Flow prediction and adjustment method and device based on generative AI quantization external factors
By combining generative AI with time-series prediction algorithms, we have achieved traffic prediction based on natural language interaction, which solves the problems of high threshold, slow response and difficulty in fusing unstructured factors in existing technologies, and provides fast, accurate traffic prediction and adaptive capabilities.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-25
- Publication Date
- 2026-04-14
AI Technical Summary
Existing traffic prediction technologies have high barriers to entry, slow response times, difficulty in integrating unstructured external factors, lack of ability to predict the future, and rigid interaction, making it impossible to quickly obtain accurate traffic data.
By employing a generative AI approach to quantify external factors, combined with Large Language Model (LLM) and time-series prediction algorithms (such as Prophet), traffic prediction results are obtained through natural language interaction, and the interaction logic is automatically adjusted to achieve the quantification and dynamic updating of unstructured events.
It achieves interactive natural language prediction with zero barriers to entry, can quickly obtain accurate network traffic prediction results, dynamically quantify the impact of sudden events, and the model has self-evolution capabilities, improving the flexibility and efficiency of business planning.
Smart Images
Figure CN121864616A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of traffic forecasting, and in particular to a method and apparatus for traffic forecasting and adjustment based on generative AI-quantified external factors. Background Technology
[0002] 1. Business Demand-Driven: With the explosive growth of cloud network services, network traffic has become extremely complex. The Network Operations Center (NOC) and marketing department currently face new and pressing needs:
[0003] Immediacy requirement: Business planning and marketing assurance require quick access to future traffic data, rather than waiting for hours of analysis reports.
[0004] Multidimensional attribution needs: Traffic is not only affected by time cycles, but also by "unstructured" external events (such as breaking news, large concerts, and e-commerce promotions). The business side urgently needs tools that can automatically quantify the impact of these "events".
[0005] 2. Pain points of existing technologies: Existing traffic prediction technologies have the following significant drawbacks:
[0006] High barriers to entry and slow response: Traditional methods rely on analysts writing Python / SQL code or operating complex BI tools, which non-technical personnel cannot use directly, and the cost of cross-departmental communication is extremely high.
[0007] Missing external factors: Traditional linear forecasting models are mainly based on historical trends and have difficulty in incorporating "unstructured" external factors. For example, the impact of trending news, sudden large-scale concerts, holiday adjustments, or specific marketing activities on traffic is difficult to automatically quantify in traditional models.
[0008] Inflexible interaction: Existing systems only display real-time or historical data, lack the ability to "predict" the future, and do not support multiple rounds of follow-up questions. Summary of the Invention
[0009] To address the aforementioned technical problems in existing traffic prediction technologies, this invention provides a method and apparatus for traffic prediction and adjustment based on generative AI to quantify external factors. It combines Large Language Model (LLM) with time-series prediction algorithms (such as Prophet) to quickly obtain accurate network traffic prediction results and automatically adjusts the interaction logic according to user needs (such as supplementing key information and explaining prediction results).
[0010] This invention relates to an intelligent network traffic prediction system.
[0011] To achieve the above objectives, the present invention adopts the following technical solution:
[0012] In one embodiment of the present invention, a method for traffic prediction and adjustment based on generative AI-quantified external factors is proposed, the method comprising:
[0013] The user inputs natural language into the fine-tuned large language model, which performs parameter parsing and / or parameter completion on the natural language to generate a structured API request.
[0014] The structured API request is input into the trained Prophet model, and the trained Prophet model loads the impact factors corresponding to the hot events parsed from the impact factor library, and calculates the traffic trend and peak data by combining them with historical traffic data.
[0015] The fine-tuned large language model reads the prediction results and converts them into natural language and a visual trend chart, which is then returned to the user.
[0016] After the forecast is completed, the actual flow rate is automatically compared with the forecast value, and the impact factors in the impact factor library are dynamically updated.
[0017] Furthermore, the Prophet model formula is as follows:
[0018]
[0019] in, Represents the predicted flow value. Represents the trend item. Represents a periodic term. Represents holidays / events. This represents the error term.
[0020] Furthermore, the algorithm for quantifying unstructured hotspot events is as follows:
[0021] Use a pre-trained text embedding model to map the text describing trending events into high-dimensional vectors. The cosine similarity is calculated between the vectors and those in the historical event database to match the historical average impact factor of this type of hot topic event. .
