A Short-Term Traffic Flow Prediction Method Based on Data Decomposition and Spatiotemporal Feature Fusion

Abstract

This invention provides a short-term traffic flow prediction method based on data decomposition and spatiotemporal feature fusion. First, urban traffic flow and related auxiliary data are collected. An improved algorithm incorporating a dynamic forgetting factor and regularization is used to decompose the traffic flow data. Data correlation analysis is then performed, and the decomposed data is integrated with the original data to construct an input matrix. A spatiotemporal feature fusion model is built to deeply mine the spatiotemporal features of the traffic flow. A robust ensemble deep model is used to initially output prediction results. Then, an optimization algorithm is used to optimize the model's hyperparameters, and finally, prediction biases are corrected. This method effectively improves the accuracy and anti-interference ability of short-term traffic flow prediction, and has stronger stability. It can provide accurate and reliable data support and decision-making reference for intelligent traffic management work such as urban smart traffic scheduling and road network traffic control.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a short-term traffic flow prediction method based on data decomposition and spatiotemporal feature fusion. Background Technology

[0002] With the simultaneous advancement of the "dual-carbon" strategy and smart city construction, traffic flow prediction has evolved from a traditional travel information service into a core upstream component of city-level traffic management, signal optimization, and carbon emission accounting. Short-term (5–30 min) traffic flow prediction, because it directly determines real-time strategies such as signal timing, guidance screen display, and public transport priority, requires significantly higher accuracy and robustness than medium- and long-term prediction. However, urban road networks exhibit "three highs and one weakness": high dynamics (morning and evening rush hours, sudden accidents), high coupling (upstream and downstream connections, regional spillover), high noise (detector drift, abnormal events), and weak prior knowledge (newly built roads, traffic restriction policies). Traditional statistical models (ARIMA, Kalman) assume linear stationarity, making it difficult to characterize the nonlinear abrupt changes in traffic flow. While single deep learning models (CNN, LSTM, GCN) can fit complex mappings, they face three major bottlenecks: (1) Input side – the original traffic flow sequence is mixed with multi-scale components such as trends, cycles, and events; (2) Structure side – strong spatiotemporal heterogeneity, neither the fixed adjacency of GCN nor the single temporal receptive field of LSTM can simultaneously take into account "spatial multi-order correlation + temporal multi-scale dependence"; (3) Output side – management departments face significant challenges in risk quantification assessment under extreme scenarios. Therefore, effectively decoupling the constituent elements of multi-source heterogeneous data at the data level and synergistically integrating spatiotemporal dual-domain feature information at the model level have become key common problems that urgently need to be solved to improve the accuracy and robustness of short-term traffic flow prediction. Summary of the Invention

[0003] Objective of the Invention: The technical problem to be solved by this invention is to address the shortcomings of existing technologies by providing a short-term traffic flow prediction method based on data decomposition and spatiotemporal feature fusion, which can provide a highly reliable decision-making basis for intelligent urban traffic management. The method includes the following steps:

[0004] Step 1: Collect historical traffic flow data by deploying detection equipment at key urban nodes;

[0005] Step 2: Obtain multi-dimensional urban data at the same time and spatial scale as historical traffic flow data, including meteorological data, macroeconomic indicators, and population dynamics data; clean, align, and merge the multi-dimensional urban data with historical traffic flow data to construct a multi-dimensional and integrated traffic flow dataset;

[0006] Step 3: Improve the recursive least squares (RLS) method by introducing a dynamic forgetting factor and regularization techniques, and integrate it with multivariate successive variational mode decomposition (MVMD) to decompose traffic flow data and obtain intrinsic mode function (IMF) components.

[0007] Step 4: Use Spearman's rank correlation coefficient to perform correlation analysis on each IMF component obtained from the decomposition in Step 3;

[0008] Step 5: By introducing the design spatial module GAT (Graph Attention Network) and the long short-term memory neural network LSTM (Long Short-Term Memory), a spatiotemporal feature fusion module based on the improved dual-feedback spatiotemporal graph convolutional network (GL-DFSTGCN) is built. The dual-feedback spatiotemporal graph convolutional network includes a spatial convolutional block based on the multiscale spatial graph convolutional network MSGCN (Multiscale Spatial Graph Convolutional Network) and a temporal convolutional block based on the Bayesian gate-atcn (Attention-based Temporal Convolutional Network). The temporal convolutional block is denoted as the Bayesian gate-atcn temporal module.

[0009] Step 6: Build an outlier robust ensemble deep OredRVFL-SLC model based on subspace local connectivity to predict traffic flow data and obtain short-term traffic flow prediction results.

[0010] Step 7: Optimize the hyperparameters of the prediction model using the improved global swarm optimization algorithm IHSO, and use the optimal parameters to predict the components. The improved global swarm optimization algorithm IHSO is based on the original global swarm optimization algorithm HSO. It adopts the best point set strategy to improve the original population initialization process and introduces the fitness balance optimization strategy in the algorithm position update stage to obtain the improved global swarm optimization algorithm IHSO.

[0011] Step 8: Further improve the stochastic vector function linking RVFL model by introducing the concept of generalized regularization, forming the GRRVFL model to correct the error of the residuals and obtain the final prediction result.

[0012] Step 1 includes:

[0013] Fixed road sections: Deploy fixed detection equipment at key nodes and key road sections of the urban road network;

[0014] Mobile devices: By using onboard GPS, Beidou dual-mode positioning terminals and onboard automatic diagnostic system OBD diagnostic interfaces deployed on taxis, buses, ride-hailing vehicles and freight vehicles, the instantaneous speed, latitude and longitude coordinates and driving direction angle of the vehicle are collected at a preset sampling frequency, and the travel speed and travel time of the road segment are estimated by map matching algorithm.

[0015] After time synchronization and outlier removal, the collected multi-source heterogeneous raw data are structured and stored according to time and spatial granularity to construct a historical traffic flow database covering multiple time scales.

[0016] Step 3 includes the following steps:

[0017] Step 3.1: Initialize all variables, including the modal components of MVMD. Center frequency and Lagrange multipliers And the RLS weights w, covariance matrix P, and regularization parameters. , where k is the mode index and c is the channel index;

[0018] Step 3.2, through harmonic shifting Gradient function L2 norm estimation The bandwidth, and the resulting cost function for:

[0019] ,

[0020] in, The analytical modulated signal is of channel number c and mode number k. Indicates will Adjust to the corresponding center frequency ; This is the regularization cost function used to estimate the bandwidth of each mode and constrain the smoothness of the signal; This indicates the derivative with respect to time t, representing the time gradient or rate of change;

[0021] Step 3.3: When constructing the variational constraint model of MVMD, the augmented Lagrangian function also needs to be adjusted accordingly:

[0022] ,

[0023] in, Let c be the number of channels and k be the number of modes at time t; Let c be the Lagrange operator for the number of channels at time t; The input data is the number of channels c at time t; is the penalty factor in the MVMD algorithm; C is the total number of channels; K is the total number of intrinsic mode functions to be obtained by decomposition; This represents the augmented Lagrangian function, which is the objective function to be minimized during MVMD optimization;

[0024] Step 3.4, in each iteration of MVMD, for each sample point of the signal Recursive execution enables real-time filtering and parameter updates; input vectors are constructed. :

[0025] ,

[0026] Where L is the filter order; T represents transpose; This represents the input column vector corresponding to the nth signal sample point. The element is from the current nth signal observation value Tracing back to One signal observation value;

[0027] Step 3.5, RLS output The estimated modes used to construct MVMD, error Feedback is fed back to parameter updates, forming a closed-loop adaptive system, and the gain vector is calculated. :

[0028] ,

[0029] in, It is the covariance matrix of the previous time step. For a dynamically adjusted forgetting factor, It is a parameter. The square of the L2 norm of the input vector. The formula is:

[0030] ,

[0031] in This represents the minimum value of the forgetting factor; This represents the maximum value of the forgetting factor; Functions for adjusting dynamic factors:

[0032] ,

[0033] in, It is a regulation function;

[0034] Step 3.6, Update the weight vector:

[0035] ,

[0036] in It is an L2 regularization term; It is iterated to the th The weight vector obtained after the step update; It is the first The Kalman gain vector of the step; Let n be the prediction error at step n;

[0037] Step 3.7, Update the covariance matrix :

[0038] ,

[0039] Where I is the identity matrix;

[0040] Step 3.8, calculate the filter output using the updated weight vector:

[0041] ,

[0042] in It is the output of the filter;

[0043] Step 3.9, calculate the error between the reference output and the actual output:

[0044] ,

[0045] in, It is a reference signal;

[0046] Step 3.10, RLS output The estimated modes used to construct MVMD are obtained by passing the error. Feedback is sent to the RLS parameter update team to determine whether all sample points have been completed. If not, return to step 3.4; if completed, perform MVMD variational optimization.

[0047] Step 3.11: Iterative update using the alternating direction method of multipliers. , and The optimal solution to the variational optimization problem is obtained, and the solution process includes mode update and center frequency update of the variational optimization problem.

[0048] Modal update to:

[0049] ,

[0050] Where n is the iteration number, n = 0, 1, 2, ..., N; N represents the maximum number of iterations in the iteration process; Indicates the first Each channel has an input signal at a frequency of [frequency]. Frequency domain estimation or observed spectral values ​​at a given location; Indicates the first The first in the passage Estimation of the amplitude coefficients of each component; Indicates the first The first in the passage Each frequency component in Frequency domain basis function or component spectrum estimation at the specified location; Indicates the first The result of the iteration is the first One modal spectrum, Indicates from the first The modality to the first Frequency domain estimates of each mode;

[0051] The center frequency has been updated to:

[0052] ,

[0053] in, For the t-th iteration, the Fourier transform of the k-th component in the c-th channel is given, where w is the angular frequency and d represents the derivative. It represents the (n+1)th moment weighted average frequency of the k-th component (k-th mode).

