Adaptive and interpretable traffic situation prediction method and system based on large language model

Abstract

The application discloses a self-adaptive and interpretable traffic situation prediction method and system based on a large language model, which comprises the following steps: obtaining historical traffic data related to a prediction task, inputting a trained time series prediction base model, generating a plurality of candidate traffic state trajectories of a prediction period based on the prediction task; obtaining context information related to the prediction period, and evaluating the candidate traffic state trajectories based on a configured large language reasoning model to obtain an optimal trajectory conforming to the context information of the prediction period; and generating a text report by using a configured large language interpretation model based on the optimal trajectory and the context information. The application performs probability prediction by using the time series prediction base model, and performs context reasoning and text generation by using the large language model, thereby significantly improving the accuracy, adaptability and interpretability of traffic prediction under abnormal events, and providing strong support for realizing intelligent traffic.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, and more specifically to an adaptive and interpretable traffic situation prediction method and system based on a large language model. Background Technology

[0002] Currently, accurate traffic forecasting is fundamental for intelligent transportation systems to achieve key functions such as congestion mitigation, dynamic route guidance, and emergency management. In recent years, deep learning methods, represented by graph neural networks (GNNs) and sequence models, have made significant progress in short-term forecasting under normal traffic conditions. These models can effectively learn the inherent spatiotemporal patterns in historical data.

[0003] However, existing technologies face three major bottlenecks in practical applications: (1) Vulnerability to abnormal events: When unconventional events such as traffic accidents, severe weather, road construction, or large-scale events occur, the distribution of traffic data changes significantly, leading to a sharp decline in the performance of prediction models that mainly rely on historical conventional patterns for training. At these very moments, accurate prediction is crucial for emergency response. (2) Limitations of contextual information fusion: Existing models usually adopt a fixed input structure, making it difficult to flexibly fuse diverse real-time contextual information. For example, dynamically integrating an accident report described in natural language or a structured construction plan into the prediction process often requires complex feature engineering or model retraining, resulting in poor adaptability. (3) Lack of interpretability and operability: The output of traditional prediction models is usually a high-dimensional numerical matrix (such as the speed or flow of each road segment in the future), which itself does not provide any analysis of the reasons behind the prediction results, nor can it be directly transformed into action guidelines that traffic managers can understand and implement. This "black box" characteristic limits its application value in practical decision support.

[0004] Therefore, existing traffic prediction technologies lack robustness and flexibility in the face of sudden scenarios and are difficult to support real-time decision-making needs. There is an urgent need for a solution that can integrate contextual information, has interpretability, and is generalizable.

[0005] On the other hand, large language models (LLMs) have demonstrated powerful natural language understanding, contextual reasoning, and generative capabilities, offering new possibilities for solving the aforementioned problems. However, despite their strong text reasoning and generation capabilities, large language models are not suitable for predicting numerical continuous variables, especially in the absence of numerical awareness, where prediction accuracy is limited. Therefore, relying solely on LLMs cannot meet the stringent requirements of traffic prediction for quantitative accuracy.

[0006] Therefore, how to effectively combine the numerical modeling capabilities of time series forecasting models with the contextual reasoning capabilities of large language models to develop a new paradigm for traffic forecasting that is accurate, adaptable, and interpretable is a technical challenge that urgently needs to be solved in this field. Summary of the Invention

[0007] In view of this, the present invention provides an adaptive and interpretable traffic situation prediction method and system based on a large language model. By combining the probabilistic prediction capability of the time series model with the contextual reasoning and generation capability of the large language model, the accuracy, adaptability and interpretability of traffic prediction under abnormal events are significantly improved, providing strong support for realizing a more efficient and resilient intelligent transportation system.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] In a first aspect, the present invention provides an adaptive and interpretable traffic situation prediction method based on a large language model, comprising the following steps:

[0010] S1. Obtain historical traffic data related to the prediction task, input it into the trained time series prediction model, and generate multiple candidate traffic state trajectories for the prediction period based on the prediction task.

[0011] S2. Obtain contextual information related to the predicted time period, and evaluate the candidate traffic state trajectory based on the configured large language inference model to obtain the optimal trajectory that conforms to the contextual information of the predicted time period.