[0022] Furthermore, the steps for constructing the impact factor library are as follows:
[0023] Using the Prophet model, a baseline model without event features is trained using historical traffic data to calculate the theoretical event-free traffic for each historical day. ;
[0024] Find all the dates of the hot events. Calculate the actual flow The ratio or difference is calculated using the following formula:
[0025]
[0026] Calculated Cluster the data according to the type of trending events and take the average value before storing it in the database.
[0027] In one embodiment of the present invention, a flow prediction and adjustment device based on generative AI quantification of external factors is also proposed, the device comprising:
[0028] The semantic interaction module, based on a large language model, is used for natural language intent recognition, parameter extraction, and response generation.
[0029] The factor quantification module is used to convert unstructured hot topics into numerical impact factors and store them in the database.
[0030] The time-series prediction engine module integrates the Prophet model, which is used to perform multivariate regression calculations based on historical traffic data and the aforementioned influencing factors to obtain traffic trends and peak data within the predicted time period.
[0031] The dynamic correction module is used to compare the actual flow with the predicted value and dynamically update the impact factors in the impact factor library.
[0032] Furthermore, the Prophet model formula is as follows:
[0033]
[0034] in, Represents the predicted flow value. Represents the trend item. Represents a periodic term. Represents holidays / events. This represents the error term.
[0035] Furthermore, the algorithm for quantifying unstructured hotspot events is as follows:
[0036] Use a pre-trained text embedding model to map the text describing trending events into high-dimensional vectors. The cosine similarity is calculated between the vectors and those in the historical event database to match the historical average impact factor of this type of hot topic event. .
[0037] Furthermore, the steps for constructing the impact factor library are as follows:
[0038] Using the Prophet model, a baseline model without event features is trained using historical traffic data to calculate the theoretical event-free traffic for each historical day. ;
[0039] Find all the dates of the hot events. Calculate the actual flow The ratio or difference is calculated using the following formula:
[0040]
[0041] Calculated Cluster the data according to the type of trending events and take the average value before storing it in the database.
[0042] In one embodiment of the present invention, a computer device is also proposed, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the aforementioned flow prediction and adjustment based on generative AI quantification of external factors.
[0043] In one embodiment of the present invention, a computer-readable storage medium is also proposed, which stores a computer program that performs flow prediction and adjustment based on generative AI quantification of external factors.
[0044] Beneficial effects:
[0045] 1. Event-driven dynamic quantification prediction mechanism: This invention innovatively uses the semantic understanding capability of a large language model to quantify unstructured hot events (such as "concerts") into computable mathematical features and dynamic weights in real time, enabling the prediction model to accurately capture the impact of sudden events and breaking through the limitation of traditional time series models that rely solely on historical data.
[0046] 2. Generative + Analytical Dual-Engine Architecture: This invention integrates the semantic interaction capabilities of a large language model with the computational accuracy of temporal models such as Prophet. The large language model is responsible for intent recognition, parameter completion, and result interpretation (Function Calling), while the temporal model is responsible for rigorous mathematical deduction. The two complement each other's strengths.
[0047] 3. Parameter Adaptive Correction Based on Posterior Data: This invention establishes an automatic closed-loop feedback mechanism. After each prediction, the actual flow rate is automatically compared with the predicted value, and the influence factors in the influence factor library are dynamically updated, realizing the self-evolution capability of the model to become more accurate with use.
[0048] 4. Zero-barrier interactive natural language prediction: This invention simplifies the prediction process from "writing code / operating BI" to "natural language dialogue". Users can dynamically adjust prediction conditions through multi-round dialogue, greatly improving the flexibility and efficiency of business planning. Attached Figure Description
[0049] Figure 1 This is a schematic diagram of the flow prediction and adjustment method for external factors based on generative AI in this invention;
[0050] Figure 2 This is a schematic diagram of the flow prediction and adjustment device based on generative AI to quantify external factors according to the present invention.
[0051] Figure 3 This is a schematic diagram of the computer device structure of the present invention. Detailed Implementation
[0052] The principles and spirit of the present invention will now be described with reference to several exemplary embodiments. It should be understood that these embodiments are provided merely to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. Rather, these embodiments are provided to make this disclosure more thorough and complete, and to fully convey the scope of this disclosure to those skilled in the art.