[0054] In step 4, the Spearman rank correlation coefficient for:

[0055] ,

[0056] in, It is the first The difference in rank (ranking) between two variables (a candidate feature variable and a target output variable) in a pair of samples.

[0057] In step 5, the improved dual-feedback spatiotemporal graph convolutional network GL-DFSTGCN performs the following steps:

[0058] Step 5.1: After performing Spearman rank correlation analysis on the original multi-source traffic flow data, a multi-source traffic flow dataset S strongly correlated with the target road segment is obtained. S is then input into the improved dual-feedback spatiotemporal graph convolutional network GL-DFSTGCN, which uses a sliding window sampling method to construct sliding samples covering different road segments and different time granularities, thus obtaining the input tensor. Where R represents the real number space, B represents the batch size of the input samples, and S represents the time step length of the historical traffic flow data obtained by sliding window sampling; This represents the number of input feature channels for each node;

[0059] Step 5.2, the improved dual-feedback spatiotemporal graph convolutional network GL-DFSTGCN models spatial dependencies and dynamic temporal patterns through a dual-branch architecture. The temporal convolutional block and the spatial convolutional block together constitute the spatiotemporal convolutional block ST-Conv block.

[0060] The improved dual-feedback spatiotemporal graph convolutional network GL-DFSTGCN includes spatial convolutional blocks based on the multiscale spatial graph convolutional network MSGCN (Multiscale Spatial Graph Convolutional Network) and temporal convolutional blocks based on the Bayesian gated attention temporal convolutional network ATCN (Attention-based Temporal Convolutional Network). The Bayesian gated attention temporal convolutional network ATCN block includes Bayesian TCN blocks, LSTM, GRU, and attention mechanisms.

[0061] Spatial features are extracted from a multi-source traffic flow dataset S using MSGCN. In MSGCN, different node aggregation orders can preserve spatial correlation information at different levels in the road network. For the i-th node The features on the l-th graph convolutional layer GCLayer are calculated using the following formula:

[0062] ,

[0063] ,

[0064] Where AGG is the aggregation function; CNN(•) represents the convolution operation for feature extraction. Representatives and nodes Corresponding slip samples, For the i-th node in the l-th graph convolutional layer GCLayer and the j-th node Edge features between (•) denotes a similarity function; W is a learnable weight matrix, which belongs to the training parameters of the graph convolutional layer GCLayer; This represents the activation function. Represents a node The set of neighboring nodes, Represents a node In the Feature representations on each graph convolutional layer Represents a node In the Feature representations on a graph convolutional layer;

[0065] Step 5.3: Construct the design space GAT module. After the graph convolutional network GCN and the design space module GAT extract spatial features from different angles through their two channels, the two feature outputs extracted by the graph convolutional network GCN and the design space module GAT are fed into the subsequent Bayesian TCN module. An attention mechanism is introduced to automatically determine the importance of adjacent nodes and calculate the attention coefficient. :

[0066] ,

[0067] in, Represents a node Relative to node Attention score Represents a node The set of adjacent nodes; Q and are the training parameters. Represents the characteristics of a node; exp is the natural exponential function; It is an activation function; Represents a node eigenvectors, Represents a node eigenvectors, Represents a node eigenvectors, Represents a node eigenvectors;

[0068] Step 5.4: After fusing the dual-channel outputs of the Graph Convolutional Network (GCN) and the Design Space Module (GAT), the final weighted output of the Multi-Scale Spatial Graph Convolutional Network (MSGCN) is obtained. Represented as: ,

[0069] in, This represents the weights corresponding to the p-order sensor output. It is the output of a P-order perceptual convolutional layer. Indicates the first The correlation between the first-order feature and all-order features; Indicates the first The convolutional layer outputs features corresponding to the order perception.

[0070] Step 5.5: Input spatial features into the Bayesian gate-atcn temporal module to analyze temporal dependencies; use the Bayesian gate-atcn temporal module to analyze the temporal dependencies of extracted features; LSTM and TCN form a dual-channel serial fusion temporal model; input the temporal correlations extracted from the Bayesian TCN block into LSTM and GRU; and use a global attention mechanism to reconstruct the output of the hidden layer, sequence elements. The unrolled convolution is calculated as follows:

[0071] ,

[0072] in, The Dilated Causal Convolution (DCC) operator; b is the size of the convolution kernel; b is the unwinding convolution operator; W is the output feature at the position after this DCC convolution operation. The value at; Input features In position The eigenvalue at that location; dr is the expansion ratio; Indicates the time position The output feature value after DCC convolution, Indicates the first convolution kernel Weight parameters for each position;

[0073] Step 5.6: Introduce Bayesian theory into the TCN framework. Given the input samples Z={X,Y} of the Bayesian TCN, the Bayesian TCN will assign all model parameters δ∈ Set from posterior distribution The random variable for sampling is determined as follows:

[0074] ;

[0075] Where X is the input feature dataset or input sample; Y is the output label or target dataset, corresponding to the true annotation and prediction target of the input sample X; It represents the explanatory power of the current parameter δ on the training data (X,Y); This represents the prior probability, indicating the probability of obtaining the model parameters before seeing the data. The prior assumptions; Indicates observed data The overall probability;

[0076] Step 5.7: Repeatedly stack the spatiotemporal convolutional blocks ST-Conv blocks constructed in steps 5.2 to 5.6 into M layers to extract higher-order nonlinear spatiotemporal features layer by layer, and then convert the output of the m layers... As the input of the (m+1)th layer, the final output is... ∈ ,in, and The first The number of channels and the length of the temporal dimension of the layer; the spatiotemporal convolutional block ST-Conv includes the multi-scale spatial graph convolutional network MSGCN and the Bayesian gate-atcn temporal module;

[0077] Step 5.8, output of the last ST-Conv block Execution-time convolution maps the historical time dimension to the prediction time step. The calculation formula is as follows:

[0078] ,

[0079] in, This represents the output feature of the last spatiotemporal convolutional block; Z represents the time step to be predicted; Z represents the final output feature tensor obtained after this temporal convolution operation. This represents the trainable convolution kernel parameters of the output layer's temporal convolution; This represents the number of channels for the output features of the M ST-Conv blocks; Represents the convolution operator in the time dimension.

[0080] Step 6 includes the following steps:

[0081] Step 6.1 transforms the problem of minimizing structural risk and empirical risk into the problem of minimizing the L2 norm output weight β, making the training error e sparse, as expressed in:

[0082] ,

[0083] Where y is the label vector of the training sample or the observation output vector; H is the input feature matrix; Indicates constraints. Indicates a To optimize the objective function that minimizes the variables;

[0084] Step 6.2: In the initial stage of the Augmented Lagrange Method (ALM) iteration, the unknowns e and β are updated to achieve minimization. The iterative steps are as follows:

[0085] ,

[0086] ,

[0087] ,

[0088] Where β is the parameter vector; e is the error vector to be optimized; To find the independent variable that minimizes the objective function; The penalty factor is The augmented Lagrangian function; H is the observation matrix, Indicates the first The parameter vector after +1 iterations. Indicates the first The error vector after +1 iterations.

[0089] This represents the penalty correction term for the constraint residuals. Indicates the first Lagrange multiplier vectors at the next iteration;

[0090] Step 6.3: Define the decision entropy of each attribute subspace based on the posterior probability of the class in the given attribute subspace. Then, iteratively update the attribute subspace. In each iteration, select the attribute subspace with the minimum decision entropy as the update candidate. When the outlier robust ensemble depth OredRVFL-SLC based on subspace local connectivity reaches a stable verification accuracy, stop the iteration and take the current attribute subspace as the optimal attribute subspace.

[0091] Step 6.4, Set the original attribute space It is divided into k disjoint subspaces: ,

[0092] Where J is the set of all original features; Represents the first in the original attribute space One characteristic, Indicates the first A feature subspace, Indicates the first The first subspace One feature;

[0093] Step 6.5, calculate the decision entropy: ; It is the first Individual indicators The corresponding decision entropy value;

[0094] Step 6.6: Sort the attribute subspaces in ascending order according to the magnitude of the decision entropy: ;

[0095] Here, J represents a single attribute subset, that is, an attribute subspace, which is a subset of attributes partitioned from the original full attribute space; Indicates the sorted order of the first... Each attribute subspace;

[0096] Step 6.7, using the initial attribute subspace Train the OredRVFL-SLC and obtain the initial validation accuracy a0; use the decision entropy as the criterion for judging the quality of attribute grouping. In each iteration, prioritize the fusion of attribute subspaces that are more conducive to the classifier's classification, then train an OredRVFL-SLC and obtain the validation accuracy. If the validation accuracy increases, determine the final attribute subspace for this iteration and proceed to the next iteration; if the combination of all attribute subspaces still cannot improve the validation accuracy after iteration, stop training and output the final OredRVFL-SLC.

[0097] Step 7 includes the following steps:

[0098] Step 7.1: Set the population size, dimension, number of iterations, and upper and lower bounds of the search space for the HSO algorithm;

[0099] Step 7.2, In the population initialization phase, introduce the optimal point set strategy to improve the algorithm initialization:

[0100] ,

[0101] ,

[0102] Where F is the population size, and M is a number greater than F. It is the generating basis of the set of best points in dimension d. It is the i-th hyperparameter value of the k-th individual. It is the k-th candidate individual (solution / search point) in the initial population.