[0012] S3. Based on the optimal trajectory and context information, generate a text report containing traffic situation analysis and actionable traffic management suggestions using the configured large language interpretation model.

[0013] Furthermore, the time series prediction basic model described in step S1 is built based on the Transformer architecture;

[0014] By segmenting and quantizing numerical traffic time-series data, the spatiotemporal dependence of traffic flow is captured, and the congestion probability and vehicle speed are predicted for the corresponding time period; thus, multiple candidate traffic state trajectories are obtained.

[0015] Furthermore, the historical traffic data mentioned in step S1 is numerical time-series data, including historical traffic flow and speed data for the predicted road segment and surrounding area; using X... {1:T} Indicates that T represents the current time;

[0016] Based on the historical traffic data, K candidate traffic state trajectories for prediction time periods are generated, using {X} k} {T+1:T+n}Let k = 1, ..., K, where k is the number of candidate trajectories and n is the prediction step size.

[0017] Furthermore, the contextual information related to the prediction period mentioned in step S2 includes structured data and unstructured data;

[0018] The structured data includes weather forecasts and planned road construction schedules;

[0019] The unstructured data includes traffic accident reports and notices of large public events in natural language format.

[0020] Furthermore, the reasoning process of the large language reasoning model described in step S2 is as follows:

[0021] The contextual information is semantically understood to analyze its causal impact on traffic flow, including the location, severity, and duration of the impact.

[0022] Based on the causal influence, the optimal trajectory is selected from the candidate traffic state trajectories of the multiple prediction time periods.

[0023] Furthermore, step S2 is expressed by the following formula:

[0024]

[0025] in, This represents the optimal trajectory that matches the context information of the prediction period; F LLM Represents the adjustment function of the large language reasoning model; {X k} {T+1:T+n} This represents the set of K candidate traffic state trajectories for the prediction period; S represents the context information related to the prediction period.

[0026] Furthermore, the process of obtaining the optimal trajectory that conforms to the context information of the prediction period in step S2 is equivalent to solving a posterior probability maximization problem with s as a condition.

[0027] The solution formula is:

[0028]

[0029] Where P(·|s) represents the posterior reasonableness score of each candidate trajectory given contextual information s, as evaluated by the large language reasoning model.

[0030] Furthermore, the actionable traffic management recommendations described in step S3 specifically include:

[0031] Suggestions for adjusting traffic light timings, traffic management or detour strategies, and the release of travel information to the public for the relevant road sections.

[0032] Secondly, this invention provides an adaptive and interpretable traffic situation prediction system based on a large language model, comprising the following modules:

[0033] Trajectory generation module: used to acquire historical traffic data related to the prediction task, input the trained time series prediction base model, and generate multiple candidate traffic state trajectories for the prediction period based on the prediction task;

[0034] Trajectory evaluation module: used to obtain contextual information related to the predicted time period, and evaluate the candidate traffic state trajectory based on the configured large language inference model to obtain the optimal trajectory that conforms to the contextual information of the predicted time period;

[0035] The trajectory traffic interpretation module is used to generate a text report containing traffic situation analysis and actionable traffic management suggestions based on the optimal trajectory and context information, using a configured big data interpretation model.

[0036] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses an adaptive and interpretable traffic situation prediction method and system based on a large language model, which has the following beneficial effects:

[0037] This invention processes high-dimensional, noisy numerical data using a time-series prediction model to generate multiple possible trajectories; then, it combines this with a large language model to perform reasoning and decision-making based on external, semantically rich contextual information. This avoids having the large language model directly handle precise numerical regression tasks that it is not good at, and also avoids the predicament of traditional models having poor generalization when faced with unseen external events; at the same time, it further improves the accuracy of anomaly event prediction.

[0038] The large language model of this invention performs analysis, reasoning, and decision interpretation through different system instructions and task prompts; it processes various structured and unstructured information through natural language prompts, and possesses zero-shot analysis and response capabilities for novel types of events not seen during training; thus enhancing the model's adaptability and generalization ability. Furthermore, it transforms pure numerical prediction into intelligent decision support with accompanying causal analysis and specific action suggestions, bridging the critical gap from "prediction" to "action," greatly enhancing the practical application value of the prediction system; and achieving interpretability and operability of predictions. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0040] Figure 1 The flowchart illustrates the adaptive and interpretable traffic situation prediction method based on a large language model, as provided in this embodiment of the invention.