[0053] Those skilled in the art will recognize that embodiments of the present invention can be implemented as an apparatus, device, device, method, or computer program product. Therefore, this disclosure can be specifically implemented in the following forms: entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.
[0054] According to an embodiment of the present invention, a traffic prediction and adjustment method based on generative AI to quantify external factors is proposed, combining Large Language Model (LLM) and time series prediction algorithms (such as Prophet) for network traffic prediction. Previously, predicting "how much base station traffic will be if a concert is held next Friday" required professional data analysts to write code and run it for several hours. Now, managers only need to ask the AI, as casually as chatting on WeChat: "Predict the traffic at the stadium next Friday, there's a concert that day." The AI immediately understands, automatically checks historical data, calculates the specific influencing factors of the "concert" event on traffic, and then tells the user the result. Furthermore, the AI can learn and adjust its prediction logic for the next time based on the accuracy of the prediction.
[0055] The principles and spirit of the present invention will be explained in detail below with reference to several representative embodiments.
[0056] Figure 1 This is a schematic diagram of the flow chart of the method for predicting and adjusting traffic based on generative AI to quantify external factors, as per the present invention. Figure 1 As shown, the specific solution is as follows:
[0057] 1. Core Algorithm Selection and Model Construction:
[0058] 1.1 Construction of Time Series Prediction Algorithm: The Prophet model is selected. The Prophet model is essentially an additive regression model, which is very suitable for processing time series data with strong periodicity (daily / weekly / yearly), holiday effects, and sudden events (added through external regression, which is used to explain the numerical variables of non-periodic and sudden impacts in traffic fluctuations), and has good interpretability.
[0059] 1.2 Prophet Model Formula: ;
[0060] : Predicted traffic value;
[0061] Trend: Used to fit non-periodic changes in traffic (such as long-term upward trends caused by user growth).
[0062] Seasonality, fitted using Fourier series, includes weekly (weekday / weekend) and annual cycles;
[0063] Holidays / Events; This is the core improvement of the present invention;
[0064] Error term.
[0065] 1.3 Unstructured Hotspot Event Quantification Algorithm: LLM Embeddings + Similarity Matching;
[0066] To address the issue that unstructured descriptions of trending events such as "concerts" and "heavy rain" cannot be directly input into the Prophet model, the following algorithm is adopted:
[0067] Algorithm: Use a pre-trained text embedding model (such as BERT or the embedding API that comes with a large model).
[0068] Process: Mapping the text describing trending events into high-dimensional vectors The cosine similarity is calculated between the vectors and those in the historical event database to match the historical average impact factor of this type of hot topic event. This represents the historical average gain or attenuation factor of a specific hot event category (such as "concert" or "major ball game") on network traffic.
[0069] The core data of the historical event database comes from the output of the "attack behavior clustering module". After each cluster analysis is completed, the clustering results are standardized into "attack event records" that can be stored and retrieved.
[0070] 2. Model training:
[0071] 2.1 Data Preparation Stage:
[0072] 2.2 Data cleaning: Denoising and missing value interpolation are performed on historical traffic data; historical traffic data is the time-series traffic data of a specific network element (Node) over a period of time (e.g., the past 90 days).
[0073] 2.3 Feature Engineering:
[0074] Time characteristics: Weekends and statutory holidays (Spring Festival, National Day) are marked;
[0075] Event characteristics: Collect trending events (concerts, events, promotions) from the past 3 years and create a table of "event description - occurrence time".
[0076] 3. The construction and training of the Impact Factor Library aims to establish the mathematical relationship between "events and flows":
[0077] 3.1 Training Steps: Baseline Calculation: Using the Prophet model, train a baseline model (excluding event features) using only historical traffic data to calculate the "theoretical no-event traffic" for each historical day. .
[0078] 3.2 Residual Analysis: Identify the dates of all hotspot events from pre-collected and processed external data, either manually or through automated systems. Calculate the actual flow The ratio or difference:
[0079]
[0080] 3.3 Clustering and Archiving: The calculated clusters are then archived. Cluster the data according to the type of trending events (e.g., concerts, e-sports, rainstorms) and take the average value before storing it in the database.
[0081] 4. Fine-tuning of Large Language Models (LLM) (SFT):
[0082] 4.1 Objective: To improve the accuracy of intent recognition and the precision of SQL / API parameter extraction.
