby alleviating the problem of excessive congestion concentration and queue spillover in urban road networks with multiple intersections.

[0096] In summary, the model-free iterative learning traffic signal control method of the present invention, which considers disturbance compensation, solves the problems of difficulty in modeling traffic systems, difficulty in accurately estimating external disturbances, and dependence on empirical setting of control initial values ​​in the prior art. Example 2 The model-free iterative learning traffic signal control method considering disturbance compensation is as follows: Step 1: Collect traffic status data and construct a network-level congestion imbalance index; Step 2: Establish the historical iteration dataset and training samples for the random forest model; Step 3: Predict the initial green light duration control value based on the random forest model; Step 4: Estimate external disturbances and generate disturbance compensation amounts based on the extended state observer; Step 5: Construct a model-free adaptive iterative learning controller and control optimization objective; Step 6: Estimate pseudo-partial derivatives online and update model-free adaptive iterative learning control parameters; Step 7: Update the green light duration control value and generate the green light duration control strategy; Step 8: Generate and execute signal timing instructions to form a closed-loop iterative control process.

[0097] Example 3 The difference from Example 2 is as follows: Step 1 is implemented as follows: Step 1.1: Determine the urban road traffic network to be controlled; wherein, the urban road traffic network includes multiple intersections, each intersection includes at least one approach link and a signal phase corresponding to the approach link; the signal phase includes east-west phase, north-south phase or other actually set release phases; Step 1.2: Collect real-time traffic data at each intersection; Specifically, real-time traffic data at each intersection is collected through a traffic detection system; the traffic detection system includes geomagnetic coil detectors, video detectors, microwave detectors, or other detection equipment capable of acquiring traffic operation status; The real-time traffic data includes the number of vehicles queuing at each intersection, the number of arriving vehicles, and the current green light duration. Step 1.3: Normalize the number of vehicles queuing at each intersection during each signal cycle to obtain the normalized congestion status of each intersection. Specifically, the normalized congestion state of the i-th intersection in the k-th signal cycle is defined as:

[0098] in, Intersection In the Normalized congestion status for each signal cycle; Intersection In the Total number of vehicles in queue per signal cycle Intersection The sum of effective storage capacities; Number the intersections; Number the signal period; Step 1.4: Based on the normalized congestion status of each intersection obtained in Step 1.3, calculate the normalized congestion difference between intersections and construct a network-level congestion imbalance index:

[0099] in, For the first Within the signal period, the first Normalized congestion differences corresponding to each intersection; and Intersections Intersection Normalized congestion status; Number the intersections; This represents the total number of crossover pairs in the network. Step 1.5: Use the network-level congestion imbalance index obtained in Step 1.4 as the output feedback signal; and input the output feedback signal, along with the current green light duration and the number of arriving vehicles, into the subsequent steps.

[0100] Example 4 The difference from Example 2 is as follows: Step 2 shall be implemented in accordance with the following: Step 2.1: Store the traffic data processed in Step 1 according to the signal cycle; specifically, the stored data includes normalized congestion status, network-level congestion imbalance index, number of arriving vehicles, and current green light duration. Step 2.2: Iterate the traffic data stored in Step 2.1 according to the iterative learning control method. After each iteration cycle, record the normalized congestion status, tracking error and green light duration control value corresponding to each time step in the current iteration cycle to form historical input data and historical output data. Step 2.3: Filter the historical iteration data; Specifically, records with tracking errors lower than the average tracking error are selected from each iteration time step as historical samples with better control effect; the historical tracking errors are sorted from smallest to largest, and samples with errors in the first 80th percentile are retained, while samples with errors in the last 20th percentile are removed or marked as abnormal samples. Step 2.4: Construct the training sample set for the random forest model; The input features of the training sample set include the number of arriving vehicles, the current green light duration, the normalized congestion status, the network-level congestion imbalance index, and the historical tracking error; the output label of the training sample set is the green light duration control value in the corresponding historical sample.

[0101] Example 5 The difference from Example 2 is as follows: Step 3 shall be implemented in accordance with the following: Step 3.1: Construct a regression-based random forest model; The random forest model consists of multiple regression decision trees, each of which is trained based on the training sample set constructed in step 2.4. Each regression decision tree establishes a regression relationship between the intersection queue length and the green light duration control based on randomly selected samples and features. Step 3.2: Combine the current number of arriving vehicles, current green light duration, normalized congestion status, network-level congestion imbalance index, and tracking error corresponding to the most recent iteration obtained in Step 1 to form the input feature vector. , This represents the number of iterations. Step 3.3, input feature vector The results are input into each trained regression decision tree to obtain the prediction results of each regression decision tree for the initial green light duration of the current iteration cycle. , For the first Each regression decision tree is used for the current input feature vector The output is the predicted green light duration control value; Step 3.4: Prediction results of each regression decision tree. By performing average integration, the initial green light duration control value for the current iteration cycle is obtained:

[0102] in, For the first The initial green light duration control value at the start of the next iteration; The total number of decision trees in the random forest; For the first Each regression decision tree is used for the current input feature vector The output is the predicted green light duration control value; Number the regression decision tree; This represents the number of iterations.