[0041] Figure 2 This is a schematic diagram illustrating the implementation process of the adaptive and interpretable traffic situation prediction method based on a large language model provided in an embodiment of the present invention.

[0042] Figure 3 The diagram shows the structure of the adaptive and interpretable traffic situation prediction system based on a large language model, as provided in this embodiment of the invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] Example 1

[0045] This invention discloses an adaptive and interpretable traffic situation prediction method based on a large language model, referring to... Figure 1 As shown, it includes the following steps:

[0046] S1. Obtain historical traffic data related to the prediction task, input it into the trained time series prediction model, and generate multiple candidate traffic state trajectories for the prediction period based on the prediction task.

[0047] S2. Obtain contextual information related to the prediction period, and evaluate candidate traffic state trajectories based on the configured large language inference model to obtain the optimal trajectory that conforms to the contextual information of the prediction period.

[0048] S3. Based on the optimal trajectory and contextual information, use the configured large language interpretation model to generate a text report containing traffic situation analysis and actionable traffic management suggestions.

[0049] This example is used to predict a planned road construction event; the traffic management center receives a notification: "On a certain city's main road, from 9:00 a.m. to 1:00 p.m. tomorrow, one inner lane of each of the eight lanes in both directions will be closed for road maintenance."

[0050] This embodiment adopts a multi-stage collaborative framework for traffic situation prediction during road maintenance. It separates numerical prediction from contextual reasoning and adopts a modular design concept. The prediction process is completed in three stages: "numerical prediction - contextual reasoning - decision generation". The structure is clear, the interpretability is strong, and the adaptability is good.

[0051] The implementation steps of this embodiment are as follows:

[0052] Step 1: Probabilistic Trajectory Generation: Based on historical traffic time-series data, a set of multiple (K, K>1) candidate future traffic state trajectories is generated through a time-series prediction model; among them, multiple candidate trajectories together constitute a probabilistic prediction of future traffic state.

[0053] Step 2: Contextual Reasoning and Trajectory Calibration: Obtain real-time contextual information related to the prediction period, and input the contextual information and the set of multiple candidate trajectories generated in Step 1 into a large language model configured for analysis and selection tasks; the large language model performs causal reasoning on the contextual information, evaluates the degree of matching between each candidate trajectory and the causal reasoning, and selects or adjusts an optimal future traffic state prediction trajectory accordingly.

[0054] Step 3: Explanatory Report and Recommendation Generation: Input the optimal predicted trajectory and contextual information determined in Step 2 into the large language model configured for explanation and generation tasks, which will generate a human-readable text report containing traffic situation analysis and actionable management recommendations.

[0055] The implementation steps of this embodiment are described in detail below:

[0056] Following step 1, first obtain historical traffic time series data.

[0057] In this embodiment, before the prediction begins (e.g., at 8:30 AM), historical traffic flow and speed data for the past few hours for the road segment and surrounding area are first acquired and compiled into a historical traffic state sequence X. {1:T} Processing is performed. Among them, X... {1:t} ={x1,...,x T}, t=1,...,T; T is the current time; x t Let t be the traffic state vector at monitoring location N at time t.

[0058] Secondly, a basic model for time-series prediction is built based on the Transformer architecture; by segmenting and quantizing numerical traffic time-series data, the spatiotemporal dependence of traffic flow is captured, and the congestion probability and vehicle speed are predicted for the corresponding time period; thus, multiple candidate traffic state trajectories are obtained.

[0059] The fundamental time-series prediction model is dedicated to extracting spatiotemporal patterns from historical traffic data and generating multiple probabilistic future trajectories, providing a decision space for contextual reasoning in subsequent large language models. Its input consists of structured numerical time-series data (speed, flow, occupancy); the output is multiple candidate future traffic state trajectories (representing uncertainty). The advantage of the fundamental time-series prediction model lies in overcoming the limitations of traditional time-series models in long-range dependency modeling and probabilistic prediction through word segmentation and the Transformer algorithm.