[0083] 4.2 Data: Construct an instruction dataset that fine-tunes the “user question - standard API payload (passing parameters)”.
[0084] 4.3 Method: The Large Language Model (LLM) was lightly tuned using the LoRA (Low-Rank Adaptation) technique to enable it to understand telecommunications-specific terms (such as "network element", "peak utilization", "ER Lang" etc.).
[0085] 5. Multimodal input and parameter parsing.
[0086] 5.1 Receiving instructions: The user inputs natural language (e.g., "Predict the traffic at Tianhe Sports Center next Friday, when there is a major ball game") into the fine-tuned large language model.
[0087] 5.2 Further, Intent Recognition and Slot Filling (LLM).
[0088] 5.3 Furthermore, the intent is identified as Traffic_Predict using a large language model.
[0089] 5.4 Further, extract entities from the original user input natural language: time = Next_Friday, location = Tianhe_Center, event description = Big_Soccer_Match.
[0090] 5.5 Further, parameter completion: If the location or time is missing, the large language model triggers the question-and-answer logic to automatically complete it. This is achieved through the large language model's dialogue management capabilities and pre-trained parameter verification rules, which essentially simulates the human question-and-answer and confirmation process.
[0091] 6. Quantification and Influence Factors of Unstructured Hot Topics:
[0092] 6.1 Further, event vectorization: Inputting "major ball game" into the embedding model generates vectors. ;vector As a semantic probe, it finds the most semantically similar historical event experience in the impact factor database, thereby providing numerical and dynamic external regression parameters for the Prophet model. .
[0093] 6.2 Further, feature matching: Search the built-in "Impact Factor Library". This library stores historical events and their corresponding impact factors.
[0094] 6.3 Furthermore, load the influencing factors: load the factors within the prediction time period. The value corresponds to the influencing factor β, and the external regression parameter (Regressor) required for Prophet is constructed to enhance the prediction of flow. The external regression parameter is expressed in the Prophet model formula as follows: .
[0095] 7. Model Inference and Prediction:
[0096] 7.1 Further, API calls: The Large Language Model (LLM) generates structured requests through Function Calling, which includes: {time_range, node_id, event_impact_factor: 1.5}; where time_range is the prediction time range, node_id is the network element or location identifier, and event_impact_factor is the quantified impact factor of unstructured hot events.
[0097] This API is the key connection point for realizing the "generative AI + analytical AI" dual-engine architecture. This API is an internal service interface used to call the backend time series prediction engine to perform actual calculations.
[0098] 7.2 Further, load data: Retrieve historical traffic data of the network element (Node) for the past N days (e.g., 90 days) from a large telecommunications network or traffic prediction system. .
[0099] 7.3 Further, Prophet fitting and prediction.
[0100] enter: +Future timestamp +External regression (obtained from 6.3);
[0101] Calculation: The Prophet model is based on the formula
[0102] Calculate the predicted sequence;
[0103] Output: Predicted curve data points and confidence intervals (upper / lower limits).
[0104] 8. Results Generation and Interaction:
[0105] 8.1 The Large Language Model (LLM) reads the predicted JSON data, combines it with the user's original question background, and converts it into natural language: "The peak traffic is expected to be XX next Friday, with a week-on-week increase of XX% due to the impact of the ball game."
[0106] 8.2 Furthermore, the front end synchronously renders a visual trend chart.
[0107] 9. Posterior correction and dynamic adjustment (training closed loop):
[0108] 9.1 Data Collection: After the forecast is completed, collect the actual traffic flow. .
[0109] 9.2 Further, deviation calculation: compare with predicted values With real traffic .
[0110] 9.3 Impact Factor Update (Retrain), if there is a bias If the absolute value of the prediction deviation (the difference between the actual flow and the predicted flow) is greater than a pre-set threshold (the upper limit of the tolerance error), then the event impact factor is updated and corrected (Retrained), and the actual event impact factor is calculated in reverse.
[0111] 9.4. Update the formula: in The learning rate; As a new impact factor, The old impact factor, is the actual impact factor, and 'a' is the learning rate, which is a hyperparameter used to control the magnitude of each correction.
[0112] 9.5 The updated impact factor is stored in the impact factor database for use in predicting similar events in the future.
