Traffic flow prediction method based on self-organizing fuzzy recursive random configuration network

Abstract

The invention discloses a traffic flow prediction method based on a self-organizing fuzzy recursive random configuration network, and the method comprises the steps: fusing multi-source traffic and environment data, and constructing an input feature vector through abnormal value processing, normalization and periodic time coding; then, a fuzzy rule antecedent structure is automatically constructed based on input features, and an independent recursive random configuration sub-reserve pool is configured for each rule and is used for depicting local nonlinear time sequence dynamics; a candidate rule generation and supervision screening mechanism is designed and introduced to screen fuzzy rules, rule processing is executed through residual projection gain, and self-organizing evolution of a network structure is achieved; and updating the weight of the output layer online by adopting a recursive least square algorithm. And finally, based on the rule triggering weight, performing weighted fusion on the output of each sub-reserve pool to obtain a traffic flow prediction result.

Description

Technical Field

[0001] This invention relates to the field of intelligent traffic flow prediction, specifically to a traffic flow prediction method based on a self-organizing fuzzy recursive random configuration network. Background Technology

[0002] Traffic flow prediction is a crucial foundational function of intelligent transportation systems, with direct applications in traffic signal optimization, congestion mitigation, emergency response, and travel services. The traffic conditions on urban roads and highways are influenced by both daily periodic patterns and are affected by unforeseen events and external environmental factors. Therefore, effective traffic flow prediction needs to simultaneously capture nonlinear temporal dependencies, adapt to environmental disturbances, and adaptively adjust the model structure as data distribution changes.

[0003] In existing technologies, classical statistical methods perform stably when handling linear and stationary time series tasks, but their ability to handle the strong nonlinearity and multi-factor coupling of traffic flow is insufficient. While deep learning models such as convolutional neural networks, recurrent neural networks, and long short-term memory networks have strong fitting capabilities, their training costs and parameter scales are often large, and their fixed model structures are not conducive to real-time online adaptive and interpretable analysis in streaming data scenarios. Fuzzy systems have good interpretability and are easier to deploy than traditional statistical methods and deep learning models, making them more suitable for large-scale, low-barrier deployments. However, single fuzzy models are limited in their performance when facing complex temporal dynamics, long- and short-term dependencies, and exogenous disturbances. Therefore, there is an urgent need for a method that can maintain the advantages of fuzzy systems while also improving the ability to capture temporal dynamics and automatically adjust its structure during operation to cope with complex traffic flow dynamics, providing a new solution for intelligent traffic flow prediction. Summary of the Invention

[0004] Purpose of the invention: To address the problems mentioned in the background art, this invention discloses a traffic flow prediction method based on a self-organizing fuzzy recursive random configuration network. By integrating multi-source traffic and environmental features, introducing fuzzy rule partitioning and recursive random configuration of sub-reservoir pools, and combining structural self-organization and online learning mechanisms, it achieves high-precision, real-time traffic flow prediction in complex traffic environments.

[0005] Technical solution:

[0006] This invention discloses a traffic flow prediction method based on a self-organizing fuzzy recursive random configuration network, the method comprising the following steps:

[0007] S1: Acquire traffic flow data and environmental data and perform preprocessing;

[0008] S2: Construct the input feature vector and input it into the traffic flow prediction model;

[0009] S3: Construct the traffic flow prediction model;

[0010] S3.1: Construct fuzzy rule antecedents using Gaussian membership functions, determine the rule center and width, and calculate membership degree and normalized triggering weight;

[0011] S3.2: Configure an independent recursive random sub-reservoir for each fuzzy rule, and update the state vector of the corresponding sub-reservoir according to the input feature vector;

[0012] S3.3: In the rule layer and recursive random configuration layer, a candidate rule generation and supervision screening mechanism is introduced to screen fuzzy rules; the effectiveness of candidate rules and their sub-reservoirs is evaluated by residual projection gain, and if the gain exceeds the threshold, they are included in the existing network; a rule evolution strategy is designed to carry out the self-organizing structure evolution of rules and their sub-reservoirs.