[0060] The data segmentation and quantization processes specifically include: numerical discretization and spatiotemporal embedding coding.

[0061] Numerical discretization transforms continuous values ​​into discrete intervals: the original traffic data is divided into B ordered intervals, for example, the vehicle speed [0,120] km / h is divided into intervals of 5 km / h; B=24, including intervals [0,5), [5,10),...,[115,120], each interval is assigned a unique ID, and a traffic semantic vocabulary is constructed; and it adapts to the discrete token processing paradigm of Transformer, preserving the numerical order ID relationship and avoiding information loss.

[0062] Next, time embedding is performed; the input timestamp t is decomposed into [hour, day of the week, whether it is a holiday], and spatial embedding is performed in the d-dimensional direction; the road segment topology map is used as input, and the graph neural network generates d-dimensional location codes; the feature concatenation result is: Token = Embedding(numerical semantics of feature interval ID ⊕ time context t ⊕ spatial location relationship).

[0063] The time-series prediction model in this embodiment does not output a single deterministic prediction, but instead generates a set {X} containing K candidate future traffic state trajectories through probabilistic sampling. k} {T+1:T+n} , where k is the index of the candidate trajectory k = 1,...,K; n is the predicted future time step. This set represents the probabilistic distribution of future traffic states, capturing the inherent uncertainty of the system.

[0064] The model in this embodiment outputs 20 candidate traffic state trajectories for the period from 9:00 AM to 2:00 PM. Among these trajectories, some show a normal post-mortem decline pattern, while others may show slightly increased congestion due to minor perturbations in historical data.

[0065] This embodiment also evaluates these trajectories; three evaluation indicators are used to evaluate the effectiveness of the spatiotemporal traffic state prediction: (1) Mean Absolute Error (MAE); (2) Mean Absolute Percentage Error (MAPE); (3) Mean Standard Error (RMSE).

[0066] By selecting short-term (15 minutes), medium-term (30 minutes), and long-term (60 minutes) prediction results, three evaluation indicators were calculated respectively, and the results are shown in Table 1:

[0067]

[0068] Table 1 Comparison of Results for Three Evaluation Indicators

[0069] It can be seen that the prediction in step 1 of this embodiment is accurate and reliable in terms of multiple evaluation indicators and multiple dimensions, laying a good foundation for subsequent steps.

[0070] Next, following step 2, refer to Figure 2 As shown, the optimal trajectory is obtained by evaluating the large language model in combination with contextual information.

[0071] This embodiment uses the real-time acquired multimodal context information s (including text, tables, etc.) related to the prediction period, along with the candidate trajectory set {X} generated in step 1. k} {T+1:T+n} The data is input into a large language model. This large language model is a pre-configured large language inference model for analysis and selection tasks; it performs causal inference on contextual information, evaluates the degree of matching between each candidate trajectory and the causal inference, and selects or adjusts an optimal future traffic state prediction trajectory accordingly.

[0072] The large language reasoning model in this embodiment serves as a context adjustment function F. LLM It performs comprehensive reasoning on the information and outputs an optimal predicted trajectory after context correction. This process can be represented as: The function F LLM The internal logic is to first understand the causal effects of the events described in s, and then evaluate each candidate trajectory X. k The degree of matching with this influence ultimately determines the most reasonable trajectory to be selected or merged.

[0073] In this embodiment, the following information is combined into a prompt and input into the large language foundation model (such as Qwen2.5-14B-Instruct).

[0074] Contextual information: "Event: Planned construction. Location: Main road X. Time: 9:00 a.m. to 1:00 p.m. Details: Closure of one inner lane on each of the eight lanes in both directions, reducing lane capacity by 25%."

[0075] Candidate trajectory summary: Provides summary information for 20 candidate trajectories, such as "Candidate 1: Congestion peak at 9:30, speed 35km / h" and "Candidate 2: Congestion continues until 11:00, minimum speed 30km / h".

[0076] Task instructions: "Based on the above event information, assess its impact on traffic and select the most likely trajectory from the following candidate routes."

[0077] The large language model makes the following inference:

[0078] “Closing 25% of the lanes occurs during the off-peak period right after the morning rush hour, but the reduced traffic capacity will create a bottleneck effect.”