[0113] In summary, this invention adopts a dual-engine architecture of "generative AI (large model) + analytical AI (time series prediction)" and realizes business requirements through the following key innovations: user input (natural language) → LLM interface layer → intelligent parameter completion → multi-factor prediction engine → result interpretation and visualization → user output (natural language + charts).
[0114] It should be noted that although the operation of the method of the present invention has been described in a specific order in the above embodiments and figures, this does not require or imply that the operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.
[0115] To provide a clearer explanation of the above-mentioned flow prediction and adjustment based on generative AI quantification of external factors, a specific embodiment is described below. However, it is worth noting that this embodiment is only for better illustrating the present invention and does not constitute an improper limitation of the present invention.
[0116] Example:
[0117] This embodiment focuses on describing the complete technical path of "large model semantic parsing -> external event quantization -> temporal model inference -> closed-loop feedback correction". The specific implementation steps are as follows:
[0118] Step 1: Multi-round dialogue and intent analysis. The user enters the command in the dialog box: Predict the traffic flow at Guangzhou Tianhe Sports Center next Friday, taking into account the national sports event.
[0119] 1. Intent Recognition: The semantic interaction module receives text and identifies the user's intent as TRAFFIC_PREDICTION (traffic prediction).
[0120] 2. Entity Extraction (Slot Filling):
[0121] Time: Recognizes "next Friday", automatically calculates the current date, and converts it to a specific date format YYYY-MM-DD (e.g., 2025-11-21).
[0122] Object (Node): Identifies "Guangzhou Tianhe Sports Center" and maps it to the network element ID GZ_TH_001 in the system through fuzzy matching.
[0123] External Events: Identify "national sporting events" as key influencing factors.
[0124] 3. Parameter integrity verification and follow-up questioning: If the user only enters "predict traffic", and the location and time are missing, the system will automatically generate a reply: "Which specific area or network element do you need to predict the traffic of at what time?" to achieve intelligent parameter completion.
[0125] The second step is the vectorization and impact factor matching of unstructured hot topics (the core step). This step solves the problem of how to input "text events" into a "mathematical model".
[0126] 1. Text Vectorization: The Embedding model is used to transform the event description "national sports event" extracted in step 1 into a high-dimensional feature vector. .
[0127] 2. Similarity Search: Perform vector search in the pre-set "Impact Factor Library" to find the event types with the highest historical similarity (such as "major e-sports competition" and "football match").
[0128] 3. Obtain the Impact Factor:
[0129] 3.1 Retrieve the historical average flow gain coefficient, i.e., the influence factor, for this type of event. .
[0130] 3.2 For example: Historical data shows that such events typically lead to a surge in uplink traffic and an increase in the impact factor. (That is, the flow rate is 1.6 times the baseline)
[0131] 3.3 Generate a structured parameter package (Payload)
[0132] {
[0133] "node_id": "GZ_TH_001",
[0134] "date": "2025-11-21",
[0135] "event_regressor": {
[0136] "type": "sports_national",
[0137] "weight": 1.6
[0138] }
[0139] }
[0140] Step 3: Multi-factor time-series inference based on the Prophet model, utilizing the Function Calling mechanism to call the backend Python prediction service, and executing the following algorithm flow:
[0141] 1. Data Loading: Read the historical traffic time-series data of network element GZ_TH_001 for the past 90 days. ;
[0142] 2. Model Building: Initialize the Prophet model and configure external regressors.
[0143] Trend: Fits the long-term linear trend resulting from user growth.
[0144] Seasonality: Fits the periodic fluctuations of "Monday to Sunday" and "24 hours" (using Fourier series).
[0145] Holidays / Events: The weight of 1.6 obtained in step 2 is used as the external regression value for that specific date and input into the model.
[0146] Prediction calculation:
[0147]
[0148]
[0149] 4. Output Results: The algorithm outputs the flow curve sequence (24-hour granularity) for the predicted date and peak data (e.g., peak 120Gbps).
[0150] Step 4: Result generation and visualization rendering.
[0151] 1. Natural Language Generation: LLM receives prediction data and generates an analysis report: "The Guangzhou Tianhe Sports Center is expected to reach a peak traffic of 120Gbps next Friday at 20:00. Due to the national sports event, this is expected to increase by 60% compared to a typical Friday (approximately 75Gbps)."