[0013] S3.4: Update the output weight parameters of each sub-reservoir;

[0014] S4: Based on the normalized triggering power of each fuzzy rule and the local output of its corresponding sub-reservoir, calculate and generate the predicted values ​​of the target traffic parameters.

[0015] Furthermore, the preprocessing process described in S1 includes: acquiring traffic flow data and environmental data, and cleaning, handling missing values, normalizing, and encoding periodic time features of the data.

[0016] Furthermore, the input feature vector described in S2 is composed of historical multi-step data of traffic flow, speed, and occupancy, along with temperature, rainfall, snowfall, cloud cover percentage, holiday identifiers, and periodically encoded time features. Among these, the historical window of the three traffic flow parameters is used to provide a dynamic input mode for the regional division of fuzzy rules, while environmental and time features are used to enhance the distinguishability of subsequent rule triggering and local dynamic modeling of sub-reservoir pools.

[0017] Furthermore, the specific steps in S3.1 include:

[0018] The membership function layer uses Gaussian membership functions to achieve a soft partition of the input space, thereby forming a fuzzy rule region suitable for local recursive modeling. The membership degree of the i-th rule in the j-th dimension is defined as follows:

[0019]

[0020] Among them, c i,d With σ i,d Let represent the center and width of the i-th rule on the d-th input dimension, respectively;

[0021] Each node in the rule layer corresponds to a fuzzy rule R. i(i = 1, 2, …, Q), where Q is the number of fuzzy rules activated in the current network, and each rule corresponds to an independent recursive sub-reservoir model, whose center vector and width vector determine the local input distribution area;

[0022] The normalization layer proportionally adjusts the trigger strength of different rules to obtain the relative contribution of the rules at the current sample. The normalized rule weights are defined as follows:

[0023]

[0024] in, Let i be the center vector of the i-th rule. It is a width vector.

[0025] Furthermore, the specific steps in S3.2 include: the recursive random configuration layer sets up a recursive sub-reservoir for each rule to capture the local nonlinear temporal dynamics within the rule's scope. Its state update is jointly determined by the random input weights, recursive weights, and leakage rate, and state stability is ensured through spectral radius constraints. The state of the recursive sub-reservoir corresponding to each rule is updated according to the following formula:

[0026]

[0027] in, The input weight matrix is ​​random. The recursive weight matrix is ​​scaled by the spectral radius such that... , Leakage rate.

[0028] Furthermore, the candidate rule generation and supervised candidate screening mechanism design described in S3.3 includes:

[0029] Calculate the current average error based on the sliding window length W:

[0030]

[0031] in, As a measure of prediction error, when an error increase is detected to exceed a set threshold, i.e. When this occurs, a candidate rule generation mechanism is triggered to address model structure deviations caused by changes in traffic patterns;

[0032] Candidate rule centers c are generated from clustering high residual samples. cand and width σ cand And randomly configure the parameter set {W} of the corresponding candidate sub-reserve pool. in,cand W r,cand , N cand ,λ cand}, to construct new structural units that allow for self-organizing expansion of the model;

[0033] For each candidate, a supervised effectiveness evaluation is performed using the residual projection gain Δ, and the residual projection gain index Δ is calculated:

[0034]

[0035] Where e is the current residual vector, P cand The prediction contribution vector of the candidate in the observation space. For numerically stable small constants, when Δ exceeds a preset threshold When this happens, the candidate fuzzy rule and its sub-reservoir are incorporated into the existing structure.

[0036] Furthermore, the rule evolution strategy design steps described in S3.3 include:

[0037] For rules with low long-term triggering, their contribution is determined by verifying the error and the deletion threshold before deletion. For rules with close center distances and highly similar local outputs, rule merging is performed to achieve continuous evolution and redundancy compression of the structure.

[0038] The specific criteria for rule merging and rule deletion are as follows:

[0039] Merge: When two rules R i R j satisfy Furthermore, its local output sequence similarity metric is greater than the threshold ρ. sim At that time, it was merged into a new rule R. ij Its center and width can be reset according to the weighted average or clustering results;

[0040] Deletion: When a rule is in consecutive T... inact Average triggering weight within a sliding window Furthermore, if the increase in verification error after deleting the rule does not exceed the preset tolerance γ, the rule and its sub-reservoir will be deleted to save computing resources.

