Traffic index prediction system and method based on multi-source fusion

Abstract

The application discloses a traffic index prediction system and method based on multi-source fusion, comprising the following steps: a dual driving architecture is constructed by fusing a traffic physical model and a machine learning model. The system comprises a data fusion and standardization module, a system dynamics modeling module, a supervised learning modeling module, a result fusion and weight self-adapting module, a rolling prediction module, a scenario simulation module and a visual output module. The system collects multi-source data such as traffic flow, trajectory and signal timing, unifies the formats and semantics, and then inputs the traffic mechanism model and the learning model for prediction; subsequently, the model output weight is dynamically adjusted according to the error and is fused; on the basis of rolling prediction, a human intervention parameter is introduced for scenario simulation; finally, the historical, predicted and simulated results are uniformly organized and output through an interface. The application can realize multi-scenario and multi-model fusion prediction of traffic operation indexes.

Description

Technical Field

[0001] This invention relates to the fields of intelligent transportation and big data analysis technology, and in particular to a traffic indicator prediction system and method based on multi-source fusion. Background Technology

[0002] With the acceleration of urbanization and the rapid development of intelligent transportation systems, urban traffic indicator prediction based on multi-source data has become an important supporting technology for traffic management, congestion mitigation, and policy decision-making. Currently, urban traffic prediction widely relies on time-series data modeling of indicators such as traffic flow, speed, and density, and uses techniques such as system dynamics or machine learning to predict traffic operation trends and support management decisions.

[0003] Existing traffic indicator prediction methods can be mainly divided into two categories: the first is purely data-driven machine learning methods, such as regression models, neural networks, and tree models. Although they perform well in terms of accuracy and deployment efficiency, they generally suffer from weak interpretability, poor model generalization ability, and difficulty in dealing with sudden interventions or strategy simulations. The second is purely mechanism-driven traffic system dynamics models, such as traffic flow conservation equations, basic graph models, and queuing theory. Although they have good physical interpretability, they have limited modeling ability and poor flexibility when faced with multi-source heterogeneous data fusion, multi-scale spatiotemporal coupling, and time-varying complex interference factors, such as signal timing, vehicle composition, and sudden events.

[0004] Furthermore, existing technologies generally lack rolling prediction mechanisms for arbitrary historical starting points and controllable scenario simulation capabilities, making it difficult to effectively support quantitative evaluation and scheme comparison of urban transportation policies. For example, in real transportation networks, users often want to set the simulation starting point based on a certain historical year and dynamically observe the evolution trend over the next few years; at the same time, they also hope to adjust traffic parameters to evaluate the effects of policy interventions, but existing models lack parameter overwrite interfaces and impact measurement logic, making it difficult to conduct systematic simulation and control simulation.

[0005] Therefore, how to provide a traffic indicator prediction system and method based on multi-source fusion is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a traffic indicator prediction system and method based on multi-source fusion. This invention comprehensively applies traffic mechanism modeling and supervised learning methods, and achieves dynamic prediction and intervention simulation of urban traffic indicators by constructing a complete process of data fusion, model prediction, result weighting, rolling calculation, and scenario parameter adjustment. This invention clearly standardizes a unified processing method for multi-source traffic data, establishes an error-driven fusion mechanism, and supports year-by-year rolling prediction based on historical starting years and multi-scenario parameter adjustment. It possesses the advantages of unified data processing standards, accurate collaborative model prediction, and strong intervention simulation capabilities.

[0007] The traffic indicator prediction system and method based on multi-source fusion according to embodiments of the present invention includes the following steps: The data fusion and standardization module is used to collect multi-source traffic data, including traffic flow data, vehicle network data, trajectory data, mobile signaling data, remote sensing data, signal timing parameters and road topology attribute data, perform time alignment and semantic mapping, and generate standardized traffic feature data. The system dynamics mechanism prediction module is used to receive standardized traffic characteristic data, establish a system dynamics model that includes traffic flow conservation relationships, basic graphical models, disturbance variables and external intervention parameters, and generate traffic indicator mechanism prediction results. The supervised machine learning prediction module is used to receive standardized traffic feature data and historical traffic indicator data, build a supervised learning model, and generate machine learning prediction results for traffic indicators. The dual-drive fusion and weight adaptive module is used to receive the traffic indicator mechanism prediction results and the traffic indicator machine learning prediction results, generate fusion weights based on the verification error, and output the traffic indicator fusion prediction results. The dynamic start year rolling prediction module is used to receive the traffic indicator fusion prediction results and historical traffic indicator data, perform training and rolling prediction according to the set start year and prediction window, and generate a rolling prediction sequence. The scenario simulation and parameter overwriting module is used to receive intervention parameters during the rolling prediction process, overwrite standardized traffic feature data, and generate a traffic indicator scenario prediction sequence. The visualization and interface module is used to receive historical traffic indicator data, rolling traffic indicator prediction sequences, and scenario-based traffic indicator prediction sequences, and to organize them in a structured manner and output them to the system interface.

[0008] Optionally, modules can be integrated using the following methods: Traffic flow data, vehicle network data, mobile trajectory data, remote sensing data, signal timing parameters, and road geometric attribute data are collected. The data are then time-aligned and semantically mapped to generate standardized traffic feature data. The system receives the standardized traffic characteristic data, establishes a system dynamics model based on traffic flow conservation relationships, basic graphical models, disturbance variables and external intervention parameters, and generates traffic indicator mechanism prediction results. Receive the standardized traffic feature data and historical traffic indicator data, construct a supervised learning model, and generate machine learning prediction results for traffic indicators; Receive the traffic indicator mechanism prediction results and traffic indicator machine learning prediction results, generate fusion weights based on the verification error, and output the traffic indicator fusion prediction results; Receive the traffic indicator fusion prediction results and historical traffic indicator data, and perform training and rolling prediction according to the set prediction start year and prediction window to generate a traffic indicator rolling prediction sequence. During the rolling prediction process, intervention parameters are received, the standardized traffic feature data is overwritten with parameters, and a traffic indicator scenario prediction sequence is generated. Receive the historical traffic indicator data, the rolling traffic indicator prediction sequence, and the scenario prediction sequence of the traffic indicator, and organize them in a structured manner and output them to the system interface.