[0152] 2. Chart rendering: The front-end interface receives data and plots two curves: one for "prediction without event baseline" and the other for "prediction with event enhancement" for users to compare intuitively.
[0153] Step 5: Closed-Loop Dynamic Adjustment Based on Posterior Data (Invention Point: Adaptive Mechanism) This embodiment includes an automated model evolution process to address the problem of "inaccurate predictions." Posterior data: Real feedback data obtained only after the event occurs, used to calibrate the model and determine the accuracy of predictions; it is the sole basis for triggering and executing adaptive corrections.
[0154] 1. Real-time monitoring: Automatically collect real traffic data for November 21, 2025, the day after the event ends. ;
[0155] 2. Deviation Calculation: Calculate the predicted peak value. Compared with the actual peak The deviation rate.
[0156] Assumption: Actual traffic surged to 150Gbps (higher than the predicted 120Gbps), a significant deviation.
[0157] 3. Weight Correction:
[0158] Determine the original impact factors It's been underestimated.
[0159] Based on the reverse calculation using the actual values, the actual impact factor of this event should be: .
[0160] 4. Library Update: Automatically update the impact factor library, adjusting the default impact factor for "National Sports Events" from 1.6 to 1.65 (using a moving average update strategy to avoid excessive impact from a single abnormal fluctuation).
[0161] Effect: When a user queries traffic for a similar event again, the model will use the revised influence factor, resulting in more accurate predictions.
[0162] This embodiment achieves a closed-loop process from natural language input to mathematical model calculation, and then to real data feedback and correction through the above steps. Its core advantage lies in using a large model to solve the problem that traditional time series forecasting cannot handle unstructured "events," and through an automated weight correction mechanism, it ensures that the forecasting model has the ability to continuously evolve in the face of a changing external environment.
[0163] Based on the same inventive concept, this invention also proposes a flow prediction and adjustment device based on generative AI to quantify external factors. The implementation of this device can refer to the implementation of the method described above, and repeated details will not be repeated. The term "module" used below can refer to a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.
[0164] Figure 2 This is a schematic diagram of the flow prediction and adjustment device based on generative AI to quantify external factors according to the present invention. Figure 2 As shown, the device includes:
[0165] The semantic interaction module 101, based on a large language model, is used for natural language intent recognition, parameter extraction, and response generation.
[0166] Factor quantification module 102 is used to convert unstructured hot topics into numerical impact factors and store them in the database;
[0167] The algorithm for quantifying unstructured hotspot events is as follows:
[0168] Use a pre-trained text embedding model to map the text describing trending events into high-dimensional vectors. The cosine similarity is calculated between the vectors and those in the historical event database to match the historical average impact factor of this type of hot topic event. .
[0169] The time-series prediction engine module 103 integrates the Prophet model, which is used to perform multivariate regression calculations based on historical traffic data and the influencing factors to obtain traffic trends and peak data within the prediction period.
[0170] The formula for the Prophet model is as follows:
[0171]
[0172] in, Represents the predicted flow value. Represents the trend item. Represents a periodic term. Represents holidays / events. This represents the error term.
[0173] The steps for constructing the impact factor library are as follows:
[0174] Using the Prophet model, a baseline model without event features is trained using historical traffic data to calculate the theoretical event-free traffic for each historical day. ;
[0175] Find all the dates of the hot events. Calculate the actual flow The ratio or difference is calculated using the following formula:
[0176]
[0177] Calculated Cluster the data according to the type of trending events and take the average value before storing it in the database.
[0178] The dynamic correction module 104 is used to compare the actual flow with the predicted value and dynamically update the impact factors in the impact factor library.
[0179] It should be noted that although several modules of the flow prediction and adjustment device based on generative AI quantification of external factors are mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of the present invention, the features and functions of two or more modules described above can be embodied in one module. Conversely, the features and functions of one module described above can be further divided and embodied by multiple modules.
[0180] Based on the aforementioned inventive concept, such as Figure 3 As shown, the present invention also proposes a computer device 200, including a memory 210, a processor 220, and a computer program 230 stored in the memory 210 and executable on the processor 220. When the processor 220 executes the computer program 230, it realizes the aforementioned flow prediction and adjustment based on generative AI quantification of external factors.