[0103] Step 7.3: Generate the initial positions of all search agents within the search space after improvement based on the optimal point set strategy. ;

[0104] Step 7.4, calculate the fitness value for each search agent. ;

[0105] Step 7.5, Selection based on adaptive simulated annealing: The selection process based on simulated annealing includes dynamic temperature updates for each iteration, as follows:

[0106] ,

[0107] in, The effective number of iterations is CR(k), which is the adaptive cooling coefficient. The temperature of the current iteration; This represents the temperature value at the start of simulated annealing, and it is usually set relatively large to give the algorithm a strong global search capability in the initial stage. Indicates the minimum temperature;

[0108] Step 7.6: If the new fitness value is better than the previous value, the algorithm runs greedily, replacing the old solution with the new one. The following criteria apply to the selection:

[0109] ,

[0110] ,

[0111] when ;

[0112] Where P is the acceptance probability of accepting the inferior solution (the new candidate solution with a worse fitness); Indicates the fitness difference. Indicates a new candidate solution The corresponding fitness value, Indicates the current old solution The corresponding fitness value, Represents a uniformly random number within the interval [0,1].

[0113] Step 7.7, calculate the root mean square of all fitness values. :

[0114] ;

[0115] Where, f( ) indicates the first The fitness value of each individual;

[0116] Step 7.8: Calculate the coefficients of the displacement vector based on the root mean square difference of each fitness value. The individual differences are: ;

[0117] in This represents the difference for the i-th individual;

[0118] Step 7.9: The difference between coefficients c and the root mean square value is positive, and the difference between coefficients c and the root mean square value is negative. The sum of the normalized absolute values ​​is 1. The expression for coefficient c is:

[0119] ;

[0120] in, This represents the result after normalizing the i-th coefficient; sign is the sign function, which checks if the input is a positive number. If the input is negative, If the input is 0, ;

[0121] Step 7.10, Location Update: The location of the searched individual is iteratively updated. A fitness balance optimization strategy is introduced during the location update phase, with the following update rules:

[0122] ,

[0123] in, It is the average fitness of the t-th generation population. It is the position of the individual with the best fitness in the population. It is the historical best position of individual i. It is the position of the least fit individual in the population. , , , These are weighting coefficients. , , , yes Random numbers between; Indicates the first The individual in the first Current position at the next iteration Indicates the first The fitness value of each individual at its current position;

[0124] Step 7.11, Adaptive Mutation: Adaptively define the mutation operator and dynamically adjust the mutation rate and mutation step size according to the current iteration.

[0125] ,

[0126] ;

[0127] in, This represents the mutation rate at the current iteration step; Indicates the initial rate of variation; Indicates the final mutation rate; Indicates the variable length of the current iteration step; Indicates the current iteration number. Indicates the maximum number of iterations. Indicates the initial variable asynchronous length. This represents the final variable asynchronous length at the end of the algorithm iteration;

[0128] Step 7.12, if a mutation occurs, add a random perturbation to the individual's location:

[0129] ,

[0130] ;

[0131] in, It is the first The optimized individual in the iteration... The positional disturbance of the generation. This represents a standard normal distribution (Gaussian distribution), which means generating a standard normal random number with a mean of 0 and a variance of 1.

[0132] Step 7.13: Search for the global optimal solution based on the position update formula to obtain the optimal parameter combination of the OredRVFL-SLC model;

[0133] Step 7.14: Determine if the maximum number of iterations has been reached. If not, return to step 7.4 until the maximum number of iterations is reached, and output the optimal parameter configuration. The OredRVFL-SLC model uses the optimal parameter configuration to predict the components and obtain the final prediction result.

[0134] Step 8 includes the following steps:

[0135] Step 8.1, Initialization. Randomly generate parameters for hidden nodes. , Output the weight matrix B and the auxiliary variable Z;

[0136] Step 8.2, construct the objective function, the calculation formula is:

[0137] ,

[0138] Where K is the hidden layer output matrix; L is the target output; and All are regularization parameters;

[0139] Step 8.3: Calculate the error between the initial predicted values ​​and the actual values ​​of the OredRVFL-SLC model, and correct it using the GRRVFL model:

[0140] ,

[0141] Where z is the error and n is the actual value. These are the initial predictions from the OredRVFL-SLC model;

[0142] Step 8.4: The GRRVFL model includes an input layer, a randomized hidden layer, and an output layer; randomly generate the weights and biases of the hidden layer nodes, and calculate the output K of the hidden layer.

[0143] ,

[0144] in, is the activation function, D is the randomly generated weight matrix, b is the bias vector, and G is the input feature matrix;

[0145] Step 8.5, learn the output weights under regularization constraints This makes the model output close to the actual value of the error term:

[0146] ,

[0147] ,

[0148] in, It is a generalized regularization term;

[0149] Step 8.6, Calculate the error prediction value :

[0150] ;

[0151] Step 8.7: Add the initial predicted value of the OredRVFL-SLC model to the corrected error predicted value to obtain the final predicted value. :

[0152] ,

[0153] in, These are the initial predictions from the OredRVFL-SLC model.

[0154] The present invention also provides an electronic device, including a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method.

[0155] The present invention also provides a storage medium storing a computer program or instructions that, when the computer program or instructions are run on a computer, execute the steps of the method described.

[0156] Beneficial Effects: 1. Traffic flow data is decomposed using an improved recursive least squares method fused with multivariate successive variational mode decomposition (RLS-MVMD). RLS-MVMD's dynamic forgetting factor and regularization mechanism adaptively handle noise and non-stationary characteristics in the data, avoiding overfitting and thus extracting cleaner traffic flow features. Spearman's rank correlation coefficient is suitable for nonlinear data and can effectively capture the correlation between traffic flow and other urban variables (such as weather events), ensuring that the input matrix contains multidimensional information. 2. An improved dual-feedback spatiotemporal graph convolutional network (improved dual-feedback spatiotemporal graph convolutional network GL-DFSTGCN) module is constructed. The GAT module focuses on spatial dependencies (such as the influence of adjacent roads), and the LSTM module handles temporal trends (such as periodic changes in traffic flow). After the three are fused, the model can more accurately model the complex spatiotemporal evolution of traffic flow, avoiding the limitations of the single perspective of traditional methods. 3. The core prediction model adopts an improved outlier robust ensemble deep model (OredRVFL-SLC). The decision-making effectiveness of attribute grouping is evaluated by defining a criterion called decision entropy, and then the optimal attribute grouping for network training is found through continuous iteration. 4. An improved global swarm optimization algorithm (IHSO) is used to optimize the model hyperparameters. This algorithm avoids getting trapped in local optima through an improved explore-exploitation balance, ensuring that model parameters (such as the weights of OredRVFL-SLC) converge quickly to the optimal state. 5. The GRRVFL error correction mechanism reduces accumulated error, resulting in more reliable final predictions. Attached Figure Description

[0157] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0158] Figure 1 A schematic diagram of the improved dual-feedback spatiotemporal graph convolutional network GL-DFSTGCN model provided by the present invention.

[0159] Figure 2 A flowchart illustrating the improved HSO algorithm (IHSO) provided by this invention.

[0160] Figure 3 The structural diagram of the edRVFL model provided by this invention.

[0161] Figure 4 This is a schematic diagram of the process for optimizing the OredRVFL-SLC model using the IHSO algorithm provided in this invention. Detailed Implementation

[0162] Combination Figure 1 , Figure 2 and, Figure 3 and Figure 4This invention provides a short-term traffic flow prediction method based on data decomposition and spatiotemporal feature fusion, comprising the following steps:

[0163] Step 1: Collect historical traffic flow data (including core indicators such as flow rate, speed, and occupancy rate) by deploying various detection devices (such as loop detectors, cameras, radar, floating car GPS, etc.) at key urban nodes (such as intersections and road sections).

[0164] Step 2: Acquire multi-dimensional urban data at the same time and spatial scale as the historical traffic flow data, including meteorological data (such as weather conditions, temperature, precipitation, and visibility), macroeconomic indicators (such as GDP and holidays), and population dynamics data (such as commuting tides and information on large-scale events). Clean, align (timestamp, spatial location matching), and merge these heterogeneous data with the historical traffic flow data acquired in Step 1 to construct a multi-dimensional and integrated traffic flow dataset.

[0165] Step 3: Improve the recursive least squares (RLS) method by introducing a dynamic forgetting factor and regularization techniques, and integrate it with multivariate variational mode decomposition (MVMD) to form a (RLS-MVMD) decomposition of the traffic flow data into a series of relatively stable, frequency band-separated intrinsic mode function (IMF) components.

[0166] Step 4: Use Spearman's rank correlation coefficient to perform correlation analysis on each IMF component obtained from the decomposition in Step 3;

[0167] Step 5: By introducing the design spatial module GAT (Graph Attention Network) and the long short-term memory neural network LSTM (Long Short-Term Memory), a spatiotemporal feature fusion module based on the improved dual-feedback spatiotemporal graph convolutional network (improved dual-feedback spatiotemporal graph convolutional network GL-DFSTGCN) is built. The dual-feedback spatiotemporal graph convolutional network is based on the spatial convolutional block of the multiscale spatial graph convolutional network MSGCN (Multiscale Spatial Graph Convolutional Network) and the temporal convolutional block of the Bayesian gate-atcn (Attention-based Temporal Convolutional Network) (Bayesian gate-atcn temporal module).

[0168] Step 6: Build an Outlier-robustted RVFL with Subspace-based Local Connections (OredRVFL-SLC) model to predict the traffic flow data and obtain short-term traffic flow prediction results.