[0009] Optional, the generation of standardized traffic characteristic data includes: Collect traffic flow data, vehicle network data, mobile trajectory data, remote sensing data, signal timing parameters and road geometric attribute data, and add timestamps and data source identifiers to each type of data to construct a multi-source traffic data set; A unified target time series is established for the multi-source traffic data set, and the timestamps in the set are compared with the target time series. Based on the comparison results, interpolation is performed on the data with inconsistent timestamps to make the processed data completely consistent with the target time series, thereby generating a time-aligned traffic data set. In the obtained time-aligned traffic data set, based on the source, physical meaning and structural label of each field in the set, the preset semantic mapping rules are invoked to perform unit conversion, format conversion, classification recoding and structural decomposition processing on the fields, and generate a semantically standardized set of traffic feature fields. The traffic feature field set is rearranged according to the time order in the traffic data set and combined into standardized traffic feature data with the same field order.

[0010] Optionally, the generation of traffic indicator mechanism prediction results includes: It receives standardized traffic feature data and extracts its corresponding spatial location and time index to form a set of traffic state variable inputs, including traffic density, traffic flow, average speed, free flow speed and congestion density fields. Based on the aforementioned set of traffic state variables, a continuous expression is constructed according to the traffic flow conservation relationship to characterize the partial derivative dynamic equation of traffic density changing with time and space. The expression is: ; in, For traffic density, For traffic flow, For spatial location, For time, This represents the partial derivative of traffic density with respect to time. This represents the partial derivative of traffic flow with respect to spatial location; Based on the traffic density and average speed fields in the input set of traffic state variables, the basic graphical model is used... The speed-density relationship defines a nonlinear expression for traffic flow and density, used to substitute the flow function into the traffic flow conservation relationship: ; in, Indicates the free flow velocity. Indicates blockage density. This indicates the inhibitory effect of relative density saturation on velocity; The replaced basic graph model is discretized, and a differential update formula is constructed based on the time step and spatial grid division to form the initial traffic state recursive function. A disturbance variable is introduced into the traffic state recursive function to express modeling error or random factors. The disturbance variable is an independent random term based on time change and is superimposed on each state update operation. The data from external input is fused with the free-flow velocity adjustment, link capacity adjustment, and signal timing parameter adjustment defined in the standardization module and embedded as external intervention parameters into the model structure to modify the recursive coefficients, boundary conditions, and velocity terms. Based on the traffic conditions, disturbance variables, and external intervention parameters of the previous time step, the traffic indicator prediction values ​​for the next time step are generated recursively at each time step, and the final traffic indicator mechanism prediction results are output.

[0011] Optionally, the generation of traffic indicator machine learning prediction results includes: Fields containing traffic volume, traffic density, average speed, queue length, delay time and congestion index are extracted from standardized traffic feature data, and a historical data set of traffic indicators is constructed by combining the corresponding time index. In the historical data set of traffic indicators, a fixed-length time window with the target prediction time as a reference is selected, continuous time segments are extracted as feature input samples, and the fields in each time segment are constructed into a set of structured feature vectors in a time stacking manner. The structured feature vectors mentioned above are used as input variables of the supervised learning model, and the traffic index field at the target prediction time is selected as the output variable of the supervised learning model. An input-output mapping relationship is constructed to generate a supervised modeling sample set. Based on the complexity and data structure of the supervised learning model, a linear regression model is selected on the supervised modeling sample set, and the input layer dimension, parameter set and output node settings are initialized; Define a loss function targeting prediction error. ,in The model parameter set is updated using gradient descent based on the training sample set. The process continues until a preset stopping condition is met, resulting in a fully trained supervised learning model. Standardized traffic feature data prior to the target prediction time are used to construct input feature vectors in the same way as training samples. These vectors are then input into the trained supervised learning model to generate machine learning prediction results for traffic indicators.

[0012] Optionally, the generation of traffic indicator fusion prediction results includes: Receive the traffic indicator mechanism prediction results and the traffic indicator machine learning prediction results, and combine their outputs under the same prediction time index to form a prediction result pair; Extract the actual observation values ​​corresponding to the above time index from the historical data of traffic indicators, form a sequence of true values ​​corresponding to the prediction results, and align them according to the index; The validation error between the traffic indicator mechanism prediction results and the actual value sequence, and the validation error between the traffic indicator machine learning prediction results and the actual value sequence are calculated separately. A set of standardized error pairs are obtained by using a unified error evaluation method. A unified error assessment method refers to using the same error calculation method when comparing the traffic indicator mechanism prediction results and the traffic indicator machine learning prediction results with the corresponding true values ​​in the historical data of traffic indicators. This usually includes the mean absolute error or the mean squared error, and the calculation is performed point by point on the same time index. This ensures that the two prediction methods are comparable in terms of error evaluation dimensions, scales and time structures, thereby providing a unified benchmark for the generation of fusion weights. Based on the standardized error, an initial fusion weight factor is generated. The side with smaller error is assigned a higher weight, and the side with larger error is assigned a lower weight. The two weight factors are then smoothed using a preset proportional adjustment parameter to generate the final fusion weight. The final fusion weights are respectively assigned to the output values ​​of the traffic indicator mechanism prediction results and the traffic indicator machine learning prediction results. The two are then weighted and combined according to a unified time index to obtain the fused traffic indicator prediction value. Arrange the fusion results under all prediction time in chronological order to form a sequence of traffic indicator fusion prediction results.

[0013] Optionally, the generation of the rolling forecast series for traffic indicators includes: The prediction start year is set as the starting point for rolling prediction. The prediction start year is selected from historical traffic indicator data, requiring that there are several consecutive years of standardized traffic characteristic data and traffic indicator fusion prediction results before it. The prediction window length is set to limit the historical time span used in each round of model training. The prediction window consists of w complete years before the prediction start date, and standardized traffic feature data and fused prediction results for the corresponding years are extracted from it. The extracted data is used as the model training input to train the system dynamics model and the supervised learning model, respectively, forming a dual-model structure for the current prediction round. The trained dual-model structure is used to predict traffic indicators at the corresponding time point of the prediction start year, generating the first rolling prediction value. The first rolling predicted value is indexed and matched with the true value in the historical data of traffic indicators. An error assessment operation is performed to generate a validation error, which is used to determine whether to update the model parameters or structure. After generating the first rolling forecast, advance the forecast year by one unit year, update the forecast window to include the time period w years before the latest forecast year, and re-extract the training data to repeat the training, forecasting and evaluation process. Arrange the rolling forecast values ​​for all years in chronological order to form a rolling forecast sequence for traffic indicators.