[0181] Based on the aforementioned inventive concept, the present invention also proposes a computer-readable storage medium storing a computer program that performs the aforementioned flow prediction and adjustment based on generative AI quantification of external factors.
[0182] The present invention proposes a method and apparatus for predicting and adjusting traffic based on generative AI quantification of external factors, which enables users to quickly obtain accurate network traffic prediction results through natural language interaction similar to daily chat, and the interaction logic will be automatically adjusted according to user needs (such as supplementing key information and explaining prediction results).
[0183] While the spirit and principles of the invention have been described with reference to several specific embodiments, it should be understood that the invention is not limited to the disclosed specific embodiments, and the division of aspects does not imply that features in these aspects cannot be combined for benefit; such division is merely for ease of description. The invention is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.
[0184] Regarding the limitation of the scope of protection of this invention, those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solution of this invention are still within the scope of protection of this invention.
Claims
1. A method for traffic prediction and adjustment based on generative AI-quantified external factors, characterized in that, The method includes: The user inputs natural language into the fine-tuned large language model, which performs parameter parsing and / or parameter completion on the natural language to generate a structured API request. The structured API request is input into the trained Prophet model, and the trained Prophet model loads the impact factors corresponding to the hot events parsed from the impact factor library, and calculates the traffic trend and peak data by combining them with historical traffic data. The fine-tuned large language model reads the prediction results and converts them into natural language and a visual trend chart, which is then returned to the user. After the forecast is completed, the actual flow rate is automatically compared with the forecast value, and the impact factors in the impact factor library are dynamically updated.
2. The method for traffic prediction and adjustment based on generative AI-quantified external factors according to claim 1, characterized in that, The Prophet model formula is as follows: in, Represents the predicted flow value. Represents the trend item. Represents a periodic term. Represents holidays / events. This represents the error term.
3. The method for traffic prediction and adjustment based on generative AI-quantified external factors according to claim 2, characterized in that, The algorithm for quantifying unstructured hotspot events is as follows: Use a pre-trained text embedding model to map the text describing trending events into high-dimensional vectors. The cosine similarity is calculated between the vectors and those in the historical event database to match the historical average impact factor of this type of hot topic event. .
4. The method for traffic prediction and adjustment based on generative AI-quantified external factors according to claim 1, characterized in that, The steps for constructing the impact factor library are as follows: Using the Prophet model, a baseline model without event features is trained using historical traffic data to calculate the theoretical event-free traffic for each historical day. ; Find all dates of hot events. Calculate the actual flow The ratio or difference is calculated using the following formula: Calculated Cluster the data according to the type of trending events and take the average value before storing it in the database.
5. A flow prediction and adjustment device based on generative AI to quantify external factors, characterized in that, The device includes: The semantic interaction module, based on a large language model, is used for natural language intent recognition, parameter extraction, and response generation. The factor quantification module is used to convert unstructured hot topics into numerical impact factors and store them in the database. The time-series prediction engine module integrates the Prophet model, which is used to perform multivariate regression calculations based on historical traffic data and the aforementioned influencing factors to obtain traffic trends and peak data within the predicted time period. The dynamic correction module is used to compare the actual flow with the predicted value and dynamically update the impact factors in the impact factor library.
6. The flow prediction and adjustment device based on generative AI quantification of external factors according to claim 5, characterized in that, The Prophet model formula is as follows: in, Represents the predicted flow value. Represents the trend item. Represents a periodic term. Represents holidays / events. This represents the error term.
7. The flow prediction and adjustment device based on generative AI quantification of external factors according to claim 6, characterized in that, The algorithm for quantifying unstructured hotspot events is as follows: Use a pre-trained text embedding model to map the text describing trending events into high-dimensional vectors. The cosine similarity is calculated between the vectors and those in the historical event database to match the historical average impact factor of this type of hot topic event. .
8. The flow prediction and adjustment device based on generative AI quantification of external factors according to claim 5, characterized in that, The steps for constructing the impact factor library are as follows: Using the Prophet model, a baseline model without event features is trained using historical traffic data to calculate the theoretical event-free traffic for each historical day. ; Find all dates of hot events. Calculate the actual flow The ratio or difference is calculated using the following formula: Calculated Cluster the data according to the type of trending events and take the average value before storing it in the database.
9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method according to any one of claims 1-4.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that performs the method according to any one of claims 1-4.