[0169] Step 7: Optimize the hyperparameters of the prediction model using the improved Holistic swarm optimization (IHSO) algorithm, and use the optimal parameters to predict the components. The IHSO algorithm is based on the original Holistic swarm optimization algorithm HSO, and improves the original population initialization process by adopting the best point set strategy. In the algorithm position update stage, a fitness balance optimization strategy is introduced to obtain the improved IHSO algorithm.

[0170] Step 8: Further improve the stochastic vector function linking RVFL model by introducing the concept of generalized regularization, forming the GRRVFL model to correct the error of the residuals and obtain the final prediction result.

[0171] Step 1 includes:

[0172] Fixed road sections: Fixed detection equipment is deployed at key nodes of the urban road network (including but not limited to signalized intersections, unsignalized intersections, expressway entrance and exit ramps, and bridge-tunnel connection sections) and key road sections (including but not limited to bottleneck sections, long downhill sections, bus lanes, and tidal flow lanes); Mobile devices: Through vehicle-mounted GPS / BeiDou dual-mode positioning terminals (Floating Car Data, FCD) and OBD (On-Board Diagnostic) interfaces deployed on taxis, buses, ride-hailing vehicles, and freight vehicles, the instantaneous speed, latitude and longitude coordinates, and driving direction angle of the vehicles are collected at a preset sampling frequency (preferably 10-60 seconds / time), and then the road segment travel speed (LinkTravelSpeed) and travel time (TravelTime) are estimated through map matching algorithms.

[0173] Step 2 includes:

[0174] After time synchronization and outlier removal, the collected multi-source heterogeneous raw data is structured and stored according to time granularity (preferably 5 minutes, 15 minutes, and 1 hour aggregation) and spatial granularity (lane level, road segment level, and region level) to build a historical traffic flow database covering multiple time scales (daily / weekly / seasonal patterns), providing training samples for subsequent prediction models.

[0175] Step 3, RLS-MVMD decomposition, includes the following steps:

[0176] Step 3.1: Initialize all variables, including the modal components of MVMD. Center frequency and Lagrange multipliers (where k is the mode index and c is the channel index), and the RLS weights w, covariance matrix P, and regularization parameters. The weights and covariance matrix of RLS are initialized to zero or small values, which facilitates rapid adaptation to signal changes during MVMD iteration;

[0177] Step 3.2, through harmonic shifting Gradient function L2 norm estimation The bandwidth, and the resulting cost function for:

[0178] ,

[0179] in, The analytical modulated signal is of channel number c and mode number k. Indicates will Adjust to the corresponding center frequency , This is the regularization cost function used to estimate the bandwidth of each mode and constrain the smoothness of the signal; This indicates the derivative with respect to time t, and represents the time gradient or rate of change.

[0180] Step 3.3: Considering the existence of multiple linear equality constraints, the augmented Lagrangian function also needs to be adjusted accordingly when constructing the variational constraint model of MVMD.

[0181] ,

[0182] in, The modulated signal is for channel number c and mode number k; The Lagrange operator for the channel number c; The input data is for the number of channels, c. is the penalty factor in the MVMD algorithm; K is the total number of eigenmode functions to be obtained from the decomposition. This represents the augmented Lagrangian function, which is the objective function to be minimized during MVMD optimization.

[0183] Step 3.4, in each iteration of MVMD, for each sample point of the signal Recursive execution enables real-time filtering and parameter updates. The RLS output is used to construct the "estimated mode" of the MVMD, and the error signal is used to adjust subsequent parameters. The input vector is constructed. :

[0184] ,

[0185] Where L is the filter order; This represents the input column vector corresponding to the nth signal sample point. The elements of this vector are the values ​​from the current nth signal observation. tracing back to Each signal observation value.

[0186] Step 3.5, RLS output The "estimated mode" used to construct MVMD, while the error Feedback is fed back to parameter updates, forming a closed-loop adaptive system. This loop is executed for all sample points until completion, calculating the gain vector:

[0187] ,

[0188] in, It is a gain vector. It is the covariance matrix of the previous time step. It is the input vector. For a dynamically adjusted forgetting factor, It is a parameter. The gain vector is the squared L2 norm of the input vector. Used to adjust the update magnitude of the weight vector. The calculation of the gain vector takes into account the forgetting factor. The influence of forgetting factor Change to a dynamic function to control the forgetting factor. With the error signal e(n) and the adjustment parameter The weights of historical data are adjusted in real time to ensure the model simultaneously meets the requirements of fast tracking speed and small stability error, thus ensuring the algorithm can adapt to changes in the input signal. The formula is:

[0189] ,

[0190] in The minimum value of the forgetting factor

[0191] The maximum value of the forgetting factor.

[0192] in Functions for adjusting dynamic factors:

[0193] ,

[0194] in, This is an adjustment function used for control. The value indirectly controls The range of variation, The larger, The faster the response to changes in the error signal; The smaller, The smoother the response to changes in error signals, the better. (Dynamic forgetting factor) Adaptive adjustment is achieved through the error signal e(n), with the nonlinear response driven by the square of the error;

[0195] Step 3.6, Update the weight vector:

[0196] ,

[0197] in This is the error signal. An L2 regularization term has been added here. To enhance the model's generalization ability and prevent overfitting, the weights are forcibly multiplied by a coefficient at each update step. This causes the absolute value of the weight to shrink proportionally; It is iterated to the th The weight vector obtained after the step update; It is the first The Kalman gain vector of the step.

[0198] Step 3.7, Update the covariance matrix:

[0199] ,

[0200] Where I is the identity matrix and the covariance matrix. The update reflects the algorithm's ability to adapt to changes in the input signal. The update process considers the gain vector. The influence of the input vector u(n) is considered, while the weights of historical data are adjusted through the forgetting factor λ.

[0201] Step 3.8, Calculate the output: Use the updated weight vector to calculate the filter's output.

[0202] ,

[0203] in It is the output of the filter; after updating the weight vector, the weight vector at the current time step is used. and input vector Calculate the output of the filter Output signal It is the result of the filter processing the input signal, used for error calculation and system feedback;

[0204] Step 3.9, calculate the error between the reference output and the actual output:

[0205] ,

[0206] in, It is an error signal, and it is a reference signal. With filter output The differences between them.

[0207] Step 3.10, RLS output The "estimated modes" used to construct MVMD are obtained through error. Feedback is sent to the RLS parameter update team to determine whether all sample points have been completed. If not, return to step 3.4; if completed, perform MVMD variational optimization.

[0208] Step 3.11: Iterative update using the alternating direction method of multipliers. , and The optimal solution to the variational optimization problem is obtained, and the solution process includes mode update and center frequency update of the variational optimization problem.

[0209] Modal update to:

[0210] ,

[0211] in, , , These are the signals after Fourier transform; n is the iteration number, n = 0, 1, 2, ..., N; N represents the maximum number of iterations during the iteration process; Indicates the first Each channel has an input signal at a frequency of [frequency]. Frequency domain estimation / observed spectral values ​​at the location; Indicates the first In the passage, the first Estimation of the amplitude coefficients of each component; Indicates the first The first in the passage Each frequency component in Frequency domain basis function or component spectrum estimation at the specified location;

[0212] Indicates the first The result of the iteration is the first Modal spectrum

[0213] Indicates from the first The modality to the first Frequency domain estimates of each mode

[0214] The center frequency has been updated to:

[0215] ,

[0216] in, For the t-th iteration, the Fourier transform of the k-th component in the c-th channel is given, where w is the angular frequency and d represents the derivative. Let C represent the (n+1)th moment weighted average frequency of the k-th component (k-th mode), and C be the total number of channels.

[0217] Step 4, the Spearman rank correlation coefficient, includes the following steps:

[0218] Spearman rank correlation coefficient ,

[0219] Among them, the correlation coefficient The larger the absolute value of the correlation coefficient (the closer the correlation coefficient is to 1 or -1), the stronger the correlation between X and Y; The closer to 0, the weaker the correlation between X and Y; among them, It is the first The difference in rank (ranking) between two variables (a candidate feature variable and a target output variable) in a pair of samples.

[0220] The improved dual-feedback spatiotemporal graph convolutional network GL-DFSTGCN model in step 5 performs the following steps:

[0221] Step 5.1: After performing Spearman rank correlation analysis on the original multi-source traffic flow data, select the multi-source traffic flow dataset that is strongly correlated with the target road segment. ;Will The improved dual-feedback spatiotemporal graph convolutional network GL-DFSTGCN, based on the proposed spatiotemporal feature fusion module, is used to construct sliding samples covering different road segments and time granularities through sliding window sampling, thus obtaining the input tensor. Where R represents the real number space, B represents the batch size of the input samples, and S represents the time step length of the historical traffic flow data obtained by sliding window sampling. This represents the number of input feature channels for each node.