[0041] Furthermore, S3.4 uses recursive least squares to update the output weight parameters of each sub-reservoir.

[0042] Furthermore, the specific steps of S4 include:

[0043] The local output of the i-th fuzzy rule at time t is calculated by the state of its corresponding recursive sub-reservoir and the output weight:

[0044]

[0045] in, This is the updated output weight vector;

[0046] The final predicted value is obtained by weighting and fusing the corresponding local outputs with the normalized trigger weights of all rules, in order to achieve interpretable partitioning modeling and combined prediction of multi-regional and multi-scale temporal dynamics. The final output predicted value can be expressed by the following formula:

[0047]

[0048] in, The trigger weights are for normalization.

[0049] Beneficial effects:

[0050] 1. This invention divides the traffic state space into several local regions with clear semantics through fuzzy rules, and configures an independent recursive random configuration sub-reservoir for each rule, so that the model can characterize the local dynamic features under different traffic conditions, thereby effectively capturing multi-scale temporal dependencies and further enhancing the interpretability of fuzzy systems in traffic flow prediction. Compared with traditional single global models, it can better adapt to the nonlinear changes in traffic flow.

[0051] 2. This invention employs randomly configured weights to construct sub-reservoirs and utilizes recursive least squares to update the output layer weights in real time, enabling the network to quickly adapt each time it receives new data, eliminating the need for overall retraining. This mechanism significantly reduces model training overhead, improves real-time performance, and makes this invention suitable for deployment in traffic management platforms and roadside edge devices. Compared to traditional deep neural networks, it offers higher computational efficiency and lower hardware requirements, further meeting the long-term online prediction needs of large-scale road networks.

[0052] 3. The present invention designs a candidate rule generation and supervised candidate screening mechanism, which can dynamically generate and combine fuzzy rules with corresponding sub-reservoir evaluation based on the sliding window error when the data distribution is drifted due to traffic patterns, weather changes or holidays. This avoids structural redundancy and enables the model to maintain appropriate complexity and good generalization ability in long-term operation, and maintain high-precision prediction performance in complex, time-varying and sudden traffic environments. Attached Figure Description

[0053] Figure 1 This is an overall flowchart of the traffic flow prediction method of the present invention.

[0054] Figure 2 This is a network structure diagram of the traffic flow prediction model of the present invention.

[0055] Figure 3 This is a comparison chart of the traffic flow prediction results of this invention with those of other models. Detailed Implementation

[0056] 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.

[0057] like Figure 1 As shown, this invention discloses a traffic flow prediction method based on a self-organizing fuzzy recursive stochastic configuration network. The proposed self-organizing fuzzy recursive stochastic configuration network employs a streaming data-driven approach for network construction and online prediction. The fuzzy rules of the network and their corresponding recursive stochastic configuration sub-reservoirs are not determined all at once, but are dynamically generated, updated, or deleted as data is gradually input over time. The network performs a prediction once at each time step when a new traffic feature vector is input, while simultaneously monitoring error changes and triggering a self-organizing process as needed to achieve timely adaptation to changes in traffic patterns, weather conditions, and periodic behaviors.

[0058] The experimental data came from the Traffic Experiment Platform at the University of Minnesota, USA. Traffic flow and environmental data collected every five minutes in July 2024 by detector number 6908 were selected. The first 70% of the data was used as the training set, and the last 30% as the test set. Traffic and environmental variables from 12 historical time steps were used as input to predict traffic flow within the next 15 minutes. The specific process is as follows:

[0059] Step 1: Data Collection and Preprocessing. Traffic flow data was obtained from the traffic experiment platform, and environmental data was obtained from the WunderGround weather website. The raw data underwent the following preprocessing:

[0060] (1) Replace missing and outlier values ​​using the historical mean method;

[0061] (2) Normalize continuous variables to the [0,1] interval using the min-max method;

[0062] (3) Use periodic encoding for time variables to convert time into:

[0063]

[0064] Where t represents the current time.