[0103] Example 6 The difference from Example 2 is as follows: Step 4 shall be implemented in accordance with the following: Step 4.1, identify external disturbances; where external disturbances include traffic flow changes caused by school dismissal, concentrated entry and exit from large parking lots, temporary road construction occupying roads, sudden traffic flow in non-intersection areas, or other non-modeling factors; Step 4.2: Employ the extended state observer method; Specifically, to measure the control effect of the network-level congestion imbalance index, a target reference sequence corresponding to the network-level congestion imbalance index is first set. ,in, The target network-level congestion imbalance index is represented as a reference value; then the extended state observer uses the network-level congestion imbalance index calculated in step 1 as the feedback output and the target network-level congestion imbalance index as the reference input to estimate external disturbances online. Step 4.3, the deviation between the actual observed network-level congestion imbalance index and the target network-level congestion imbalance index is shown below:

[0104] in, For the first In the nth iteration The perturbation estimate for each time step; For the first The perturbation estimate of the previous time step in the next iteration; For observer gain; For the first In the nth iteration The network-level congestion imbalance index is calculated at each time step based on the normalized congestion status. For the first The target network-level congestion imbalance index corresponding to the next iteration; Numbering the time steps in the iteration process; Step 4.4: Apply saturation constraints to the disturbance estimates obtained in Step 4.3 to obtain the disturbance estimates after saturation constraint processing. The details are as follows:

[0105] in, This is the saturation constant for the disturbance estimate; when the disturbance estimate is greater than... At that time, the disturbance estimate is restricted to When the disturbance estimate is less than At that time, the disturbance estimate is restricted to .

Claims

1. A model-free iterative learning traffic signal control method considering disturbance compensation, characterized in that, Specifically: Step 1: Collect traffic status data and construct a network-level congestion imbalance index; Step 2: Establish the historical iteration dataset and training samples for the random forest model; Step 3: Predict the initial green light duration control value based on the random forest model; Step 4: Estimate external disturbances and generate disturbance compensation amounts based on the extended state observer; Step 5: Construct a model-free adaptive iterative learning controller and control optimization objective; Step 6: Estimate pseudo-partial derivatives online and update model-free adaptive iterative learning control parameters; Step 7: Update the green light duration control value and generate the green light duration control strategy; Step 8: Generate and execute signal timing instructions to form a closed-loop iterative control process.

2. The model-free iterative learning traffic signal control method considering disturbance compensation according to claim 1, characterized in that, Step 1 is implemented as follows: Step 1.1: Determine the urban road traffic network to be controlled; wherein, the urban road traffic network includes multiple intersections, each intersection includes at least one approach link and a signal phase corresponding to the approach link; the signal phase includes east-west phase, north-south phase or other actually set release phases; Step 1.2: Collect real-time traffic data at each intersection; Specifically, real-time traffic data at each intersection is collected through a traffic detection system; the traffic detection system includes geomagnetic coil detectors, video detectors, microwave detectors, or other detection equipment capable of acquiring traffic operation status. The real-time traffic data includes the number of vehicles queuing at each intersection, the number of arriving vehicles, and the current green light duration. Step 1.3: Normalize the number of vehicles queuing at each intersection during each signal cycle to obtain the normalized congestion status of each intersection. Specifically, the normalized congestion state of the i-th intersection in the k-th signal cycle is defined as: in, Intersection In the Normalized congestion status for each signal cycle; Intersection In the Total number of vehicles in queue per signal cycle Intersection The sum of effective storage capacities; Number the intersections; Number the signal period; Step 1.4: Based on the normalized congestion status of each intersection obtained in Step 1.3, calculate the normalized congestion difference between intersections and construct a network-level congestion imbalance index: in, For the first Within the signal period, the first Normalized congestion differences corresponding to each intersection; and Intersections Intersection Normalized congestion status; Number the intersections; This represents the total number of crossover pairs in the network. Step 1.5: Use the network-level congestion imbalance index obtained in Step 1.4 as the output feedback signal; and input the output feedback signal, along with the current green light duration and the number of arriving vehicles, into the subsequent steps.