[0014] Optionally, the generation of traffic indicator scenario prediction sequences includes: Receive the rolling prediction sequence of traffic indicators, extract the standardized traffic feature data under the corresponding time index, and use it as the baseline input for scenario simulation. Set the target scenario conditions and access the intervention parameter set, which includes free flow speed adjustment, link capacity adjustment and signal timing parameter adjustment. Each intervention parameter has a structure field identifier and a quantitative value. Based on the structural field identifier in the intervention parameter set, the corresponding field position is found in the standardized traffic feature data input to the scenario simulation baseline, and the parameter overwrite operation is performed. The parameter overwrite operation is to replace the original field value with the specified intervention value to generate updated traffic feature input data. The updated traffic feature input data is input into the trained system dynamics model and supervised learning model. Without changing the model structure and parameters, only the input features are replaced, and a complete prediction process is executed to generate traffic indicator prediction values ​​under intervention conditions. The above traffic indicator predictions are compared with the rolling predictions under the corresponding time index for structural consistency, and then organized according to the setting order of the intervention scenarios to generate scenario predictions. Arrange all the scenario prediction values ​​under the time index in chronological order to form a traffic indicator scenario prediction sequence.

[0015] Optional, structured organization and system interface outputs include: Receive historical traffic indicator data, parse it according to its time index and indicator fields, and form historical data structure units; Receive the rolling forecast sequence of traffic indicators, and organize the rolling forecast values ​​into rolling forecast data structure units based on the time index and indicator fields consistent with historical data; Receive traffic indicator scenario prediction sequences and construct scenario prediction data structure units using the same time index and indicator fields; The above-mentioned historical data structure unit, rolling prediction data structure unit, and scenario prediction data structure unit are merged according to a unified field order and a unified time index to generate a structured output dataset. The structured output dataset is encapsulated into a data format that can be called by the system interface. The output content in the specified format is generated according to the interface call requirements, and the calling and data reading capabilities are provided to the outside world through the system interface.

[0016] The beneficial effects of this invention are: (1) This invention constructs a dual-drive traffic prediction system consisting of data fusion and standardization, system dynamics model, supervised learning model, fusion prediction mechanism, rolling prediction control, scenario parameter simulation and structured output, and realizes multi-model collaborative prediction of traffic flow, density and speed indicators. It overcomes the defects of single model structure and lack of adaptability of prediction results in the prior art, and has a systematic prediction capability with clear structure and clear functional coupling.

[0017] (2) This invention adopts unified time alignment and semantic mapping rules, and for the first time systematically standardizes the preprocessing process of multi-source traffic data, ensuring the uniformity of heterogeneous data such as traffic flow, trajectory, and remote sensing at the field, time granularity and physical semantic level, and solving the problem of unstable model performance caused by inconsistent data input in existing methods.

[0018] (3) By introducing an error-driven fusion weight mechanism, this invention dynamically adjusts the contribution ratio between the system dynamics prediction results and the supervised learning prediction results, thereby achieving adaptive fusion in the prediction stage and effectively improving the accuracy and robustness of multi-model combined prediction.

[0019] (4) This invention proposes a rolling prediction mechanism based on the prediction start year and sliding window control, which supports prediction from any year and automatically adjusts the model parameters in combination with each round of error assessment, so as to realize continuous modeling and dynamic updating of the long-term traffic evolution trend of the city.

[0020] (5) This invention further supports the integration of traffic policy intervention parameters during the prediction process. Standardized input data fields can be flexibly modified through parameter overwriting, and scenario simulation can be completed without changing the model structure. This solves the problem that existing models are difficult to perform multi-strategy comparative analysis.

[0021] (6) The present invention ultimately organizes historical data, rolling prediction results and scenario simulation results in a structured manner, and outputs them in a standardized manner through the system interface, providing standardized, unified and callable data support for the upper-level traffic management platform or visualization system, thereby improving the system integration efficiency and actual deployment availability. Attached Figure Description

[0022] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the traffic indicator prediction system and method based on multi-source fusion proposed in this invention; Figure 2 This is a modeling structure diagram of the system dynamics mechanism prediction process proposed in this invention. Detailed Implementation

[0023] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0024] refer to Figures 1-2 A traffic indicator prediction system and method based on multi-source fusion includes the following steps: The data fusion and standardization module is used to collect multi-source traffic data, including traffic flow data, vehicle network data, trajectory data, mobile signaling data, remote sensing data, signal timing parameters and road topology attribute data, perform time alignment and semantic mapping, and generate standardized traffic feature data. The system dynamics mechanism prediction module is used to receive standardized traffic characteristic data, establish a system dynamics model that includes traffic flow conservation relationships, basic graphical models, disturbance variables and external intervention parameters, and generate traffic indicator mechanism prediction results. The supervised machine learning prediction module is used to receive standardized traffic feature data and historical traffic indicator data, build a supervised learning model, and generate machine learning prediction results for traffic indicators. The dual-drive fusion and weight adaptive module is used to receive the traffic indicator mechanism prediction results and the traffic indicator machine learning prediction results, generate fusion weights based on the verification error, and output the traffic indicator fusion prediction results. The dynamic start year rolling prediction module is used to receive the traffic indicator fusion prediction results and historical traffic indicator data, perform training and rolling prediction according to the set start year and prediction window, and generate a rolling prediction sequence. The scenario simulation and parameter overwriting module is used to receive intervention parameters during the rolling prediction process, overwrite standardized traffic feature data, and generate a traffic indicator scenario prediction sequence. The visualization and interface module is used to receive historical traffic indicator data, rolling traffic indicator prediction sequences, and scenario-based traffic indicator prediction sequences, and to organize them in a structured manner and output them to the system interface.

[0025] In this embodiment, the modules are connected through the following method: Traffic flow data, vehicle network data, mobile trajectory data, remote sensing data, signal timing parameters, and road geometric attribute data are collected. The data are then time-aligned and semantically mapped to generate standardized traffic feature data. The system receives the standardized traffic characteristic data, establishes a system dynamics model based on traffic flow conservation relationships, basic graphical models, disturbance variables and external intervention parameters, and generates traffic indicator mechanism prediction results. Receive the standardized traffic feature data and historical traffic indicator data, construct a supervised learning model, and generate machine learning prediction results for traffic indicators; Receive the traffic indicator mechanism prediction results and traffic indicator machine learning prediction results, generate fusion weights based on the verification error, and output the traffic indicator fusion prediction results; Receive the traffic indicator fusion prediction results and historical traffic indicator data, and perform training and rolling prediction according to the set prediction start year and prediction window to generate a traffic indicator rolling prediction sequence. During the rolling prediction process, intervention parameters are received, the standardized traffic feature data is overwritten with parameters, and a traffic indicator scenario prediction sequence is generated. Receive the historical traffic indicator data, the rolling traffic indicator prediction sequence, and the scenario prediction sequence of the traffic indicator, and organize them in a structured manner and output them to the system interface.