[0222] Step 5.2: The improved dual-feedback spatiotemporal graph convolutional network GL-DFSTGCN models complex spatial dependencies and dynamic temporal patterns through a dual-branch architecture. The improved GL-DFSTGCN includes spatial convolutional blocks based on the Multiscale Spatial Graph Convolutional Network (MSGCN) and temporal convolutional blocks based on the Bayesian gated attention temporal convolutional network (ATCN). The ATCN block incorporates Bayesian TCN, LSTM, GRU, and attention mechanisms. MSGCN is used to extract rich spatial features from multi-source traffic signals S (such as flow / speed at different road segments and time granularities). In MSGCN, different node aggregation orders can preserve spatial correlation information at different levels of the road network. For node v i The features on the l-th graph convolutional layer (GCLayer) are calculated using the following formula:

[0223] ,

[0224] ,

[0225] Where AGG() is the aggregation function that fuses neighborhood features; CNN(•) represents the convolution operation for feature extraction. Representatives and nodes The corresponding slip sample. For the l-th GCLayer node and The edge features between them, (•) denotes a similarity function; W is a learnable weight matrix, which belongs to the training parameters of the graph convolutional layer GCLayer. This represents the activation function. Represents a node The set of neighboring nodes, Represents a node In the Feature representations on each graph convolutional layer Represents a node In the Feature representations on a graph convolutional layer;

[0226] Step 5.3: To form a dual-channel spatial convolution with GCN, a GAT module is constructed to enhance local node interactions through dynamic weighting. After the graph convolutional network GCN and the design space module GAT extract spatial features from different angles in their two channels, the two feature outputs from GCN and GAT are fed into the subsequent Bayesian TCN module. The GAT method, which automatically learns the importance of each neighboring node, better reflects the connection strength between neighboring nodes. An attention mechanism is introduced to automatically determine the importance of neighboring nodes and calculate the attention coefficient. :

[0227] ,

[0228] in, Represents a node Relative to node Attention score Represents a node The set of adjacent nodes. Q and are the training parameters. Representing the characteristics of a node; It is an activation function. Represents a node eigenvectors, Represents a node eigenvectors, Represents a node eigenvectors, Represents a node eigenvectors;

[0229] Step 5.4: After fusing the dual-channel outputs of the Graph Convolutional Network (GCN) and the Design Space Module (GAT), the final weighted output of the Multi-Scale Spatial Graph Convolutional Network (MSGCN) is obtained. Represented as: ,

[0230] in, This represents the weights corresponding to the p-order sensor output. It is the output of a P-order perceptual convolutional layer. By calculating the first The correlation between the first-order feature and all-order features is analyzed, and the feature weights are dynamically adjusted. Indicates the first The convolutional layer outputs features corresponding to the order perception.

[0231] Step 5.5: Then, the spatial features are input into the Bayesian gate-atcn temporal module to analyze temporal dependencies. The Bayesian gate-atcn is used to further analyze the temporal dependencies of the extracted features, mainly including Bayesian TCN, LSTM, GRU, and attention mechanisms. LSTM and TCN form a dual-channel serial fusion temporal model. The temporal correlations extracted from the Bayesian TCN block are input into LSTM and GRU, and the output of the hidden layer is reconstructed using a global attention mechanism. Sequence elements. The unrolled convolution is calculated as follows:

[0232] ,

[0233] The DCC (Dilated Causal Convolution) operator is used; dr is the dilation ratio; τ is the kernel size; b is the unwinding operator; and W is the output feature at the position after this DCC convolution operation. The value at; Input features In position The eigenvalue at that location. Indicates the time position The output feature value after DCC convolution. Indicates the first convolution kernel Weight parameters for each position.

[0234] Step 5.6 introduces Bayesian theory into the TCN framework. Given the input samples Z={X,Y} of the Bayesian TCN, in order to quantify the uncertainty, the Bayesian TCN quantifies all model parameters δ∈ Let the random variable be sampled from the posterior distribution P(δ|X,Y), and determine it as follows:

[0235] ;

[0236] Where X is the input feature dataset / input sample; Y is the output label / target dataset, corresponding to the true annotation and prediction target of the input sample X. It represents the explanatory power of the current parameter δ on the training data (X,Y). This represents the prior probability, indicating the probability of obtaining the model parameters before seeing the data. The prior assumptions. Indicates observed data The overall probability.

[0237] Step 5.7: Repeatedly stack the spatiotemporal convolutional blocks ST-Conv (Multi-Scale Spatial Graph Convolutional Network MSGCN and Bayesian gate-atcn temporal module) constructed in steps 5.2 to 5.6 into M layers to extract higher-order nonlinear spatiotemporal features layer by layer. The output of the m layers... As the input to the (m+1)th layer, the final output is ∈ :

[0238] in, and The first The number of channels and the length of the time dimension of the layer.

[0239] Step 5.8, output of the last ST-Conv block Execution-time convolution maps the historical time dimension to the prediction time step. The calculation formula is as follows:

[0240] ,

[0241] in, This represents the output feature of the last spatiotemporal convolutional block; Z represents the time step to be predicted; Z represents the final output feature tensor obtained after this temporal convolution operation. This represents the trainable convolution kernel parameters of the output layer's temporal convolution; This represents the number of channels for the output features of M ST-Conv blocks. Represents the convolution operator in the time dimension.

[0242] Step 6 involves constructing the OredRVFL-SLC model, including the following steps:

[0243] Step 6.1, the ORedRVFL (Outlier-robustted RVFL) algorithm is an improved version of the edRVFL algorithm, mainly utilizing regularization and normalization to improve the robustness and generalization ability of the model. Combining statistical theory, the problem of minimizing structural risk and empirical risk is transformed into the problem of minimizing the l2 norm output weight β, making the training error e sparse, as expressed below:

[0244] ,

[0245] Where y is the label vector / observation output vector of the training sample; H is the input feature matrix. Indicates constraints. Indicates a The objective function is to minimize the variables.

[0246] Step 6.2: In the initial stage of the Augmented Lagrange Method (ALM) iteration, the unknowns e and β are updated to achieve minimization. The iterative steps are as follows:

[0247] ,

[0248] ,

[0249] ,

[0250] Where β is the parameter vector; e is the error vector to be optimized; To find the independent variable that minimizes the objective function; The penalty factor is The augmented Lagrangian function; H is the observation matrix. Indicates the first The parameter vector after +1 iterations. Indicates the first The error vector after +1 iterations. This represents the penalty correction term for the constraint residual. Indicates the first The Lagrange multiplier vector at the next iteration.

[0251] Step 6.3: This scheme uses a novel and efficient decision entropy criterion to divide the original attribute space into disjoint attribute subspaces, where different attribute subspaces correspond to their respective hidden layer weights. First, the decision entropy of each attribute subspace is defined based on the posterior probability of the class in a given attribute subspace. Then, the attribute subspaces are iteratively updated to improve the validation accuracy of the outlier robust ensemble depth OredRVFL-SLC based on subspace local connectivity. In each iteration, the attribute subspace with the minimum decision entropy is selected as the update candidate. When OredRVFL-SLC reaches a stable validation accuracy, the iteration stops and the current attribute subspace is taken as the optimal attribute subspace.

[0252] Step 6.4, Set the original attribute space It is divided into k disjoint subspaces: ,

[0253] Among them, the sum of the number of attributes in all attribute subspaces is the original number of attributes; the original attribute space can be obtained by merging all attribute subspaces; any two attribute subspaces intersect to form an empty set, that is, attributes will not be repeatedly assigned to different attribute subspaces; J is the entire original complete attribute (feature) space, which is the set of all original features; A is a single attribute (feature) in the original attribute space. Represents the first in the original attribute space One characteristic. Indicates the first A feature subspace. Indicates the first The first subspace One characteristic.

[0254] Step 6.5, calculate the decision entropy: ; It is the first Individual indicators The corresponding decision entropy value.

[0255] Step 6.6: Sort the attribute subspaces in ascending order according to the magnitude of the decision entropy: ;

[0256] Here, J represents a single attribute subset, that is, an attribute subspace, which is a subset of attributes partitioned from the original full attribute space; Indicates the sorted order of the first... Each attribute subspace.

[0257] Step 6.7, using the initial attribute subspace Train the OredRVFL-SLC and obtain the initial validation accuracy a0. Use decision entropy as the criterion for judging the quality of attribute grouping. In each iteration, prioritize merging attribute subspaces that are more conducive to the classifier's classification, then train an OredRVFL-SLC and obtain the validation accuracy. If the validation accuracy increases, determine the final attribute subspace for this iteration and proceed to the next iteration. If combining all attribute subspaces still cannot improve the validation accuracy, stop training and output the final OredRVFL-SLC.

[0258] The improved global swarm optimization algorithm IHSO in step 7 includes the following steps:

[0259] Step 7.1: Set the population size, dimension, number of iterations, and upper and lower bounds of the search space for the HSO algorithm;

[0260] Step 7.2, In the population initialization phase, introduce the optimal point set strategy to improve the algorithm initialization:

[0261] ,

[0262] ,

[0263] Where F is the population size, and M is a number greater than F. It is the generating basis of the set of best points in dimension d. It is the i-th hyperparameter value of the k-th individual. It is the k-th candidate individual (solution / search point) in the initial population.

[0264] Step 7.3: Generate the initial positions of all search agents within the search space after improvement based on the optimal point set strategy. ;

[0265] Step 7.4, calculate the fitness value for each search agent. ;

[0266] Step 7.5, Selection based on adaptive simulated annealing: The selection process based on simulated annealing (SA) includes dynamic temperature updates for each iteration. The updates are as follows:

[0267] ,

[0268] in, The effective number of iterations is CR(k), which is the adaptive cooling coefficient. The temperature for the current iteration. This represents the temperature value at the start of the simulated annealing process. It is usually set to a relatively large value to give the algorithm a strong global search capability in the initial stage. This indicates the minimum temperature.

[0269] Step 7.6: If the new fitness value is better than the previous value, the algorithm runs greedily, replacing the old solution with the new one. The following criteria apply to the selection:

[0270] ,

[0271] ,

[0272] ;

[0273] Where P is the acceptance probability of accepting the inferior solution (the new candidate solution with a worse fitness). This represents the fitness difference. Indicates a new candidate solution The corresponding fitness value. Indicates the current old solution The corresponding fitness value. It represents a uniform random number within the interval [0,1].

[0274] Step 7.7, calculate the root mean square (RMS) of all fitness values:

[0275] ,

[0276] in, f( is the root mean square of the fitness values ​​of all individuals) ) indicates the first The fitness value of each individual.