3. The model-free iterative learning traffic signal control method considering disturbance compensation according to claim 2, characterized in that, Step 2 shall be implemented in accordance with the following: Step 2.1: Store the traffic data processed in Step 1 according to the signal cycle; specifically, the stored data includes normalized congestion status, network-level congestion imbalance index, number of arriving vehicles, and current green light duration. Step 2.2: Iterate the traffic data stored in Step 2.1 according to the iterative learning control method. After each iteration cycle, record the normalized congestion status, tracking error and green light duration control value corresponding to each time step in the current iteration cycle to form historical input data and historical output data. Step 2.3: Filter the historical iteration data; Specifically, records with tracking errors lower than the average tracking error are selected from each iteration time step as historical samples with better control effect; the historical tracking errors are sorted from smallest to largest, and samples with errors in the first 80th percentile are retained, while samples with errors in the last 20th percentile are removed or marked as abnormal samples. Step 2.4: Construct the training sample set for the random forest model; The input features of the training sample set include the number of arriving vehicles, the current green light duration, the normalized congestion status, the network-level congestion imbalance index, and the historical tracking error; the output label of the training sample set is the green light duration control value in the corresponding historical sample.

4. The model-free iterative learning traffic signal control method considering disturbance compensation according to claim 3, characterized in that, Step 3 shall be implemented in accordance with the following: Step 3.1: Construct a regression-based random forest model; The random forest model consists of multiple regression decision trees, each of which is trained based on the training sample set constructed in step 2.

4. Each regression decision tree establishes a regression relationship between the intersection queue length and the green light duration control based on randomly selected samples and features. Step 3.2: Combine the current number of arriving vehicles, current green light duration, normalized congestion status, network-level congestion imbalance index, and tracking error corresponding to the most recent iteration obtained in Step 1 to form the input feature vector. , This represents the number of iterations. Step 3.3, input feature vector The results are input into each trained regression decision tree to obtain the prediction results of each regression decision tree for the initial green light duration of the current iteration cycle. , For the first Each regression decision tree is used for the current input feature vector The output is the predicted green light duration control value; Step 3.4: Prediction results of each regression decision tree. By performing average integration, the initial green light duration control value for the current iteration cycle is obtained: in, For the first The initial green light duration control value at the start of the next iteration; The total number of decision trees in the random forest; For the first Each regression decision tree is used for the current input feature vector The output is the predicted green light duration control value; Number the regression decision tree; This represents the number of iterations.

5. The model-free iterative learning traffic signal control method considering disturbance compensation according to claim 4, characterized in that, Step 4 shall be implemented in accordance with the following: Step 4.1, identify external disturbances; where external disturbances include traffic flow changes caused by school dismissal, concentrated entry and exit from large parking lots, temporary road construction occupying roads, sudden traffic flow in non-intersection areas, or other non-modeling factors; Step 4.2: Employ the extended state observer method; Specifically, to measure the control effect of the network-level congestion imbalance index, a target reference sequence corresponding to the network-level congestion imbalance index is first set. ,in, The target network-level congestion imbalance index is represented as a reference value; then the extended state observer uses the network-level congestion imbalance index calculated in step 1 as the feedback output and the target network-level congestion imbalance index as the reference input to estimate external disturbances online. Step 4.3, the deviation between the actual observed network-level congestion imbalance index and the target network-level congestion imbalance index is shown below: in, For the first In the nth iteration The perturbation estimate for each time step; For the first The perturbation estimate of the previous time step in the next iteration; For observer gain; For the first In the nth iteration The network-level congestion imbalance index is calculated at each time step based on the normalized congestion status. For the first The target network-level congestion imbalance index corresponding to the next iteration; Numbering the time steps in the iteration process; Step 4.4: Apply saturation constraints to the disturbance estimates obtained in Step 4.3 to obtain the disturbance estimates after saturation constraint processing. The details are as follows: in, This is the saturation constant for the disturbance estimate; when the disturbance estimate is greater than... At that time, the disturbance estimate is restricted to When the disturbance estimate is less than At that time, the disturbance estimate is restricted to .

6. The model-free iterative learning traffic signal control method considering disturbance compensation according to claim 5, characterized in that, Step 5 shall be implemented in accordance with the following: Step 5.1, define the iterative increment of the green light duration control quantity, the specific expression is as follows: in, No. In the nth iteration The change in green light duration control amount at each time step relative to the previous time step; For the first In the nth iteration The green light duration control quantity for each time step; For the first The green light duration control value of the previous time step in the next iteration; Step 5.2, define the normalized congestion difference tracking error, the specific expression of which is as follows: in, For the first In the nth iteration Normalized congestion difference tracking error at each time step; For the first The target network-level congestion imbalance index corresponding to the next iteration; For the first In the nth iteration The network-level congestion imbalance index is calculated at each time step based on the normalized congestion status. Step 5.3: Based on the iterative increment of the green light duration control quantity obtained in Step 5.1, the tracking error obtained in Step 5.2, and the disturbance estimation result obtained in Step 4, construct the objective function for the green light duration control quantity: in, The objective function for controlling the green light duration; This is a pseudo-partial derivative used to characterize the impact of changes in green light duration control on changes in normalized congestion differences; The disturbance estimate obtained from step 4 is input into the controller; To control the weight of the smoothing term, it is used to limit the variation of the green light duration control amount between adjacent time steps; This is an energy consumption penalty factor used to limit the adjustment range of the green light duration control. Step 5.4, adjust the objective function from step 5.3 with respect to the control input. Taking the partial derivative and setting it to zero, we obtain the optimal control law for the green light duration control variable: In formula (9), the first term controls the degree of inheritance of the green light duration of the current iteration relative to the green light duration of the previous time step; the second term is used to correct the green light duration based on the normalized congestion difference tracking error and the disturbance estimate; the meanings of the remaining symbols are the same as those in formula (8).