[0026] In this embodiment, the generation of standardized traffic characteristic data includes: Collect traffic flow data, vehicle network data, mobile trajectory data, remote sensing data, signal timing parameters and road geometric attribute data, and add timestamps and data source identifiers to each type of data to construct a multi-source traffic data set; A unified target time series is established for the multi-source traffic data set, and the timestamps in the set are compared with the target time series. Based on the comparison results, interpolation is performed on the data with inconsistent timestamps to make the processed data completely consistent with the target time series, thereby generating a time-aligned traffic data set. For traffic flow data, vehicle network data, mobile trajectory data, remote sensing data, signal timing parameters, and road geometric attribute data with inconsistent timestamps, after establishing a correspondence with the target time series, the data values ​​of the previous and next valid time nodes are selected. Based on their relative positions in time, the data values ​​of the missing time nodes between them are calculated in a linear manner, and the calculation results are filled into the corresponding positions. This ensures that this type of data has continuous numerical output at each time node of the target time series, thereby forming a complete time-aligned traffic data set. In the obtained time-aligned traffic data set, based on the source, physical meaning and structural label of each field in the set, the preset semantic mapping rules are invoked to perform unit conversion, format conversion, classification recoding and structural decomposition processing on the fields, and generate a semantically standardized set of traffic feature fields. The preset semantic mapping rules consist of a set of transformation logics for each field in traffic flow data, vehicle network data, mobile trajectory data, remote sensing data, signal timing parameters and road geometric attribute data. Specifically, they include the mapping relationship between fields and target feature names, field unit specifications, classification recoding methods, numerical interval classification logic and spatial attribute label reconstruction rules, which are used to convert the original fields after time alignment into traffic feature fields with consistent structure and clear semantics. For time-aligned traffic data fields, based on the recording method of the fields in the original data, the fields involving physical quantities are converted to units according to a unified dimension; fields recorded in string or combined form, such as time, date, or location, are parsed and reconstructed according to a unified format; fields represented in text or enumeration form, such as vehicle type, road grade, or phase type, are recoded according to a preset classification table; information fields containing multiple values ​​or composite structures are decomposed into independent subfields according to their internal structure order, so that each field has a clear numerical form or classification form, which is convenient for constructing a semantically standardized set of traffic feature fields; The traffic feature field set is rearranged according to the time order in the traffic data set and combined into standardized traffic feature data with the same field order.

[0027] In this embodiment, the generation of traffic indicator mechanism prediction results includes: It receives standardized traffic feature data and extracts its corresponding spatial location and time index to form a set of traffic state variable inputs, including traffic density, traffic flow, average speed, free flow speed and congestion density fields. Based on the aforementioned set of traffic state variables, a continuous expression is constructed according to the traffic flow conservation relationship to characterize the partial derivative dynamic equation of traffic density changing with time and space. The expression is: ; in, For traffic density, For traffic flow, For spatial location, For time, This represents the partial derivative of traffic density with respect to time. This represents the partial derivative of traffic flow with respect to spatial location; Based on the traffic density and average speed fields in the input set of traffic state variables, the basic graphical model is used... The speed-density relationship defines a nonlinear expression for traffic flow and density, used to substitute the flow function into the traffic flow conservation relationship: ; in, Indicates the free flow velocity. Indicates blockage density. This indicates the inhibitory effect of relative density saturation on velocity; The replaced basic graph model is discretized, and a differential update formula is constructed based on the time step and spatial grid division to form the initial traffic state recursive function. For continuous partial differential equations established based on traffic flow conservation relations and basic graphical models, a preset time step and spatial step are used to divide continuous time variables and spatial location variables into discrete grid points. At each time step t, the traffic density and flow values ​​at the current time and the previous time are used, combined with the state of adjacent spatial grid points, and the derivative terms are replaced according to the explicit difference or implicit difference scheme to construct the recursive expression of traffic state variables. Thus, the original continuous model is transformed into a discrete dynamic model suitable for computer iterative solution, and serves as the core function for mechanism prediction. A disturbance variable is introduced into the traffic state recursive function to express modeling error or random factors. The disturbance variable is an independent random term based on time change and is superimposed on each state update operation. The data from external input is fused with the free-flow velocity adjustment, link capacity adjustment, and signal timing parameter adjustment defined in the standardization module and embedded as external intervention parameters into the model structure to modify the recursive coefficients, boundary conditions, and velocity terms. Free-flow velocity adjustment refers to the increase or decrease of the free-flow velocity parameter in the model, used to simulate changes in road capacity under traffic improvement or restriction conditions; link capacity adjustment is used to adjust the maximum number of vehicles that can pass through the road, reflecting capacity changes caused by construction, accidents, or reconstruction and expansion; signal timing parameter adjustment refers to the modification of control variables such as signal phase, green ratio, or cycle length, used to simulate changes in control timing under intervention strategies. These three types of parameters are injected into the state update function during model execution, having a structural impact on key state quantities such as speed, density, or flow, and are used to generate scenario prediction sequences. Based on the traffic conditions, disturbance variables, and external intervention parameters of the previous time step, the traffic indicator prediction values ​​for the next time step are generated recursively at each time step, and the final traffic indicator mechanism prediction results are output.