[0277] Step 7.8: Calculate the coefficients of the displacement vector based on the root mean square difference of each fitness value. The individual differences are: ;

[0278] in This represents the difference for the i-th individual.

[0279] Step 7.9: The difference between coefficients c and the root mean square value is positive, and the difference between coefficients c and the root mean square value is negative. The sum of the normalized absolute values ​​is 1. The expression for coefficient c is: ;

[0280] in, This represents the result after normalizing the i-th coefficient; sign is the sign function, which checks if the input is a positive number. If the input is negative, If the input is 0, ;

[0281] Step 7.10, Position Update: The position of each individual is iteratively updated. Each individual's position is influenced by the weighted sum of the differences between its current position and the positions of all other individuals. A fitness balance optimization strategy is introduced during the position update phase, with the following update rule: ,

[0282] in, It is the average fitness of the t-th generation population. It is the position of the individual with the best fitness in the population. It is the historical best position of individual i. It is the position of the least fit individual in the population. , , , These are all weighting coefficients. , , , yes Random numbers between; Indicates the first The individual in the first The current position at the next iteration. Indicates the first The fitness value of an individual at its current position.

[0283] Step 7.11, Adaptive Mutation: To further improve exploration capabilities, an adaptive mutation operator is defined, dynamically adjusting the mutation rate and mutation step size based on the current iteration.

[0284] ,

[0285] ;

[0286] in, This represents the mutation rate at the current iteration step; Indicates the initial rate of variation; σ represents the final mutation rate; σ represents the variable time length. This indicates the variable length of the current iteration step. This indicates the current iteration number. This indicates the maximum number of iterations. Indicates the initial variable asynchronous length. This represents the final variable asynchronous length at the end of the algorithm iteration.

[0287] Step 7.12, if a mutation occurs, add a random perturbation to the individual's location:

[0288] ,

[0289] ;

[0290] in, It is the first The optimized individual in the iteration... The positional disturbance of the generation. This represents a standard normal distribution (Gaussian distribution), which means generating a standard normal random number with a mean of 0 and a variance of 1.

[0291] Step 7.13: Search for the global optimal solution based on the position update formula to obtain the optimal parameter combination of the OredRVFL-SLC model;

[0292] Step 7.14: Determine if the maximum number of iterations has been reached. If not, return to step 7.4 until the maximum number of iterations is reached, and output the optimal parameter configuration. The OredRVFL-SLC model uses the optimal parameter configuration to predict the components and obtain the final prediction result.

[0293] Step 8, which employs GRRVFL error correction, includes the following steps:

[0294] Step 8.1, Initialization. Randomly generate parameters for hidden nodes. , Output the weight matrix B and the auxiliary variable Z;

[0295] Step 8.2, construct the objective function, the calculation formula is:

[0296] ,

[0297] Where K is the hidden layer output matrix; L is the target output; and These are all regularization parameters.

[0298] Step 8.3: Calculate the error between the initial predicted values ​​and the actual values ​​of the OredRVFL-SLC model, and correct it using the GRRVFL model:

[0299] ,

[0300] Where z is the error and n is the actual value. These are the initial predictions from the OredRVFL-SLC model;

[0301] Step 8.4: The GRRVFL model consists of an input layer, a randomized hidden layer, and an output layer; the weights and biases of the hidden layer nodes are randomly generated, and the output H of the hidden layer is calculated.

[0302] ,

[0303] in, is the activation function, D is the randomly generated weight matrix, b is the bias vector, and G is the input feature matrix;

[0304] Step 8.5, learn the output weights under regularization constraints This makes the model output close to the actual value of the error term:

[0305] ,

[0306] ,

[0307] in, It is the vector of actual error values; It is a generalized regularization term;

[0308] Step 8.6: Calculate the error prediction value using the learned output weights. :

[0309] ;

[0310] Step 8.7: Add the initial predicted value of the OredRVFL-SLC model to the corrected error predicted value to obtain the final predicted value:

[0311] ,

[0312] in, This is the final predicted value. These are the initial predictions from the OredRVFL-SLC model. It is the error prediction value after cointegration test.

[0313] This embodiment uses the PeMS04 dataset from the California, USA traffic sensor network. This dataset consists of 307 sensors that collect data every 5 minutes, covering a 59-day period. Each station collects 16,992 data points for validation. In the feature preprocessing stage, multi-source heterogeneous data alignment is first performed: based on timestamp synchronization and spatial grid coding (GeoHash-7 precision), fixed detector data, floating car data, and external variables (meteorological data: temperature, humidity, precipitation, visibility; calendar data: holiday markers, weekday / weekend labels, large event levels; macroeconomic indicators: regional GDP growth rate, vehicle ownership) are spatiotemporally registered. The recursive least squares (RLS) algorithm is improved by introducing a dynamic forgetting factor and regularization techniques, and then fused with multivariate successive variational mode decomposition (MVMD) to form the RLS-MVMD algorithm for deep signal decoupling of the cleaned traffic flow data. This process adaptively decomposes the original traffic flow sequence into a series of frequency-band separated intrinsic mode functions (IMFs), encompassing low-frequency components reflecting long-term trend evolution, mid-frequency components capturing periodic rhythms, and high-frequency noise components characterizing random disturbances. Based on this, Spearman's rank correlation coefficient is used to perform nonlinear correlation analysis between each IMF component and the target road segment traffic flow data. A threshold ρ > 0.65 is set to eliminate redundant components that are weakly correlated with the target flow or contain only noise, thus selecting effective IMF components with high correlation. The selected IMF components and the original traffic flow data are combined to form a multi-channel input matrix, which serves as the high-gain input vector for the subsequent spatiotemporal feature fusion module. This effectively reduces the curse of dimensionality while retaining crucial multi-scale spatiotemporal correlation information.

[0314] At the model building level, this embodiment establishes a collaborative architecture between an outlier robust ensemble deep stochastic vector function linking network (OredRVFL-SLC) driven by an improved global swarm optimization algorithm (IHSO) and a spatiotemporal feature fusion module (an improved dual-feedback spatiotemporal graph convolutional network GL-DFSTGCN). The improved spatiotemporal feature fusion module, GL-DFSTGCN, employs a dual-branch encoder structure: the spatial branch uses a multi-scale spatial graph convolutional network (MSGCN) with a 3rd-order adjacency matrix, fusing a graph convolutional network (GCN) and a graph attention network (GAT) with dual channels. The GAT features an 8-head attention mechanism, dynamically capturing spatial heterogeneity dependencies between road segments by calculating attention coefficients between nodes. The temporal branch uses a Bayesiangate-ATCN with 3 layers of dilated causal convolutions (DCC) with dilation rates d_r of 1, 2, and 4, and a kernel size τ=3. It also introduces a Long Short-Term Memory (LSTM, 128 hidden units) network and a gated recurrent unit (GRU, 64 hidden units) to form a dual-channel serial fusion, reconstructing the hidden layer output using a global attention mechanism. The spatiotemporal convolutional blocks (ST-Conv) are repeatedly stacked M=2 layers, ultimately mapped to prediction strides (15 minutes, 30 minutes, 45 minutes multi-step prediction) through temporal convolution.

[0315] In the residual correction stage, a Generalized Regularized Stochastic Vector Function Link Network (GRRVFL) is introduced to perform secondary learning on the prediction residuals of OredRVFL-SLC. GRRVFL is configured with 400 hidden nodes and adopts L1 / L2 hybrid regularization (elastic network, α=0.5, λ=0.001). The output weights are constrained by the generalized regularization term to establish a nonlinear mapping between the residuals and input features. The final prediction value is the sum of the initial prediction value of OredRVFL-SLC and the residual compensation value of GRRVFL.

[0316] Experimental results show that, under extremely short input conditions relying only on the first 15 minutes of historical data (3 time steps), and under harsh interference scenarios such as a 20% initial bias error of the simulated detector (simulated sensor drift) and a sudden traffic accident (40% reduction in traffic capacity), this method exhibits excellent tracking accuracy and convergence robustness. In the 15-minute prediction task on the California PeMS04 dataset, compared to the benchmark methods STSGCN (Spatio-Temporal Graph Convolutional Network) and DCRNN (Diffusion Convolutional Recurrent Neural Network), this embodiment, based on RLS-MVMD signal decoupling and Spearman rank correlation feature selection, combined with the improved dual-feedback spatio-temporal graph convolutional network GL-DFSTGCN spatio-temporal feature fusion and the IHSO-optimized OredRVFL-SLC ensemble prediction architecture, significantly reduced the mean absolute percentage error (MAPE) from 11.2% to 4.05%, the root mean square error (RMSE) from 42.5veh / 5min to 11.8veh / 5min, and kept the mean absolute error (MAE) within 9.2veh / 5min.

[0317] Especially during the sensitive phase of sudden traffic flow polarization caused by traffic accidents, this method, through the accurate extraction of dynamic load jump components by RLS-MVMD and the improved dual-feedback spatiotemporal graph convolutional network GL-DFSTGCN for modeling spatial spillover effects, reduces the maximum estimation error (MaxError) from 68.2 veh / 5min in the benchmark method to 22.4 veh / 5min. As shown in Table 1, the experiments fully demonstrate that the combined application of RLS-MVMD feature decoupling, Spearman nonlinear filtering, and OredRVFL-SLC integrated prediction achieves highly reliable and real-time tracking of short-term traffic flow without the need for complex post-error compensation.