[0079] "Congestion will be significantly higher than on a normal workday, vehicle speeds will drop dramatically, and this will last throughout the entire construction period."

[0080] "The predictions of Candidates 1 and 2 are too optimistic and do not fully reflect the serious impact of lane closures."

[0081] Based on this reasoning, the large language model in this embodiment selects or synthesizes a new trajectory. This trajectory shows that the vehicle speed drops sharply from 9:00 to below 20 km / h and continues to slowly recover until after 1:00 PM. This trajectory is the optimal predicted trajectory.

[0082] Finally, according to step 3, refer to Figure 2 As shown, the optimal predicted trajectory and context information determined in step 2 are input into a large language model configured for interpretation and generation tasks, which generates a human-readable text report containing traffic situation analysis and actionable management recommendations.

[0083] This embodiment will use the optimal predicted trajectory determined in step 2. The contextual information s that led to this adjustment is then input into the large language model again. This large language model has been configured via a prompt to perform both interpretation and generation tasks. The model is instructed to generate two parts: (i) a concise, human-readable traffic situation analysis report explaining the expected traffic conditions and their causes (e.g., "Due to the accident on G6 Expressway, congestion on this section is expected to worsen in the next hour, with average vehicle speeds dropping below 20 km / h"); and (ii) a series of specific and actionable traffic management recommendations (e.g., "It is recommended to activate control measures on the upstream ramp of the accident and guide vehicles to detour via the XX exit using information boards").

[0084] In this embodiment, the optimal predicted trajectory and construction information obtained in step 2 are input into the big data interpretation model, and the instruction is given: "Based on the prediction results and construction events, generate a traffic situation report and management recommendations." The big data interpretation model outputs the following text report:

[0085] Situation Analysis Report: "Warning: Main Road X will have some lanes closed for construction from 9:00 AM to 1:00 PM, which is expected to cause severe congestion. Forecast data shows that the average speed on this section of road will drop below 20 km / h during this time, and delays may exceed 30 minutes. The congestion is expected to spread upstream to the Y intersection."

[0086] Actionable management suggestions:

[0087] Public information release: "It is recommended to immediately issue warnings through traffic radio and mobile map apps to remind drivers to plan detour routes in advance."

[0088] Traffic management: "Consider implementing temporary traffic control at the Z intersection upstream of the main road X to guide some traffic flow onto the parallel alternative route E."

[0089] Signal Coordination and Control: "It is recommended to adjust the signal timing scheme of each intersection along the alternative route E, and increase the green light time to accommodate the additional detour traffic."

[0090] Through this embodiment, the method of the present invention successfully transforms a simple construction notice into an accurate and adaptive prediction of future traffic conditions, and ultimately generates decision support information with high interpretability and practical guidance value, fully demonstrating the technical concept and superior effects of the present invention.

[0091] Example 2

[0092] This invention discloses an adaptive and interpretable traffic situation prediction system based on a large language model, referring to... Figure 3 As shown, it includes the following modules:

[0093] Trajectory generation module: used to acquire historical traffic data related to the prediction task, input the trained time series prediction base model, and generate multiple candidate traffic state trajectories for the prediction period based on the prediction task;

[0094] Trajectory evaluation module: used to obtain contextual information related to the predicted time period, and evaluate the candidate traffic state trajectory based on the configured large language inference model to obtain the optimal trajectory that conforms to the contextual information of the predicted time period;

[0095] The trajectory traffic interpretation module is used to generate a text report containing traffic situation analysis and actionable traffic management suggestions based on the optimal trajectory and context information, using a configured big data interpretation model.

[0096] The prediction system in this embodiment employs a unique task decoupling and collaborative design. First, the complex traffic prediction task is decomposed into two parts through a trajectory generation module: a time-series prediction base model processes high-dimensional, noisy numerical data, generating multiple possibilities; then, a large language model, based on external, semantically rich contextual information, performs reasoning and decision-making within these possibilities. This avoids having the large model directly handle precise numerical regression tasks, which it is not adept at, and also avoids the predicament of traditional models exhibiting poor generalization when facing unseen external events.