[0028] In this embodiment, the generation of traffic indicator machine learning prediction results includes: Fields containing traffic volume, traffic density, average speed, queue length, delay time and congestion index are extracted from standardized traffic feature data, and a historical data set of traffic indicators is constructed by combining the corresponding time index. In the historical data set of traffic indicators, a fixed-length time window with the target prediction time as a reference is selected, continuous time segments are extracted as feature input samples, and the fields in each time segment are constructed into a set of structured feature vectors in a time stacking manner. Selecting a fixed-length time window with the target prediction time as a reference means that when constructing supervised learning input features, the time point corresponding to the traffic indicator to be predicted is used as a reference, and a set number of continuous time steps are traced back. Fields such as traffic volume, traffic density, average speed, delay time, queue length and congestion index are extracted sequentially from the standardized traffic feature data at each moment within the window, and they are arranged in chronological order to form input samples, so that each predicted value corresponds to a unique set of equal-length historical input feature vectors. The structured feature vectors mentioned above are used as input variables of the supervised learning model, and the traffic index field at the target prediction time is selected as the output variable of the supervised learning model. An input-output mapping relationship is constructed to generate a supervised modeling sample set. Based on the complexity and data structure of the supervised learning model, a linear regression model is selected on the supervised modeling sample set, and the input layer dimension, parameter set and output node settings are initialized; Based on the complexity and data structure of supervised learning models, during the training process, different types of model structures with different expressive capabilities are selected according to data characteristics such as the dimensionality of the input feature vector, the proportion of continuous features, the number of categorical variables, and the sample size. These include linear models for processing low-dimensional continuous data, ensemble tree models for processing nonlinear relationships and data structures with discrete variables, or feedforward neural network models with multiple layers of neurons for sample sets with a large proportion of high-dimensional and nonlinear features, thereby matching the requirements of data structure characteristics and model fitting capabilities. Define a loss function targeting prediction error. ,in The model parameter set is updated using gradient descent based on the training sample set. The process continues until a preset stopping condition is met, resulting in a fully trained supervised learning model. Standardized traffic feature data prior to the target prediction time are used to construct input feature vectors in the same way as training samples. These vectors are then input into the trained supervised learning model to generate machine learning prediction results for traffic indicators.

[0029] In this embodiment, the generation of traffic indicator fusion prediction results includes: Receive the traffic indicator mechanism prediction results and the traffic indicator machine learning prediction results, and combine their outputs under the same prediction time index to form a prediction result pair; Extract the actual observation values ​​corresponding to the above time index from the historical data of traffic indicators, form a sequence of true values ​​corresponding to the prediction results, and align them according to the index; The validation error between the traffic indicator mechanism prediction results and the actual value sequence, and the validation error between the traffic indicator machine learning prediction results and the actual value sequence are calculated separately. A set of standardized error pairs are obtained by using a unified error evaluation method. A unified error assessment method refers to using the same error calculation method when comparing the traffic indicator mechanism prediction results and the traffic indicator machine learning prediction results with the corresponding true values ​​in the historical data of traffic indicators. This usually includes the mean absolute error or the mean squared error, and the calculation is performed point by point on the same time index. This ensures that the two prediction methods are comparable in terms of error evaluation dimensions, scales and time structures, thereby providing a unified benchmark for the generation of fusion weights. Based on the standardized error, an initial fusion weight factor is generated. The side with smaller error is assigned a higher weight, and the side with larger error is assigned a lower weight. The two weight factors are then smoothed using a preset proportional adjustment parameter to generate the final fusion weight. Assigning higher weights to the side with smaller errors and lower weights to the side with larger errors means that after obtaining the verification errors corresponding to the traffic indicator mechanism prediction results and the traffic indicator machine learning prediction results, the two errors are compared with each other, and weight values ​​for the two prediction methods are generated according to the inverse proportional allocation principle based on the relationship between the error magnitudes. This ensures that the prediction method with smaller errors receives higher weight values ​​in the weight allocation, and the prediction method with larger errors receives lower weight values. The two weights are then normalized using a preset weight smoothing parameter to keep the total weight consistent for subsequent weighted fusion calculations. The preset scaling parameter is a numerical factor used to control the degree of weight difference between the two prediction methods in the fusion weight allocation. When the verification error difference is large, this parameter can smoothly adjust the initial weights to prevent the weights from being extremely biased towards a single prediction source. The scaling parameter is introduced during the fusion weight calculation process to limit the maximum weight difference, making the fusion weights more stable and adaptable to dynamic error fluctuations. The final fusion weights are respectively assigned to the output values ​​of the traffic indicator mechanism prediction results and the traffic indicator machine learning prediction results. The two are then weighted and combined according to a unified time index to obtain the fused traffic indicator prediction value. Weighted combination refers to multiplying the traffic indicator mechanism prediction result and the traffic indicator machine learning prediction result by their respective fusion weights under a unified prediction time index, and then summing the products to generate the fused prediction value of the traffic indicator at that time point. The fused prediction values ​​of all time points are arranged in chronological order to form a complete fused prediction sequence of traffic indicators. Arrange the fusion results under all prediction time in chronological order to form a sequence of traffic indicator fusion prediction results.

[0030] In this embodiment, the generation of the rolling prediction sequence for traffic indicators includes: The prediction start year is set as the starting point for rolling prediction. The prediction start year is selected from historical traffic indicator data, requiring that there are several consecutive years of standardized traffic characteristic data and traffic indicator fusion prediction results before it. The prediction window length is set to limit the historical time span used in each round of model training. The prediction window consists of w complete years before the prediction start date, and standardized traffic feature data and fused prediction results for the corresponding years are extracted from it. The extracted data is used as the model training input to train the system dynamics model and the supervised learning model, respectively, forming a dual-model structure for the current prediction round. The trained dual-model structure is used to predict traffic indicators at the corresponding time point of the prediction start year, generating the first rolling prediction value. The first rolling predicted value is indexed and matched with the true value in the historical data of traffic indicators. An error assessment operation is performed to generate a validation error, which is used to determine whether to update the model parameters or structure. The error assessment operation refers to aligning the current forecast value of traffic indicators with the actual observed values ​​of traffic indicators in the historical data of the current year after each round of rolling forecast. Based on a unified error calculation method, such as mean absolute error or mean squared error, the difference between the forecast value and the actual value is calculated item by item to generate verification error, which supports the subsequent model update or fusion weight allocation process. After generating the first rolling forecast, advance the forecast year by one unit year, update the forecast window to include the time period w years before the latest forecast year, and re-extract the training data to repeat the training, forecasting and evaluation process. The repeated training, prediction, and evaluation process refers to the process in rolling prediction where, as the prediction window moves forward, updated training data is extracted again, and a new round of parameter training is performed on the system dynamics model and the supervised learning model. After training, the updated model is used to predict the new target year, and the predicted value is compared with the historical true value to evaluate the error and form a new verification error. Arrange the rolling forecast values ​​for all years in chronological order to form a rolling forecast sequence for traffic indicators.