[0318] Table 1

[0319]

[0320] This invention provides a short-term traffic flow prediction method based on data decomposition and spatiotemporal feature fusion. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A short-term traffic flow prediction method based on data decomposition and spatiotemporal feature fusion, characterized in that, Includes the following steps: Step 1: Collect historical traffic flow data by deploying detection equipment at key urban nodes; Step 2: Obtain multi-dimensional urban data at the same spatial scale and contemporaneous with historical traffic flow data, including meteorological data, macroeconomic indicators, and population dynamics data; The city's multi-dimensional data and historical traffic flow data are cleaned, aligned, and integrated to construct a multi-dimensional and integrated traffic flow dataset; Step 3: Improve the recursive least squares (RLS) method by introducing a dynamic forgetting factor and regularization techniques, and integrate it with multivariate successive variational mode decomposition (MVMD) to decompose traffic flow data and obtain intrinsic mode function (IMF) components. Step 4: Use Spearman's rank correlation coefficient to perform correlation analysis on each IMF component obtained from the decomposition in Step 3; Step 5: By introducing the design spatial module GAT and the long short-term memory neural network LSTM, a spatiotemporal feature fusion module based on the improved dual-feedback spatiotemporal graph convolutional network GL-DFSTGCN is built. The dual-feedback spatiotemporal graph convolutional network includes a spatial convolutional block based on the multi-scale spatial graph convolutional network MSGCN and a temporal convolutional block based on the Bayesian gate attention temporal convolutional network ATCN. The temporal convolutional block is denoted as the Bayesian gate-atcn temporal module. Step 6: Build an outlier robust ensemble deep OredRVFL-SLC model based on subspace local connectivity to predict traffic flow data and obtain short-term traffic flow prediction results. Step 7: Optimize the hyperparameters of the prediction model using the improved global swarm optimization algorithm IHSO, and use the optimal parameters to predict the components. The improved global swarm optimization algorithm IHSO is based on the original global swarm optimization algorithm HSO. It adopts the best point set strategy to improve the original population initialization process and introduces the fitness balance optimization strategy in the algorithm position update stage to obtain the improved global swarm optimization algorithm IHSO. Step 8: Further improve the stochastic vector function linked RVFL model by introducing the concept of generalized regularization, forming the GRRVFL model to correct the error of the residuals and obtain the final prediction result.

2. The method according to claim 1, characterized in that, Step 1 includes: Fixed road sections: Deploy fixed detection equipment at key nodes and key road sections of the urban road network; Mobile devices: By using onboard GPS, Beidou dual-mode positioning terminals and onboard automatic diagnostic system OBD diagnostic interfaces deployed on taxis, buses, ride-hailing vehicles and freight vehicles, the instantaneous speed, latitude and longitude coordinates and driving direction angle of the vehicle are collected at a preset sampling frequency, and the travel speed and travel time of the road segment are estimated by map matching algorithm. After time synchronization and outlier removal, the collected multi-source heterogeneous raw data are structured and stored according to time and spatial granularity to construct a historical traffic flow database covering multiple time scales.

3. The method according to claim 2, characterized in that, Step 3 includes the following steps: Step 3.1: Initialize all variables, including the modal components of MVMD. Center frequency and Lagrange multipliers And the RLS weights w, covariance matrix P, and regularization parameters. , where k is the mode index and c is the channel index; Step 3.2, through harmonic shifting Gradient function L2 norm estimation The bandwidth, and the resulting cost function for: ， in, The analytical modulated signal is of channel number c and mode number k. Indicates will Adjust to the corresponding center frequency ; This is the regularization cost function used to estimate the bandwidth of each mode and constrain the smoothness of the signal; This indicates the derivative with respect to time t, representing the time gradient or rate of change; Step 3.3: When constructing the variational constraint model of MVMD, the augmented Lagrangian function also needs to be adjusted accordingly. ， in, Let c be the number of channels and k be the number of modes at time t; Let c be the Lagrange operator for the number of channels at time t; The input data is the number of channels c at time t; is the penalty factor in the MVMD algorithm; C is the total number of channels; K is the total number of intrinsic mode functions to be obtained by decomposition; Represents the augmented Lagrange function; Step 3.4, in each iteration of MVMD, for each sample point of the signal Recursive execution enables real-time filtering and parameter updates; input vectors are constructed. : ， Where L is the filter order; T represents transpose; This represents the input column vector corresponding to the nth signal sample point. The element is from the current nth signal observation value Tracing back to One signal observation value; Step 3.5, RLS output The estimated modes used to construct MVMD, error Feedback is fed back to parameter updates, forming a closed-loop adaptive system, and the gain vector is calculated. : ， in, It is the covariance matrix of the previous time step. For a dynamically adjusted forgetting factor, It is a parameter. The square of the L2 norm of the input vector. The formula is: ， in This represents the minimum value of the forgetting factor; This represents the maximum value of the forgetting factor; Functions for adjusting dynamic factors: ， in, For adjustment functions; Step 3.6, Update the weight vector: ， in It is an L2 regularization term; It is iterated to the th The weight vector obtained after the step update; It is the first The Kalman gain vector of the step; Let n be the prediction error at step n; Step 3.7, Update the covariance matrix : ， Where I is the identity matrix; Step 3.8, calculate the filter output using the updated weight vector: ， in It is the output of the filter; Step 3.9, calculate the error between the reference output and the actual output: ， in, It is a reference signal; Step 3.10, RLS output The estimated modes used to construct MVMD are obtained by passing the error. Feedback is sent to the RLS parameter update team to determine whether all sample points have been completed. If not, return to step 3.4; if completed, perform MVMD variational optimization. Step 3.11: Iterative update using the alternating direction method of multipliers. , and The optimal solution to the variational optimization problem is obtained, and the solution process includes mode update and center frequency update of the variational optimization problem. Modal update is: ， Where n is the iteration number, n = 0, 1, 2, ..., N; N represents the maximum number of iterations in the iteration process; Indicates the first Each channel has an input signal at a frequency of [frequency]. Frequency domain estimation or observed spectral values ​​at a given location; Indicates the first The first in the passage Estimation of the amplitude coefficients of each component; Indicates the first The first in the passage Each frequency component in Frequency domain basis function or component spectrum estimation at the specified location; Indicates the first The iteration yielded the... One modal spectrum, Indicates from the first The modality to the first Frequency domain estimates of each mode; The center frequency has been updated to: ， in, For the t-th iteration, the Fourier transform of the k-th component in the c-th channel is given, where w is the angular frequency and d represents the derivative. It represents the (n+1)th moment weighted average frequency of the k-th component.

4. The method according to claim 3, characterized in that, In step 4, the Spearman rank correlation coefficient for: ， in, It is the first The difference between the ranks of two variables in a pair of samples.

5. The method according to claim 4, characterized in that, In step 5, the improved dual-feedback spatiotemporal graph convolutional network GL-DFSTGCN performs the following steps: Step 5.1: After performing Spearman rank correlation analysis on the original multi-source traffic flow data, a multi-source traffic flow dataset S strongly correlated with the target road segment is obtained. S is then input into the improved dual-feedback spatiotemporal graph convolutional network GL-DFSTGCN, which uses a sliding window sampling method to construct sliding samples covering different road segments and different time granularities, thus obtaining the input tensor. Where R represents the real number space, B represents the batch size of the input samples, and S represents the time step length of the historical traffic flow data obtained by sliding window sampling; This represents the number of input feature channels for each node; Step 5.2, the improved dual-feedback spatiotemporal graph convolutional network GL-DFSTGCN models spatial dependencies and dynamic temporal patterns through a dual-branch architecture. The temporal convolutional block and the spatial convolutional block together constitute the spatiotemporal convolutional block ST-Conv block. The improved dual-feedback spatiotemporal graph convolutional network GL-DFSTGCN includes a spatial convolutional block based on the multi-scale spatial graph convolutional network MSGCN and a temporal convolutional block based on the Bayesian gated attention temporal convolutional network ATCN. The ATCN block includes a Bayesian TCN block, LSTM, GRU and attention mechanism. Spatial features are extracted from a multi-source traffic flow dataset S using MSGCN. In MSGCN, different node aggregation orders can preserve spatial association information at different levels of the road network. For the i-th node The features on the l-th graph convolutional layer GCLayer are calculated using the following formula: ， ， Where AGG is the aggregation function; CNN(•) represents the convolution operation for feature extraction. Representatives and nodes Corresponding slip samples, For the i-th node in the l-th graph convolutional layer GCLayer and the j-th node Edge features between (•) denotes a similarity function; W is a learnable weight matrix. This represents the activation function. Represents a node The set of neighboring nodes, Represents a node In the Feature representations on each graph convolutional layer Represents a node In the Feature representations on a graph convolutional layer; Step 5.3: Construct the design space GAT module. After the graph convolutional network GCN and the design space module GAT extract spatial features from different angles through their two channels, the two feature outputs extracted by the graph convolutional network GCN and the design space module GAT are fed into the subsequent Bayesian TCN module. An attention mechanism is introduced to automatically determine the importance of adjacent nodes and calculate the attention coefficient. : ， in, Represents a node Relative to node Attention score Represents a node The set of adjacent nodes; Q and are the training parameters. Represents the characteristics of a node; exp is the natural exponential function; It is an activation function; Represents a node eigenvectors, Represents a node eigenvectors, Represents a node eigenvectors, Represents a node eigenvectors; Step 5.4: After fusing the dual-channel outputs of the Graph Convolutional Network (GCN) and the Design Space Module (GAT), the final weighted output of the Multi-Scale Spatial Graph Convolutional Network (MSGCN) is obtained. Represented as: , in, This represents the weights corresponding to the p-order sensor output. It is the output of a P-order perceptual convolutional layer. Indicates the first The correlation between the first-order feature and all-order features; Indicates the first The convolutional layer outputs features corresponding to the order perception; Step 5.5: Input spatial features into the Bayesian gate-atcn temporal module to analyze temporal dependencies; use the Bayesian gate-atcn temporal module to analyze the temporal dependencies of extracted features; LSTM and TCN form a dual-channel serial fusion temporal model; input the temporal correlations extracted from the Bayesian TCN block into LSTM and GRU; and use a global attention mechanism to reconstruct the output of the hidden layer, sequence elements. The unrolled convolution is calculated as follows: ， in, To expand the causal convolution DCC operator; b is the size of the convolution kernel; b is the unwinding convolution operator; W is the position of the output feature after the DCC convolution operation. The value at; Input features In position The eigenvalue at that location; dr is the expansion ratio; Indicates the time position The output feature value after DCC convolution, Indicates the first convolution kernel Weight parameters for each position; Step 5.6: Introduce Bayesian theory into the TCN framework. Given the input samples Z={X,Y} of the Bayesian TCN, the Bayesian TCN will assign all model parameters δ∈ Set from posterior distribution The random variable for sampling is determined as follows: ； Where X is the input feature dataset or input sample; Y is the output label or target dataset, corresponding to the true annotation and prediction target of the input sample X; It represents the explanatory power of the current parameter δ on the training data (X,Y); This indicates that the model parameters were considered before the data was seen. The prior assumptions; Indicates observed data The overall probability; Step 5.7: Repeatedly stack the spatiotemporal convolutional blocks ST-Conv blocks constructed in steps 5.2 to 5.6 into M layers to extract higher-order nonlinear spatiotemporal features layer by layer, and then convert the output of the m layers... As the input of the (m+1)th layer, the final output is... ∈ ,in, and The first The number of channels and the length of the temporal dimension of the layer; the spatiotemporal convolutional block ST-Conv includes the multi-scale spatial graph convolutional network MSGCN and the Bayesian gate-atcn temporal module; Step 5.8, output of the last ST-Conv block Execution-time convolution maps the historical time dimension to the prediction time step. The calculation formula is as follows: ， in, This represents the output feature of the last spatiotemporal convolutional block; Z represents the time step to be predicted; Z represents the final output feature tensor obtained after this temporal convolution operation. This represents the trainable convolution kernel parameters of the output layer's temporal convolution; This represents the number of channels for the output features of the M ST-Conv blocks; Represents the convolution operator in the time dimension.