[0065] (4) Time-align the processed environmental data with the traffic flow data.

[0066] Step 2: Input Feature Vector Construction. Construct a historical window, combining historical multi-step data on traffic flow, speed, and occupancy. Combine temperature, rainfall, snowfall, cloud cover percentage, and holiday / weekday data. The final input feature vector can be represented as:

[0067]

[0068] Among them, h Q (t)=[Q(tk),…,Q(t-1)],h V (t)=[V(tk),…,V(t-1)],h O (t)=[O(tk),…,O(t-1)] represents the historical window characteristics of traffic flow, speed, and occupancy, respectively, where k is the length of the historical window. Temp(t), Rain(t), Snow(t), and Cloud(t) represent the temperature, rainfall, snowfall, and cloud cover percentage at time t, respectively. weekday(t) and Holiday(t) represent whether it is a weekday and a holiday at time t, respectively, and are represented by binary variables of 0 / 1.

[0069] The input feature vector can then be further represented as:

[0070]

[0071] Step 3: Calculate the traffic flow prediction model based on the self-organizing fuzzy recursive randomized network (SO-F-RSCN model is shown below), the model structure is as follows. Figure 2 As shown, the designed SO-F-RSCN model has a six-layer structure, including an input layer, a membership function layer, a normalization layer, a recursive random allocation layer, and an output layer.

[0072] Step 3.1: Fuzzy Rule Construction. The initial number of fuzzy rules is set to 0. The data input to the model is not trained all at once, but in batches over time steps. When the data from the first time step begins training, the first fuzzy rule is generated. Fuzzy C-means clustering is used to obtain the centers of the membership functions of the rule antecedents. At this point, Gaussian membership functions are used to construct the fuzzy rule antecedent structure. The membership degree of the fuzzy rule is defined as:

[0073]

[0074] Among them, c i and σ i These represent the rule center and width parameters, respectively.

[0075] After obtaining the membership degree of the current rule, its corresponding normalized triggering weight can be expressed as:

[0076]

[0077] When training begins with the data from the second time step, the data from the first two time steps collectively determine the fuzzy rules. At this point, the number of fuzzy rules increases to two. The rule centers are updated using fuzzy C-means clustering, and the membership degrees and normalized weights are updated sequentially. When training begins with the data from the third time step, the data from the first three time steps collectively determine the fuzzy rules. Based on the set rules growth, merging, and deletion conditions, corresponding fuzzy rule self-organization operations are performed, and the above process is repeated. The remaining input time steps are trained according to the above logic until all data is trained.

[0078] Step 3.2: Recursive Random Configuration Sub-Reserve Pool Construction and State Update. Each fuzzy rule in the preceding steps corresponds to an independent recursive sub-reserve pool, used to characterize local traffic dynamics. The state of the recursive sub-reserve pool corresponding to each rule is updated using the following formula:

[0079]

[0080] in, The input weight matrix is ​​random. The recursive weight matrix is ​​scaled by the spectral radius such that... , Leakage rate.

[0081] During training, the data at each time step not only self-organizes the change in the number of fuzzy rules, but also the number of nodes in each corresponding independent recursive sub-reservoir adjusts self-organizingly based on the current data.

[0082] Step 3.3: Self-organizing structure optimization. The model introduces a candidate rule generation and supervised screening mechanism to filter fuzzy rules. Based on the supervised gain evaluation results, it will perform self-organizing structure optimization by adding, merging, or deleting rules. The specific self-organizing optimization process is as follows:

[0083] (1) Use a sliding window to monitor the current prediction error;

[0084] (2) If the error continues to increase, candidate rules and sub-reservoirs are generated from the high residual samples;

[0085] (3) Calculate the residual projection gain Δ of the candidate structure. If Δ exceeds the threshold, add the rule to the model.

[0086] (4) Rules with consistently low activation rates will be deleted;

[0087] (5) If the scope of two rules overlaps and the consequent outputs are similar, then the rules are merged.