7. The model-free iterative learning traffic signal control method considering disturbance compensation according to claim 6, characterized in that, Step 6 shall be implemented in accordance with the following: Step 6.1, due to the pseudo-partial derivatives Since it does not rely on precise traffic flow models in advance, the pseudo-partial derivatives are estimated online based on the changes in green light duration control and network-level congestion imbalance indicators in step 5, thus constructing a cost function for online pseudo-partial derivative estimation. : in, For the first In the nth iteration The increment of the network-level congestion imbalance index calculated based on the normalized congestion status at each time step; The increment for controlling the green light duration; This is the estimate of the disturbance from the previous time step; This is the pseudo-partial derivative estimate from the previous time step; To update the smoothing factor, which is used to limit abrupt changes in pseudo-partial derivative estimates; Step 6.2: By minimizing the pseudo-partial derivative estimation cost function in Step 6.1, the pseudo-partial derivative estimate is obtained according to the following update formula: in, For the first In the nth iteration The pseudo-partial derivative estimates at each time step; To adjust the step size; This represents the observation increment used to characterize the changes in normalized congestion differences during the current iteration.

8. The model-free iterative learning traffic signal control method considering disturbance compensation according to claim 7, characterized in that, Step 7 shall be implemented in accordance with the following: Step 7.1, use the initial green light duration control value obtained in Step 3. The disturbance estimate obtained in step 4 The normalized congestion difference tracking error obtained in step 5.2 The pseudo-partial derivative estimation results obtained in step 6 are input together into the model-free adaptive iterative learning controller. Step 7.2, considering the disturbance estimation, update the green light duration control quantity in the following concise form: in, For the first In the nth iteration The green light duration control quantity for each time step; For the first The green light duration control value of the previous time step in the next iteration; This is a historical retention factor used to adjust the degree to which the current green light duration control value inherits from the historical green light duration control value; This is the error amplification factor, used to adjust the sensitivity of the control strategy to the response of tracking error and disturbance compensation results; To normalize the tracking error for congestion differences; This is the disturbance estimate after saturation constraint processing; Step 7.3: Perform boundary constraint processing on the green light duration control quantity obtained in Step 7.2 to ensure that the green light duration meets the safe operation requirements of the actual traffic signal controller. Specifically, the green light duration will be limited to the minimum green light duration. With maximum green light duration Within the constraints; when the calculated green light duration control value is less than At that time, take When the calculated green light duration control value is greater than At that time, take ; in, The minimum green light duration is preset; The preset maximum green light duration; Step 7.4: Determine the green light duration control quantity after boundary constraint processing as the green light duration control strategy for the corresponding signal phase of each intersection; the control strategy includes the green light duration of the east-west phase and the corresponding green light duration of the north-south phase, or includes the green light duration of other phases configured by the actual traffic signal controller.

9. The model-free iterative learning traffic signal control method considering disturbance compensation according to claim 8, characterized in that, Step 8 shall be implemented in accordance with the following: Step 8.1: Based on the green light duration control strategy for the corresponding signal phases of each intersection obtained in Step 7, generate signal timing instructions; wherein, the signal timing instructions include the intersection number, signal phase number, green light duration control amount, and corresponding execution signal cycle; Step 8.2: Send the signal timing command generated in step 8.1 to the signal controller of the corresponding intersection, and the signal controller will adjust the green light duration of each phase. Step 8.3: After the signal controller adjusts the green light duration of each phase, real-time traffic status data of each intersection is collected again, and the data is returned to Step 1 for normalization processing and network-level congestion imbalance index calculation. Step 8.4: Based on the newly collected traffic state data, repeat steps 2 to 8 to form a closed-loop control process of "traffic state collection - historical data update - initial control quantity prediction - disturbance estimation compensation - model-free iterative learning control - timing instruction execution", so that the network-level congestion imbalance index gradually converges to the target network-level congestion imbalance index.

Images