6. The method according to claim 5, characterized in that, Step 6 includes the following steps: Step 6.1 transforms the problem of minimizing structural risk and empirical risk into the problem of minimizing the L2 norm output weight β, making the training error e sparse, as expressed in: ， Where y is the label vector of the training sample or the observation output vector; H is the input feature matrix; Indicates constraints. Indicates a To optimize the objective function that minimizes the variables; Step 6.2: In the initial stage of the Augmented Lagrange Method (ALM) iteration, the unknowns e and β are updated to achieve minimization. The iterative steps are as follows: ， ， ， Where β is the parameter vector; e is the error vector to be optimized; To find the independent variable that minimizes the objective function; The penalty factor is The augmented Lagrangian function; H is the observation matrix, Indicates the first The parameter vector after +1 iterations. Indicates the first Error vector updated after +1 iterations This represents the penalty correction term for the constraint residuals. Indicates the first Lagrange multiplier vectors at the next iteration; Step 6.3: Define the decision entropy of each attribute subspace based on the posterior probability of the class in the given attribute subspace. Then, iteratively update the attribute subspace. In each iteration, select the attribute subspace with the minimum decision entropy as the update candidate. When the outlier robust ensemble depth OredRVFL-SLC based on subspace local connectivity reaches a stable verification accuracy, stop the iteration and take the current attribute subspace as the optimal attribute subspace. Step 6.4, Set the original attribute space It is divided into k disjoint subspaces: ， Where J is the set of all original features; Represents the first in the original attribute space One characteristic, Indicates the first A feature subspace, Indicates the first The first subspace One feature; Step 6.5, calculate the decision entropy: ; It is the first Individual indicators The corresponding decision entropy value; Step 6.6: Sort the attribute subspaces in ascending order according to the magnitude of the decision entropy: ; Where J represents a single subset of attributes; Indicates the sorted order of the first... Each attribute subspace; Step 6.7, using the initial attribute subspace Train the OredRVFL-SLC and obtain the initial validation accuracy a0; use the decision entropy as the criterion for judging the quality of attribute grouping. In each iteration, prioritize the fusion of attribute subspaces that are more conducive to the classifier's classification, then train an OredRVFL-SLC and obtain the validation accuracy. If the validation accuracy increases, determine the final attribute subspace for this iteration and proceed to the next iteration; if the combination of all attribute subspaces still cannot improve the validation accuracy after iteration, stop training and output the final OredRVFL-SLC.

7. The method according to claim 6, characterized in that, Step 7 includes the following steps: Step 7.1: Set the population size, dimension, number of iterations, and upper and lower bounds of the search space for the HSO algorithm; Step 7.2, In the population initialization phase, introduce the optimal point set strategy to improve the algorithm initialization: ， ， Where F is the population size, and M is a number greater than F. It is the generating basis of the set of best points in dimension d. It is the i-th hyperparameter value of the k-th individual. It is the k-th candidate individual in the initialization population; Step 7.3: Generate the initial positions of all search agents within the search space after improvement based on the optimal point set strategy. ; Step 7.4, calculate the fitness value for each search agent. ; Step 7.5, Selection based on adaptive simulated annealing: The selection process based on simulated annealing includes dynamic temperature updates for each iteration, as follows: ， in, The effective number of iterations is CR(k), which is the adaptive cooling coefficient. The temperature of the current iteration; This represents the temperature value at the start of the simulated annealing. Indicates the minimum temperature; Step 7.6: If the new fitness value is better than the previous value, the algorithm runs greedily, replacing the old solution with the new one. The following criteria apply to the selection: ， ， when ; Where P is the acceptance probability of accepting the difference solution; Indicates the fitness difference. Indicates a new candidate solution The corresponding fitness value, Indicates the current old solution The corresponding fitness value, Represents a uniformly random number within the interval [0,1]. Step 7.7, calculate the root mean square of all fitness values. : ； Where, f( ) indicates the first The fitness value of each individual; Step 7.8: Calculate the coefficients of the displacement vector based on the root mean square difference of each fitness value. The individual differences are: ; in This represents the difference for the i-th individual; Step 7.9: The difference between coefficients c and the root mean square value is positive, and the difference between coefficients c and the root mean square value is negative. The sum of the normalized absolute values ​​is 1. The expression for coefficient c is: ； in, This represents the result after normalizing the i-th coefficient; sign is the sign function, which checks if the input is a positive number. If the input is negative, If the input is 0, ; Step 7.10, Location Update: The location of the searched individual is iteratively updated. A fitness balance optimization strategy is introduced during the location update phase, with the following update rules: ， in, It is the average fitness of the t-th generation population. It is the position of the individual with the best fitness in the population. It is the historical best position of individual i. It is the position of the least fit individual in the population. , , , These are weighting coefficients. , , , yes Random numbers between; Indicates the first The individual in the first Current position at the next iteration Indicates the first The fitness value of each individual at its current position; Step 7.11, Adaptive Mutation: Adaptively define the mutation operator and dynamically adjust the mutation rate and mutation step size according to the current iteration. ， ； in, This represents the mutation rate at the current iteration step; Indicates the initial rate of variation; Indicates the final mutation rate; Indicates the variable length of the current iteration step; Indicates the current iteration number. Indicates the maximum number of iterations. Indicates the initial variable asynchronous length. This represents the final variable asynchronous length at the end of the algorithm iteration; Step 7.12, if a mutation occurs, add a random perturbation to the individual's location: ， ； in, It is the first The optimized individual in the iteration... The positional disturbance of the generation. Represents a standard normal distribution; Step 7.13: Search for the global optimal solution based on the position update formula to obtain the optimal parameter combination of the OredRVFL-SLC model; Step 7.14: Determine if the maximum number of iterations has been reached. If not, return to step 7.4 until the maximum number of iterations is reached, and output the optimal parameter configuration. The OredRVFL-SLC model uses the optimal parameter configuration to predict the components and obtain the final prediction result.

8. The method according to claim 7, characterized in that, Step 8 includes the following steps: Step 8.1, Initialization: Randomly generate parameters for hidden nodes. , Output the weight matrix B and the auxiliary variable Z; Step 8.2, construct the objective function, the calculation formula is: ， Where K is the hidden layer output matrix; L is the target output; and All are regularization parameters; Step 8.3: Calculate the error between the initial predicted values ​​and the actual values ​​of the OredRVFL-SLC model, and correct it using the GRRVFL model: ， Where z is the error and n is the actual value. These are the initial predictions from the OredRVFL-SLC model; Step 8.4: The GRRVFL model includes an input layer, a randomized hidden layer, and an output layer; randomly generate the weights and biases of the hidden layer nodes, and calculate the output K of the hidden layer. ， in, is the activation function, D is the randomly generated weight matrix, b is the bias vector, and G is the input feature matrix; Step 8.5, learn the output weights under regularization constraints This makes the model output close to the actual value of the error term: ， ， in, It is a generalized regularization term; Step 8.6, Calculate the predicted error value : ； Step 8.7: Add the initial predicted value of the OredRVFL-SLC model to the corrected error predicted value to obtain the final predicted value. : ， in, These are the initial predictions from the OredRVFL-SLC model.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 8.

10. A storage medium, characterized in that, It stores a computer program or instructions that, when run on a computer, perform the steps of the method as described in any one of claims 1 to 8.

Images