[0088] The candidate rule generation and supervised candidate screening process includes:

[0089] Calculate the current average error based on the sliding window length W:

[0090]

[0091] in, This is a measure of prediction error. When an error increase is detected to exceed a set threshold, i.e. When this occurs, the candidate rule generation mechanism is triggered.

[0092] Generate candidate rule centers c from high-error samples by clustering. cand and width σ cand And randomly configure the parameter set {W} of the corresponding candidate sub-reserve pool. in,cand W r,cand , N cand ,λ cand}, to construct new structural units that allow for self-organizing expansion of the model.

[0093] For each candidate, a supervised effectiveness evaluation is performed using the residual projection gain Δ, and the residual projection gain index Δ is calculated:

[0094]

[0095] Where e is the current residual vector, P cand The prediction contribution vector of the candidate in the observation space. It is a small numerically stable constant. When When this happens, the candidate fuzzy rule and its sub-reservoir are incorporated into the existing structure.

[0096] Rule merging and deletion are performed according to the following rule evolution strategy:

[0097] Rule merging: When two rules R i R j satisfy Furthermore, its local output sequence similarity metric is greater than the threshold ρ. sim At that time, it was merged into a new rule R. ij Its center and width can be reset according to the weighted average or clustering results;

[0098] Rule deletion: When a rule is deleted in consecutive T... inact Average triggering weight within a sliding window Furthermore, if the increase in verification error after deleting the rule does not exceed the preset tolerance γ, the rule and its sub-reservoir will be deleted to save computing resources.

[0099] Step 3.4: Update output weight parameters. Output weight W o,i The Recursive Least Squares (RLS) algorithm is used for updating. First, the prediction error at the current time step is calculated:

[0100]

[0101] Among them, y true (t) represents the actual traffic flow at the current moment. This represents the predicted traffic flow at the current moment.

[0102] Based on the local recursive state vector r i (t) and the covariance matrix P of the previous time step i (t-1) Calculate the gain vector:

[0103]

[0104] Where, λ R ∈(0,1] represents the forgetting factor.

[0105] The output weights are then updated as follows:

[0106]

[0107] Among them, K i e(t) is the gain vector, and e(t) is the prediction error.

[0108] The covariance matrix at each time step is updated as follows:

[0109]

[0110] Step 4: Output value calculation. The local output of the sub-reservoir corresponding to the i-th rule is:

[0111]

[0112] Based on the triggering power of each fuzzy rule and the local output of the corresponding sub-reservoir, the final traffic flow prediction value is calculated as follows:

[0113]

[0114] in, The trigger weights are for normalization.

[0115] Step 5: Model Training and Evaluation. The experiment was conducted using MATLAB software. The data used was downloaded from the Traffic Experiment Platform at the University of Minnesota, specifically traffic parameter data recorded every 5 minutes in July 2024 by detector number 6908, including traffic flow, speed, and occupancy. Simultaneously, relevant environmental data for the corresponding coordinates of the detector, including temperature, rainfall, snowfall, and cloud cover, were retrieved and downloaded from the WunderGround weather website. After preprocessing both sets of data, time alignment was performed to obtain a complete data table. A total of 2976 valid data sets were ultimately obtained, divided into training and test sets in a 7:3 ratio. The data from the first 12 time steps were used as model input to predict traffic flow for the next 3 time steps, i.e., the next 15 minutes.

[0116] Data is input gradually over time, and after the model is trained according to the aforementioned steps, the final output is obtained. To verify the model's performance, the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination R are introduced. 2 Three widely used evaluation indicators. The calculation methods for each evaluation indicator are as follows:

[0117]

[0118]

[0119]

[0120] in, Indicates the number of test samples. Indicates actual traffic flow. This indicates the predicted traffic flow. This represents the average actual traffic flow. These metrics can be used to evaluate the model's goodness of fit and predictive ability. Specifically, the smaller the values ​​of RMSE and MAE, the better the model's predictive performance; the closer R² is to 1, the stronger the model's explanatory power.