[0097] Secondly, contextual reasoning and trajectory calibration are performed through the trajectory evaluation module. The large language model is used as an analytical reasoning engine, and its input is many-to-many trajectories and multi-source information. The task objective is to converge to an optimal solution.

[0098] Finally, the trajectory traffic interpretation module generates explanatory reports and suggestions. The large language model is used as a narrative generation engine with a single input (an optimal trajectory). The task objective is to generate explanatory and instructive texts in a divergent manner.

[0099] The large language model in this embodiment uses the same LLM base, but achieves a clear functional distinction through different system instructions and task prompts. By positioning the dual functions of LLM, it together constitutes the complete technical chain of this invention from accurate prediction to intelligent decision-making.

[0100] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the system apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0101] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A self-adaptive and interpretable traffic situation prediction method based on a large language model, characterized in that, The method comprises the following steps: S1, obtaining historical traffic data related to a prediction task, inputting a trained time series prediction base model, and generating candidate traffic state trajectories for multiple prediction periods based on the prediction task; S2, obtaining context information related to the prediction period, and evaluating the candidate traffic state trajectories based on a configured large language reasoning model to obtain an optimal trajectory that meets the context information of the prediction period; S3, based on the optimal trajectory and context information, generating a text report containing traffic situation analysis and actionable traffic management recommendations using a configured large language interpretation model; In step S2, the reasoning process of the large language reasoning model is: performing semantic understanding on the context information and analyzing its causal impact on traffic flow, including the location, severity, and duration of the impact; selecting the optimal trajectory from the multiple candidate traffic state trajectories based on the causal impact; The step S2 is expressed by the formula: wherein, represents an optimal trajectory that fits the prediction horizon context information; represents an adjustment function of the large language reasoning model; represents K a set of candidate traffic state trajectories for a prediction horizon; represents context information related to a prediction horizon; The process of deriving the optimal trajectory that fits the prediction period context information in step S2 is equivalent to solving a posterior probability maximization problem with the condition The solution formula is: wherein, represents the posterior plausibility score of each candidate trajectory given the context information the posterior plausibility score of each candidate trajectory given the context information 2. The large language model based adaptive and interpretable traffic situation prediction method of claim 1, wherein, The time series prediction base model in step S1 is based on a Transformer architecture; By tokenizing and quantizing numerical traffic time series data, capturing traffic flow spatio-temporal dependence, and predicting congestion probability and vehicle speed for the corresponding time period, multiple candidate traffic state trajectories are obtained.

3. The large language model based adaptive and interpretable traffic situation prediction method of claim 1, wherein, The historical traffic data in step S1 is numerical time series data, including historical traffic flow and speed data of the predicted road section and the surrounding area; the historical traffic data is preprocessed by using denotes, t is the current time; generated based on the historical traffic data K a candidate trajectory of traffic states for a prediction horizon, using representations; wherein, is the number of candidate trajectories; is the prediction step.

4. The large language model based adaptive and interpretable traffic situation prediction method of claim 1, wherein, In step S2, the context information related to the prediction period includes structured data and unstructured data; The structured data includes weather forecasts and scheduled road construction plans; The unstructured data includes traffic accident reports and large public event notifications in natural language form.

5. The large language model based adaptive and interpretable traffic situation prediction method of claim 1, wherein, The actionable traffic management recommendations in step S3 include: signal timing adjustment recommendations for corresponding road segments, traffic diversion or detour strategy recommendations, and public travel information release recommendations.

6. The adaptive and interpretable traffic situation prediction system based on large language model, adopting the adaptive and interpretable traffic situation prediction method based on large language model according to any one of claims 1-5, characterized in that, The method comprises the following modules: Trajectory generation module: used to obtain historical traffic data related to a prediction task, input a trained time series prediction base model, and generate multiple candidate traffic state trajectories for a prediction period based on the prediction task; Trajectory evaluation module: used to obtain context information related to the prediction period, and evaluate the candidate traffic state trajectories based on a configured large language reasoning model to obtain an optimal trajectory that meets the context information of the prediction period; Trajectory traffic interpretation module: used to generate a text report containing traffic situation analysis and actionable traffic management recommendations using a configured large language interpretation model based on the optimal trajectory and context information.

Images