[0031] In this embodiment, the generation of the traffic indicator scenario prediction sequence includes: Receive the rolling prediction sequence of traffic indicators, extract the standardized traffic feature data under the corresponding time index, and use it as the baseline input for scenario simulation. Set the target scenario conditions and access the intervention parameter set, which includes free flow speed adjustment, link capacity adjustment and signal timing parameter adjustment. Each intervention parameter has a structure field identifier and a quantitative value. Based on the structural field identifier in the intervention parameter set, the corresponding field position is found in the standardized traffic feature data input to the scenario simulation baseline, and the parameter overwrite operation is performed. The parameter overwrite operation is to replace the original field value with the specified intervention value to generate updated traffic feature input data. The updated traffic feature input data is input into the trained system dynamics model and supervised learning model. Without changing the model structure and parameters, only the input features are replaced, and a complete prediction process is executed to generate traffic indicator prediction values ​​under intervention conditions. The above traffic indicator predictions are compared with the rolling predictions under the corresponding time index for structural consistency, and then organized according to the setting order of the intervention scenarios to generate scenario predictions. Structural consistency comparison refers to comparing the generated traffic indicator scenario prediction values ​​with the rolling traffic indicator prediction values ​​output by the rolling prediction module at the data structure level under the same time index to ensure that the two correspond one-to-one in terms of the number of fields, field names, time labels, and data arrangement order, and to avoid data processing conflicts caused by differences in fields or inconsistencies in order. Arrange all the scenario prediction values ​​under the time index in chronological order to form a traffic indicator scenario prediction sequence.

[0032] In this embodiment, the structured organization and system interface output includes: Receive historical traffic indicator data, parse it according to its time index and indicator fields, and form historical data structure units; Receive the rolling forecast sequence of traffic indicators, and organize the rolling forecast values ​​into rolling forecast data structure units based on the time index and indicator fields consistent with historical data; Receive traffic indicator scenario prediction sequences and construct scenario prediction data structure units using the same time index and indicator fields; The above-mentioned historical data structure unit, rolling prediction data structure unit, and scenario prediction data structure unit are merged according to a unified field order and a unified time index to generate a structured output dataset. The structured output dataset is encapsulated into a data format that can be called by the system interface. The output content in the specified format is generated according to the interface call requirements, and the calling and data reading capabilities are provided to the outside world through the system interface.

[0033] Example 1: To verify the feasibility of this invention in practice, it was applied to an urban transportation system. Faced with complex operational states, a changing external environment, and uncertain policy influences, relying solely on a single prediction model or static rules is insufficient to achieve high-precision dynamic prediction of traffic indicators. To address these issues, a dual-drive fusion prediction method for transportation systems based on multi-source data fusion is proposed, and a complete system architecture including data preprocessing, model fusion, rolling computation, scenario simulation, and output interfaces is constructed. In practical applications, this system can receive historical data input, complete standardization processing, and collaboratively predict future traffic operation indicators through system dynamics mechanisms and supervised learning models. It also allows users to set intervention scenario variables and overwrite input data, enabling strategy evaluation and result output.

[0034] In actual deployment, the system extracts multi-source traffic data from historical databases, including road traffic monitoring data, road network structure attributes, vehicle-to-everything (V2X) data, remote sensing image-derived variables, demographic information, industrial structure indicators, and urban traffic management parameters. Each data type is appended with a data source identifier and timestamp. Time alignment processing is performed on the collected data to construct a unified target timeline. Linear interpolation is used to complete missing time points, ensuring consistency across all fields in the target time series. The system performs semantic mapping operations, performing unit conversion, classification recoding, format parsing, and field reconstruction based on the physical meaning, unit type, and structural classification rules of the fields, ultimately generating standardized traffic feature data.

[0035] In the modeling phase, the system constructs both a system dynamics model and a supervised learning model. During system dynamics modeling, based on traffic flow conservation, the relationship between traffic density and traffic flow is expressed using continuous partial derivative equations. A speed suppression term is defined using the Greenshields speed-density relationship from the basic graphical model, and an iterative function for traffic states is constructed. The model is discretized spatially at the road link level and updates state variables at fixed time steps in the time dimension, completing the recursive process of traffic indicators through a differential iterative formula. Furthermore, disturbance variables are embedded as modeling error terms, receiving adjustments for free-flow speed, signal period, and capacity from the external intervention parameter set to overwrite the state update coefficients.

[0036] The system extracts variables such as traffic volume, speed, density, congestion index, and queue length from standardized data, constructs time-series segments by dividing the data into time windows, and generates the structured feature inputs required for the supervised learning model. Using the index values ​​at the target prediction time point as the supervised output variables, an input-output mapping relationship is constructed. Linear regression and XGBoost models are trained and validated on the sample set. The mean squared error loss function is used, and parameter optimization is performed through gradient descent to train the supervised learning model.

[0037] The system generates prediction results through fusion. It compares the system dynamics prediction results with the supervised learning prediction results using a unified index against historical true values, calculates their respective errors, calculates the fusion weights using the inverse proportion of standardized errors, adjusts parameters according to a preset ratio to generate the final weighting coefficients, and outputs the fused value. During the rolling prediction process, the system sets the starting prediction year and the training window span, extracts training data from historical time periods, reconstructs the model structure, and executes training predictions. The system refreshes the model input and window data for each year, outputting the corresponding prediction results. The rolling prediction results for all years are sequentially arranged to form a complete rolling prediction sequence.

[0038] During the simulation phase, users can define intervention scenarios and input a set of intervention parameters, such as increasing net immigration rate, optimizing signal timing, and increasing investment in road network construction. After receiving the parameters, the system locates the corresponding field in the standardized data and directly performs a numerical replacement operation. It does not change the model structure or parameters, only modifying the input vector content. The updated input is then re-introduced into the system dynamics model and the supervised learning model for a complete prediction process. The predicted values ​​generated under the intervention scenario are compared with the original rolling predicted values ​​for structural consistency. All scenario simulation results are output in a set order, ultimately forming a traffic indicator scenario prediction sequence.

[0039] The following is a partial output of data in this embodiment, used to demonstrate the model's operation process and fusion strategy mechanism in real traffic data prediction tasks.

[0040] Table 1. Historical Data Sample of Traffic Indicators (Unit: 10,000 ton-kilometers) ; Table 2 Comparison of Model Predictions and Fusion Output (Unit: 10,000 tons / kilometer) ; Table 3. Results of the intervention scenario simulation (net immigration rate +6%, capacity +10%) ; Based on the data in the table above, technical feature analysis can be conducted from multiple dimensions: Table 1 shows the key input features used to build the prediction model, including variables such as resident population, GDP, and average flow velocity. These variables, after initial standardization and semantic mapping, form traffic feature input data with a consistent structure. By observing the changes in these variables over different periods, it can be seen that there is a trend of growth and correlation between the indicators, which provides a modeling foundation for supervised learning models and system dynamics models.