[0121] To further verify the effectiveness of the present invention,

[0122] The same data was also used in experiments on the F-RSCN, F-SCN, and SCN models, and the experimental results are compared as follows: Figure 3 The prediction results for each of the three time steps are shown.

[0123] Table 1

[0124]

[0125] Table 1 summarizes the evaluation metrics of the four models. The model proposed in this invention achieves the best performance in this traffic flow prediction task. Its RMSE, MAE, and R² are 18.5623, 13.6686, and 0.9686, respectively, significantly outperforming the three comparative models: F-RSCN, F-SCN, and SCN. In particular, compared with the traditional SCN, the error is reduced by more than 30%, and it also shows a significant improvement compared with F-RSCN, which has a recursive structure. This indicates that the self-organizing mechanism and supervised candidate selection mechanism introduced in this invention effectively enhance the model's temporal expressiveness and structural adaptability.

[0126] from Figure 3 As can be observed from the prediction curves, the model of this invention maintains a high degree of fit with the real data at peaks, fluctuation ranges, and abrupt change points, and can more accurately capture the local dynamic characteristics of traffic flow, while other models all exhibit peak bias or lag to varying degrees. Furthermore, the training time of the model of this invention is 6.44 seconds, which is shorter than that of the fixed-structure F-RSCN model, verifying that the self-organizing structure evolution strategy has higher training efficiency while ensuring prediction accuracy.

[0127] In summary, the experimental results show that the model of the present invention outperforms existing methods in terms of prediction accuracy, fitting stability, and computational efficiency, further demonstrating its effectiveness and technical advantages in real-world traffic flow prediction scenarios.

[0128] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent transformations or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A traffic flow prediction method based on a self-organizing fuzzy recursive stochastic configuration network, characterized in that, The method includes the following steps: S1: Acquire traffic flow data and environmental data and perform preprocessing; S2: Construct the input feature vector and input it into the traffic flow prediction model; S3: Construct the traffic flow prediction model; S3.1: Construct fuzzy rule antecedents using Gaussian membership functions, determine the rule center and width, and calculate membership degree and normalized triggering weight; S3.2: Configure an independent recursive random sub-reservoir for each fuzzy rule, and update the state vector of the corresponding sub-reservoir according to the input feature vector; S3.3: In the rule layer and recursive random configuration layer, a candidate rule generation and supervision screening mechanism is introduced to screen fuzzy rules; the effectiveness of candidate rules and their sub-reservoirs is evaluated by residual projection gain, and if the gain exceeds the threshold, they are included in the existing network; a rule evolution strategy is designed to carry out the self-organizing structure evolution of rules and their sub-reservoirs. S3.4: Update the output weight parameters of each sub-reservoir; S4: Based on the normalized triggering power of each fuzzy rule and the local output of its corresponding sub-reservoir, calculate and generate the predicted values ​​of the target traffic parameters.

2. The traffic flow prediction method based on a self-organizing fuzzy recursive random configuration network according to claim 1, characterized in that, The preprocessing process described in S1 includes: acquiring traffic flow data and environmental data, and cleaning, handling missing values, normalizing, and encoding periodic time features of the data.

3. The traffic flow prediction method based on a self-organizing fuzzy recursive random configuration network according to claim 1, characterized in that, The input feature vector described in S2 is composed of historical multi-step data of traffic flow, speed, and occupancy, along with temperature, rainfall, snowfall, cloud cover percentage, holiday identifiers, and periodically encoded time features. Among these, the historical window of the three traffic flow parameters is used to provide a dynamic input mode for the regional division of fuzzy rules, while environmental and time features are used to enhance the distinguishability of subsequent rule triggering and local dynamic modeling of sub-reservoir pools.