[0041] Table 2 compares the actual values ​​of traffic indicators with the model outputs. The error fluctuations between the predictions of the system dynamics model and the machine learning model and the actual values ​​are controlled within a reasonable range. The final fused value, through an error weighting strategy, achieves a more accurate and realistic output than a single model. The relative error of the fused value is compressed to less than 0.1%, fully demonstrating the accuracy optimization capability of the fusion mechanism. Furthermore, the continuity and stability of the fused predictions across different years also illustrate the effectiveness of the rolling training window mechanism.

[0042] Table 3 further illustrates the scenario prediction results generated by the system after incorporating intervention parameters (such as increased net inflow rate and road capacity adjustment). While maintaining the original model structure and parameters, the system can accurately reflect the potential changing trends of traffic indicators in the simulation dimension simply by overriding the input fields. Compared with the rolling prediction values, the scenario values ​​after intervention show a significant increase, and the increase is within the expected range of traffic management, reflecting the system's functional completeness and sensitivity for policy simulation and evaluation.

[0043] Through multi-source data fusion, model collaboration, error-driven weighting, sliding window rolling training, and intervention parameter injection mechanisms, the prediction system can flexibly respond to actual input fluctuations while maintaining a stable structure, improving the model's adaptability and controllability in medium- and long-term traffic evolution prediction. The entire system does not rely on a specific model architecture; instead, it uses data structures and standard processes as its main framework, possessing good generalization and scenario portability capabilities, making it suitable for various types of traffic indicator prediction needs.

[0044] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A traffic index prediction system based on multi-source fusion, characterized in that, include: The data fusion and standardization module is used to collect multi-source traffic data, perform time alignment and semantic mapping, and generate standardized traffic feature data. The system dynamics mechanism prediction module is used to receive standardized traffic characteristic data, establish a system dynamics model that includes traffic flow conservation relationships, basic graphical models, disturbance variables and external intervention parameters, and generate traffic indicator mechanism prediction results. The supervised machine learning prediction module is used to receive standardized traffic feature data and historical traffic indicator data, build a supervised learning model, and generate machine learning prediction results for traffic indicators. The dual-drive fusion and weight adaptive module is used to receive the traffic indicator mechanism prediction results and the traffic indicator machine learning prediction results, generate fusion weights based on the verification error, and output the traffic indicator fusion prediction results. The dynamic start year rolling prediction module is used to receive the traffic indicator fusion prediction results and historical traffic indicator data, perform training and rolling prediction according to the set start year and prediction window, and generate a rolling prediction sequence. The scenario simulation and parameter overwriting module is used to receive intervention parameters during the rolling prediction process, overwrite standardized traffic feature data, and generate a traffic indicator scenario prediction sequence. The visualization and interface module is used to receive historical traffic indicator data, rolling traffic indicator prediction sequences, and scenario-based traffic indicator prediction sequences, and to organize them in a structured manner and output them to the system interface.

2. The traffic index prediction method based on multi-source fusion, characterized in that, The modules are connected in the following way: Traffic flow data, vehicle network data, mobile trajectory data, remote sensing data, signal timing parameters, and road geometric attribute data are collected. The data are then time-aligned and semantically mapped to generate standardized traffic feature data. The system receives the standardized traffic characteristic data, establishes a system dynamics model based on traffic flow conservation relationships, basic graphical models, disturbance variables and external intervention parameters, and generates traffic indicator mechanism prediction results. Receive the standardized traffic feature data and historical traffic indicator data, construct a supervised learning model, and generate machine learning prediction results for traffic indicators; Receive the traffic indicator mechanism prediction results and traffic indicator machine learning prediction results, generate fusion weights based on the verification error, and output the traffic indicator fusion prediction results; Receive the traffic indicator fusion prediction results and historical traffic indicator data, and perform training and rolling prediction according to the set prediction start year and prediction window to generate a traffic indicator rolling prediction sequence. During the rolling prediction process, intervention parameters are received, the standardized traffic feature data is overwritten with parameters, and a traffic indicator scenario prediction sequence is generated. Receive the historical traffic indicator data, the rolling traffic indicator prediction sequence, and the scenario prediction sequence of the traffic indicator, and organize them in a structured manner and output them to the system interface. 3.The traffic index prediction method based on multi-source fusion according to claim 2, characterized in that, The generation of standardized traffic feature data includes: Collect traffic flow data, vehicle network data, mobile trajectory data, remote sensing data, signal timing parameters and road geometric attribute data, and add timestamps and data source identifiers to each type of data to construct a multi-source traffic data set; A unified target time series is established for the multi-source traffic data set, and the timestamps in the set are compared with the target time series. Based on the comparison results, interpolation is performed on data with inconsistent timestamps to generate a time-aligned traffic data set. In the obtained time-aligned traffic data set, based on the source, physical meaning and structural label of each field in the set, the preset semantic mapping rules are invoked to perform unit conversion, format conversion, classification recoding and structural decomposition processing on the fields, and generate a semantically standardized set of traffic feature fields. The traffic feature field set is rearranged according to the time order in the traffic data set and combined into standardized traffic feature data with the same field order. 4.The traffic index prediction method based on multi-source fusion according to claim 2, characterized in that, The generation of traffic indicator mechanism prediction results includes: It receives standardized traffic feature data and extracts its corresponding spatial location and time index to form a set of traffic state variable inputs, including traffic density, traffic flow, average speed, free flow speed and congestion density fields. Based on the set of traffic state variables, a continuous expression is constructed according to the traffic flow conservation relationship to characterize the partial derivative dynamic equation of traffic density changing with time and space. Based on the traffic density and average speed fields in the input set of traffic state variables, the basic graphical model is used... The speed-density relationship defines a nonlinear expression for traffic flow and density, used to substitute the flow function into the traffic flow conservation relationship; The replaced basic graph model is discretized, and a differential update formula is constructed based on the time step and spatial grid division to form the initial traffic state recursive function. A disturbance variable is introduced into the traffic state recursive function. The disturbance variable is an independent random term that changes over time and is superimposed on each state update operation. The data from external input is fused with the free-flow velocity adjustment, link capacity adjustment, and signal timing parameter adjustment defined in the standardization module and embedded as external intervention parameters into the model structure to modify the recursive coefficients, boundary conditions, and velocity terms. Based on the traffic conditions, disturbance variables, and external intervention parameters of the previous time step, the traffic indicator prediction values ​​for the next time step are generated recursively at each time step, and the final traffic indicator mechanism prediction results are output.