4. The traffic flow prediction method based on a self-organizing fuzzy recursive random configuration network according to claim 1, characterized in that, The specific steps in S3.1 include: The membership function layer uses Gaussian membership functions to achieve a soft partition of the input space, thereby forming a fuzzy rule region suitable for local recursive modeling. The membership degree of the i-th rule in the j-th dimension is defined as follows: ； Among them, c i,d With σ i,d Let represent the center and width of the i-th rule on the d-th input dimension, respectively; Each node in the rule layer corresponds to a fuzzy rule R. i (i = 1, 2, …, Q), where Q is the number of fuzzy rules activated in the current network, and each rule corresponds to an independent recursive sub-reservoir model, whose center vector and width vector determine the local input distribution area; The normalization layer proportionally adjusts the trigger strength of different rules to obtain the relative contribution of the rules at the current sample. The normalized rule weights are defined as follows: ； in, Let i be the center vector of the i-th rule. It is a width vector.

5. The traffic flow prediction method based on a self-organizing fuzzy recursive random configuration network according to claim 4, characterized in that, The specific steps in S3.2 include: The recursive random configuration layer sets up a recursive sub-reservoir for each rule to capture the local nonlinear temporal dynamics within the rule's scope. Its state update is jointly determined by the random input weights, recursive weights, and leakage rate, and state stability is ensured through spectral radius constraints. The state of the recursive sub-reservoir corresponding to each rule is updated using the following formula: ； in, The input weight matrix is ​​random. The recursive weight matrix is ​​scaled by the spectral radius such that... , Leakage rate.

6. The traffic flow prediction method based on a self-organizing fuzzy recursive random configuration network according to claim 5, characterized in that, The candidate rule generation and supervised candidate screening mechanism design described in S3.3 includes: Calculate the current average error based on the sliding window length W: ； in, As a measure of prediction error, when an error increase is detected to exceed a set threshold, i.e. When this occurs, a candidate rule generation mechanism is triggered to address model structure deviations caused by changes in traffic patterns; Candidate rule centers c are generated from clustering high residual samples. cand and width σ cand And randomly configure the parameter set {W} of the corresponding candidate sub-reserve pool. in,cand W r,cand , N cand ,λ cand }, to construct new structural units that allow for self-organizing expansion of the model; For each candidate, a supervised effectiveness evaluation is performed using the residual projection gain Δ, and the residual projection gain index Δ is calculated: ； Where e is the current residual vector, P cand The prediction contribution vector of the candidate in the observation space. For numerically stable small constants, when Δ exceeds a preset threshold When this happens, the candidate fuzzy rule and its sub-reservoir are incorporated into the existing structure.

7. The traffic flow prediction method based on a self-organizing fuzzy recursive random configuration network according to claim 6, characterized in that, The rule evolution strategy design steps described in S3.3 include: For rules with low long-term triggering, their contribution is determined by verifying the error and the deletion threshold before deletion. For rules with close center distances and highly similar local outputs, rule merging is performed to achieve continuous evolution and redundancy compression of the structure. The specific criteria for rule merging and rule deletion are as follows: Merge: When two rules R i R j satisfy Furthermore, its local output sequence similarity metric is greater than the threshold ρ. sim At that time, it was merged into a new rule R. ij Its center and width can be reset according to the weighted average or clustering results; Deletion: When a rule is in consecutive T... inact Average triggering weight within a sliding window Furthermore, if the increase in verification error after deleting the rule does not exceed the preset tolerance γ, the rule and its sub-reservoir will be deleted to save computing resources.

8. The traffic flow prediction method based on a self-organizing fuzzy recursive random configuration network according to claim 1, characterized in that, S3.4 The output weight parameters of each sub-reservoir are updated using the recursive least squares method.

9. The traffic flow prediction method based on a self-organizing fuzzy recursive random configuration network according to claim 1, characterized in that, The specific steps of S4 include: The local output of the i-th fuzzy rule at time t is calculated by the state of its corresponding recursive sub-reservoir and the output weight: ； in, This is the updated output weight vector; The final predicted value is obtained by weighting and fusing the corresponding local outputs with the normalized trigger weights of all rules, in order to achieve interpretable partitioning modeling and combined prediction of multi-regional and multi-scale temporal dynamics. The final output predicted value can be expressed by the following formula: ； in, The trigger weights are for normalization.

Images