5. The traffic indicator prediction method based on multi-source fusion according to claim 2, characterized in that, The generation of traffic indicator machine learning prediction results includes: Fields containing traffic volume, traffic density, average speed, queue length, delay time and congestion index are extracted from standardized traffic feature data, and a historical data set of traffic indicators is constructed by combining the corresponding time index. In the historical data set of traffic indicators, a fixed-length time window with the target prediction time as a reference is selected, continuous time segments are extracted as feature input samples, and the fields in each time segment are constructed into a set of structured feature vectors in a time stacking manner. The structured feature vectors mentioned above are used as input variables of the supervised learning model, and the traffic index field at the target prediction time is selected as the output variable of the supervised learning model. An input-output mapping relationship is constructed to generate a supervised modeling sample set. Based on the complexity and data structure of the supervised learning model, a linear regression model is selected on the supervised modeling sample set, and the input layer dimension, parameter set and output node settings are initialized; Define a loss function targeting prediction error. ,in The model parameter set is updated using gradient descent based on the training sample set. The process continues until a preset stopping condition is met, resulting in a fully trained supervised learning model. Standardized traffic feature data prior to the target prediction time are used to construct input feature vectors in the same way as training samples. These vectors are then input into the trained supervised learning model to generate machine learning prediction results for traffic indicators. 6.The traffic index prediction method based on multi-source fusion according to claim 2, characterized in that, The generation of traffic indicator fusion prediction results includes: Receive the traffic indicator mechanism prediction results and the traffic indicator machine learning prediction results, and combine their outputs under the same prediction time index to form a prediction result pair; Extract the actual observation values ​​corresponding to the above time index from the historical data of traffic indicators, form a sequence of true values ​​corresponding to the prediction results, and align them according to the index; The validation error between the traffic indicator mechanism prediction results and the actual value sequence, and the validation error between the traffic indicator machine learning prediction results and the actual value sequence are calculated separately. A set of standardized error pairs are obtained by using a unified error evaluation method. A unified error assessment method refers to using the same error calculation method when comparing the traffic indicator mechanism prediction results and the traffic indicator machine learning prediction results with the corresponding true values ​​in the historical data of traffic indicators. This usually includes the mean absolute error or the mean squared error, and the calculation is performed point by point on the same time index. This ensures that the two prediction methods are comparable in terms of error evaluation dimensions, scales and time structures, thereby providing a unified benchmark for the generation of fusion weights. Based on the standardized error, an initial fusion weight factor is generated. The side with smaller error is assigned a higher weight, and the side with larger error is assigned a lower weight. The two weight factors are then smoothed using a preset proportional adjustment parameter to generate the final fusion weight. The final fusion weights are respectively assigned to the output values ​​of the traffic indicator mechanism prediction results and the traffic indicator machine learning prediction results. The two are then weighted and combined according to a unified time index to obtain the fused traffic indicator prediction value. Arrange the fusion results under all prediction time in chronological order to form a sequence of traffic indicator fusion prediction results. 7.The traffic index prediction method based on multi-source fusion according to claim 2, characterized in that, The generation of the rolling forecast series for traffic indicators includes: The prediction start year is set as the starting point for rolling prediction. The prediction start year is selected from historical traffic indicator data, requiring that there are several consecutive years of standardized traffic characteristic data and traffic indicator fusion prediction results before it. The prediction window length is set to limit the historical time span used in each round of model training. The prediction window consists of w complete years before the prediction start date, and standardized traffic feature data and fused prediction results for the corresponding years are extracted from it. The extracted data is used as the model training input to train the system dynamics model and the supervised learning model, respectively, forming a dual-model structure for the current prediction round. The trained dual-model structure is used to predict traffic indicators at the corresponding time point of the prediction start year, generating the first rolling prediction value. The first rolling predicted value is indexed and matched with the true value in the historical data of traffic indicators. An error assessment operation is performed to generate a validation error, which is used to determine whether to update the model parameters or structure. After generating the first rolling forecast, advance the forecast year by one unit year, update the forecast window to include the time period w years before the latest forecast year, and re-extract the training data to repeat the training, forecasting and evaluation process. All the rolling forecast values ​​for each year are arranged in chronological order to form a rolling forecast sequence for traffic indicators. 8.The traffic index prediction method based on multi-source fusion according to claim 2, characterized in that, The generation of traffic indicator scenario prediction sequences includes: Receive the rolling prediction sequence of traffic indicators, extract the standardized traffic feature data under the corresponding time index, and use it as the baseline input for scenario simulation. Set the target scenario conditions and access the intervention parameter set, which includes free flow speed adjustment, link capacity adjustment and signal timing parameter adjustment. Each intervention parameter has a structure field identifier and a quantitative value. Based on the structural field identifier in the intervention parameter set, the corresponding field position is found in the standardized traffic feature data input to the scenario simulation baseline, and the parameter overwrite operation is performed. The parameter overwrite operation is to replace the original field value with the specified intervention value to generate updated traffic feature input data. The updated traffic feature input data is input into the trained system dynamics model and supervised learning model. Without changing the model structure and parameters, only the input features are replaced, and a complete prediction process is executed to generate traffic indicator prediction values ​​under intervention conditions. The above traffic indicator predictions are compared with the rolling predictions under the corresponding time index for structural consistency, and then organized according to the setting order of the intervention scenarios to generate scenario predictions. Arrange all the scenario prediction values ​​under the time index in chronological order to form a traffic indicator scenario prediction sequence. 9.The traffic index prediction method based on multi-source fusion according to claim 2, characterized in that, Structured organization and system interface outputs include: Receive historical traffic indicator data, parse it according to its time index and indicator fields, and form historical data structure units; Receive the rolling forecast sequence of traffic indicators, and organize the rolling forecast values ​​into rolling forecast data structure units based on the time index and indicator fields consistent with historical data; Receive traffic indicator scenario prediction sequences and construct scenario prediction data structure units according to the same time index and indicator fields; The above-mentioned historical data structure unit, rolling prediction data structure unit, and scenario prediction data structure unit are merged according to a unified field order and a unified time index to generate a structured output dataset. The structured output dataset is encapsulated into a data format that can be called by the system interface. The output content in the specified format is generated according to the interface call requirements, and the calling and data reading capabilities are provided to the outside world through the system interface.

Images